/*
- * Copyright (c) 2000-2010 Apple Inc. All rights reserved.
+ * Copyright (c) 2000-2016 Apple Inc. All rights reserved.
*
* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
- *
+ *
* This file contains Original Code and/or Modifications of Original Code
* as defined in and that are subject to the Apple Public Source License
* Version 2.0 (the 'License'). You may not use this file except in
* unlawful or unlicensed copies of an Apple operating system, or to
* circumvent, violate, or enable the circumvention or violation of, any
* terms of an Apple operating system software license agreement.
- *
+ *
* Please obtain a copy of the License at
* http://www.opensource.apple.com/apsl/ and read it before using this file.
- *
+ *
* The Original Code and all software distributed under the License are
* distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
* EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
* FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
* Please see the License for the specific language governing rights and
* limitations under the License.
- *
+ *
* @APPLE_OSREFERENCE_LICENSE_HEADER_END@
*/
/*
* @OSF_FREE_COPYRIGHT@
*/
-/*
+/*
* Mach Operating System
* Copyright (c) 1991,1990,1989,1988,1987 Carnegie Mellon University
* All Rights Reserved.
- *
+ *
* Permission to use, copy, modify and distribute this software and its
* documentation is hereby granted, provided that both the copyright
* notice and this permission notice appear in all copies of the
* software, derivative works or modified versions, and any portions
* thereof, and that both notices appear in supporting documentation.
- *
+ *
* CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
* CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR
* ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
- *
+ *
* Carnegie Mellon requests users of this software to return to
- *
+ *
* Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
* School of Computer Science
* Carnegie Mellon University
* Pittsburgh PA 15213-3890
- *
+ *
* any improvements or extensions that they make and grant Carnegie Mellon
* the rights to redistribute these changes.
*/
#include <machine/machine_routines.h>
#include <machine/sched_param.h>
#include <machine/machine_cpu.h>
-#include <machine/machlimits.h>
+#include <machine/limits.h>
+#include <machine/atomic.h>
+
+#include <machine/commpage.h>
#include <kern/kern_types.h>
+#include <kern/backtrace.h>
#include <kern/clock.h>
-#include <kern/counters.h>
#include <kern/cpu_number.h>
#include <kern/cpu_data.h>
+#include <kern/smp.h>
#include <kern/debug.h>
-#include <kern/lock.h>
#include <kern/macro_help.h>
#include <kern/machine.h>
#include <kern/misc_protos.h>
+#if MONOTONIC
+#include <kern/monotonic.h>
+#endif /* MONOTONIC */
#include <kern/processor.h>
#include <kern/queue.h>
#include <kern/sched.h>
#include <kern/sched_prim.h>
+#include <kern/sfi.h>
#include <kern/syscall_subr.h>
#include <kern/task.h>
#include <kern/thread.h>
-#include <kern/wait_queue.h>
#include <kern/ledger.h>
+#include <kern/timer_queue.h>
+#include <kern/waitq.h>
+#include <kern/policy_internal.h>
+#include <kern/cpu_quiesce.h>
#include <vm/pmap.h>
#include <vm/vm_kern.h>
#include <vm/vm_map.h>
+#include <vm/vm_pageout.h>
#include <mach/sdt.h>
+#include <mach/mach_host.h>
+#include <mach/host_info.h>
#include <sys/kdebug.h>
-
+#include <kperf/kperf.h>
+#include <kern/kpc.h>
+#include <san/kasan.h>
#include <kern/pms.h>
+#include <kern/host.h>
+#include <stdatomic.h>
-struct rt_queue rt_runq;
-#define RT_RUNQ ((processor_t)-1)
-decl_simple_lock_data(static,rt_lock);
-
-#if defined(CONFIG_SCHED_TRADITIONAL) || defined(CONFIG_SCHED_PROTO) || defined(CONFIG_SCHED_GRRR) || defined(CONFIG_SCHED_FIXEDPRIORITY)
-static struct fairshare_queue fs_runq;
-#define FS_RUNQ ((processor_t)-2)
-decl_simple_lock_data(static,fs_lock);
-#endif
-
-#define DEFAULT_PREEMPTION_RATE 100 /* (1/s) */
-int default_preemption_rate = DEFAULT_PREEMPTION_RATE;
-
-#define DEFAULT_BG_PREEMPTION_RATE 400 /* (1/s) */
-int default_bg_preemption_rate = DEFAULT_BG_PREEMPTION_RATE;
+struct sched_statistics PERCPU_DATA(sched_stats);
+bool sched_stats_active;
-#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_TRADITIONAL)
-
-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_TRADITIONAL */
-
-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_TRADITIONAL)
-
-unsigned sched_tick;
-uint32_t sched_tick_interval;
-
-uint32_t sched_pri_shift = INT8_MAX;
-uint32_t sched_fixed_shift;
-
-static boolean_t sched_traditional_use_pset_runqueue = FALSE;
-
-__attribute__((always_inline))
-static inline run_queue_t runq_for_processor(processor_t processor)
+int
+rt_runq_count(processor_set_t pset)
{
- if (sched_traditional_use_pset_runqueue)
- return &processor->processor_set->pset_runq;
- else
- return &processor->runq;
+ return atomic_load_explicit(&SCHED(rt_runq)(pset)->count, memory_order_relaxed);
}
-__attribute__((always_inline))
-static inline void runq_consider_incr_bound_count(processor_t processor, thread_t thread)
+void
+rt_runq_count_incr(processor_set_t pset)
{
- if (thread->bound_processor == PROCESSOR_NULL)
- return;
-
- assert(thread->bound_processor == processor);
-
- if (sched_traditional_use_pset_runqueue)
- processor->processor_set->pset_runq_bound_count++;
-
- processor->runq_bound_count++;
+ atomic_fetch_add_explicit(&SCHED(rt_runq)(pset)->count, 1, memory_order_relaxed);
}
-__attribute__((always_inline))
-static inline void runq_consider_decr_bound_count(processor_t processor, thread_t thread)
+void
+rt_runq_count_decr(processor_set_t pset)
{
- if (thread->bound_processor == PROCESSOR_NULL)
- return;
-
- assert(thread->bound_processor == processor);
-
- if (sched_traditional_use_pset_runqueue)
- processor->processor_set->pset_runq_bound_count--;
-
- processor->runq_bound_count--;
+ atomic_fetch_sub_explicit(&SCHED(rt_runq)(pset)->count, 1, memory_order_relaxed);
}
-#endif /* CONFIG_SCHED_TRADITIONAL */
+#define DEFAULT_PREEMPTION_RATE 100 /* (1/s) */
+TUNABLE(int, default_preemption_rate, "preempt", DEFAULT_PREEMPTION_RATE);
-uint64_t sched_one_second_interval;
+#define DEFAULT_BG_PREEMPTION_RATE 400 /* (1/s) */
+TUNABLE(int, default_bg_preemption_rate, "bg_preempt", DEFAULT_BG_PREEMPTION_RATE);
-uint32_t sched_run_count, sched_share_count;
-uint32_t sched_load_average, sched_mach_factor;
+#define MAX_UNSAFE_QUANTA 800
+TUNABLE(int, max_unsafe_quanta, "unsafe", MAX_UNSAFE_QUANTA);
-/* Forwards */
-
-#if defined(CONFIG_SCHED_TRADITIONAL)
+#define MAX_POLL_QUANTA 2
+TUNABLE(int, max_poll_quanta, "poll", MAX_POLL_QUANTA);
-static void load_shift_init(void) __attribute__((section("__TEXT, initcode")));
-static void preempt_pri_init(void) __attribute__((section("__TEXT, initcode")));
+#define SCHED_POLL_YIELD_SHIFT 4 /* 1/16 */
+int sched_poll_yield_shift = SCHED_POLL_YIELD_SHIFT;
-#endif /* CONFIG_SCHED_TRADITIONAL */
+uint64_t max_poll_computation;
-static thread_t thread_select(
- thread_t thread,
- processor_t processor);
-
-#if CONFIG_SCHED_IDLE_IN_PLACE
-static thread_t thread_select_idle(
- thread_t thread,
- processor_t processor);
-#endif
+uint64_t max_unsafe_computation;
+uint64_t sched_safe_duration;
-thread_t processor_idle(
- thread_t thread,
- processor_t processor);
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
-#if defined(CONFIG_SCHED_TRADITIONAL)
+uint32_t std_quantum;
+uint32_t min_std_quantum;
+uint32_t bg_quantum;
-static thread_t steal_thread(
- processor_set_t pset);
+uint32_t std_quantum_us;
+uint32_t bg_quantum_us;
-static thread_t steal_thread_disabled(
- processor_set_t pset) __attribute__((unused));
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+uint32_t thread_depress_time;
+uint32_t default_timeshare_computation;
+uint32_t default_timeshare_constraint;
-static thread_t steal_processor_thread(
- processor_t processor);
+uint32_t max_rt_quantum;
+uint32_t min_rt_quantum;
-static void thread_update_scan(void);
+uint32_t rt_constraint_threshold;
-static void processor_setrun(
- processor_t processor,
- thread_t thread,
- integer_t options);
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
-static boolean_t
-processor_enqueue(
- processor_t processor,
- thread_t thread,
- integer_t options);
+unsigned sched_tick;
+uint32_t sched_tick_interval;
-static boolean_t
-processor_queue_remove(
- processor_t processor,
- thread_t thread);
+/* Timeshare load calculation interval (15ms) */
+uint32_t sched_load_compute_interval_us = 15000;
+uint64_t sched_load_compute_interval_abs;
+static _Atomic uint64_t sched_load_compute_deadline;
-static boolean_t processor_queue_empty(processor_t processor);
+uint32_t sched_pri_shifts[TH_BUCKET_MAX];
+uint32_t sched_fixed_shift;
-static boolean_t priority_is_urgent(int priority);
+uint32_t sched_decay_usage_age_factor = 1; /* accelerate 5/8^n usage aging */
-static ast_t processor_csw_check(processor_t processor);
+/* 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;
-static boolean_t processor_queue_has_priority(processor_t processor,
- int priority,
- boolean_t gte);
+/* 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 */
-static boolean_t should_current_thread_rechoose_processor(processor_t processor);
+uint64_t timer_deadline_tracking_bin_1;
+uint64_t timer_deadline_tracking_bin_2;
-static int sched_traditional_processor_runq_count(processor_t processor);
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
-static boolean_t sched_traditional_with_pset_runqueue_processor_queue_empty(processor_t processor);
+thread_t sched_maintenance_thread;
-static uint64_t sched_traditional_processor_runq_stats_count_sum(processor_t processor);
+/* interrupts disabled lock to guard recommended cores state */
+decl_simple_lock_data(static, sched_recommended_cores_lock);
+static uint64_t usercontrol_requested_recommended_cores = ALL_CORES_RECOMMENDED;
+static void sched_update_recommended_cores(uint64_t recommended_cores);
-static uint64_t sched_traditional_with_pset_runqueue_processor_runq_stats_count_sum(processor_t processor);
-#endif
-
+#if __arm__ || __arm64__
+static void sched_recommended_cores_maintenance(void);
+uint64_t perfcontrol_failsafe_starvation_threshold;
+extern char *proc_name_address(struct proc *p);
+#endif /* __arm__ || __arm64__ */
-#if defined(CONFIG_SCHED_TRADITIONAL)
+uint64_t sched_one_second_interval;
+boolean_t allow_direct_handoff = TRUE;
-static void
-sched_traditional_init(void);
+/* Forwards */
-static void
-sched_traditional_timebase_init(void);
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
-static void
-sched_traditional_processor_init(processor_t processor);
+static void load_shift_init(void);
+static void preempt_pri_init(void);
-static void
-sched_traditional_pset_init(processor_set_t pset);
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
-static void
-sched_traditional_with_pset_runqueue_init(void);
+thread_t processor_idle(
+ thread_t thread,
+ processor_t processor);
-#endif
+static ast_t
+csw_check_locked(
+ thread_t thread,
+ processor_t processor,
+ processor_set_t pset,
+ ast_t check_reason);
-static void
-sched_realtime_init(void) __attribute__((section("__TEXT, initcode")));
+static void processor_setrun(
+ processor_t processor,
+ thread_t thread,
+ integer_t options);
static void
sched_realtime_timebase_init(void);
-#if defined(CONFIG_SCHED_TRADITIONAL)
static void
-sched_traditional_tick_continue(void);
-
-static uint32_t
-sched_traditional_initial_quantum_size(thread_t thread);
-
-static sched_mode_t
-sched_traditional_initial_thread_sched_mode(task_t parent_task);
+sched_timer_deadline_tracking_init(void);
-static boolean_t
-sched_traditional_supports_timeshare_mode(void);
-
-static thread_t
-sched_traditional_choose_thread(
- processor_t processor,
- int priority);
-
-#endif
-
-#if DEBUG
+#if DEBUG
extern int debug_task;
#define TLOG(a, fmt, args...) if(debug_task & a) kprintf(fmt, ## args)
#else
#define TLOG(a, fmt, args...) do {} while (0)
#endif
-#if DEBUG
-static
-boolean_t thread_runnable(
- thread_t thread);
-
-#endif /*DEBUG*/
-
-/*
- * 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
- *
- */
-
-#if defined(CONFIG_SCHED_TRADITIONAL)
-int8_t sched_load_shifts[NRQS];
-int sched_preempt_pri[NRQBM];
-#endif
-
+static processor_t
+thread_bind_internal(
+ thread_t thread,
+ processor_t processor);
-#if defined(CONFIG_SCHED_TRADITIONAL)
-
-const struct sched_dispatch_table sched_traditional_dispatch = {
- sched_traditional_init,
- sched_traditional_timebase_init,
- sched_traditional_processor_init,
- sched_traditional_pset_init,
- sched_traditional_tick_continue,
- sched_traditional_choose_thread,
- steal_thread,
- compute_priority,
- choose_processor,
- processor_enqueue,
- processor_queue_shutdown,
- processor_queue_remove,
- processor_queue_empty,
- priority_is_urgent,
- processor_csw_check,
- processor_queue_has_priority,
- sched_traditional_initial_quantum_size,
- sched_traditional_initial_thread_sched_mode,
- sched_traditional_supports_timeshare_mode,
- can_update_priority,
- update_priority,
- lightweight_update_priority,
- sched_traditional_quantum_expire,
- should_current_thread_rechoose_processor,
- sched_traditional_processor_runq_count,
- sched_traditional_processor_runq_stats_count_sum,
- sched_traditional_fairshare_init,
- sched_traditional_fairshare_runq_count,
- sched_traditional_fairshare_runq_stats_count_sum,
- sched_traditional_fairshare_enqueue,
- sched_traditional_fairshare_dequeue,
- sched_traditional_fairshare_queue_remove,
- TRUE /* direct_dispatch_to_idle_processors */
-};
-
-const struct sched_dispatch_table sched_traditional_with_pset_runqueue_dispatch = {
- sched_traditional_with_pset_runqueue_init,
- sched_traditional_timebase_init,
- sched_traditional_processor_init,
- sched_traditional_pset_init,
- sched_traditional_tick_continue,
- sched_traditional_choose_thread,
- steal_thread,
- compute_priority,
- choose_processor,
- processor_enqueue,
- processor_queue_shutdown,
- processor_queue_remove,
- sched_traditional_with_pset_runqueue_processor_queue_empty,
- priority_is_urgent,
- processor_csw_check,
- processor_queue_has_priority,
- sched_traditional_initial_quantum_size,
- sched_traditional_initial_thread_sched_mode,
- sched_traditional_supports_timeshare_mode,
- can_update_priority,
- update_priority,
- lightweight_update_priority,
- sched_traditional_quantum_expire,
- should_current_thread_rechoose_processor,
- sched_traditional_processor_runq_count,
- sched_traditional_with_pset_runqueue_processor_runq_stats_count_sum,
- sched_traditional_fairshare_init,
- sched_traditional_fairshare_runq_count,
- sched_traditional_fairshare_runq_stats_count_sum,
- sched_traditional_fairshare_enqueue,
- sched_traditional_fairshare_dequeue,
- sched_traditional_fairshare_queue_remove,
- FALSE /* direct_dispatch_to_idle_processors */
-};
-
-#endif
+static void
+sched_vm_group_maintenance(void);
-const struct sched_dispatch_table *sched_current_dispatch = NULL;
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+int8_t sched_load_shifts[NRQS];
+bitmap_t sched_preempt_pri[BITMAP_LEN(NRQS_MAX)];
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
/*
* Statically allocate a buffer to hold the longest possible
* can run on multiple devices with different schedulers configured
* in the device tree.
*/
-#define SCHED_STRING_MAX_LENGTH (48)
-
char sched_string[SCHED_STRING_MAX_LENGTH];
-static enum sched_enum _sched_enum = sched_enum_unknown;
+
+uint32_t sched_debug_flags = SCHED_DEBUG_FLAG_CHOOSE_PROCESSOR_TRACEPOINTS;
+
+/* Global flag which indicates whether Background Stepper Context is enabled */
+static int cpu_throttle_enabled = 1;
void
sched_init(void)
{
- 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 (strlen(sched_arg) > 0) {
- if (0) {
- /* Allow pattern below */
-#if defined(CONFIG_SCHED_TRADITIONAL)
- } else if (0 == strcmp(sched_arg, kSchedTraditionalString)) {
- sched_current_dispatch = &sched_traditional_dispatch;
- _sched_enum = sched_enum_traditional;
- strlcpy(sched_string, kSchedTraditionalString, sizeof(sched_string));
- kprintf("Scheduler: Runtime selection of %s\n", kSchedTraditionalString);
- } else if (0 == strcmp(sched_arg, kSchedTraditionalWithPsetRunqueueString)) {
- sched_current_dispatch = &sched_traditional_with_pset_runqueue_dispatch;
- _sched_enum = sched_enum_traditional_with_pset_runqueue;
- strlcpy(sched_string, kSchedTraditionalWithPsetRunqueueString, sizeof(sched_string));
- kprintf("Scheduler: Runtime selection of %s\n", kSchedTraditionalWithPsetRunqueueString);
-#endif
-#if defined(CONFIG_SCHED_PROTO)
- } else if (0 == strcmp(sched_arg, kSchedProtoString)) {
- sched_current_dispatch = &sched_proto_dispatch;
- _sched_enum = sched_enum_proto;
- strlcpy(sched_string, kSchedProtoString, sizeof(sched_string));
- kprintf("Scheduler: Runtime selection of %s\n", kSchedProtoString);
-#endif
-#if defined(CONFIG_SCHED_GRRR)
- } else if (0 == strcmp(sched_arg, kSchedGRRRString)) {
- sched_current_dispatch = &sched_grrr_dispatch;
- _sched_enum = sched_enum_grrr;
- strlcpy(sched_string, kSchedGRRRString, sizeof(sched_string));
- kprintf("Scheduler: Runtime selection of %s\n", kSchedGRRRString);
-#endif
-#if defined(CONFIG_SCHED_FIXEDPRIORITY)
- } else if (0 == strcmp(sched_arg, kSchedFixedPriorityString)) {
- sched_current_dispatch = &sched_fixedpriority_dispatch;
- _sched_enum = sched_enum_fixedpriority;
- strlcpy(sched_string, kSchedFixedPriorityString, sizeof(sched_string));
- kprintf("Scheduler: Runtime selection of %s\n", kSchedFixedPriorityString);
- } else if (0 == strcmp(sched_arg, kSchedFixedPriorityWithPsetRunqueueString)) {
- sched_current_dispatch = &sched_fixedpriority_with_pset_runqueue_dispatch;
- _sched_enum = sched_enum_fixedpriority_with_pset_runqueue;
- strlcpy(sched_string, kSchedFixedPriorityWithPsetRunqueueString, sizeof(sched_string));
- kprintf("Scheduler: Runtime selection of %s\n", kSchedFixedPriorityWithPsetRunqueueString);
-#endif
- } else {
- panic("Unrecognized scheduler algorithm: %s", sched_arg);
+ boolean_t direct_handoff = FALSE;
+ kprintf("Scheduler: Default of %s\n", SCHED(sched_name));
+
+ if (!PE_parse_boot_argn("sched_pri_decay_limit", &sched_pri_decay_band_limit, sizeof(sched_pri_decay_band_limit))) {
+ /* No boot-args, check in device tree */
+ if (!PE_get_default("kern.sched_pri_decay_limit",
+ &sched_pri_decay_band_limit,
+ sizeof(sched_pri_decay_band_limit))) {
+ /* Allow decay all the way to normal limits */
+ sched_pri_decay_band_limit = DEFAULT_DECAY_BAND_LIMIT;
}
- } else {
-#if defined(CONFIG_SCHED_TRADITIONAL)
- sched_current_dispatch = &sched_traditional_dispatch;
- _sched_enum = sched_enum_traditional;
- strlcpy(sched_string, kSchedTraditionalString, sizeof(sched_string));
- kprintf("Scheduler: Default of %s\n", kSchedTraditionalString);
-#elif defined(CONFIG_SCHED_PROTO)
- sched_current_dispatch = &sched_proto_dispatch;
- _sched_enum = sched_enum_proto;
- strlcpy(sched_string, kSchedProtoString, sizeof(sched_string));
- kprintf("Scheduler: Default of %s\n", kSchedProtoString);
-#elif defined(CONFIG_SCHED_GRRR)
- sched_current_dispatch = &sched_grrr_dispatch;
- _sched_enum = sched_enum_grrr;
- strlcpy(sched_string, kSchedGRRRString, sizeof(sched_string));
- kprintf("Scheduler: Default of %s\n", kSchedGRRRString);
-#elif defined(CONFIG_SCHED_FIXEDPRIORITY)
- sched_current_dispatch = &sched_fixedpriority_dispatch;
- _sched_enum = sched_enum_fixedpriority;
- strlcpy(sched_string, kSchedFixedPriorityString, sizeof(sched_string));
- kprintf("Scheduler: Default of %s\n", kSchedFixedPriorityString);
-#else
-#error No default scheduler implementation
-#endif
}
-
+
+ kprintf("Setting scheduler priority decay band limit %d\n", sched_pri_decay_band_limit);
+
+ if (PE_parse_boot_argn("sched_debug", &sched_debug_flags, sizeof(sched_debug_flags))) {
+ kprintf("Scheduler: Debug flags 0x%08x\n", sched_debug_flags);
+ }
+ strlcpy(sched_string, SCHED(sched_name), sizeof(sched_string));
+
+ cpu_quiescent_counter_init();
+
SCHED(init)();
- SCHED(fairshare_init)();
- sched_realtime_init();
- ast_init();
-
+ SCHED(rt_init)(&pset0);
+ sched_timer_deadline_tracking_init();
+
SCHED(pset_init)(&pset0);
SCHED(processor_init)(master_processor);
+
+ if (PE_parse_boot_argn("direct_handoff", &direct_handoff, sizeof(direct_handoff))) {
+ allow_direct_handoff = direct_handoff;
+ }
}
void
sched_timebase_init(void)
{
- uint64_t abstime;
-
+ 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_TRADITIONAL)
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
-static void
-sched_traditional_init(void)
+void
+sched_timeshare_init(void)
{
/*
* Calculate the timeslicing quantum
* in us.
*/
- if (default_preemption_rate < 1)
+ if (default_preemption_rate < 1) {
default_preemption_rate = DEFAULT_PREEMPTION_RATE;
+ }
std_quantum_us = (1000 * 1000) / default_preemption_rate;
printf("standard timeslicing quantum is %d us\n", std_quantum_us);
- if (default_bg_preemption_rate < 1)
+ if (default_bg_preemption_rate < 1) {
default_bg_preemption_rate = DEFAULT_BG_PREEMPTION_RATE;
+ }
bg_quantum_us = (1000 * 1000) / default_bg_preemption_rate;
printf("standard background quantum is %d us\n", bg_quantum_us);
sched_tick = 0;
}
-static void
-sched_traditional_timebase_init(void)
+void
+sched_timeshare_timebase_init(void)
{
- uint64_t abstime;
- uint32_t shift;
+ uint64_t abstime;
+ uint32_t shift;
/* standard timeslicing quantum */
clock_interval_to_absolutetime_interval(
- std_quantum_us, NSEC_PER_USEC, &abstime);
+ std_quantum_us, NSEC_PER_USEC, &abstime);
assert((abstime >> 32) == 0 && (uint32_t)abstime != 0);
std_quantum = (uint32_t)abstime;
/* quantum for background tasks */
clock_interval_to_absolutetime_interval(
- bg_quantum_us, NSEC_PER_USEC, &abstime);
+ 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);
+ NSEC_PER_USEC, &abstime);
assert((abstime >> 32) == 0 && (uint32_t)abstime != 0);
sched_tick_interval = (uint32_t)abstime;
+ /* timeshare load calculation interval & deadline initialization */
+ clock_interval_to_absolutetime_interval(sched_load_compute_interval_us, NSEC_PER_USEC, &sched_load_compute_interval_abs);
+ os_atomic_init(&sched_load_compute_deadline, sched_load_compute_interval_abs);
+
/*
* Compute conversion factor from usage to
* timesharing priorities with 5/8 ** n aging.
*/
abstime = (abstime * 5) / 3;
- for (shift = 0; abstime > BASEPRI_DEFAULT; ++shift)
+ for (shift = 0; abstime > BASEPRI_DEFAULT; ++shift) {
abstime >>= 1;
+ }
sched_fixed_shift = shift;
- max_unsafe_computation = max_unsafe_quanta * std_quantum;
- sched_safe_duration = 2 * max_unsafe_quanta * std_quantum;
-
- max_poll_computation = max_poll_quanta * std_quantum;
+ for (uint32_t i = 0; i < TH_BUCKET_MAX; i++) {
+ sched_pri_shifts[i] = INT8_MAX;
+ }
+
+ max_unsafe_computation = ((uint64_t)max_unsafe_quanta) * std_quantum;
+ sched_safe_duration = 2 * ((uint64_t)max_unsafe_quanta) * std_quantum;
+
+ max_poll_computation = ((uint64_t)max_poll_quanta) * std_quantum;
thread_depress_time = 1 * std_quantum;
default_timeshare_computation = std_quantum / 2;
default_timeshare_constraint = std_quantum;
+#if __arm__ || __arm64__
+ perfcontrol_failsafe_starvation_threshold = (2 * sched_tick_interval);
+#endif /* __arm__ || __arm64__ */
}
-static void
-sched_traditional_processor_init(processor_t processor)
-{
- if (!sched_traditional_use_pset_runqueue) {
- run_queue_init(&processor->runq);
- }
- processor->runq_bound_count = 0;
-}
-
-static void
-sched_traditional_pset_init(processor_set_t pset)
-{
- if (sched_traditional_use_pset_runqueue) {
- run_queue_init(&pset->pset_runq);
- }
- pset->pset_runq_bound_count = 0;
-}
-
-static void
-sched_traditional_with_pset_runqueue_init(void)
-{
- sched_traditional_init();
- sched_traditional_use_pset_runqueue = TRUE;
-}
-
-#endif /* CONFIG_SCHED_TRADITIONAL */
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
-#if defined(CONFIG_SCHED_TRADITIONAL) || defined(CONFIG_SCHED_PROTO) || defined(CONFIG_SCHED_GRRR) || defined(CONFIG_SCHED_FIXEDPRIORITY)
void
-sched_traditional_fairshare_init(void)
-{
- simple_lock_init(&fs_lock, 0);
-
- fs_runq.count = 0;
- queue_init(&fs_runq.queue);
-}
-#endif
-
-static void
-sched_realtime_init(void)
+pset_rt_init(processor_set_t pset)
{
- simple_lock_init(&rt_lock, 0);
-
- rt_runq.count = 0;
- queue_init(&rt_runq.queue);
+ os_atomic_init(&pset->rt_runq.count, 0);
+ queue_init(&pset->rt_runq.queue);
+ memset(&pset->rt_runq.runq_stats, 0, sizeof pset->rt_runq.runq_stats);
}
static void
/* maximum rt computation (50 ms) */
clock_interval_to_absolutetime_interval(
- 50, 1000*NSEC_PER_USEC, &abstime);
+ 50, 1000 * NSEC_PER_USEC, &abstime);
assert((abstime >> 32) == 0 && (uint32_t)abstime != 0);
max_rt_quantum = (uint32_t)abstime;
+ /* constraint threshold for sending backup IPIs (4 ms) */
+ clock_interval_to_absolutetime_interval(4, NSEC_PER_MSEC, &abstime);
+ assert((abstime >> 32) == 0 && (uint32_t)abstime != 0);
+ rt_constraint_threshold = (uint32_t)abstime;
+}
+
+void
+sched_check_spill(processor_set_t pset, thread_t thread)
+{
+ (void)pset;
+ (void)thread;
+
+ return;
+}
+
+bool
+sched_thread_should_yield(processor_t processor, thread_t thread)
+{
+ (void)thread;
+
+ return !SCHED(processor_queue_empty)(processor) || rt_runq_count(processor->processor_set) > 0;
+}
+
+/* Default implementations of .steal_thread_enabled */
+bool
+sched_steal_thread_DISABLED(processor_set_t pset)
+{
+ (void)pset;
+ return false;
+}
+
+bool
+sched_steal_thread_enabled(processor_set_t pset)
+{
+ return bit_count(pset->node->pset_map) > 1;
}
-#if defined(CONFIG_SCHED_TRADITIONAL)
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
/*
* Set up values for timeshare
static void
load_shift_init(void)
{
- int8_t k, *p = sched_load_shifts;
- uint32_t i, j;
+ int8_t k, *p = sched_load_shifts;
+ uint32_t i, j;
+
+ uint32_t sched_decay_penalty = 1;
+
+ if (PE_parse_boot_argn("sched_decay_penalty", &sched_decay_penalty, sizeof(sched_decay_penalty))) {
+ kprintf("Overriding scheduler decay penalty %u\n", sched_decay_penalty);
+ }
+
+ if (PE_parse_boot_argn("sched_decay_usage_age_factor", &sched_decay_usage_age_factor, sizeof(sched_decay_usage_age_factor))) {
+ kprintf("Overriding scheduler decay usage age factor %u\n", sched_decay_usage_age_factor);
+ }
+
+ if (sched_decay_penalty == 0) {
+ /*
+ * There is no penalty for timeshare threads for using too much
+ * CPU, so set all load shifts to INT8_MIN. Even under high load,
+ * sched_pri_shift will be >INT8_MAX, and there will be no
+ * penalty applied to threads (nor will sched_usage be updated per
+ * thread).
+ */
+ for (i = 0; i < NRQS; i++) {
+ sched_load_shifts[i] = INT8_MIN;
+ }
+
+ return;
+ }
*p++ = INT8_MIN; *p++ = 0;
- for (i = j = 2, k = 1; i < NRQS; ++k) {
- for (j <<= 1; i < j; ++i)
+ /*
+ * 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)
{
- int i, *p = sched_preempt_pri;
+ bitmap_t *p = sched_preempt_pri;
- for (i = BASEPRI_FOREGROUND + 1; i < MINPRI_KERNEL; ++i)
- setbit(i, p);
+ for (int i = BASEPRI_FOREGROUND; i < MINPRI_KERNEL; ++i) {
+ bitmap_set(p, i);
+ }
- for (i = BASEPRI_PREEMPT; i <= MAXPRI; ++i)
- setbit(i, p);
+ for (int i = BASEPRI_PREEMPT; i <= MAXPRI; ++i) {
+ bitmap_set(p, i);
+ }
}
-#endif /* CONFIG_SCHED_TRADITIONAL */
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
/*
* Thread wait timer expiration.
*/
void
thread_timer_expire(
- void *p0,
- __unused void *p1)
+ void *p0,
+ __unused void *p1)
{
- thread_t thread = p0;
- spl_t s;
+ thread_t thread = p0;
+ spl_t s;
+
+ assert_thread_magic(thread);
s = splsched();
thread_lock(thread);
splx(s);
}
-#ifndef __LP64__
-
-/*
- * 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().
- */
-void
-thread_set_timer(
- uint32_t interval,
- uint32_t scale_factor)
-{
- thread_t thread = current_thread();
- uint64_t deadline;
- spl_t s;
-
- s = splsched();
- thread_lock(thread);
- if ((thread->state & TH_WAIT) != 0) {
- clock_interval_to_deadline(interval, scale_factor, &deadline);
- if (!timer_call_enter(&thread->wait_timer, deadline, thread->sched_pri >= BASEPRI_RTQUEUES ? TIMER_CALL_CRITICAL : 0))
- thread->wait_timer_active++;
- thread->wait_timer_is_set = TRUE;
- }
- thread_unlock(thread);
- splx(s);
-}
-
-void
-thread_set_timer_deadline(
- uint64_t deadline)
-{
- thread_t thread = current_thread();
- spl_t s;
-
- s = splsched();
- thread_lock(thread);
- if ((thread->state & TH_WAIT) != 0) {
- if (!timer_call_enter(&thread->wait_timer, deadline, thread->sched_pri >= BASEPRI_RTQUEUES ? TIMER_CALL_CRITICAL : 0))
- thread->wait_timer_active++;
- thread->wait_timer_is_set = TRUE;
- }
- thread_unlock(thread);
- splx(s);
-}
-
-void
-thread_cancel_timer(void)
-{
- thread_t thread = current_thread();
- spl_t s;
-
- s = splsched();
- thread_lock(thread);
- if (thread->wait_timer_is_set) {
- if (timer_call_cancel(&thread->wait_timer))
- thread->wait_timer_active--;
- thread->wait_timer_is_set = FALSE;
- }
- thread_unlock(thread);
- splx(s);
-}
-
-#endif /* __LP64__ */
-
/*
* thread_unblock:
*
* Unblock thread on wake up.
*
- * Returns TRUE if the thread is still running.
+ * Returns TRUE if the thread should now be placed on the runqueue.
*
* Thread must be locked.
+ *
+ * Called at splsched().
*/
boolean_t
thread_unblock(
- thread_t thread,
- wait_result_t wresult)
+ thread_t thread,
+ wait_result_t wresult)
{
- boolean_t result = FALSE;
+ boolean_t ready_for_runq = FALSE;
+ thread_t cthread = current_thread();
+ uint32_t new_run_count;
+ int old_thread_state;
/*
* Set wait_result.
* Cancel pending wait timer.
*/
if (thread->wait_timer_is_set) {
- if (timer_call_cancel(&thread->wait_timer))
+ if (timer_call_cancel(&thread->wait_timer)) {
thread->wait_timer_active--;
+ }
thread->wait_timer_is_set = FALSE;
}
+ boolean_t aticontext, pidle;
+ ml_get_power_state(&aticontext, &pidle);
+
/*
* Update scheduling state: not waiting,
* set running.
*/
- thread->state &= ~(TH_WAIT|TH_UNINT);
+ old_thread_state = thread->state;
+ thread->state = (old_thread_state | TH_RUN) &
+ ~(TH_WAIT | TH_UNINT | TH_WAIT_REPORT);
- if (!(thread->state & TH_RUN)) {
- thread->state |= TH_RUN;
+ if ((old_thread_state & TH_RUN) == 0) {
+ uint64_t ctime = mach_approximate_time();
+ thread->last_made_runnable_time = thread->last_basepri_change_time = ctime;
+ timer_start(&thread->runnable_timer, ctime);
+
+ ready_for_runq = TRUE;
- (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread);
+ if (old_thread_state & TH_WAIT_REPORT) {
+ (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread);
+ }
+
+ /* Update the runnable thread count */
+ new_run_count = SCHED(run_count_incr)(thread);
+#if CONFIG_SCHED_AUTO_JOIN
+ if (aticontext == FALSE && work_interval_should_propagate(cthread, thread)) {
+ work_interval_auto_join_propagate(cthread, thread);
+ }
+#endif /*CONFIG_SCHED_AUTO_JOIN */
+ } else {
/*
- * Update run counts.
+ * Either the thread is idling in place on another processor,
+ * or it hasn't finished context switching yet.
*/
- sched_run_incr();
- if (thread->sched_mode == TH_MODE_TIMESHARE)
- sched_share_incr();
- }
- else {
+ assert((thread->state & TH_IDLE) == 0);
/*
- * Signal if idling on another processor.
+ * The run count is only dropped after the context switch completes
+ * and the thread is still waiting, so we should not run_incr here
*/
-#if CONFIG_SCHED_IDLE_IN_PLACE
- if (thread->state & TH_IDLE) {
- processor_t processor = thread->last_processor;
-
- if (processor != current_processor())
- machine_signal_idle(processor);
- }
-#else
- assert((thread->state & TH_IDLE) == 0);
-#endif
-
- result = TRUE;
+ new_run_count = os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed);
}
/*
* Calculate deadline for real-time threads.
*/
if (thread->sched_mode == TH_MODE_REALTIME) {
- thread->realtime.deadline = mach_absolute_time();
- thread->realtime.deadline += thread->realtime.constraint;
+ uint64_t ctime;
+
+ ctime = mach_absolute_time();
+ thread->realtime.deadline = thread->realtime.constraint + ctime;
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SET_RT_DEADLINE) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->realtime.deadline, thread->realtime.computation, 0);
}
/*
* Clear old quantum, fail-safe computation, etc.
*/
- thread->current_quantum = 0;
+ thread->quantum_remaining = 0;
thread->computation_metered = 0;
thread->reason = AST_NONE;
+ thread->block_hint = kThreadWaitNone;
- /* Event should only be triggered if thread is not already running */
- if (result == FALSE) {
- 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, 0, 0);
+ /* 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.
+ */
+
+ if (__improbable(aticontext && !(thread_get_tag_internal(thread) & THREAD_TAG_CALLOUT))) {
+ DTRACE_SCHED2(iwakeup, struct thread *, thread, struct proc *, thread->task->bsd_info);
+
+ uint64_t ttd = current_processor()->timer_call_ttd;
+
+ if (ttd) {
+ if (ttd <= timer_deadline_tracking_bin_1) {
+ thread->thread_timer_wakeups_bin_1++;
+ } else if (ttd <= timer_deadline_tracking_bin_2) {
+ thread->thread_timer_wakeups_bin_2++;
+ }
+ }
+
+ ledger_credit_thread(thread, thread->t_ledger,
+ task_ledgers.interrupt_wakeups, 1);
+ if (pidle) {
+ ledger_credit_thread(thread, thread->t_ledger,
+ task_ledgers.platform_idle_wakeups, 1);
+ }
+ } else if (thread_get_tag_internal(cthread) & THREAD_TAG_CALLOUT) {
+ /* TODO: what about an interrupt that does a wake taken on a callout thread? */
+ if (cthread->callout_woken_from_icontext) {
+ ledger_credit_thread(thread, thread->t_ledger,
+ task_ledgers.interrupt_wakeups, 1);
+ thread->thread_callout_interrupt_wakeups++;
+
+ if (cthread->callout_woken_from_platform_idle) {
+ ledger_credit_thread(thread, thread->t_ledger,
+ task_ledgers.platform_idle_wakeups, 1);
+ thread->thread_callout_platform_idle_wakeups++;
+ }
+
+ cthread->callout_woke_thread = TRUE;
+ }
+ }
+
+ if (thread_get_tag_internal(thread) & THREAD_TAG_CALLOUT) {
+ thread->callout_woken_from_icontext = !!aticontext;
+ thread->callout_woken_from_platform_idle = !!pidle;
+ thread->callout_woke_thread = FALSE;
}
+#if KPERF
+ if (ready_for_runq) {
+ kperf_make_runnable(thread, aticontext);
+ }
+#endif /* KPERF */
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_MAKE_RUNNABLE) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->sched_pri, thread->wait_result,
+ sched_run_buckets[TH_BUCKET_RUN], 0);
+
DTRACE_SCHED2(wakeup, struct thread *, thread, struct proc *, thread->task->bsd_info);
- return (result);
+ return ready_for_runq;
+}
+
+/*
+ * Routine: thread_allowed_for_handoff
+ * Purpose:
+ * Check if the thread is allowed for handoff operation
+ * Conditions:
+ * thread lock held, IPC locks may be held.
+ * TODO: In future, do not allow handoff if threads have different cluster
+ * recommendations.
+ */
+boolean_t
+thread_allowed_for_handoff(
+ thread_t thread)
+{
+ thread_t self = current_thread();
+
+ if (allow_direct_handoff &&
+ thread->sched_mode == TH_MODE_REALTIME &&
+ self->sched_mode == TH_MODE_REALTIME) {
+ return TRUE;
+ }
+
+ return FALSE;
}
/*
* Conditions:
* thread lock held, IPC locks may be held.
* thread must have been pulled from wait queue under same lock hold.
- * Returns:
+ * thread must have been waiting
+ * Returns:
* KERN_SUCCESS - Thread was set running
- * KERN_NOT_WAITING - Thread was not waiting
+ *
+ * TODO: This should return void
*/
kern_return_t
thread_go(
- thread_t thread,
- wait_result_t wresult)
+ thread_t thread,
+ wait_result_t wresult,
+ waitq_options_t option)
{
+ thread_t self = current_thread();
+
+ assert_thread_magic(thread);
+
assert(thread->at_safe_point == FALSE);
assert(thread->wait_event == NO_EVENT64);
- assert(thread->wait_queue == WAIT_QUEUE_NULL);
+ assert(thread->waitq == NULL);
+
+ assert(!(thread->state & (TH_TERMINATE | TH_TERMINATE2)));
+ assert(thread->state & TH_WAIT);
- if ((thread->state & (TH_WAIT|TH_TERMINATE)) == TH_WAIT) {
- if (!thread_unblock(thread, wresult))
- thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
- return (KERN_SUCCESS);
+ if (thread_unblock(thread, wresult)) {
+#if SCHED_TRACE_THREAD_WAKEUPS
+ backtrace(&thread->thread_wakeup_bt[0],
+ (sizeof(thread->thread_wakeup_bt) / sizeof(uintptr_t)), NULL);
+#endif
+ if ((option & WQ_OPTION_HANDOFF) &&
+ thread_allowed_for_handoff(thread)) {
+ thread_reference(thread);
+ assert(self->handoff_thread == NULL);
+ self->handoff_thread = thread;
+ } else {
+ thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
+ }
}
- return (KERN_NOT_WAITING);
+ return KERN_SUCCESS;
}
/*
__private_extern__
wait_result_t
thread_mark_wait_locked(
- thread_t thread,
- wait_interrupt_t interruptible)
+ thread_t thread,
+ wait_interrupt_t interruptible_orig)
{
- boolean_t at_safe_point;
+ boolean_t at_safe_point;
+ wait_interrupt_t interruptible = interruptible_orig;
- assert(thread == current_thread());
+ if (thread->state & TH_IDLE) {
+ panic("Invalid attempt to wait while running the idle thread");
+ }
+
+ assert(!(thread->state & (TH_WAIT | TH_IDLE | TH_UNINT | TH_TERMINATE2 | TH_WAIT_REPORT)));
/*
* The thread may have certain types of interrupts/aborts masked
* are OK, we have to honor mask settings (outer-scoped code may
* not be able to handle aborts at the moment).
*/
- if (interruptible > (thread->options & TH_OPT_INTMASK))
+ interruptible &= TH_OPT_INTMASK;
+ if (interruptible > (thread->options & TH_OPT_INTMASK)) {
interruptible = thread->options & TH_OPT_INTMASK;
+ }
at_safe_point = (interruptible == THREAD_ABORTSAFE);
- if ( interruptible == THREAD_UNINT ||
- !(thread->sched_flags & TH_SFLAG_ABORT) ||
- (!at_safe_point &&
- (thread->sched_flags & TH_SFLAG_ABORTSAFELY))) {
-
- if ( !(thread->state & TH_TERMINATE))
+ if (interruptible == THREAD_UNINT ||
+ !(thread->sched_flags & TH_SFLAG_ABORT) ||
+ (!at_safe_point &&
+ (thread->sched_flags & TH_SFLAG_ABORTSAFELY))) {
+ if (!(thread->state & TH_TERMINATE)) {
DTRACE_SCHED(sleep);
+ }
- thread->state |= (interruptible) ? TH_WAIT : (TH_WAIT | TH_UNINT);
+ int state_bits = TH_WAIT;
+ if (!interruptible) {
+ state_bits |= TH_UNINT;
+ }
+ if (thread->sched_call) {
+ wait_interrupt_t mask = THREAD_WAIT_NOREPORT_USER;
+ if (is_kerneltask(thread->task)) {
+ mask = THREAD_WAIT_NOREPORT_KERNEL;
+ }
+ if ((interruptible_orig & mask) == 0) {
+ state_bits |= TH_WAIT_REPORT;
+ }
+ }
+ thread->state |= state_bits;
thread->at_safe_point = at_safe_point;
- return (thread->wait_result = THREAD_WAITING);
+
+ /* TODO: pass this through assert_wait instead, have
+ * assert_wait just take a struct as an argument */
+ assert(!thread->block_hint);
+ thread->block_hint = thread->pending_block_hint;
+ thread->pending_block_hint = kThreadWaitNone;
+
+ return thread->wait_result = THREAD_WAITING;
+ } else {
+ if (thread->sched_flags & TH_SFLAG_ABORTSAFELY) {
+ thread->sched_flags &= ~TH_SFLAG_ABORTED_MASK;
+ }
}
- else
- if (thread->sched_flags & TH_SFLAG_ABORTSAFELY)
- thread->sched_flags &= ~TH_SFLAG_ABORTED_MASK;
+ thread->pending_block_hint = kThreadWaitNone;
- return (thread->wait_result = THREAD_INTERRUPTED);
+ return thread->wait_result = THREAD_INTERRUPTED;
}
/*
* Returns:
* The old interrupt level for the thread.
*/
-__private_extern__
+__private_extern__
wait_interrupt_t
thread_interrupt_level(
wait_interrupt_t new_level)
return result;
}
-/*
- * Check to see if an assert wait is possible, without actually doing one.
- * This is used by debug code in locks and elsewhere to verify that it is
- * always OK to block when trying to take a blocking lock (since waiting
- * for the actual assert_wait to catch the case may make it hard to detect
- * this case.
- */
-boolean_t
-assert_wait_possible(void)
-{
-
- thread_t thread;
-
-#if DEBUG
- if(debug_mode) return TRUE; /* Always succeed in debug mode */
-#endif
-
- thread = current_thread();
-
- return (thread == NULL || wait_queue_assert_possible(thread));
-}
-
/*
* assert_wait:
*
*/
wait_result_t
assert_wait(
- event_t event,
- wait_interrupt_t interruptible)
+ event_t event,
+ wait_interrupt_t interruptible)
{
- register wait_queue_t wq;
- register int index;
-
- assert(event != NO_EVENT);
+ 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(event), 0, 0, 0, 0);
+ 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);
+}
- index = wait_hash(event);
- wq = &wait_queues[index];
- return wait_queue_assert_wait(wq, event, interruptible, 0);
+/*
+ * assert_wait_queue:
+ *
+ * Return the global waitq for the specified event
+ */
+struct waitq *
+assert_wait_queue(
+ event_t event)
+{
+ return global_eventq(event);
}
wait_result_t
assert_wait_timeout(
- event_t event,
- wait_interrupt_t interruptible,
- uint32_t interval,
- uint32_t scale_factor)
+ event_t event,
+ wait_interrupt_t interruptible,
+ uint32_t interval,
+ uint32_t scale_factor)
{
- thread_t thread = current_thread();
- wait_result_t wresult;
- wait_queue_t wqueue;
- uint64_t deadline;
- spl_t s;
+ thread_t thread = current_thread();
+ wait_result_t wresult;
+ uint64_t deadline;
+ spl_t s;
- assert(event != NO_EVENT);
- wqueue = &wait_queues[wait_hash(event)];
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
+ }
+
+ struct waitq *waitq;
+ waitq = global_eventq(event);
s = splsched();
- wait_queue_lock(wqueue);
- thread_lock(thread);
+ waitq_lock(waitq);
clock_interval_to_deadline(interval, scale_factor, &deadline);
-
+
KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
- MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT)|DBG_FUNC_NONE,
- VM_KERNEL_UNSLIDE(event), interruptible, deadline, 0, 0);
-
- wresult = wait_queue_assert_wait64_locked(wqueue, CAST_DOWN(event64_t, event),
- interruptible, deadline, thread);
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE,
+ VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0);
- thread_unlock(thread);
- wait_queue_unlock(wqueue);
- splx(s);
+ wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event),
+ interruptible,
+ TIMEOUT_URGENCY_SYS_NORMAL,
+ deadline, TIMEOUT_NO_LEEWAY,
+ thread);
- return (wresult);
+ waitq_unlock(waitq);
+ splx(s);
+ return wresult;
}
wait_result_t
-assert_wait_deadline(
- event_t event,
- wait_interrupt_t interruptible,
- uint64_t deadline)
+assert_wait_timeout_with_leeway(
+ event_t event,
+ wait_interrupt_t interruptible,
+ wait_timeout_urgency_t urgency,
+ uint32_t interval,
+ uint32_t leeway,
+ uint32_t scale_factor)
{
- thread_t thread = current_thread();
- wait_result_t wresult;
- wait_queue_t wqueue;
- spl_t s;
+ thread_t thread = current_thread();
+ wait_result_t wresult;
+ uint64_t deadline;
+ uint64_t abstime;
+ uint64_t slop;
+ uint64_t now;
+ spl_t s;
+
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
+ }
+
+ now = mach_absolute_time();
+ clock_interval_to_absolutetime_interval(interval, scale_factor, &abstime);
+ deadline = now + abstime;
- assert(event != NO_EVENT);
- wqueue = &wait_queues[wait_hash(event)];
+ clock_interval_to_absolutetime_interval(leeway, scale_factor, &slop);
+
+ struct waitq *waitq;
+ waitq = global_eventq(event);
s = splsched();
- wait_queue_lock(wqueue);
- thread_lock(thread);
+ waitq_lock(waitq);
KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
- MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT)|DBG_FUNC_NONE,
- VM_KERNEL_UNSLIDE(event), interruptible, deadline, 0, 0);
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE,
+ VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0);
- wresult = wait_queue_assert_wait64_locked(wqueue, CAST_DOWN(event64_t,event),
- interruptible, deadline, thread);
+ wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event),
+ interruptible,
+ urgency, deadline, slop,
+ thread);
- thread_unlock(thread);
- wait_queue_unlock(wqueue);
+ waitq_unlock(waitq);
splx(s);
-
- return (wresult);
+ return wresult;
}
-/*
- * thread_sleep_fast_usimple_lock:
- *
- * Cause the current thread to wait until the specified event
- * occurs. The specified simple_lock is unlocked before releasing
- * the cpu and re-acquired as part of waking up.
- *
- * This is the simple lock sleep interface for components that use a
- * faster version of simple_lock() than is provided by usimple_lock().
- */
-__private_extern__ wait_result_t
-thread_sleep_fast_usimple_lock(
- event_t event,
- simple_lock_t lock,
- wait_interrupt_t interruptible)
+wait_result_t
+assert_wait_deadline(
+ event_t event,
+ wait_interrupt_t interruptible,
+ uint64_t deadline)
{
- wait_result_t res;
+ thread_t thread = current_thread();
+ wait_result_t wresult;
+ spl_t s;
- res = assert_wait(event, interruptible);
- if (res == THREAD_WAITING) {
- simple_unlock(lock);
- res = thread_block(THREAD_CONTINUE_NULL);
- simple_lock(lock);
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
}
- return res;
-}
+ struct waitq *waitq;
+ waitq = global_eventq(event);
+
+ s = splsched();
+ waitq_lock(waitq);
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE,
+ VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0);
+
+ wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event),
+ interruptible,
+ TIMEOUT_URGENCY_SYS_NORMAL, deadline,
+ TIMEOUT_NO_LEEWAY, thread);
+ waitq_unlock(waitq);
+ splx(s);
+ return wresult;
+}
-/*
- * thread_sleep_usimple_lock:
- *
- * Cause the current thread to wait until the specified event
- * occurs. The specified usimple_lock is unlocked before releasing
- * the cpu and re-acquired as part of waking up.
- *
- * This is the simple lock sleep interface for components where
- * simple_lock() is defined in terms of usimple_lock().
- */
wait_result_t
-thread_sleep_usimple_lock(
- event_t event,
- usimple_lock_t lock,
- wait_interrupt_t interruptible)
+assert_wait_deadline_with_leeway(
+ event_t event,
+ wait_interrupt_t interruptible,
+ wait_timeout_urgency_t urgency,
+ uint64_t deadline,
+ uint64_t leeway)
{
- wait_result_t res;
+ thread_t thread = current_thread();
+ wait_result_t wresult;
+ spl_t s;
- res = assert_wait(event, interruptible);
- if (res == THREAD_WAITING) {
- usimple_unlock(lock);
- res = thread_block(THREAD_CONTINUE_NULL);
- usimple_lock(lock);
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
}
- return res;
+
+ struct waitq *waitq;
+ waitq = global_eventq(event);
+
+ s = splsched();
+ waitq_lock(waitq);
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAIT) | DBG_FUNC_NONE,
+ VM_KERNEL_UNSLIDE_OR_PERM(event), interruptible, deadline, 0, 0);
+
+ wresult = waitq_assert_wait64_locked(waitq, CAST_EVENT64_T(event),
+ interruptible,
+ urgency, deadline, leeway,
+ thread);
+ waitq_unlock(waitq);
+ splx(s);
+ return wresult;
}
/*
- * thread_sleep_lock_write:
+ * thread_isoncpu:
+ *
+ * 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.
*
- * Cause the current thread to wait until the specified event
- * occurs. The specified (write) lock is unlocked before releasing
- * the cpu. The (write) lock will be re-acquired before returning.
+ * Thread locked, returns the same way. While locked, fields
+ * like "state" cannot change. "runq" can change only from set to unset.
*/
-wait_result_t
-thread_sleep_lock_write(
- event_t event,
- lock_t *lock,
- wait_interrupt_t interruptible)
+static inline boolean_t
+thread_isoncpu(thread_t thread)
{
- wait_result_t res;
+ /* Not running or runnable */
+ if (!(thread->state & TH_RUN)) {
+ return FALSE;
+ }
+
+ /* 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;
+ }
- res = assert_wait(event, interruptible);
- if (res == THREAD_WAITING) {
- lock_write_done(lock);
- res = thread_block(THREAD_CONTINUE_NULL);
- lock_write(lock);
+ /*
+ * 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;
}
- return res;
+
+ /*
+ * Thread must be running on a processor, or
+ * about to run, or just did run. In all these
+ * cases, an AST to the processor is needed
+ * to guarantee that the thread is kicked out
+ * of userspace and the processor has
+ * context switched (and saved register state).
+ */
+ return TRUE;
}
/*
* thread_stop:
*
* Force a preemption point for a thread and wait
- * for it to stop running. Arbitrates access among
+ * for it to stop running on a CPU. If a stronger
+ * guarantee is requested, wait until no longer
+ * runnable. Arbitrates access among
* multiple stop requests. (released by unstop)
*
* The thread must enter a wait state and stop via a
*/
boolean_t
thread_stop(
- thread_t thread)
+ thread_t thread,
+ boolean_t until_not_runnable)
{
- wait_result_t wresult;
- spl_t s = splsched();
+ wait_result_t wresult;
+ spl_t s = splsched();
+ boolean_t oncpu;
wake_lock(thread);
thread_lock(thread);
wake_unlock(thread);
splx(s);
- if (wresult == THREAD_WAITING)
+ if (wresult == THREAD_WAITING) {
wresult = thread_block(THREAD_CONTINUE_NULL);
+ }
- if (wresult != THREAD_AWAKENED)
- return (FALSE);
+ if (wresult != THREAD_AWAKENED) {
+ return FALSE;
+ }
s = splsched();
wake_lock(thread);
thread->state |= TH_SUSP;
- while (thread->state & TH_RUN) {
- processor_t processor = thread->last_processor;
+ while ((oncpu = thread_isoncpu(thread)) ||
+ (until_not_runnable && (thread->state & TH_RUN))) {
+ processor_t processor;
- if (processor != PROCESSOR_NULL && processor->active_thread == thread)
+ if (oncpu) {
+ assert(thread->state & TH_RUN);
+ processor = thread->chosen_processor;
cause_ast_check(processor);
+ }
thread->wake_active = TRUE;
thread_unlock(thread);
wake_unlock(thread);
splx(s);
- if (wresult == THREAD_WAITING)
+ if (wresult == THREAD_WAITING) {
wresult = thread_block(THREAD_CONTINUE_NULL);
+ }
if (wresult != THREAD_AWAKENED) {
thread_unstop(thread);
- return (FALSE);
+ return FALSE;
}
s = splsched();
wake_unlock(thread);
splx(s);
- return (TRUE);
+ /*
+ * We return with the thread unlocked. To prevent it from
+ * transitioning to a runnable state (or from TH_RUN to
+ * being on the CPU), the caller must ensure the thread
+ * is stopped via an external means (such as an AST)
+ */
+
+ return TRUE;
}
/*
*/
void
thread_unstop(
- thread_t thread)
+ thread_t thread)
{
- spl_t s = splsched();
+ spl_t s = splsched();
wake_lock(thread);
thread_lock(thread);
- if ((thread->state & (TH_RUN|TH_WAIT|TH_SUSP)) == TH_SUSP) {
- thread->state &= ~TH_SUSP;
- thread_unblock(thread, THREAD_AWAKENED);
+ assert((thread->state & (TH_RUN | TH_WAIT | TH_SUSP)) != TH_SUSP);
- thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
- }
- else
if (thread->state & TH_SUSP) {
thread->state &= ~TH_SUSP;
splx(s);
}
-/*
- * Thread locked, returns the same way
- */
-static inline boolean_t
-thread_isoncpu(thread_t thread)
-{
- processor_t processor = thread->last_processor;
-
- return ((processor != PROCESSOR_NULL) && (processor->active_thread == thread));
-}
/*
* thread_wait:
*
*/
void
thread_wait(
- thread_t thread,
- boolean_t until_not_runnable)
+ thread_t thread,
+ boolean_t until_not_runnable)
{
- wait_result_t wresult;
- boolean_t oncpu;
- processor_t processor;
- spl_t s = splsched();
+ wait_result_t wresult;
+ boolean_t oncpu;
+ processor_t processor;
+ spl_t 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
+ * 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 ((oncpu = thread_isoncpu(thread)) ||
+ (until_not_runnable && (thread->state & TH_RUN))) {
if (oncpu) {
assert(thread->state & TH_RUN);
- processor = thread->last_processor;
+ processor = thread->chosen_processor;
cause_ast_check(processor);
}
wake_unlock(thread);
splx(s);
- if (wresult == THREAD_WAITING)
+ if (wresult == THREAD_WAITING) {
thread_block(THREAD_CONTINUE_NULL);
+ }
s = splsched();
wake_lock(thread);
*/
__private_extern__ kern_return_t
clear_wait_internal(
- thread_t thread,
- wait_result_t wresult)
+ thread_t thread,
+ wait_result_t wresult)
{
- wait_queue_t wq = thread->wait_queue;
- uint32_t i = LockTimeOut;
+ uint32_t i = LockTimeOutUsec;
+ struct waitq *waitq = thread->waitq;
do {
- if (wresult == THREAD_INTERRUPTED && (thread->state & TH_UNINT))
- return (KERN_FAILURE);
+ if (wresult == THREAD_INTERRUPTED && (thread->state & TH_UNINT)) {
+ return KERN_FAILURE;
+ }
- if (wq != WAIT_QUEUE_NULL) {
- if (wait_queue_lock_try(wq)) {
- wait_queue_pull_thread_locked(wq, thread, TRUE);
- /* wait queue unlocked, thread still locked */
- }
- else {
+ if (waitq != NULL) {
+ if (!waitq_pull_thread_locked(waitq, thread)) {
thread_unlock(thread);
delay(1);
-
+ if (i > 0 && !machine_timeout_suspended()) {
+ i--;
+ }
thread_lock(thread);
- if (wq != thread->wait_queue)
- return (KERN_NOT_WAITING);
-
+ if (waitq != thread->waitq) {
+ return KERN_NOT_WAITING;
+ }
continue;
}
}
- return (thread_go(thread, wresult));
- } while ((--i > 0) || machine_timeout_suspended());
+ /* TODO: Can we instead assert TH_TERMINATE is not set? */
+ if ((thread->state & (TH_WAIT | TH_TERMINATE)) == TH_WAIT) {
+ return thread_go(thread, wresult, WQ_OPTION_NONE);
+ } else {
+ return KERN_NOT_WAITING;
+ }
+ } while (i > 0);
panic("clear_wait_internal: deadlock: thread=%p, wq=%p, cpu=%d\n",
- thread, wq, cpu_number());
+ thread, waitq, cpu_number());
- return (KERN_FAILURE);
+ return KERN_FAILURE;
}
*/
kern_return_t
clear_wait(
- thread_t thread,
- wait_result_t result)
+ thread_t thread,
+ wait_result_t result)
{
kern_return_t ret;
- spl_t s;
+ spl_t s;
s = splsched();
thread_lock(thread);
*/
kern_return_t
thread_wakeup_prim(
- event_t event,
- boolean_t one_thread,
- wait_result_t result)
+ event_t event,
+ boolean_t one_thread,
+ wait_result_t result)
{
- return (thread_wakeup_prim_internal(event, one_thread, result, -1));
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
+ }
+
+ struct waitq *wq = global_eventq(event);
+
+ if (one_thread) {
+ return waitq_wakeup64_one(wq, CAST_EVENT64_T(event), result, WAITQ_ALL_PRIORITIES);
+ } else {
+ return waitq_wakeup64_all(wq, CAST_EVENT64_T(event), result, WAITQ_ALL_PRIORITIES);
+ }
}
+/*
+ * Wakeup a specified thread if and only if it's waiting for this event
+ */
+kern_return_t
+thread_wakeup_thread(
+ event_t event,
+ thread_t thread)
+{
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
+ }
+
+ if (__improbable(thread == THREAD_NULL)) {
+ panic("%s() called with THREAD_NULL", __func__);
+ }
+
+ struct waitq *wq = global_eventq(event);
+
+ return waitq_wakeup64_thread(wq, CAST_EVENT64_T(event), thread, THREAD_AWAKENED);
+}
+/*
+ * Wakeup a thread waiting on an event and promote it to a priority.
+ *
+ * Requires woken thread to un-promote itself when done.
+ */
kern_return_t
-thread_wakeup_prim_internal(
- event_t event,
- boolean_t one_thread,
- wait_result_t result,
- int priority)
+thread_wakeup_one_with_pri(
+ event_t event,
+ int priority)
+{
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
+ }
+
+ struct waitq *wq = global_eventq(event);
+
+ return waitq_wakeup64_one(wq, CAST_EVENT64_T(event), THREAD_AWAKENED, priority);
+}
+
+/*
+ * Wakeup a thread waiting on an event,
+ * promote it to a priority,
+ * and return a reference to the woken thread.
+ *
+ * Requires woken thread to un-promote itself when done.
+ */
+thread_t
+thread_wakeup_identify(event_t event,
+ int priority)
{
- register wait_queue_t wq;
- register int index;
+ if (__improbable(event == NO_EVENT)) {
+ panic("%s() called with NO_EVENT", __func__);
+ }
- index = wait_hash(event);
- wq = &wait_queues[index];
- if (one_thread)
- return (wait_queue_wakeup_one(wq, event, result, priority));
- else
- return (wait_queue_wakeup_all(wq, event, result));
+ struct waitq *wq = global_eventq(event);
+
+ return waitq_wakeup64_identify(wq, CAST_EVENT64_T(event), THREAD_AWAKENED, priority);
}
/*
* thread_bind:
*
* Force the current thread to execute on the specified processor.
+ * Takes effect after the next thread_block().
*
* Returns the previous binding. PROCESSOR_NULL means
* not bound.
*/
processor_t
thread_bind(
- processor_t processor)
+ processor_t processor)
{
- thread_t self = current_thread();
- processor_t prev;
- spl_t s;
+ thread_t self = current_thread();
+ processor_t prev;
+ spl_t s;
s = splsched();
thread_lock(self);
- prev = self->bound_processor;
- self->bound_processor = processor;
+ prev = thread_bind_internal(self, processor);
thread_unlock(self);
splx(s);
- return (prev);
+ return prev;
}
/*
- * thread_select:
+ * thread_bind_internal:
*
- * Select a new thread for the current processor to execute.
+ * If the specified thread is not the current thread, and it is currently
+ * running on another CPU, a remote AST must be sent to that CPU to cause
+ * the thread to migrate to its bound processor. Otherwise, the migration
+ * will occur at the next quantum expiration or blocking point.
*
- * May select the current thread, which must be locked.
+ * When the thread is the current thread, and explicit thread_block() should
+ * be used to force the current processor to context switch away and
+ * let the thread migrate to the bound processor.
+ *
+ * Thread must be locked, and at splsched.
*/
-static thread_t
-thread_select(
- thread_t thread,
- processor_t processor)
+
+static processor_t
+thread_bind_internal(
+ thread_t thread,
+ processor_t processor)
{
- processor_set_t pset = processor->processor_set;
- thread_t new_thread = THREAD_NULL;
- boolean_t inactive_state;
+ processor_t prev;
- assert(processor == current_processor());
+ /* <rdar://problem/15102234> */
+ 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);
- do {
- /*
- * Update the priority.
- */
- if (SCHED(can_update_priority)(thread))
- SCHED(update_priority)(thread);
-
- processor->current_pri = thread->sched_pri;
- processor->current_thmode = thread->sched_mode;
+ KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_THREAD_BIND), thread_tid(thread), processor ? (uintptr_t)processor->cpu_id : (uintptr_t)-1, 0, 0, 0);
- pset_lock(pset);
+ prev = thread->bound_processor;
+ thread->bound_processor = processor;
- assert(pset->low_count);
- assert(pset->low_pri);
+ return prev;
+}
- inactive_state = processor->state != PROCESSOR_SHUTDOWN && machine_processor_is_inactive(processor);
+/*
+ * 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).
+ */
- simple_lock(&rt_lock);
+/*
+ * 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;
- /*
- * Test to see if the current thread should continue
- * to run on this processor. Must be runnable, and not
- * bound to a different processor, nor be in the wrong
- * processor set.
- */
- if ( ((thread->state & ~TH_SUSP) == TH_RUN) &&
- (thread->sched_pri >= BASEPRI_RTQUEUES ||
- processor->processor_meta == PROCESSOR_META_NULL ||
- processor->processor_meta->primary == processor) &&
- (thread->bound_processor == PROCESSOR_NULL ||
- thread->bound_processor == processor) &&
- (thread->affinity_set == AFFINITY_SET_NULL ||
- thread->affinity_set->aset_pset == pset) ) {
- if ( thread->sched_pri >= BASEPRI_RTQUEUES &&
- first_timeslice(processor) ) {
- if (rt_runq.count > 0) {
- register queue_t q;
-
- q = &rt_runq.queue;
- if (((thread_t)q->next)->realtime.deadline <
- processor->deadline) {
- thread = (thread_t)dequeue_head(q);
- thread->runq = PROCESSOR_NULL;
- SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count);
- rt_runq.count--;
- }
- }
+void
+thread_vm_bind_group_add(void)
+{
+ thread_t self = current_thread();
- simple_unlock(&rt_lock);
+ thread_reference_internal(self);
+ self->options |= TH_OPT_SCHED_VM_GROUP;
- processor->deadline = thread->realtime.deadline;
+ simple_lock(&sched_vm_group_list_lock, LCK_GRP_NULL);
+ assert(sched_vm_group_thread_count < MAX_VM_BIND_GROUP_COUNT);
+ sched_vm_group_thread_list[sched_vm_group_thread_count++] = self;
+ simple_unlock(&sched_vm_group_list_lock);
- pset_unlock(pset);
+ thread_bind(master_processor);
- return (thread);
- }
+ /* 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, LCK_GRP_NULL);
- if (!inactive_state && (thread->sched_mode != TH_MODE_FAIRSHARE || SCHED(fairshare_runq_count)() == 0) && (rt_runq.count == 0 || BASEPRI_RTQUEUES < thread->sched_pri) &&
- (new_thread = SCHED(choose_thread)(processor, thread->sched_mode == TH_MODE_FAIRSHARE ? MINPRI : thread->sched_pri)) == THREAD_NULL) {
+ s = splsched();
- simple_unlock(&rt_lock);
+ for (i = 0; i < sched_vm_group_thread_count; i++) {
+ thread_t thread = sched_vm_group_thread_list[i];
+ assert(thread != THREAD_NULL);
+ thread_lock(thread);
+ if ((thread->state & (TH_RUN | TH_WAIT)) == TH_RUN) {
+ if (thread->runq != PROCESSOR_NULL && thread->last_made_runnable_time < longtime) {
+ high_latency_observed = TRUE;
+ } else if (thread->runq == PROCESSOR_NULL) {
+ /* There are some cases where a thread be transitiong that also fall into this case */
+ runnable_and_not_on_runq_observed = TRUE;
+ }
+ }
+ thread_unlock(thread);
- /* I am the highest priority runnable (non-idle) thread */
+ if (high_latency_observed && runnable_and_not_on_runq_observed) {
+ /* All the things we are looking for are true, stop looking */
+ break;
+ }
+ }
- pset_pri_hint(pset, processor, processor->current_pri);
+ splx(s);
- pset_count_hint(pset, processor, SCHED(processor_runq_count)(processor));
+ if (sched_vm_group_temporarily_unbound) {
+ /* If we turned off binding, make sure everything is OK before rebinding */
+ if (!high_latency_observed) {
+ /* rebind */
+ bind_target_changed = TRUE;
+ bind_target = master_processor;
+ sched_vm_group_temporarily_unbound = FALSE; /* might be reset to TRUE if change cannot be completed */
+ }
+ } else {
+ /*
+ * Check if we're in a bad state, which is defined by high
+ * latency with no core currently executing a thread. If a
+ * single thread is making progress on a CPU, that means the
+ * binding concept to reduce parallelism is working as
+ * designed.
+ */
+ if (high_latency_observed && !runnable_and_not_on_runq_observed) {
+ /* unbind */
+ bind_target_changed = TRUE;
+ bind_target = PROCESSOR_NULL;
+ sched_vm_group_temporarily_unbound = TRUE;
+ }
+ }
- processor->deadline = UINT64_MAX;
+ if (bind_target_changed) {
+ s = splsched();
+ for (i = 0; i < sched_vm_group_thread_count; i++) {
+ thread_t thread = sched_vm_group_thread_list[i];
+ boolean_t removed;
+ assert(thread != THREAD_NULL);
- pset_unlock(pset);
+ thread_lock(thread);
+ removed = thread_run_queue_remove(thread);
+ if (removed || ((thread->state & (TH_RUN | TH_WAIT)) == TH_WAIT)) {
+ thread_bind_internal(thread, bind_target);
+ } else {
+ /*
+ * Thread was in the middle of being context-switched-to,
+ * or was in the process of blocking. To avoid switching the bind
+ * state out mid-flight, defer the change if possible.
+ */
+ if (bind_target == PROCESSOR_NULL) {
+ thread_bind_internal(thread, bind_target);
+ } else {
+ sched_vm_group_temporarily_unbound = TRUE; /* next pass will try again */
+ }
+ }
- return (thread);
+ if (removed) {
+ thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ);
}
+ thread_unlock(thread);
}
+ splx(s);
+ }
- if (new_thread != THREAD_NULL ||
- (SCHED(processor_queue_has_priority)(processor, rt_runq.count == 0 ? IDLEPRI : BASEPRI_RTQUEUES, TRUE) &&
- (new_thread = SCHED(choose_thread)(processor, MINPRI)) != THREAD_NULL)) {
- simple_unlock(&rt_lock);
-
- if (!inactive_state) {
- pset_pri_hint(pset, processor, new_thread->sched_pri);
+ simple_unlock(&sched_vm_group_list_lock);
+}
- pset_count_hint(pset, processor, SCHED(processor_runq_count)(processor));
- }
+/* 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.
+ */
- processor->deadline = UINT64_MAX;
- pset_unlock(pset);
+#if (DEVELOPMENT || DEBUG)
+int sched_smt_balance = 1;
+#endif
- return (new_thread);
- }
+/* Invoked with pset locked, returns with pset unlocked */
+void
+sched_SMT_balance(processor_t cprocessor, processor_set_t cpset)
+{
+ processor_t ast_processor = NULL;
- if (rt_runq.count > 0) {
- thread = (thread_t)dequeue_head(&rt_runq.queue);
+#if (DEVELOPMENT || DEBUG)
+ if (__improbable(sched_smt_balance == 0)) {
+ goto smt_balance_exit;
+ }
+#endif
- thread->runq = PROCESSOR_NULL;
- SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count);
- rt_runq.count--;
+ assert(cprocessor == current_processor());
+ if (cprocessor->is_SMT == FALSE) {
+ goto smt_balance_exit;
+ }
- simple_unlock(&rt_lock);
+ processor_t sib_processor = cprocessor->processor_secondary ? cprocessor->processor_secondary : cprocessor->processor_primary;
- processor->deadline = thread->realtime.deadline;
- pset_unlock(pset);
+ /* Determine if both this processor and its sibling are idle,
+ * indicating an SMT rebalancing opportunity.
+ */
+ if (sib_processor->state != PROCESSOR_IDLE) {
+ goto smt_balance_exit;
+ }
- return (thread);
+ processor_t sprocessor;
+
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ uint64_t running_secondary_map = (cpset->cpu_state_map[PROCESSOR_RUNNING] &
+ ~cpset->primary_map);
+ for (int cpuid = lsb_first(running_secondary_map); cpuid >= 0; cpuid = lsb_next(running_secondary_map, cpuid)) {
+ sprocessor = processor_array[cpuid];
+ if ((sprocessor->processor_primary->state == PROCESSOR_RUNNING) &&
+ (sprocessor->current_pri < BASEPRI_RTQUEUES)) {
+ ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL);
+ if (ipi_type != SCHED_IPI_NONE) {
+ assert(sprocessor != cprocessor);
+ ast_processor = sprocessor;
+ break;
+ }
}
+ }
- simple_unlock(&rt_lock);
+smt_balance_exit:
+ pset_unlock(cpset);
- /* No realtime threads and no normal threads on the per-processor
- * runqueue. Finally check for global fairshare threads.
- */
- if ((new_thread = SCHED(fairshare_dequeue)()) != THREAD_NULL) {
+ if (ast_processor) {
+ KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_SMT_BALANCE), ast_processor->cpu_id, ast_processor->state, ast_processor->processor_primary->state, 0, 0);
+ sched_ipi_perform(ast_processor, ipi_type);
+ }
+}
- processor->deadline = UINT64_MAX;
- pset_unlock(pset);
-
- return (new_thread);
- }
-
- processor->deadline = UINT64_MAX;
+static cpumap_t
+pset_available_cpumap(processor_set_t pset)
+{
+ return (pset->cpu_state_map[PROCESSOR_IDLE] | pset->cpu_state_map[PROCESSOR_DISPATCHING] | pset->cpu_state_map[PROCESSOR_RUNNING]) &
+ pset->recommended_bitmask;
+}
- /*
- * Set processor inactive based on
- * indication from the platform code.
- */
- if (inactive_state) {
- if (processor->state == PROCESSOR_RUNNING)
- remqueue((queue_entry_t)processor);
- else
- if (processor->state == PROCESSOR_IDLE)
- remqueue((queue_entry_t)processor);
+static cpumap_t
+pset_available_but_not_running_cpumap(processor_set_t pset)
+{
+ return (pset->cpu_state_map[PROCESSOR_IDLE] | pset->cpu_state_map[PROCESSOR_DISPATCHING]) &
+ pset->recommended_bitmask;
+}
- processor->state = PROCESSOR_INACTIVE;
+bool
+pset_has_stealable_threads(processor_set_t pset)
+{
+ pset_assert_locked(pset);
- pset_unlock(pset);
+ cpumap_t avail_map = pset_available_but_not_running_cpumap(pset);
+ /*
+ * Secondary CPUs never steal, so allow stealing of threads if there are more threads than
+ * available primary CPUs
+ */
+ avail_map &= pset->primary_map;
- return (processor->idle_thread);
- }
+ return (pset->pset_runq.count > 0) && ((pset->pset_runq.count + rt_runq_count(pset)) > bit_count(avail_map));
+}
- /*
- * No runnable threads, attempt to steal
- * from other processors.
- */
- new_thread = SCHED(steal_thread)(pset);
- if (new_thread != THREAD_NULL) {
- return (new_thread);
- }
+/*
+ * Called with pset locked, on a processor that is committing to run a new thread
+ * Will transition an idle or dispatching processor to running as it picks up
+ * the first new thread from the idle thread.
+ */
+static void
+pset_commit_processor_to_new_thread(processor_set_t pset, processor_t processor, thread_t new_thread)
+{
+ pset_assert_locked(pset);
+
+ if (processor->state == PROCESSOR_DISPATCHING || processor->state == PROCESSOR_IDLE) {
+ assert(current_thread() == processor->idle_thread);
/*
- * If other threads have appeared, shortcut
- * around again.
+ * Dispatching processor is now committed to running new_thread,
+ * so change its state to PROCESSOR_RUNNING.
*/
- if (!SCHED(processor_queue_empty)(processor) || rt_runq.count > 0 || SCHED(fairshare_runq_count)() > 0)
- continue;
+ pset_update_processor_state(pset, processor, PROCESSOR_RUNNING);
+ } else {
+ assert((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_SHUTDOWN));
+ }
- pset_lock(pset);
+ processor_state_update_from_thread(processor, new_thread);
- /*
- * Nothing is runnable, so set this processor idle if it
- * was running.
- */
- if (processor->state == PROCESSOR_RUNNING) {
- remqueue((queue_entry_t)processor);
- processor->state = PROCESSOR_IDLE;
-
- if (processor->processor_meta == PROCESSOR_META_NULL || processor->processor_meta->primary == processor) {
- enqueue_head(&pset->idle_queue, (queue_entry_t)processor);
- pset_pri_init_hint(pset, processor);
- pset_count_init_hint(pset, processor);
+ if (new_thread->sched_pri >= BASEPRI_RTQUEUES) {
+ bit_set(pset->realtime_map, processor->cpu_id);
+ } else {
+ bit_clear(pset->realtime_map, processor->cpu_id);
+ }
+
+ pset_node_t node = pset->node;
+
+ if (bit_count(node->pset_map) == 1) {
+ /* Node has only a single pset, so skip node pset map updates */
+ return;
+ }
+
+ cpumap_t avail_map = pset_available_cpumap(pset);
+
+ if (new_thread->sched_pri >= BASEPRI_RTQUEUES) {
+ if ((avail_map & pset->realtime_map) == avail_map) {
+ /* No more non-RT CPUs in this pset */
+ atomic_bit_clear(&node->pset_non_rt_map, pset->pset_id, memory_order_relaxed);
+ }
+ avail_map &= pset->primary_map;
+ if ((avail_map & pset->realtime_map) == avail_map) {
+ /* No more non-RT primary CPUs in this pset */
+ atomic_bit_clear(&node->pset_non_rt_primary_map, pset->pset_id, memory_order_relaxed);
+ }
+ } else {
+ if ((avail_map & pset->realtime_map) != avail_map) {
+ if (!bit_test(atomic_load(&node->pset_non_rt_map), pset->pset_id)) {
+ atomic_bit_set(&node->pset_non_rt_map, pset->pset_id, memory_order_relaxed);
}
- else {
- enqueue_head(&processor->processor_meta->idle_queue, (queue_entry_t)processor);
- pset_unlock(pset);
- return (processor->idle_thread);
+ }
+ avail_map &= pset->primary_map;
+ if ((avail_map & pset->realtime_map) != avail_map) {
+ if (!bit_test(atomic_load(&node->pset_non_rt_primary_map), pset->pset_id)) {
+ atomic_bit_set(&node->pset_non_rt_primary_map, pset->pset_id, memory_order_relaxed);
}
}
+ }
+}
- pset_unlock(pset);
+static processor_t choose_processor_for_realtime_thread(processor_set_t pset, processor_t skip_processor, bool consider_secondaries);
+static bool all_available_primaries_are_running_realtime_threads(processor_set_t pset);
+#if defined(__x86_64__)
+static bool these_processors_are_running_realtime_threads(processor_set_t pset, uint64_t these_map);
+#endif
+static bool sched_ok_to_run_realtime_thread(processor_set_t pset, processor_t processor);
+static bool processor_is_fast_track_candidate_for_realtime_thread(processor_set_t pset, processor_t processor);
+int sched_allow_rt_smt = 1;
+int sched_avoid_cpu0 = 1;
+
+/*
+ * 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;
-#if CONFIG_SCHED_IDLE_IN_PLACE
+ assert(processor == current_processor());
+ assert((thread->state & (TH_RUN | TH_TERMINATE2)) == TH_RUN);
+
+ do {
/*
- * Choose idle thread if fast idle is not possible.
+ * Update the priority.
*/
- 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);
+ if (SCHED(can_update_priority)(thread)) {
+ SCHED(update_priority)(thread);
+ }
+
+ pset_lock(pset);
+
+ processor_state_update_from_thread(processor, thread);
+
+restart:
+ /* Acknowledge any pending IPIs here with pset lock held */
+ bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ bit_clear(pset->pending_deferred_AST_cpu_mask, processor->cpu_id);
+#endif
+
+ bool secondary_can_only_run_realtime_thread = false;
+
+ assert(processor->state != PROCESSOR_OFF_LINE);
+
+ if (!processor->is_recommended) {
+ /*
+ * The performance controller has provided a hint to not dispatch more threads,
+ * unless they are bound to us (and thus we are the only option
+ */
+ if (!SCHED(processor_bound_count)(processor)) {
+ goto idle;
+ }
+ } else if (processor->processor_primary != processor) {
+ /*
+ * Should this secondary SMT processor attempt to find work? For pset runqueue systems,
+ * we should look for work only under the same conditions that choose_processor()
+ * would have assigned work, which is when all primary processors have been assigned work.
+ *
+ * An exception is that bound threads are dispatched to a processor without going through
+ * choose_processor(), so in those cases we should continue trying to dequeue work.
+ */
+ if (!SCHED(processor_bound_count)(processor)) {
+ if ((pset->recommended_bitmask & pset->primary_map & pset->cpu_state_map[PROCESSOR_IDLE]) != 0) {
+ goto idle;
+ }
+
+ /*
+ * TODO: What if a secondary core beat an idle primary to waking up from an IPI?
+ * Should it dequeue immediately, or spin waiting for the primary to wake up?
+ */
+
+ /* There are no idle primaries */
+
+ if (processor->processor_primary->current_pri >= BASEPRI_RTQUEUES) {
+ bool secondary_can_run_realtime_thread = sched_allow_rt_smt && rt_runq_count(pset) && all_available_primaries_are_running_realtime_threads(pset);
+ if (!secondary_can_run_realtime_thread) {
+ goto idle;
+ }
+ secondary_can_only_run_realtime_thread = true;
+ }
+ }
+ }
/*
- * Perform idling activities directly without a
- * context switch. Return dispatched thread,
- * else check again for a runnable thread.
+ * 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 = thread_select_idle(thread, processor);
-#else /* !CONFIG_SCHED_IDLE_IN_PLACE */
-
+ /* i.e. not waiting, not TH_SUSP'ed */
+ bool still_running = ((thread->state & (TH_TERMINATE | TH_IDLE | TH_WAIT | TH_RUN | TH_SUSP)) == TH_RUN);
+
/*
- * 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.
+ * Threads running on SMT processors are forced to context switch. Don't rebalance realtime threads.
+ * TODO: This should check if it's worth it to rebalance, i.e. 'are there any idle primary processors'
+ * <rdar://problem/47907700>
+ *
+ * A yielding thread shouldn't be forced to context switch.
*/
- return (processor->idle_thread);
-#endif /* !CONFIG_SCHED_IDLE_IN_PLACE */
+ bool is_yielding = (*reason & AST_YIELD) == AST_YIELD;
- } while (new_thread == THREAD_NULL);
+ bool needs_smt_rebalance = !is_yielding && thread->sched_pri < BASEPRI_RTQUEUES && processor->processor_primary != processor;
- return (new_thread);
-}
+ bool affinity_mismatch = thread->affinity_set != AFFINITY_SET_NULL && thread->affinity_set->aset_pset != pset;
-#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;
+ bool bound_elsewhere = thread->bound_processor != PROCESSOR_NULL && thread->bound_processor != processor;
- if (thread->sched_mode == TH_MODE_TIMESHARE)
- sched_share_decr();
- sched_run_decr();
+ bool avoid_processor = !is_yielding && SCHED(avoid_processor_enabled) && SCHED(thread_avoid_processor)(processor, thread);
- thread->state |= TH_IDLE;
- processor->current_pri = IDLEPRI;
- processor->current_thmode = TH_MODE_NONE;
+ if (still_running && !needs_smt_rebalance && !affinity_mismatch && !bound_elsewhere && !avoid_processor) {
+ /*
+ * This thread is eligible to keep running on this processor.
+ *
+ * RT threads with un-expired quantum stay on processor,
+ * unless there's a valid RT thread with an earlier deadline.
+ */
+ if (thread->sched_pri >= BASEPRI_RTQUEUES && processor->first_timeslice) {
+ if (rt_runq_count(pset) > 0) {
+ thread_t next_rt = qe_queue_first(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links);
+
+ if (next_rt->realtime.deadline < processor->deadline &&
+ (next_rt->bound_processor == PROCESSOR_NULL ||
+ next_rt->bound_processor == processor)) {
+ /* The next RT thread is better, so pick it off the runqueue. */
+ goto pick_new_rt_thread;
+ }
+ }
- /* Reload precise timing global policy to thread-local policy */
- thread->precise_user_kernel_time = use_precise_user_kernel_time(thread);
-
- thread_unlock(thread);
+ /* This is still the best RT thread to run. */
+ processor->deadline = thread->realtime.deadline;
- /*
- * Switch execution timing to processor idle thread.
- */
- processor->last_dispatch = mach_absolute_time();
- 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;
+ sched_update_pset_load_average(pset, 0);
- /*
- * Cancel the quantum timer while idling.
- */
- timer_call_cancel(&processor->quantum_timer);
- processor->timeslice = 0;
+ processor_t next_rt_processor = PROCESSOR_NULL;
+ sched_ipi_type_t next_rt_ipi_type = SCHED_IPI_NONE;
- (*thread->sched_call)(SCHED_CALL_BLOCK, thread);
+ if (rt_runq_count(pset) - bit_count(pset->pending_AST_URGENT_cpu_mask) > 0) {
+ next_rt_processor = choose_processor_for_realtime_thread(pset, processor, true);
+ if (next_rt_processor) {
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)0, (uintptr_t)-4, next_rt_processor->cpu_id, next_rt_processor->state, 0);
+ if (next_rt_processor->state == PROCESSOR_IDLE) {
+ pset_update_processor_state(pset, next_rt_processor, PROCESSOR_DISPATCHING);
+ }
+ next_rt_ipi_type = sched_ipi_action(next_rt_processor, NULL, false, SCHED_IPI_EVENT_PREEMPT);
+ }
+ }
+ pset_unlock(pset);
- thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0);
+ if (next_rt_processor) {
+ sched_ipi_perform(next_rt_processor, next_rt_ipi_type);
+ }
- /*
- * Enable interrupts and perform idling activities. No
- * preemption due to TH_IDLE being set.
- */
- spllo(); new_thread = processor_idle(thread, processor);
+ return thread;
+ }
- /*
- * Return at splsched.
- */
- (*thread->sched_call)(SCHED_CALL_UNBLOCK, thread);
+ if ((rt_runq_count(pset) == 0) &&
+ SCHED(processor_queue_has_priority)(processor, thread->sched_pri, TRUE) == FALSE) {
+ /* This thread is still the highest priority runnable (non-idle) thread */
+ processor->deadline = UINT64_MAX;
- thread_lock(thread);
+ sched_update_pset_load_average(pset, 0);
+ pset_unlock(pset);
- /*
- * If we idled in place, simulate a context switch back
- * to the original priority of the thread so that the
- * platform layer cannot distinguish this from a true
- * switch to the idle thread.
- */
- if (thread->sched_mode == TH_MODE_REALTIME)
- thread_tell_urgency(THREAD_URGENCY_REAL_TIME, thread->realtime.period, thread->realtime.deadline);
- /* Identify non-promoted threads which have requested a
- * "background" priority.
- */
- else if ((thread->sched_pri <= MAXPRI_THROTTLE) &&
- (thread->priority <= MAXPRI_THROTTLE))
- thread_tell_urgency(THREAD_URGENCY_BACKGROUND, thread->sched_pri, thread->priority);
- else
- thread_tell_urgency(THREAD_URGENCY_NORMAL, thread->sched_pri, thread->priority);
+ return thread;
+ }
+ } else {
+ /*
+ * This processor must context switch.
+ * If it's due to a rebalance, we should aggressively find this thread a new home.
+ */
+ if (needs_smt_rebalance || affinity_mismatch || bound_elsewhere || avoid_processor) {
+ *reason |= AST_REBALANCE;
+ }
+ }
- /*
- * If awakened, switch to thread timer and start a new quantum.
- * Otherwise skip; we will context switch to another thread or return here.
- */
- if (!(thread->state & TH_WAIT)) {
- processor->last_dispatch = mach_absolute_time();
- thread_timer_event(processor->last_dispatch, &thread->system_timer);
- PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer;
+ bool secondary_forced_idle = ((processor->processor_secondary != PROCESSOR_NULL) &&
+ (thread_no_smt(thread) || (thread->sched_pri >= BASEPRI_RTQUEUES)) &&
+ (processor->processor_secondary->state == PROCESSOR_IDLE));
- thread_quantum_init(thread);
- thread->last_quantum_refill_time = processor->last_dispatch;
+ /* OK, so we're not going to run the current thread. Look at the RT queue. */
+ bool ok_to_run_realtime_thread = sched_ok_to_run_realtime_thread(pset, processor);
+ if ((rt_runq_count(pset) > 0) && ok_to_run_realtime_thread) {
+ thread_t next_rt = qe_queue_first(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links);
- processor->quantum_end = processor->last_dispatch + thread->current_quantum;
- timer_call_enter1(&processor->quantum_timer, thread, processor->quantum_end, TIMER_CALL_CRITICAL);
- processor->timeslice = 1;
+ if (__probable((next_rt->bound_processor == PROCESSOR_NULL ||
+ (next_rt->bound_processor == processor)))) {
+pick_new_rt_thread:
+ new_thread = qe_dequeue_head(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links);
- thread->computation_epoch = processor->last_dispatch;
- }
+ new_thread->runq = PROCESSOR_NULL;
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_decr(pset);
- thread->state &= ~TH_IDLE;
+ processor->deadline = new_thread->realtime.deadline;
- sched_run_incr();
- if (thread->sched_mode == TH_MODE_TIMESHARE)
- sched_share_incr();
+ pset_commit_processor_to_new_thread(pset, processor, new_thread);
- return (new_thread);
-}
-#endif /* CONFIG_SCHED_IDLE_IN_PLACE */
+ sched_update_pset_load_average(pset, 0);
-#if defined(CONFIG_SCHED_TRADITIONAL)
-static thread_t
-sched_traditional_choose_thread(
- processor_t processor,
- int priority)
-{
- thread_t thread;
-
- thread = choose_thread(processor, runq_for_processor(processor), priority);
- if (thread != THREAD_NULL) {
- runq_consider_decr_bound_count(processor, thread);
- }
-
- return thread;
-}
+ processor_t ast_processor = PROCESSOR_NULL;
+ processor_t next_rt_processor = PROCESSOR_NULL;
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ sched_ipi_type_t next_rt_ipi_type = SCHED_IPI_NONE;
-#endif /* defined(CONFIG_SCHED_TRADITIONAL) */
+ if (processor->processor_secondary != NULL) {
+ processor_t sprocessor = processor->processor_secondary;
+ if ((sprocessor->state == PROCESSOR_RUNNING) || (sprocessor->state == PROCESSOR_DISPATCHING)) {
+ ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL);
+ ast_processor = sprocessor;
+ }
+ }
+ if (rt_runq_count(pset) - bit_count(pset->pending_AST_URGENT_cpu_mask) > 0) {
+ next_rt_processor = choose_processor_for_realtime_thread(pset, processor, true);
+ if (next_rt_processor) {
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)0, (uintptr_t)-5, next_rt_processor->cpu_id, next_rt_processor->state, 0);
+ if (next_rt_processor->state == PROCESSOR_IDLE) {
+ pset_update_processor_state(pset, next_rt_processor, PROCESSOR_DISPATCHING);
+ }
+ next_rt_ipi_type = sched_ipi_action(next_rt_processor, NULL, false, SCHED_IPI_EVENT_PREEMPT);
+ }
+ }
+ pset_unlock(pset);
-#if defined(CONFIG_SCHED_TRADITIONAL) || defined(CONFIG_SCHED_FIXEDPRIORITY)
+ if (ast_processor) {
+ sched_ipi_perform(ast_processor, ipi_type);
+ }
-/*
- * choose_thread:
- *
- * Locate a thread to execute from the processor run queue
- * and return it. Only choose a thread with greater or equal
- * priority.
- *
- * Associated pset must be locked. Returns THREAD_NULL
- * on failure.
- */
-thread_t
-choose_thread(
- processor_t processor,
- run_queue_t rq,
- int priority)
-{
- queue_t queue = rq->queues + rq->highq;
- int pri = rq->highq, count = rq->count;
- thread_t thread;
-
- while (count > 0 && pri >= priority) {
- thread = (thread_t)queue_first(queue);
- while (!queue_end(queue, (queue_entry_t)thread)) {
- if (thread->bound_processor == PROCESSOR_NULL ||
- thread->bound_processor == processor) {
- remqueue((queue_entry_t)thread);
-
- thread->runq = PROCESSOR_NULL;
- SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
- rq->count--;
- if (SCHED(priority_is_urgent)(pri)) {
- rq->urgency--; assert(rq->urgency >= 0);
+ if (next_rt_processor) {
+ sched_ipi_perform(next_rt_processor, next_rt_ipi_type);
}
- if (queue_empty(queue)) {
- if (pri != IDLEPRI)
- clrbit(MAXPRI - pri, rq->bitmap);
- rq->highq = MAXPRI - ffsbit(rq->bitmap);
+
+ return new_thread;
+ }
+ }
+ if (secondary_can_only_run_realtime_thread) {
+ goto idle;
+ }
+
+ processor->deadline = UINT64_MAX;
+
+ /* No RT threads, so let's look at the regular threads. */
+ if ((new_thread = SCHED(choose_thread)(processor, MINPRI, *reason)) != THREAD_NULL) {
+ pset_commit_processor_to_new_thread(pset, processor, new_thread);
+ sched_update_pset_load_average(pset, 0);
+
+ processor_t ast_processor = PROCESSOR_NULL;
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+
+ processor_t sprocessor = processor->processor_secondary;
+ if ((sprocessor != NULL) && (sprocessor->state == PROCESSOR_RUNNING)) {
+ if (thread_no_smt(new_thread)) {
+ ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL);
+ ast_processor = sprocessor;
}
+ } else if (secondary_forced_idle && !thread_no_smt(new_thread) && pset_has_stealable_threads(pset)) {
+ pset_update_processor_state(pset, sprocessor, PROCESSOR_DISPATCHING);
+ ipi_type = sched_ipi_action(sprocessor, NULL, true, SCHED_IPI_EVENT_PREEMPT);
+ ast_processor = sprocessor;
+ }
+ pset_unlock(pset);
- return (thread);
+ if (ast_processor) {
+ sched_ipi_perform(ast_processor, ipi_type);
}
- count--;
+ return new_thread;
+ }
- thread = (thread_t)queue_next((queue_entry_t)thread);
+ if (processor->must_idle) {
+ processor->must_idle = false;
+ goto idle;
}
- queue--; pri--;
- }
+ if (SCHED(steal_thread_enabled)(pset) && (processor->processor_primary == processor)) {
+ /*
+ * No runnable threads, attempt to steal
+ * from other processors. Returns with pset lock dropped.
+ */
- return (THREAD_NULL);
-}
+ if ((new_thread = SCHED(steal_thread)(pset)) != THREAD_NULL) {
+ /*
+ * Avoid taking the pset_lock unless it is necessary to change state.
+ * It's safe to read processor->state here, as only the current processor can change state
+ * from this point (interrupts are disabled and this processor is committed to run new_thread).
+ */
+ if (processor->state == PROCESSOR_DISPATCHING || processor->state == PROCESSOR_IDLE) {
+ pset_lock(pset);
+ pset_commit_processor_to_new_thread(pset, processor, new_thread);
+ pset_unlock(pset);
+ } else {
+ assert((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_SHUTDOWN));
+ processor_state_update_from_thread(processor, new_thread);
+ }
+
+ return new_thread;
+ }
+
+ /*
+ * If other threads have appeared, shortcut
+ * around again.
+ */
+ if (!SCHED(processor_queue_empty)(processor) || (ok_to_run_realtime_thread && (rt_runq_count(pset) > 0))) {
+ continue;
+ }
+
+ pset_lock(pset);
+
+ /* Someone selected this processor while we had dropped the lock */
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id)) {
+ goto restart;
+ }
+ }
+
+idle:
+ /*
+ * Nothing is runnable, so set this processor idle if it
+ * was running.
+ */
+ if ((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_DISPATCHING)) {
+ pset_update_processor_state(pset, processor, PROCESSOR_IDLE);
+ processor_state_update_idle(processor);
+ }
-#endif /* defined(CONFIG_SCHED_TRADITIONAL) || defined(CONFIG_SCHED_FIXEDPRIORITY) */
+ /* Invoked with pset locked, returns with pset unlocked */
+ SCHED(processor_balance)(processor, pset);
+
+ new_thread = processor->idle_thread;
+ } while (new_thread == THREAD_NULL);
+
+ return new_thread;
+}
/*
- * Perform a context switch and start executing the new thread.
+ * thread_invoke
*
- * Returns FALSE on failure, and the thread is re-dispatched.
+ * Called at splsched with neither thread locked.
*
- * Called at splsched.
+ * Perform a context switch and start executing the new thread.
+ *
+ * Returns FALSE when the context switch didn't happen.
+ * The reference to the new thread is still consumed.
+ *
+ * "self" is what is currently running on the processor,
+ * "thread" is the new thread to context switch to
+ * (which may be the same thread in some cases)
*/
-
-#define funnel_release_check(thread, debug) \
-MACRO_BEGIN \
- if ((thread)->funnel_state & TH_FN_OWNED) { \
- (thread)->funnel_state = TH_FN_REFUNNEL; \
- KERNEL_DEBUG(0x603242c | DBG_FUNC_NONE, \
- (thread)->funnel_lock, (debug), 0, 0, 0); \
- funnel_unlock((thread)->funnel_lock); \
- } \
-MACRO_END
-
-#define funnel_refunnel_check(thread, debug) \
-MACRO_BEGIN \
- if ((thread)->funnel_state & TH_FN_REFUNNEL) { \
- kern_return_t result = (thread)->wait_result; \
- \
- (thread)->funnel_state = 0; \
- KERNEL_DEBUG(0x6032428 | DBG_FUNC_NONE, \
- (thread)->funnel_lock, (debug), 0, 0, 0); \
- funnel_lock((thread)->funnel_lock); \
- KERNEL_DEBUG(0x6032430 | DBG_FUNC_NONE, \
- (thread)->funnel_lock, (debug), 0, 0, 0); \
- (thread)->funnel_state = TH_FN_OWNED; \
- (thread)->wait_result = result; \
- } \
-MACRO_END
-
static boolean_t
thread_invoke(
- register thread_t self,
- register thread_t thread,
- ast_t reason)
+ thread_t self,
+ thread_t thread,
+ ast_t reason)
{
- thread_continue_t continuation = self->continuation;
- void *parameter = self->parameter;
- processor_t processor;
-
- if (get_preemption_level() != 0) {
+ if (__improbable(get_preemption_level() != 0)) {
int pl = get_preemption_level();
panic("thread_invoke: preemption_level %d, possible cause: %s",
pl, (pl < 0 ? "unlocking an unlocked mutex or spinlock" :
- "blocking while holding a spinlock, or within interrupt context"));
+ "blocking while holding a spinlock, or within interrupt context"));
+ }
+
+ thread_continue_t continuation = self->continuation;
+ void *parameter = self->parameter;
+ processor_t processor;
+
+ uint64_t ctime = mach_absolute_time();
+
+#ifdef CONFIG_MACH_APPROXIMATE_TIME
+ commpage_update_mach_approximate_time(ctime);
+#endif
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+ if (!((thread->state & TH_IDLE) != 0 ||
+ ((reason & AST_HANDOFF) && self->sched_mode == TH_MODE_REALTIME))) {
+ sched_timeshare_consider_maintenance(ctime);
}
+#endif
+
+#if MONOTONIC
+ mt_sched_update(self);
+#endif /* MONOTONIC */
+ assert_thread_magic(self);
assert(self == current_thread());
+ assert(self->runq == PROCESSOR_NULL);
+ assert((self->state & (TH_RUN | TH_TERMINATE2)) == TH_RUN);
- /*
- * Mark thread interruptible.
- */
thread_lock(thread);
- thread->state &= ~TH_UNINT;
-#if DEBUG
- assert(thread_runnable(thread));
-#endif
+ assert_thread_magic(thread);
+ assert((thread->state & (TH_RUN | TH_WAIT | TH_UNINT | TH_TERMINATE | TH_TERMINATE2)) == TH_RUN);
+ assert(thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == current_processor());
+ assert(thread->runq == PROCESSOR_NULL);
/* Reload precise timing global policy to thread-local policy */
thread->precise_user_kernel_time = use_precise_user_kernel_time(thread);
-
+
+ /* Update SFI class based on other factors */
+ thread->sfi_class = sfi_thread_classify(thread);
+
+ /* Update the same_pri_latency for the thread (used by perfcontrol callouts) */
+ thread->same_pri_latency = ctime - thread->last_basepri_change_time;
/*
- * Allow time constraint threads to hang onto
- * a stack.
+ * In case a base_pri update happened between the timestamp and
+ * taking the thread lock
*/
- if ((self->sched_mode == TH_MODE_REALTIME) && !self->reserved_stack)
+ if (ctime <= thread->last_basepri_change_time) {
+ thread->same_pri_latency = ctime - thread->last_made_runnable_time;
+ }
+
+ /* Allow realtime threads to hang onto a stack. */
+ if ((self->sched_mode == TH_MODE_REALTIME) && !self->reserved_stack) {
self->reserved_stack = self->kernel_stack;
+ }
+
+ /* Prepare for spin debugging */
+#if INTERRUPT_MASKED_DEBUG
+ ml_spin_debug_clear(thread);
+#endif
if (continuation != NULL) {
if (!thread->kernel_stack) {
* check to see whether we can exchange it with
* that of the other thread.
*/
- if (self->kernel_stack == self->reserved_stack && !thread->reserved_stack)
+ if (self->kernel_stack == self->reserved_stack && !thread->reserved_stack) {
goto need_stack;
+ }
/*
* Context switch by performing a stack handoff.
+ * Requires both threads to be parked in a continuation.
*/
continuation = thread->continuation;
parameter = thread->parameter;
processor = current_processor();
processor->active_thread = thread;
- processor->current_pri = thread->sched_pri;
- processor->current_thmode = thread->sched_mode;
+ processor_state_update_from_thread(processor, thread);
+
if (thread->last_processor != processor && thread->last_processor != NULL) {
- if (thread->last_processor->processor_set != processor->processor_set)
+ if (thread->last_processor->processor_set != processor->processor_set) {
thread->ps_switch++;
+ }
thread->p_switch++;
}
thread->last_processor = processor;
thread->c_switch++;
ast_context(thread);
+
thread_unlock(thread);
self->reason = reason;
- processor->last_dispatch = mach_absolute_time();
- self->last_run_time = processor->last_dispatch;
- thread_timer_event(processor->last_dispatch, &thread->system_timer);
- PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer;
+ processor->last_dispatch = ctime;
+ self->last_run_time = ctime;
+ processor_timer_switch_thread(ctime, &thread->system_timer);
+ timer_update(&thread->runnable_timer, ctime);
+ processor->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),
- processor->last_dispatch,
- PROCESSOR_DATA(processor, current_state));
+ timer_update(processor->current_state, ctime);
}
-
+
KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
- MACHDBG_CODE(DBG_MACH_SCHED, MACH_STACK_HANDOFF)|DBG_FUNC_NONE,
- self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+ 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 != NULL)) {
- 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);
+ 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);
+#if KPERF
+ kperf_off_cpu(self);
+#endif /* KPERF */
+
+ /*
+ * This is where we actually switch thread identity,
+ * and address space if required. However, register
+ * state is not switched - this routine leaves the
+ * stack and register state active on the current CPU.
+ */
TLOG(1, "thread_invoke: calling stack_handoff\n");
stack_handoff(self, thread);
+ /* 'self' is now off core */
+ assert(thread == current_thread_volatile());
+
DTRACE_SCHED(on__cpu);
+#if KPERF
+ kperf_on_cpu(thread, continuation, NULL);
+#endif /* KPERF */
+
thread_dispatch(self, thread);
+#if KASAN
+ /* Old thread's stack has been moved to the new thread, so explicitly
+ * unpoison it. */
+ kasan_unpoison_stack(thread->kernel_stack, kernel_stack_size);
+#endif
+
thread->continuation = thread->parameter = NULL;
- counter(c_thread_invoke_hits++);
+ boolean_t enable_interrupts = TRUE;
- funnel_refunnel_check(thread, 2);
- (void) spllo();
+ /* idle thread needs to stay interrupts-disabled */
+ if ((thread->state & TH_IDLE)) {
+ enable_interrupts = FALSE;
+ }
assert(continuation);
- call_continuation(continuation, parameter, thread->wait_result);
+ call_continuation(continuation, parameter,
+ thread->wait_result, enable_interrupts);
/*NOTREACHED*/
- }
- else if (thread == self) {
+ } else if (thread == self) {
/* same thread but with continuation */
ast_context(self);
- counter(++c_thread_invoke_same);
+
thread_unlock(self);
+#if KPERF
+ kperf_on_cpu(thread, continuation, NULL);
+#endif /* KPERF */
+
KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
- MACHDBG_CODE(DBG_MACH_SCHED,MACH_SCHED) | DBG_FUNC_NONE,
- self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED) | DBG_FUNC_NONE,
+ self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+
+#if KASAN
+ /* stack handoff to self - no thread_dispatch(), so clear the stack
+ * and free the fakestack directly */
+ kasan_fakestack_drop(self);
+ kasan_fakestack_gc(self);
+ kasan_unpoison_stack(self->kernel_stack, kernel_stack_size);
+#endif
self->continuation = self->parameter = NULL;
- funnel_refunnel_check(self, 3);
- (void) spllo();
+ boolean_t enable_interrupts = TRUE;
+
+ /* idle thread needs to stay interrupts-disabled */
+ if ((self->state & TH_IDLE)) {
+ enable_interrupts = FALSE;
+ }
- call_continuation(continuation, parameter, self->wait_result);
+ call_continuation(continuation, parameter,
+ self->wait_result, enable_interrupts);
/*NOTREACHED*/
}
- }
- else {
+ } else {
/*
* Check that the other thread has a stack
*/
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);
+ return FALSE;
}
- }
- else if (thread == self) {
+ } 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);
+ 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);
+ return TRUE;
}
}
*/
processor = current_processor();
processor->active_thread = thread;
- processor->current_pri = thread->sched_pri;
- processor->current_thmode = thread->sched_mode;
+ processor_state_update_from_thread(processor, thread);
+
if (thread->last_processor != processor && thread->last_processor != NULL) {
- if (thread->last_processor->processor_set != processor->processor_set)
+ if (thread->last_processor->processor_set != processor->processor_set) {
thread->ps_switch++;
+ }
thread->p_switch++;
}
thread->last_processor = processor;
thread->c_switch++;
ast_context(thread);
- thread_unlock(thread);
- counter(c_thread_invoke_csw++);
+ thread_unlock(thread);
- assert(self->runq == PROCESSOR_NULL);
self->reason = reason;
- processor->last_dispatch = mach_absolute_time();
- self->last_run_time = processor->last_dispatch;
- thread_timer_event(processor->last_dispatch, &thread->system_timer);
- PROCESSOR_DATA(processor, kernel_timer) = &thread->system_timer;
+ processor->last_dispatch = ctime;
+ self->last_run_time = ctime;
+ processor_timer_switch_thread(ctime, &thread->system_timer);
+ timer_update(&thread->runnable_timer, ctime);
+ processor->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),
- processor->last_dispatch,
- PROCESSOR_DATA(processor, current_state));
+ timer_update(processor->current_state, ctime);
}
-
KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
- MACHDBG_CODE(DBG_MACH_SCHED,MACH_SCHED) | DBG_FUNC_NONE,
- self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+ 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)) {
- 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);
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MOVED) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)thread->chosen_processor->cpu_id, 0, 0, 0);
}
DTRACE_SCHED2(off__cpu, struct thread *, thread, struct proc *, thread->task->bsd_info);
SCHED_STATS_CSW(processor, self->reason, self->sched_pri, thread->sched_pri);
+#if KPERF
+ kperf_off_cpu(self);
+#endif /* KPERF */
+
/*
* This is where we actually switch register context,
* and address space if required. We will next run
* as a result of a subsequent context switch.
+ *
+ * Once registers are switched and the processor is running "thread",
+ * the stack variables and non-volatile registers will contain whatever
+ * was there the last time that thread blocked. No local variables should
+ * be used after this point, except for the special case of "thread", which
+ * the platform layer returns as the previous thread running on the processor
+ * via the function call ABI as a return register, and "self", which may have
+ * been stored on the stack or a non-volatile register, but a stale idea of
+ * what was on the CPU is newly-accurate because that thread is again
+ * running on the CPU.
+ *
+ * If one of the threads is using a continuation, thread_continue
+ * is used to stitch up its context.
+ *
+ * If we are invoking a thread which is resuming from a continuation,
+ * the CPU will invoke thread_continue next.
+ *
+ * If the current thread is parking in a continuation, then its state
+ * won't be saved and the stack will be discarded. When the stack is
+ * re-allocated, it will be configured to resume from thread_continue.
*/
assert(continuation == self->continuation);
thread = machine_switch_context(self, continuation, thread);
- 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);
+ assert(self == current_thread_volatile());
+ TLOG(1, "thread_invoke: returning machine_switch_context: self %p continuation %p thread %p\n", self, continuation, thread);
- /*
- * We have been resumed and are set to run.
- */
- thread_dispatch(thread, self);
+ assert(continuation == NULL && self->continuation == NULL);
- if (continuation) {
- self->continuation = self->parameter = NULL;
+ DTRACE_SCHED(on__cpu);
- funnel_refunnel_check(self, 3);
- (void) spllo();
+#if KPERF
+ kperf_on_cpu(self, NULL, __builtin_frame_address(0));
+#endif /* KPERF */
- call_continuation(continuation, parameter, self->wait_result);
- /*NOTREACHED*/
- }
+ /* We have been resumed and are set to run. */
+ thread_dispatch(thread, self);
- return (TRUE);
+ return TRUE;
}
+#if defined(CONFIG_SCHED_DEFERRED_AST)
/*
- * thread_dispatch:
+ * pset_cancel_deferred_dispatch:
*
- * Handle threads at context switch. Re-dispatch other thread
- * if still running, otherwise update run state and perform
+ * Cancels all ASTs that we can cancel for the given processor set
+ * if the current processor is running the last runnable thread in the
+ * system.
+ *
+ * This function assumes the current thread is runnable. This must
+ * be called with the pset unlocked.
+ */
+static void
+pset_cancel_deferred_dispatch(
+ processor_set_t pset,
+ processor_t processor)
+{
+ processor_t active_processor = NULL;
+ uint32_t sampled_sched_run_count;
+
+ pset_lock(pset);
+ sampled_sched_run_count = os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed);
+
+ /*
+ * 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)) {
+ uint64_t dispatching_map = (pset->cpu_state_map[PROCESSOR_DISPATCHING] &
+ pset->pending_deferred_AST_cpu_mask &
+ ~pset->pending_AST_URGENT_cpu_mask);
+ for (int cpuid = lsb_first(dispatching_map); cpuid >= 0; cpuid = lsb_next(dispatching_map, cpuid)) {
+ active_processor = processor_array[cpuid];
+ /*
+ * If a processor is DISPATCHING, it could be because of
+ * a cancelable signal.
+ *
+ * IF the processor is not our
+ * current processor (the current processor should not
+ * be DISPATCHING, so this is a bit paranoid), AND there
+ * is a cancelable signal pending on the processor, AND
+ * there is no non-cancelable signal pending (as there is
+ * no point trying to backtrack on bringing the processor
+ * up if a signal we cannot cancel is outstanding), THEN
+ * it should make sense to roll back the processor state
+ * to the IDLE state.
+ *
+ * If the racey nature of this approach (as the signal
+ * will be arbitrated by hardware, and can fire as we
+ * roll back state) results in the core responding
+ * despite being pushed back to the IDLE state, it
+ * should be no different than if the core took some
+ * interrupt while IDLE.
+ */
+ if (active_processor != processor) {
+ /*
+ * Squash all of the processor state back to some
+ * reasonable facsimile of PROCESSOR_IDLE.
+ */
+
+ processor_state_update_idle(active_processor);
+ active_processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, active_processor, PROCESSOR_IDLE);
+ bit_clear(pset->pending_deferred_AST_cpu_mask, active_processor->cpu_id);
+ machine_signal_idle_cancel(active_processor);
+ }
+ }
+ }
+
+ pset_unlock(pset);
+}
+#else
+/* We don't support deferred ASTs; everything is candycanes and sunshine. */
+#endif
+
+static void
+thread_csw_callout(
+ thread_t old,
+ thread_t new,
+ uint64_t timestamp)
+{
+ perfcontrol_event event = (new->state & TH_IDLE) ? IDLE : CONTEXT_SWITCH;
+ uint64_t same_pri_latency = (new->state & TH_IDLE) ? 0 : new->same_pri_latency;
+ machine_switch_perfcontrol_context(event, timestamp, 0,
+ same_pri_latency, old, new);
+}
+
+
+/*
+ * thread_dispatch:
+ *
+ * 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)
+ thread_t thread,
+ thread_t self)
{
- processor_t processor = self->last_processor;
+ processor_t processor = self->last_processor;
+ bool was_idle = false;
+
+ assert(processor == current_processor());
+ assert(self == current_thread_volatile());
+ assert(thread != self);
if (thread != THREAD_NULL) {
+ /*
+ * Do the perfcontrol callout for context switch.
+ * The reason we do this here is:
+ * - thread_dispatch() is called from various places that are not
+ * the direct context switch path for eg. processor shutdown etc.
+ * So adding the callout here covers all those cases.
+ * - We want this callout as early as possible to be close
+ * to the timestamp taken in thread_invoke()
+ * - We want to avoid holding the thread lock while doing the
+ * callout
+ * - We do not want to callout if "thread" is NULL.
+ */
+ thread_csw_callout(thread, self, processor->last_dispatch);
+
+#if KASAN
+ if (thread->continuation != NULL) {
+ /*
+ * Thread has a continuation and the normal stack is going away.
+ * Unpoison the stack and mark all fakestack objects as unused.
+ */
+ kasan_fakestack_drop(thread);
+ if (thread->kernel_stack) {
+ kasan_unpoison_stack(thread->kernel_stack, kernel_stack_size);
+ }
+ }
+
+ /*
+ * Free all unused fakestack objects.
+ */
+ kasan_fakestack_gc(thread);
+#endif
+
/*
* If blocked at a continuation, discard
* the stack.
*/
- if (thread->continuation != NULL && thread->kernel_stack != 0)
+ if (thread->continuation != NULL && thread->kernel_stack != 0) {
stack_free(thread);
+ }
- if (!(thread->state & TH_IDLE)) {
+ if (thread->state & TH_IDLE) {
+ was_idle = true;
+ 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)
+ if (processor->quantum_end > processor->last_dispatch) {
remainder = processor->quantum_end -
processor->last_dispatch;
+ }
- consumed = thread->current_quantum - remainder;
+ consumed = thread->quantum_remaining - remainder;
- if ((thread->reason & AST_LEDGER) == 0)
+ if ((thread->reason & AST_LEDGER) == 0) {
/*
- * Bill CPU time to both the individual thread
- * and the task.
+ * Bill CPU time to both the task and
+ * the individual thread.
*/
- ledger_credit(thread->t_ledger,
+ ledger_credit_thread(thread, thread->t_ledger,
task_ledgers.cpu_time, consumed);
- ledger_credit(thread->t_threadledger,
+ ledger_credit_thread(thread, thread->t_threadledger,
thread_ledgers.cpu_time, consumed);
+ if (thread->t_bankledger) {
+ ledger_credit_thread(thread, thread->t_bankledger,
+ bank_ledgers.cpu_time,
+ (consumed - thread->t_deduct_bank_ledger_time));
+ }
+ thread->t_deduct_bank_ledger_time = 0;
+ if (consumed > 0) {
+ /*
+ * This should never be negative, but in traces we are seeing some instances
+ * of consumed being negative.
+ * <rdar://problem/57782596> thread_dispatch() thread CPU consumed calculation sometimes results in negative value
+ */
+ sched_update_pset_avg_execution_time(current_processor()->processor_set, consumed, processor->last_dispatch, thread->th_sched_bucket);
+ }
+ }
wake_lock(thread);
thread_lock(thread);
/*
- * Compute remainder of current quantum.
+ * Apply a priority floor if the thread holds a kernel resource
+ * Do this before checking starting_pri to avoid overpenalizing
+ * repeated rwlock blockers.
+ */
+ if (__improbable(thread->rwlock_count != 0)) {
+ lck_rw_set_promotion_locked(thread);
+ }
+
+ boolean_t keep_quantum = processor->first_timeslice;
+
+ /*
+ * Treat a thread which has dropped priority since it got on core
+ * as having expired its quantum.
*/
- if (first_timeslice(processor) &&
- processor->quantum_end > processor->last_dispatch)
- thread->current_quantum = (uint32_t)remainder;
- else
- thread->current_quantum = 0;
+ if (processor->starting_pri > thread->sched_pri) {
+ keep_quantum = FALSE;
+ }
+
+ /* Compute remainder of current quantum. */
+ if (keep_quantum &&
+ processor->quantum_end > processor->last_dispatch) {
+ thread->quantum_remaining = (uint32_t)remainder;
+ } else {
+ thread->quantum_remaining = 0;
+ }
if (thread->sched_mode == TH_MODE_REALTIME) {
/*
* Cancel the deadline if the thread has
* consumed the entire quantum.
*/
- if (thread->current_quantum == 0) {
+ if (thread->quantum_remaining == 0) {
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_CANCEL_RT_DEADLINE) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->realtime.deadline, thread->realtime.computation, 0);
thread->realtime.deadline = UINT64_MAX;
- thread->reason |= AST_QUANTUM;
}
} else {
-#if defined(CONFIG_SCHED_TRADITIONAL)
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
/*
* For non-realtime threads treat a tiny
* remaining quantum as an expired quantum
* but include what's left next time.
*/
- if (thread->current_quantum < min_std_quantum) {
+ if (thread->quantum_remaining < min_std_quantum) {
thread->reason |= AST_QUANTUM;
- thread->current_quantum += SCHED(initial_quantum_size)(thread);
+ thread->quantum_remaining += SCHED(initial_quantum_size)(thread);
}
-#endif
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
}
/*
* If we are doing a direct handoff then
* take the remainder of the quantum.
*/
- if ((thread->reason & (AST_HANDOFF|AST_QUANTUM)) == AST_HANDOFF) {
- self->current_quantum = thread->current_quantum;
+ if ((thread->reason & (AST_HANDOFF | AST_QUANTUM)) == AST_HANDOFF) {
+ self->quantum_remaining = thread->quantum_remaining;
thread->reason |= AST_QUANTUM;
- thread->current_quantum = 0;
+ thread->quantum_remaining = 0;
+ } else {
+#if defined(CONFIG_SCHED_MULTIQ)
+ if (SCHED(sched_groups_enabled) &&
+ thread->sched_group == self->sched_group) {
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_QUANTUM_HANDOFF),
+ self->reason, (uintptr_t)thread_tid(thread),
+ self->quantum_remaining, thread->quantum_remaining, 0);
+
+ self->quantum_remaining = thread->quantum_remaining;
+ thread->quantum_remaining = 0;
+ /* Don't set AST_QUANTUM here - old thread might still want to preempt someone else */
+ }
+#endif /* defined(CONFIG_SCHED_MULTIQ) */
}
thread->computation_metered += (processor->last_dispatch - thread->computation_epoch);
if (!(thread->state & TH_WAIT)) {
/*
- * Still running.
+ * Still runnable.
*/
- if (thread->reason & AST_QUANTUM)
- thread_setrun(thread, SCHED_TAILQ);
- else
- if (thread->reason & AST_PREEMPT)
- thread_setrun(thread, SCHED_HEADQ);
- else
- thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
+ thread->last_made_runnable_time = thread->last_basepri_change_time = processor->last_dispatch;
+
+ machine_thread_going_off_core(thread, FALSE, processor->last_dispatch, TRUE);
+
+ ast_t reason = thread->reason;
+ sched_options_t options = SCHED_NONE;
+
+ if (reason & AST_REBALANCE) {
+ options |= SCHED_REBALANCE;
+ if (reason & AST_QUANTUM) {
+ /*
+ * Having gone to the trouble of forcing this thread off a less preferred core,
+ * we should force the preferable core to reschedule immediately to give this
+ * thread a chance to run instead of just sitting on the run queue where
+ * it may just be stolen back by the idle core we just forced it off.
+ * But only do this at the end of a quantum to prevent cascading effects.
+ */
+ options |= SCHED_PREEMPT;
+ }
+ }
- thread->reason = AST_NONE;
+ if (reason & AST_QUANTUM) {
+ options |= SCHED_TAILQ;
+ } else if (reason & AST_PREEMPT) {
+ options |= SCHED_HEADQ;
+ } else {
+ options |= (SCHED_PREEMPT | SCHED_TAILQ);
+ }
+
+ thread_setrun(thread, options);
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_DISPATCH) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->reason, thread->state,
+ sched_run_buckets[TH_BUCKET_RUN], 0);
if (thread->wake_active) {
thread->wake_active = FALSE;
thread_unlock(thread);
thread_wakeup(&thread->wake_active);
- }
- else
+ } else {
thread_unlock(thread);
+ }
wake_unlock(thread);
- }
- else {
+ } else {
/*
* Waiting.
*/
boolean_t should_terminate = FALSE;
+ uint32_t new_run_count;
+ int thread_state = thread->state;
/* Only the first call to thread_dispatch
* after explicit termination should add
* the thread to the termination queue
*/
- if ((thread->state & (TH_TERMINATE|TH_TERMINATE2)) == TH_TERMINATE) {
+ if ((thread_state & (TH_TERMINATE | TH_TERMINATE2)) == TH_TERMINATE) {
should_terminate = TRUE;
- thread->state |= TH_TERMINATE2;
+ thread_state |= TH_TERMINATE2;
+ }
+
+ timer_stop(&thread->runnable_timer, processor->last_dispatch);
+
+ thread_state &= ~TH_RUN;
+ thread->state = thread_state;
+
+ thread->last_made_runnable_time = thread->last_basepri_change_time = THREAD_NOT_RUNNABLE;
+ thread->chosen_processor = PROCESSOR_NULL;
+
+ new_run_count = SCHED(run_count_decr)(thread);
+
+#if CONFIG_SCHED_AUTO_JOIN
+ if ((thread->sched_flags & TH_SFLAG_THREAD_GROUP_AUTO_JOIN) != 0) {
+ work_interval_auto_join_unwind(thread);
}
+#endif /* CONFIG_SCHED_AUTO_JOIN */
- thread->state &= ~TH_RUN;
+#if CONFIG_SCHED_SFI
+ if (thread->reason & AST_SFI) {
+ thread->wait_sfi_begin_time = processor->last_dispatch;
+ }
+#endif
+ machine_thread_going_off_core(thread, should_terminate, processor->last_dispatch, FALSE);
- if (thread->sched_mode == TH_MODE_TIMESHARE)
- sched_share_decr();
- sched_run_decr();
+ 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_state & TH_WAIT_REPORT) {
+ (*thread->sched_call)(SCHED_CALL_BLOCK, thread);
+ }
if (thread->wake_active) {
thread->wake_active = FALSE;
thread_unlock(thread);
thread_wakeup(&thread->wake_active);
- }
- else
+ } else {
thread_unlock(thread);
+ }
wake_unlock(thread);
- if (should_terminate)
+ if (should_terminate) {
thread_terminate_enqueue(thread);
+ }
}
}
+ /*
+ * The thread could have been added to the termination queue, so it's
+ * unsafe to use after this point.
+ */
+ thread = THREAD_NULL;
}
+ int urgency = THREAD_URGENCY_NONE;
+ uint64_t latency = 0;
+
+ /* Update (new) current thread and reprogram running timers */
+ thread_lock(self);
+
if (!(self->state & TH_IDLE)) {
+ uint64_t arg1, arg2;
+
+#if CONFIG_SCHED_SFI
+ ast_t new_ast;
+
+ new_ast = sfi_thread_needs_ast(self, NULL);
+
+ if (new_ast != AST_NONE) {
+ ast_on(new_ast);
+ }
+#endif
+
+ assertf(processor->last_dispatch >= self->last_made_runnable_time,
+ "Non-monotonic time? dispatch at 0x%llx, runnable at 0x%llx",
+ processor->last_dispatch, self->last_made_runnable_time);
+
+ assert(self->last_made_runnable_time <= self->last_basepri_change_time);
+
+ latency = processor->last_dispatch - self->last_made_runnable_time;
+ assert(latency >= self->same_pri_latency);
+
+ urgency = thread_get_urgency(self, &arg1, &arg2);
+
+ thread_tell_urgency(urgency, arg1, arg2, latency, self);
- if (self->sched_mode == TH_MODE_REALTIME)
- thread_tell_urgency(THREAD_URGENCY_REAL_TIME, self->realtime.period, self->realtime.deadline);
- /* Identify non-promoted threads which have requested a
- * "background" priority.
- */
- else if ((self->sched_pri <= MAXPRI_THROTTLE) &&
- (self->priority <= MAXPRI_THROTTLE))
- thread_tell_urgency(THREAD_URGENCY_BACKGROUND, self->sched_pri, self->priority);
- else
- thread_tell_urgency(THREAD_URGENCY_NORMAL, self->sched_pri, self->priority);
/*
* Get a new quantum if none remaining.
*/
- if (self->current_quantum == 0) {
+ if (self->quantum_remaining == 0) {
thread_quantum_init(self);
- self->last_quantum_refill_time = processor->last_dispatch;
}
/*
* Set up quantum timer and timeslice.
*/
- processor->quantum_end = (processor->last_dispatch + self->current_quantum);
- timer_call_enter1(&processor->quantum_timer, self, processor->quantum_end, TIMER_CALL_CRITICAL);
-
- processor->timeslice = 1;
+ processor->quantum_end = processor->last_dispatch +
+ self->quantum_remaining;
- self->computation_epoch = processor->last_dispatch;
+ running_timer_setup(processor, RUNNING_TIMER_QUANTUM, self,
+ processor->quantum_end, processor->last_dispatch);
+ if (was_idle) {
+ /*
+ * kperf's running timer is active whenever the idle thread for a
+ * CPU is not running.
+ */
+ kperf_running_setup(processor, processor->last_dispatch);
+ }
+ running_timers_activate(processor);
+ processor->first_timeslice = TRUE;
+ } else {
+ running_timers_deactivate(processor);
+ processor->first_timeslice = FALSE;
+ thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0, 0, self);
}
- else {
- timer_call_cancel(&processor->quantum_timer);
- processor->timeslice = 0;
- thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0);
- }
-}
+ assert(self->block_hint == kThreadWaitNone);
+ self->computation_epoch = processor->last_dispatch;
+ self->reason = AST_NONE;
+ processor->starting_pri = self->sched_pri;
-#include <libkern/OSDebug.h>
+ thread_unlock(self);
-uint32_t kdebug_thread_block = 0;
+ machine_thread_going_on_core(self, urgency, latency, self->same_pri_latency,
+ processor->last_dispatch);
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ /*
+ * TODO: Can we state that redispatching our old thread is also
+ * uninteresting?
+ */
+ if ((os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed) == 1) && !(self->state & TH_IDLE)) {
+ pset_cancel_deferred_dispatch(processor->processor_set, processor);
+ }
+#endif
+}
/*
* thread_block_reason:
* 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_continue_t continuation,
+ void *parameter,
+ ast_t reason)
{
- register thread_t self = current_thread();
- register processor_t processor;
- register thread_t new_thread;
- spl_t s;
-
- counter(++c_thread_block_calls);
+ thread_t self = current_thread();
+ processor_t processor;
+ thread_t new_thread;
+ spl_t s;
s = splsched();
- if (!(reason & AST_PREEMPT))
- funnel_release_check(self, 2);
-
processor = current_processor();
/* If we're explicitly yielding, force a subsequent quantum */
- if (reason & AST_YIELD)
- processor->timeslice = 0;
+ if (reason & AST_YIELD) {
+ processor->first_timeslice = FALSE;
+ }
/* We're handling all scheduling AST's */
ast_off(AST_SCHEDULING);
+#if PROC_REF_DEBUG
+ if ((continuation != NULL) && (self->task != kernel_task)) {
+ if (uthread_get_proc_refcount(self->uthread) != 0) {
+ panic("thread_block_reason with continuation uthread %p with uu_proc_refcount != 0", self->uthread);
+ }
+ }
+#endif
+
self->continuation = continuation;
self->parameter = parameter;
- if (__improbable(kdebug_thread_block && kdebug_enable && self->state != TH_RUN)) {
+ 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);
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_BLOCK),
+ reason, VM_KERNEL_UNSLIDE(continuation), 0, 0, 0);
}
do {
thread_lock(self);
- new_thread = thread_select(self, processor);
+ new_thread = thread_select(self, processor, &reason);
thread_unlock(self);
} while (!thread_invoke(self, new_thread, reason));
- funnel_refunnel_check(self, 5);
splx(s);
- return (self->wait_result);
+ return self->wait_result;
}
/*
*/
wait_result_t
thread_block(
- thread_continue_t continuation)
+ thread_continue_t continuation)
{
return thread_block_reason(continuation, NULL, AST_NONE);
}
wait_result_t
thread_block_parameter(
- thread_continue_t continuation,
- void *parameter)
+ thread_continue_t continuation,
+ void *parameter)
{
return thread_block_reason(continuation, parameter, AST_NONE);
}
*/
int
thread_run(
- thread_t self,
- thread_continue_t continuation,
- void *parameter,
- thread_t new_thread)
+ thread_t self,
+ thread_continue_t continuation,
+ void *parameter,
+ thread_t new_thread)
{
- ast_t handoff = AST_HANDOFF;
+ ast_t reason = AST_NONE;
+
+ if ((self->state & TH_IDLE) == 0) {
+ reason = AST_HANDOFF;
+ }
- funnel_release_check(self, 3);
+ /*
+ * If this thread hadn't been setrun'ed, it
+ * might not have a chosen processor, so give it one
+ */
+ if (new_thread->chosen_processor == NULL) {
+ new_thread->chosen_processor = current_processor();
+ }
self->continuation = continuation;
self->parameter = parameter;
- while (!thread_invoke(self, new_thread, handoff)) {
- processor_t processor = current_processor();
+ while (!thread_invoke(self, new_thread, reason)) {
+ /* the handoff failed, so we have to fall back to the normal block path */
+ processor_t processor = current_processor();
+
+ reason = AST_NONE;
thread_lock(self);
- new_thread = thread_select(self, processor);
+ new_thread = thread_select(self, processor, &reason);
thread_unlock(self);
- handoff = AST_NONE;
}
- funnel_refunnel_check(self, 6);
-
- return (self->wait_result);
+ return self->wait_result;
}
/*
*
* Called at splsched when a thread first receives
* a new stack after a continuation.
+ *
+ * Called with THREAD_NULL as the old thread when
+ * invoked by machine_load_context.
*/
void
thread_continue(
- register thread_t thread)
+ thread_t thread)
{
- register thread_t self = current_thread();
- register thread_continue_t continuation;
- register void *parameter;
+ thread_t self = current_thread();
+ thread_continue_t continuation;
+ void *parameter;
DTRACE_SCHED(on__cpu);
continuation = self->continuation;
parameter = self->parameter;
+ assert(continuation != NULL);
+
+#if KPERF
+ kperf_on_cpu(self, continuation, NULL);
+#endif
+
thread_dispatch(thread, self);
self->continuation = self->parameter = NULL;
- funnel_refunnel_check(self, 4);
+#if INTERRUPT_MASKED_DEBUG
+ /* Reset interrupt-masked spin debugging timeout */
+ ml_spin_debug_clear(self);
+#endif
+
+ TLOG(1, "thread_continue: calling call_continuation\n");
- if (thread != THREAD_NULL)
- (void)spllo();
+ boolean_t enable_interrupts = TRUE;
- TLOG(1, "thread_continue: calling call_continuation \n");
- call_continuation(continuation, parameter, self->wait_result);
+ /* bootstrap thread, idle thread need to stay interrupts-disabled */
+ if (thread == THREAD_NULL || (self->state & TH_IDLE)) {
+ enable_interrupts = FALSE;
+ }
+
+ call_continuation(continuation, parameter, self->wait_result, enable_interrupts);
/*NOTREACHED*/
}
thread_quantum_init(thread_t thread)
{
if (thread->sched_mode == TH_MODE_REALTIME) {
- thread->current_quantum = thread->realtime.computation;
+ thread->quantum_remaining = thread->realtime.computation;
} else {
- thread->current_quantum = SCHED(initial_quantum_size)(thread);
+ thread->quantum_remaining = SCHED(initial_quantum_size)(thread);
}
}
-#if defined(CONFIG_SCHED_TRADITIONAL)
-static uint32_t
-sched_traditional_initial_quantum_size(thread_t thread)
+uint32_t
+sched_timeshare_initial_quantum_size(thread_t thread)
{
- if ((thread == THREAD_NULL) || thread->priority > MAXPRI_THROTTLE)
- return std_quantum;
- else
+ if ((thread != THREAD_NULL) && thread->th_sched_bucket == TH_BUCKET_SHARE_BG) {
return bg_quantum;
+ } else {
+ return std_quantum;
+ }
}
-static sched_mode_t
-sched_traditional_initial_thread_sched_mode(task_t parent_task)
-{
- if (parent_task == kernel_task)
- return TH_MODE_FIXED;
- else
- return TH_MODE_TIMESHARE;
-}
-
-static boolean_t
-sched_traditional_supports_timeshare_mode(void)
-{
- return TRUE;
-}
-
-#endif /* CONFIG_SCHED_TRADITIONAL */
-
/*
* run_queue_init:
*
*/
void
run_queue_init(
- run_queue_t rq)
+ run_queue_t rq)
{
- int i;
-
- rq->highq = IDLEPRI;
- for (i = 0; i < NRQBM; i++)
+ rq->highq = NOPRI;
+ for (u_int i = 0; i < BITMAP_LEN(NRQS); i++) {
rq->bitmap[i] = 0;
- setbit(MAXPRI - IDLEPRI, rq->bitmap);
- rq->urgency = rq->count = 0;
- for (i = 0; i < NRQS; i++)
- queue_init(&rq->queues[i]);
-}
-
-#if defined(CONFIG_SCHED_TRADITIONAL) || defined(CONFIG_SCHED_PROTO) || defined(CONFIG_SCHED_GRRR) || defined(CONFIG_SCHED_FIXEDPRIORITY)
-int
-sched_traditional_fairshare_runq_count(void)
-{
- return fs_runq.count;
-}
-
-uint64_t
-sched_traditional_fairshare_runq_stats_count_sum(void)
-{
- return fs_runq.runq_stats.count_sum;
-}
-
-void
-sched_traditional_fairshare_enqueue(thread_t thread)
-{
- queue_t queue = &fs_runq.queue;
-
- simple_lock(&fs_lock);
-
- enqueue_tail(queue, (queue_entry_t)thread);
-
- thread->runq = FS_RUNQ;
- SCHED_STATS_RUNQ_CHANGE(&fs_runq.runq_stats, fs_runq.count);
- fs_runq.count++;
-
- simple_unlock(&fs_lock);
-}
-
-thread_t
-sched_traditional_fairshare_dequeue(void)
-{
- thread_t thread;
-
- simple_lock(&fs_lock);
- if (fs_runq.count > 0) {
- thread = (thread_t)dequeue_head(&fs_runq.queue);
-
- thread->runq = PROCESSOR_NULL;
- SCHED_STATS_RUNQ_CHANGE(&fs_runq.runq_stats, fs_runq.count);
- fs_runq.count--;
-
- simple_unlock(&fs_lock);
-
- return (thread);
}
- simple_unlock(&fs_lock);
-
- return THREAD_NULL;
-}
-
-boolean_t
-sched_traditional_fairshare_queue_remove(thread_t thread)
-{
- queue_t q;
-
- simple_lock(&fs_lock);
- q = &fs_runq.queue;
-
- if (FS_RUNQ == thread->runq) {
- remqueue((queue_entry_t)thread);
- SCHED_STATS_RUNQ_CHANGE(&fs_runq.runq_stats, fs_runq.count);
- fs_runq.count--;
-
- thread->runq = PROCESSOR_NULL;
- simple_unlock(&fs_lock);
- return (TRUE);
+ rq->urgency = rq->count = 0;
+ for (int i = 0; i < NRQS; i++) {
+ circle_queue_init(&rq->queues[i]);
}
- else {
- /*
- * The thread left the run queue before we could
- * lock the run queue.
- */
- assert(thread->runq == PROCESSOR_NULL);
- simple_unlock(&fs_lock);
- return (FALSE);
- }
}
-#endif /* defined(CONFIG_SCHED_TRADITIONAL) || defined(CONFIG_SCHED_PROTO) || defined(CONFIG_SCHED_GRRR) || defined(CONFIG_SCHED_FIXEDPRIORITY) */
-
/*
* run_queue_dequeue:
*
*/
thread_t
run_queue_dequeue(
- run_queue_t rq,
- integer_t options)
+ run_queue_t rq,
+ sched_options_t options)
{
- thread_t thread;
- queue_t queue = rq->queues + rq->highq;
+ thread_t thread;
+ circle_queue_t queue = &rq->queues[rq->highq];
if (options & SCHED_HEADQ) {
- thread = (thread_t)dequeue_head(queue);
- }
- else {
- thread = (thread_t)dequeue_tail(queue);
+ thread = cqe_dequeue_head(queue, struct thread, runq_links);
+ } else {
+ thread = cqe_dequeue_tail(queue, struct thread, runq_links);
}
+ assert(thread != THREAD_NULL);
+ assert_thread_magic(thread);
+
thread->runq = PROCESSOR_NULL;
SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
rq->count--;
if (SCHED(priority_is_urgent)(rq->highq)) {
rq->urgency--; assert(rq->urgency >= 0);
}
- if (queue_empty(queue)) {
- if (rq->highq != IDLEPRI)
- clrbit(MAXPRI - rq->highq, rq->bitmap);
- rq->highq = MAXPRI - ffsbit(rq->bitmap);
+ if (circle_queue_empty(queue)) {
+ bitmap_clear(rq->bitmap, rq->highq);
+ rq->highq = bitmap_first(rq->bitmap, NRQS);
}
- return (thread);
+ return thread;
}
/*
*/
boolean_t
run_queue_enqueue(
- run_queue_t rq,
- thread_t thread,
- integer_t options)
+ run_queue_t rq,
+ thread_t thread,
+ sched_options_t options)
{
- queue_t queue = rq->queues + thread->sched_pri;
- boolean_t result = FALSE;
-
- if (queue_empty(queue)) {
- enqueue_tail(queue, (queue_entry_t)thread);
-
- setbit(MAXPRI - thread->sched_pri, rq->bitmap);
+ circle_queue_t queue = &rq->queues[thread->sched_pri];
+ boolean_t result = FALSE;
+
+ assert_thread_magic(thread);
+
+ if (circle_queue_empty(queue)) {
+ circle_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) {
+ circle_enqueue_tail(queue, &thread->runq_links);
+ } else {
+ circle_enqueue_head(queue, &thread->runq_links);
+ }
}
- else
- if (options & SCHED_TAILQ)
- enqueue_tail(queue, (queue_entry_t)thread);
- else
- enqueue_head(queue, (queue_entry_t)thread);
-
- if (SCHED(priority_is_urgent)(thread->sched_pri))
+ if (SCHED(priority_is_urgent)(thread->sched_pri)) {
rq->urgency++;
+ }
SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
rq->count++;
-
- return (result);
-
+
+ return result;
}
/*
*/
void
run_queue_remove(
- run_queue_t rq,
- thread_t thread)
+ run_queue_t rq,
+ thread_t thread)
{
+ circle_queue_t queue = &rq->queues[thread->sched_pri];
- remqueue((queue_entry_t)thread);
+ assert(thread->runq != PROCESSOR_NULL);
+ assert_thread_magic(thread);
+
+ circle_dequeue(queue, &thread->runq_links);
SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
rq->count--;
if (SCHED(priority_is_urgent)(thread->sched_pri)) {
rq->urgency--; assert(rq->urgency >= 0);
}
-
- if (queue_empty(rq->queues + thread->sched_pri)) {
+
+ if (circle_queue_empty(queue)) {
/* update run queue status */
- if (thread->sched_pri != IDLEPRI)
- clrbit(MAXPRI - thread->sched_pri, rq->bitmap);
- rq->highq = MAXPRI - ffsbit(rq->bitmap);
+ bitmap_clear(rq->bitmap, thread->sched_pri);
+ rq->highq = bitmap_first(rq->bitmap, NRQS);
}
-
+
thread->runq = PROCESSOR_NULL;
}
/*
- * fairshare_setrun:
+ * run_queue_peek
*
- * Dispatch a thread for round-robin execution.
+ * Peek at the runq and return the highest
+ * priority thread from the runq.
*
- * Thread must be locked. Associated pset must
- * be locked, and is returned unlocked.
+ * The run queue must be locked.
*/
-static void
-fairshare_setrun(
- processor_t processor,
- thread_t thread)
+thread_t
+run_queue_peek(
+ run_queue_t rq)
{
- processor_set_t pset = processor->processor_set;
-
- thread->chosen_processor = processor;
-
- SCHED(fairshare_enqueue)(thread);
-
- if (processor != current_processor())
- machine_signal_idle(processor);
+ if (rq->count > 0) {
+ circle_queue_t queue = &rq->queues[rq->highq];
+ thread_t thread = cqe_queue_first(queue, struct thread, runq_links);
+ assert_thread_magic(thread);
+ return thread;
+ } else {
+ return THREAD_NULL;
+ }
+}
- pset_unlock(pset);
+rt_queue_t
+sched_rtlocal_runq(processor_set_t pset)
+{
+ return &pset->rt_runq;
+}
+void
+sched_rtlocal_init(processor_set_t pset)
+{
+ pset_rt_init(pset);
}
-/*
- * realtime_queue_insert:
- *
- * Enqueue a thread for realtime execution.
- */
-static boolean_t
-realtime_queue_insert(
- thread_t thread)
+void
+sched_rtlocal_queue_shutdown(processor_t processor)
{
- queue_t queue = &rt_runq.queue;
- uint64_t deadline = thread->realtime.deadline;
- boolean_t preempt = FALSE;
+ processor_set_t pset = processor->processor_set;
+ thread_t thread;
+ queue_head_t tqueue;
- simple_lock(&rt_lock);
+ pset_lock(pset);
- if (queue_empty(queue)) {
- enqueue_tail(queue, (queue_entry_t)thread);
- preempt = TRUE;
+ /* We only need to migrate threads if this is the last active or last recommended processor in the pset */
+ if ((pset->online_processor_count > 0) && pset_is_recommended(pset)) {
+ pset_unlock(pset);
+ return;
}
- else {
- register thread_t entry = (thread_t)queue_first(queue);
- while (TRUE) {
- if ( queue_end(queue, (queue_entry_t)entry) ||
- deadline < entry->realtime.deadline ) {
- entry = (thread_t)queue_prev((queue_entry_t)entry);
- break;
- }
+ queue_init(&tqueue);
- entry = (thread_t)queue_next((queue_entry_t)entry);
- }
+ while (rt_runq_count(pset) > 0) {
+ thread = qe_dequeue_head(&pset->rt_runq.queue, struct thread, runq_links);
+ thread->runq = PROCESSOR_NULL;
+ SCHED_STATS_RUNQ_CHANGE(&pset->rt_runq.runq_stats, rt_runq_count(pset));
+ rt_runq_count_decr(pset);
+ enqueue_tail(&tqueue, &thread->runq_links);
+ }
+ sched_update_pset_load_average(pset, 0);
+ pset_unlock(pset);
+
+ qe_foreach_element_safe(thread, &tqueue, runq_links) {
+ remqueue(&thread->runq_links);
- if ((queue_entry_t)entry == queue)
- preempt = TRUE;
+ thread_lock(thread);
+
+ thread_setrun(thread, SCHED_TAILQ);
- insque((queue_entry_t)thread, (queue_entry_t)entry);
+ thread_unlock(thread);
}
+}
- thread->runq = RT_RUNQ;
- SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count);
- rt_runq.count++;
+/* Assumes RT lock is not held, and acquires splsched/rt_lock itself */
+void
+sched_rtlocal_runq_scan(sched_update_scan_context_t scan_context)
+{
+ thread_t thread;
- simple_unlock(&rt_lock);
+ pset_node_t node = &pset_node0;
+ processor_set_t pset = node->psets;
- return (preempt);
-}
+ spl_t s = splsched();
+ do {
+ while (pset != NULL) {
+ pset_lock(pset);
-/*
- * 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;
-
- thread->chosen_processor = processor;
+ qe_foreach_element_safe(thread, &pset->rt_runq.queue, runq_links) {
+ if (thread->last_made_runnable_time < scan_context->earliest_rt_make_runnable_time) {
+ scan_context->earliest_rt_make_runnable_time = thread->last_made_runnable_time;
+ }
+ }
- /*
- * Dispatch directly onto idle processor.
- */
- if ( (thread->bound_processor == processor)
- && processor->state == PROCESSOR_IDLE) {
- remqueue((queue_entry_t)processor);
- enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
-
- processor->next_thread = thread;
- processor->deadline = thread->realtime.deadline;
- processor->state = PROCESSOR_DISPATCHING;
- pset_unlock(pset);
+ pset_unlock(pset);
- if (processor != current_processor())
- machine_signal_idle(processor);
- return;
- }
+ pset = pset->pset_list;
+ }
+ } while (((node = node->node_list) != NULL) && ((pset = node->psets) != NULL));
+ splx(s);
+}
- if (realtime_queue_insert(thread)) {
- int prstate = processor->state;
- if (processor == current_processor())
- ast_on(AST_PREEMPT | AST_URGENT);
- else if ((prstate == PROCESSOR_IDLE) || (prstate == PROCESSOR_DISPATCHING))
- machine_signal_idle(processor);
- else
- cause_ast_check(processor);
- }
+int64_t
+sched_rtlocal_runq_count_sum(void)
+{
+ pset_node_t node = &pset_node0;
+ processor_set_t pset = node->psets;
+ int64_t count = 0;
- pset_unlock(pset);
-}
+ do {
+ while (pset != NULL) {
+ count += pset->rt_runq.runq_stats.count_sum;
-#if defined(CONFIG_SCHED_TRADITIONAL)
+ pset = pset->pset_list;
+ }
+ } while (((node = node->node_list) != NULL) && ((pset = node->psets) != NULL));
-static boolean_t
-priority_is_urgent(int priority)
-{
- return testbit(priority, sched_preempt_pri) ? TRUE : FALSE;
+ return count;
}
/*
- * processor_enqueue:
- *
- * Enqueue thread on a processor run queue. Thread must be locked,
- * and not already be on a run queue.
- *
- * Returns TRUE if a preemption is indicated based on the state
- * of the run queue.
+ * realtime_queue_insert:
*
- * The run queue must be locked (see thread_run_queue_remove()
- * for more info).
+ * Enqueue a thread for realtime execution.
*/
static boolean_t
-processor_enqueue(
- processor_t processor,
- thread_t thread,
- integer_t options)
-{
- run_queue_t rq = runq_for_processor(processor);
- boolean_t result;
-
- result = run_queue_enqueue(rq, thread, options);
+realtime_queue_insert(processor_t processor, processor_set_t pset, thread_t thread)
+{
+ queue_t queue = &SCHED(rt_runq)(pset)->queue;
+ uint64_t deadline = thread->realtime.deadline;
+ boolean_t preempt = FALSE;
+
+ pset_assert_locked(pset);
+
+ if (queue_empty(queue)) {
+ enqueue_tail(queue, &thread->runq_links);
+ preempt = TRUE;
+ } else {
+ /* Insert into rt_runq in thread deadline order */
+ queue_entry_t iter;
+ qe_foreach(iter, queue) {
+ thread_t iter_thread = qe_element(iter, struct thread, runq_links);
+ assert_thread_magic(iter_thread);
+
+ if (deadline < iter_thread->realtime.deadline) {
+ if (iter == queue_first(queue)) {
+ preempt = TRUE;
+ }
+ insque(&thread->runq_links, queue_prev(iter));
+ break;
+ } else if (iter == queue_last(queue)) {
+ enqueue_tail(queue, &thread->runq_links);
+ break;
+ }
+ }
+ }
+
thread->runq = processor;
- runq_consider_incr_bound_count(processor, thread);
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_incr(pset);
+
+ return preempt;
+}
+
+#define MAX_BACKUP_PROCESSORS 7
+#if defined(__x86_64__)
+#define DEFAULT_BACKUP_PROCESSORS 1
+#else
+#define DEFAULT_BACKUP_PROCESSORS 0
+#endif
- return (result);
+int sched_rt_n_backup_processors = DEFAULT_BACKUP_PROCESSORS;
+
+int
+sched_get_rt_n_backup_processors(void)
+{
+ return sched_rt_n_backup_processors;
}
-#endif /* CONFIG_SCHED_TRADITIONAL */
+void
+sched_set_rt_n_backup_processors(int n)
+{
+ if (n < 0) {
+ n = 0;
+ } else if (n > MAX_BACKUP_PROCESSORS) {
+ n = MAX_BACKUP_PROCESSORS;
+ }
+
+ sched_rt_n_backup_processors = n;
+}
/*
- * processor_setrun:
+ * realtime_setrun:
*
- * Dispatch a thread for execution on a
- * processor.
+ * Dispatch a thread for realtime execution.
*
* Thread must be locked. Associated pset must
* be locked, and is returned unlocked.
*/
static void
-processor_setrun(
- processor_t processor,
- thread_t thread,
- integer_t options)
+realtime_setrun(
+ processor_t chosen_processor,
+ thread_t thread)
{
- processor_set_t pset = processor->processor_set;
- ast_t preempt;
+ processor_set_t pset = chosen_processor->processor_set;
+ pset_assert_locked(pset);
+ ast_t preempt;
- thread->chosen_processor = processor;
+ int n_backup = 0;
- /*
- * Dispatch directly onto idle processor.
- */
- if ( (SCHED(direct_dispatch_to_idle_processors) ||
- thread->bound_processor == processor)
- && processor->state == PROCESSOR_IDLE) {
- remqueue((queue_entry_t)processor);
- enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
+ if (thread->realtime.constraint <= rt_constraint_threshold) {
+ n_backup = sched_rt_n_backup_processors;
+ }
+ assert((n_backup >= 0) && (n_backup <= MAX_BACKUP_PROCESSORS));
- processor->next_thread = thread;
- processor->deadline = UINT64_MAX;
- processor->state = PROCESSOR_DISPATCHING;
- pset_unlock(pset);
+ sched_ipi_type_t ipi_type[MAX_BACKUP_PROCESSORS + 1] = {};
+ processor_t ipi_processor[MAX_BACKUP_PROCESSORS + 1] = {};
- if (processor != current_processor())
- machine_signal_idle(processor);
- return;
- }
+ thread->chosen_processor = chosen_processor;
- /*
- * Set preemption mode.
- */
- 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->priority)
- preempt = AST_NONE;
- else
- preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE;
+ /* <rdar://problem/15102234> */
+ assert(thread->bound_processor == PROCESSOR_NULL);
- if (!SCHED(processor_enqueue)(processor, thread, options))
- preempt = AST_NONE;
+ realtime_queue_insert(chosen_processor, pset, thread);
- if (preempt != AST_NONE) {
- if (processor == current_processor()) {
- if (csw_check(processor) != AST_NONE)
- ast_on(preempt);
+ processor_t processor = chosen_processor;
+ bool chosen_process_is_secondary = chosen_processor->processor_primary != chosen_processor;
+
+ int count = 0;
+ for (int i = 0; i <= n_backup; i++) {
+ if (i > 0) {
+ processor = choose_processor_for_realtime_thread(pset, chosen_processor, chosen_process_is_secondary);
+ if ((processor == PROCESSOR_NULL) || (sched_avoid_cpu0 && (processor->cpu_id == 0))) {
+ break;
+ }
+ 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)-3, processor->cpu_id, processor->state, 0);
}
- else
- if ( processor->state == PROCESSOR_IDLE || processor->state == PROCESSOR_DISPATCHING) {
- machine_signal_idle(processor);
+ ipi_type[i] = SCHED_IPI_NONE;
+ ipi_processor[i] = processor;
+ count++;
+
+ 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;
}
- else
- if ( (processor->state == PROCESSOR_RUNNING ||
- processor->state == PROCESSOR_SHUTDOWN) &&
- (thread->sched_pri >= processor->current_pri ||
- processor->current_thmode == TH_MODE_FAIRSHARE)) {
- cause_ast_check(processor);
+
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = thread->realtime.deadline;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+ if (processor == current_processor()) {
+ ast_on(preempt);
+
+ if ((preempt & AST_URGENT) == AST_URGENT) {
+ bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
+
+ if ((preempt & AST_PREEMPT) == AST_PREEMPT) {
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ }
+ } else {
+ ipi_type[i] = sched_ipi_action(processor, thread, true, SCHED_IPI_EVENT_PREEMPT);
+ }
+ } else if (processor->state == PROCESSOR_DISPATCHING) {
+ if ((processor->current_pri < thread->sched_pri) || (processor->deadline > thread->realtime.deadline)) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = thread->realtime.deadline;
+ }
+ } else {
+ if (processor == current_processor()) {
+ ast_on(preempt);
+
+ if ((preempt & AST_URGENT) == AST_URGENT) {
+ bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
+
+ if ((preempt & AST_PREEMPT) == AST_PREEMPT) {
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ }
+ } else {
+ ipi_type[i] = sched_ipi_action(processor, thread, false, SCHED_IPI_EVENT_PREEMPT);
+ }
+ }
+ } else {
+ /* Selected processor was too busy, just keep thread enqueued and let other processors drain it naturally. */
}
}
- else
- if ( processor->state == PROCESSOR_SHUTDOWN &&
- thread->sched_pri >= processor->current_pri ) {
- cause_ast_check(processor);
- }
- else
- if ( processor->state == PROCESSOR_IDLE &&
- processor != current_processor() ) {
- machine_signal_idle(processor);
- }
pset_unlock(pset);
+
+ assert((count > 0) && (count <= (n_backup + 1)));
+ for (int i = 0; i < count; i++) {
+ assert(ipi_processor[i] != PROCESSOR_NULL);
+ sched_ipi_perform(ipi_processor[i], ipi_type[i]);
+ }
}
-#if defined(CONFIG_SCHED_TRADITIONAL)
-static boolean_t
-processor_queue_empty(processor_t processor)
+sched_ipi_type_t
+sched_ipi_deferred_policy(processor_set_t pset, processor_t dst,
+ __unused sched_ipi_event_t event)
{
- return runq_for_processor(processor)->count == 0;
-
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ if (!bit_test(pset->pending_deferred_AST_cpu_mask, dst->cpu_id)) {
+ return SCHED_IPI_DEFERRED;
+ }
+#else /* CONFIG_SCHED_DEFERRED_AST */
+ panic("Request for deferred IPI on an unsupported platform; pset: %p CPU: %d", pset, dst->cpu_id);
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+ return SCHED_IPI_NONE;
}
-static boolean_t
-sched_traditional_with_pset_runqueue_processor_queue_empty(processor_t processor)
+sched_ipi_type_t
+sched_ipi_action(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event)
{
- processor_set_t pset = processor->processor_set;
- int count = runq_for_processor(processor)->count;
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ assert(dst != NULL);
- /*
- * The pset runq contains the count of all runnable threads
- * for all processors in the pset. However, for threads that
- * are bound to another processor, the current "processor"
- * is not eligible to execute the thread. So we only
- * include bound threads that our bound to the current
- * "processor". This allows the processor to idle when the
- * count of eligible threads drops to 0, even if there's
- * a runnable thread bound to a different processor in the
- * shared runq.
- */
+ processor_set_t pset = dst->processor_set;
+ if (current_processor() == dst) {
+ return SCHED_IPI_NONE;
+ }
- count -= pset->pset_runq_bound_count;
- count += processor->runq_bound_count;
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, dst->cpu_id)) {
+ return SCHED_IPI_NONE;
+ }
- return count == 0;
+ ipi_type = SCHED(ipi_policy)(dst, thread, dst_idle, event);
+ switch (ipi_type) {
+ case SCHED_IPI_NONE:
+ return SCHED_IPI_NONE;
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ case SCHED_IPI_DEFERRED:
+ bit_set(pset->pending_deferred_AST_cpu_mask, dst->cpu_id);
+ break;
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+ default:
+ bit_set(pset->pending_AST_URGENT_cpu_mask, dst->cpu_id);
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, dst->cpu_id);
+ break;
+ }
+ return ipi_type;
}
-static ast_t
-processor_csw_check(processor_t processor)
+sched_ipi_type_t
+sched_ipi_policy(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event)
{
- run_queue_t runq;
- boolean_t has_higher;
-
- assert(processor->active_thread != NULL);
-
- runq = runq_for_processor(processor);
- if (first_timeslice(processor)) {
- has_higher = (runq->highq > processor->current_pri);
- } else {
- has_higher = (runq->highq >= processor->current_pri);
- }
- if (has_higher) {
- if (runq->urgency > 0)
- return (AST_PREEMPT | AST_URGENT);
-
- if (processor->active_thread && thread_eager_preemption(processor->active_thread))
- return (AST_PREEMPT | AST_URGENT);
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ boolean_t deferred_ipi_supported = false;
+ processor_set_t pset = dst->processor_set;
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ deferred_ipi_supported = true;
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+
+ switch (event) {
+ case SCHED_IPI_EVENT_SPILL:
+ case SCHED_IPI_EVENT_SMT_REBAL:
+ case SCHED_IPI_EVENT_REBALANCE:
+ case SCHED_IPI_EVENT_BOUND_THR:
+ /*
+ * The spill, SMT rebalance, rebalance and the bound thread
+ * scenarios use immediate IPIs always.
+ */
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ case SCHED_IPI_EVENT_PREEMPT:
+ /* In the preemption case, use immediate IPIs for RT threads */
+ if (thread && (thread->sched_pri >= BASEPRI_RTQUEUES)) {
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ }
- return AST_PREEMPT;
+ /*
+ * For Non-RT threads preemption,
+ * If the core is active, use immediate IPIs.
+ * If the core is idle, use deferred IPIs if supported; otherwise immediate IPI.
+ */
+ if (deferred_ipi_supported && dst_idle) {
+ return sched_ipi_deferred_policy(pset, dst, event);
+ }
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ default:
+ panic("Unrecognized scheduler IPI event type %d", event);
}
-
- return AST_NONE;
+ assert(ipi_type != SCHED_IPI_NONE);
+ return ipi_type;
}
-static boolean_t
-processor_queue_has_priority(processor_t processor,
- int priority,
- boolean_t gte)
+void
+sched_ipi_perform(processor_t dst, sched_ipi_type_t ipi)
{
- if (gte)
- return runq_for_processor(processor)->highq >= priority;
- else
- return runq_for_processor(processor)->highq > priority;
+ switch (ipi) {
+ case SCHED_IPI_NONE:
+ break;
+ case SCHED_IPI_IDLE:
+ machine_signal_idle(dst);
+ break;
+ case SCHED_IPI_IMMEDIATE:
+ cause_ast_check(dst);
+ break;
+ case SCHED_IPI_DEFERRED:
+ machine_signal_idle_deferred(dst);
+ break;
+ default:
+ panic("Unrecognized scheduler IPI type: %d", ipi);
+ }
}
-static boolean_t
-should_current_thread_rechoose_processor(processor_t processor)
-{
- return (processor->current_pri < BASEPRI_RTQUEUES
- && processor->processor_meta != PROCESSOR_META_NULL
- && processor->processor_meta->primary != processor);
-}
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
-static int
-sched_traditional_processor_runq_count(processor_t processor)
+boolean_t
+priority_is_urgent(int priority)
{
- return runq_for_processor(processor)->count;
+ return bitmap_test(sched_preempt_pri, priority) ? TRUE : FALSE;
}
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
-static uint64_t
-sched_traditional_processor_runq_stats_count_sum(processor_t processor)
+/*
+ * 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)
{
- return runq_for_processor(processor)->runq_stats.count_sum;
-}
+ processor_set_t pset = processor->processor_set;
+ pset_assert_locked(pset);
+ ast_t preempt;
+ enum { eExitIdle, eInterruptRunning, eDoNothing } ipi_action = eDoNothing;
-static uint64_t
-sched_traditional_with_pset_runqueue_processor_runq_stats_count_sum(processor_t processor)
-{
- if (processor->cpu_id == processor->processor_set->cpu_set_low)
- return runq_for_processor(processor)->runq_stats.count_sum;
- else
- return 0ULL;
-}
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+
+ thread->chosen_processor = processor;
+
+ /*
+ * Set preemption mode.
+ */
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ /* TODO: Do we need to care about urgency (see rdar://problem/20136239)? */
+#endif
+ if (SCHED(priority_is_urgent)(thread->sched_pri) && thread->sched_pri > processor->current_pri) {
+ preempt = (AST_PREEMPT | AST_URGENT);
+ } else if (processor->current_is_eagerpreempt) {
+ preempt = (AST_PREEMPT | AST_URGENT);
+ } else if ((thread->sched_mode == TH_MODE_TIMESHARE) && (thread->sched_pri < thread->base_pri)) {
+ if (SCHED(priority_is_urgent)(thread->base_pri) && thread->sched_pri > processor->current_pri) {
+ preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE;
+ } else {
+ preempt = AST_NONE;
+ }
+ } else {
+ preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE;
+ }
+
+ if ((options & (SCHED_PREEMPT | SCHED_REBALANCE)) == (SCHED_PREEMPT | SCHED_REBALANCE)) {
+ /*
+ * Having gone to the trouble of forcing this thread off a less preferred core,
+ * we should force the preferable core to reschedule immediately to give this
+ * thread a chance to run instead of just sitting on the run queue where
+ * it may just be stolen back by the idle core we just forced it off.
+ */
+ preempt |= AST_PREEMPT;
+ }
+
+ SCHED(processor_enqueue)(processor, thread, options);
+ sched_update_pset_load_average(pset, 0);
+
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+ ipi_action = eExitIdle;
+ } else if (processor->state == PROCESSOR_DISPATCHING) {
+ if (processor->current_pri < thread->sched_pri) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ }
+ } else if ((processor->state == PROCESSOR_RUNNING ||
+ processor->state == PROCESSOR_SHUTDOWN) &&
+ (thread->sched_pri >= processor->current_pri)) {
+ ipi_action = eInterruptRunning;
+ }
+ } else {
+ /*
+ * New thread is not important enough to preempt what is running, but
+ * special processor states may need special handling
+ */
+ if (processor->state == PROCESSOR_SHUTDOWN &&
+ thread->sched_pri >= processor->current_pri) {
+ ipi_action = eInterruptRunning;
+ } else if (processor->state == PROCESSOR_IDLE) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+
+ ipi_action = eExitIdle;
+ }
+ }
+
+ if (ipi_action != eDoNothing) {
+ if (processor == current_processor()) {
+ if ((preempt = csw_check_locked(processor->active_thread, processor, pset, AST_NONE)) != AST_NONE) {
+ ast_on(preempt);
+ }
-#endif /* CONFIG_SCHED_TRADITIONAL */
+ if ((preempt & AST_URGENT) == AST_URGENT) {
+ bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ } else {
+ bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
-#define next_pset(p) (((p)->pset_list != PROCESSOR_SET_NULL)? (p)->pset_list: (p)->node->psets)
+ if ((preempt & AST_PREEMPT) == AST_PREEMPT) {
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ } else {
+ bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ }
+ } else {
+ sched_ipi_event_t event = (options & SCHED_REBALANCE) ? SCHED_IPI_EVENT_REBALANCE : SCHED_IPI_EVENT_PREEMPT;
+ ipi_type = sched_ipi_action(processor, thread, (ipi_action == eExitIdle), event);
+ }
+ }
+ pset_unlock(pset);
+ sched_ipi_perform(processor, ipi_type);
+}
/*
* choose_next_pset:
*/
static processor_set_t
choose_next_pset(
- processor_set_t pset)
+ processor_set_t pset)
{
- processor_set_t nset = pset;
+ processor_set_t nset = pset;
do {
nset = next_pset(nset);
} while (nset->online_processor_count < 1 && nset != pset);
- return (nset);
+ return nset;
+}
+
+inline static processor_set_t
+change_locked_pset(processor_set_t current_pset, processor_set_t new_pset)
+{
+ if (current_pset != new_pset) {
+ pset_unlock(current_pset);
+ pset_lock(new_pset);
+ }
+
+ return new_pset;
}
/*
*/
processor_t
choose_processor(
- processor_set_t pset,
- processor_t processor,
- thread_t thread)
-{
- processor_set_t nset, cset = pset;
- processor_meta_t pmeta = PROCESSOR_META_NULL;
- processor_t mprocessor;
-
+ processor_set_t starting_pset,
+ processor_t processor,
+ thread_t thread)
+{
+ processor_set_t pset = starting_pset;
+ processor_set_t nset;
+
+ assert(thread->sched_pri <= BASEPRI_RTQUEUES);
+
/*
- * Prefer the hinted processor, when appropriate.
+ * Prefer the hinted processor, when appropriate.
*/
+ /* Fold last processor hint from secondary processor to its primary */
if (processor != PROCESSOR_NULL) {
- if (processor->processor_meta != PROCESSOR_META_NULL)
- processor = processor->processor_meta->primary;
+ processor = processor->processor_primary;
}
- mprocessor = machine_choose_processor(pset, processor);
- if (mprocessor != PROCESSOR_NULL)
- processor = mprocessor;
+ /*
+ * 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->state == PROCESSOR_INACTIVE ||
- processor->state == PROCESSOR_SHUTDOWN ||
- processor->state == PROCESSOR_OFF_LINE)
+ if (processor->processor_set != pset) {
processor = PROCESSOR_NULL;
- else
- if (processor->state == PROCESSOR_IDLE ||
- ((thread->sched_pri >= BASEPRI_RTQUEUES) &&
- (processor->current_pri < BASEPRI_RTQUEUES)))
- return (processor);
+ } 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.
+ */
+ if ((thread->sched_pri < BASEPRI_RTQUEUES) || processor_is_fast_track_candidate_for_realtime_thread(pset, processor)) {
+ return processor;
+ }
+ processor = PROCESSOR_NULL;
+ break;
+ 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_is_fast_track_candidate_for_realtime_thread(pset, processor)) {
+ 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.
+ * 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).
*/
- do {
- /*
- * Choose an idle processor.
- */
- if (!queue_empty(&cset->idle_queue))
- return ((processor_t)queue_first(&cset->idle_queue));
- if (thread->sched_pri >= BASEPRI_RTQUEUES) {
- integer_t lowest_priority = MAXPRI + 1;
- integer_t lowest_unpaired = MAXPRI + 1;
- uint64_t furthest_deadline = 1;
- processor_t lp_processor = PROCESSOR_NULL;
- processor_t lp_unpaired = PROCESSOR_NULL;
- processor_t fd_processor = PROCESSOR_NULL;
-
- lp_processor = cset->low_pri;
- /* Consider hinted processor */
- if (lp_processor != PROCESSOR_NULL &&
- ((lp_processor->processor_meta == PROCESSOR_META_NULL) ||
- ((lp_processor == lp_processor->processor_meta->primary) &&
- !queue_empty(&lp_processor->processor_meta->idle_queue))) &&
- lp_processor->state != PROCESSOR_INACTIVE &&
- lp_processor->state != PROCESSOR_SHUTDOWN &&
- lp_processor->state != PROCESSOR_OFF_LINE &&
- (lp_processor->current_pri < thread->sched_pri))
- return lp_processor;
-
- processor = (processor_t)queue_first(&cset->active_queue);
- while (!queue_end(&cset->active_queue, (queue_entry_t)processor)) {
- /* Discover the processor executing the
- * thread with the lowest priority within
- * this pset, or the one with the furthest
- * deadline
- */
- integer_t cpri = processor->current_pri;
- if (cpri < lowest_priority) {
- lowest_priority = cpri;
- lp_processor = processor;
- }
+ integer_t lowest_priority = MAXPRI + 1;
+ integer_t lowest_secondary_priority = MAXPRI + 1;
+ integer_t lowest_unpaired_primary_priority = MAXPRI + 1;
+ integer_t lowest_idle_secondary_priority = MAXPRI + 1;
+ integer_t lowest_count = INT_MAX;
+ uint64_t furthest_deadline = 1;
+ processor_t lp_processor = PROCESSOR_NULL;
+ processor_t lp_unpaired_primary_processor = PROCESSOR_NULL;
+ processor_t lp_idle_secondary_processor = PROCESSOR_NULL;
+ processor_t lp_paired_secondary_processor = PROCESSOR_NULL;
+ processor_t lc_processor = PROCESSOR_NULL;
+ processor_t fd_processor = PROCESSOR_NULL;
- if ((cpri >= BASEPRI_RTQUEUES) && (processor->deadline > furthest_deadline)) {
- furthest_deadline = processor->deadline;
- fd_processor = processor;
- }
+ 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->processor_meta != PROCESSOR_META_NULL &&
- !queue_empty(&processor->processor_meta->idle_queue)) {
- if (cpri < lowest_unpaired) {
- lowest_unpaired = cpri;
- lp_unpaired = processor;
- pmeta = processor->processor_meta;
- }
- else
- if (pmeta == PROCESSOR_META_NULL)
- pmeta = processor->processor_meta;
+ if (processor->current_pri >= BASEPRI_RTQUEUES) {
+ furthest_deadline = processor->deadline;
+ fd_processor = processor;
+ }
+
+ lowest_count = SCHED(processor_runq_count)(processor);
+ lc_processor = processor;
+ }
+
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+ pset_node_t node = pset->node;
+ int consider_secondaries = (!pset->is_SMT) || (bit_count(node->pset_map) == 1) || (node->pset_non_rt_primary_map == 0);
+ for (; consider_secondaries < 2; consider_secondaries++) {
+ pset = change_locked_pset(pset, starting_pset);
+ do {
+ processor = choose_processor_for_realtime_thread(pset, PROCESSOR_NULL, consider_secondaries);
+ if (processor) {
+ return processor;
}
- processor = (processor_t)queue_next((queue_entry_t)processor);
- }
- if (thread->sched_pri > lowest_unpaired)
- return lp_unpaired;
+ /* NRG Collect processor stats for furthest deadline etc. here */
- if (pmeta != PROCESSOR_META_NULL)
- return ((processor_t)queue_first(&pmeta->idle_queue));
- if (thread->sched_pri > lowest_priority)
- return lp_processor;
- if (thread->realtime.deadline < furthest_deadline)
- return fd_processor;
+ nset = next_pset(pset);
- processor = PROCESSOR_NULL;
+ if (nset != starting_pset) {
+ pset = change_locked_pset(pset, nset);
+ }
+ } while (nset != starting_pset);
}
- else {
- /*
- * Check any hinted processors in the processor set if available.
- */
- if (cset->low_pri != PROCESSOR_NULL && cset->low_pri->state != PROCESSOR_INACTIVE &&
- cset->low_pri->state != PROCESSOR_SHUTDOWN && cset->low_pri->state != PROCESSOR_OFF_LINE &&
- (processor == PROCESSOR_NULL ||
- (thread->sched_pri > BASEPRI_DEFAULT && cset->low_pri->current_pri < thread->sched_pri))) {
- processor = cset->low_pri;
- }
- else
- if (cset->low_count != PROCESSOR_NULL && cset->low_count->state != PROCESSOR_INACTIVE &&
- cset->low_count->state != PROCESSOR_SHUTDOWN && cset->low_count->state != PROCESSOR_OFF_LINE &&
- (processor == PROCESSOR_NULL || (thread->sched_pri <= BASEPRI_DEFAULT &&
- SCHED(processor_runq_count)(cset->low_count) < SCHED(processor_runq_count)(processor)))) {
- processor = cset->low_count;
+ /* Or we could just let it change to starting_pset in the loop above */
+ pset = change_locked_pset(pset, starting_pset);
+ }
+
+ do {
+ /*
+ * Choose an idle processor, in pset traversal order
+ */
+
+ uint64_t idle_primary_map = (pset->cpu_state_map[PROCESSOR_IDLE] &
+ pset->primary_map &
+ pset->recommended_bitmask);
+
+ /* there shouldn't be a pending AST if the processor is idle */
+ assert((idle_primary_map & pset->pending_AST_URGENT_cpu_mask) == 0);
+
+ int cpuid = lsb_first(idle_primary_map);
+ if (cpuid >= 0) {
+ processor = processor_array[cpuid];
+ return processor;
+ }
+
+ /*
+ * Otherwise, enumerate active and idle processors to find primary candidates
+ * with lower priority/etc.
+ */
+
+ uint64_t active_map = ((pset->cpu_state_map[PROCESSOR_RUNNING] | pset->cpu_state_map[PROCESSOR_DISPATCHING]) &
+ pset->recommended_bitmask &
+ ~pset->pending_AST_URGENT_cpu_mask);
+
+ if (SCHED(priority_is_urgent)(thread->sched_pri) == FALSE) {
+ active_map &= ~pset->pending_AST_PREEMPT_cpu_mask;
+ }
+
+ active_map = bit_ror64(active_map, (pset->last_chosen + 1));
+ for (int rotid = lsb_first(active_map); rotid >= 0; rotid = lsb_next(active_map, rotid)) {
+ cpuid = ((rotid + pset->last_chosen + 1) & 63);
+ processor = processor_array[cpuid];
+
+ integer_t cpri = processor->current_pri;
+ processor_t primary = processor->processor_primary;
+ if (primary != processor) {
+ /* If primary is running a NO_SMT thread, don't choose its secondary */
+ if (!((primary->state == PROCESSOR_RUNNING) && processor_active_thread_no_smt(primary))) {
+ if (cpri < lowest_secondary_priority) {
+ lowest_secondary_priority = cpri;
+ lp_paired_secondary_processor = processor;
+ }
+ }
+ } else {
+ if (cpri < lowest_priority) {
+ lowest_priority = cpri;
+ lp_processor = processor;
+ }
}
- /*
- * Otherwise, choose an available processor in the set.
- */
- if (processor == PROCESSOR_NULL) {
- processor = (processor_t)dequeue_head(&cset->active_queue);
- if (processor != PROCESSOR_NULL)
- enqueue_tail(&cset->active_queue, (queue_entry_t)processor);
+ if ((cpri >= BASEPRI_RTQUEUES) && (processor->deadline > furthest_deadline)) {
+ furthest_deadline = processor->deadline;
+ fd_processor = processor;
}
- if (processor != PROCESSOR_NULL && pmeta == PROCESSOR_META_NULL) {
- if (processor->processor_meta != PROCESSOR_META_NULL &&
- !queue_empty(&processor->processor_meta->idle_queue))
- pmeta = processor->processor_meta;
+ integer_t ccount = SCHED(processor_runq_count)(processor);
+ if (ccount < lowest_count) {
+ lowest_count = ccount;
+ lc_processor = processor;
}
}
/*
- * Move onto the next processor set.
+ * For SMT configs, these idle secondary processors must have active primary. Otherwise
+ * the idle primary would have short-circuited the loop above
*/
- nset = next_pset(cset);
+ uint64_t idle_secondary_map = (pset->cpu_state_map[PROCESSOR_IDLE] &
+ ~pset->primary_map &
+ pset->recommended_bitmask);
- if (nset != pset) {
- pset_unlock(cset);
+ /* there shouldn't be a pending AST if the processor is idle */
+ assert((idle_secondary_map & pset->pending_AST_URGENT_cpu_mask) == 0);
+ assert((idle_secondary_map & pset->pending_AST_PREEMPT_cpu_mask) == 0);
- cset = nset;
- pset_lock(cset);
- }
- } while (nset != pset);
+ for (cpuid = lsb_first(idle_secondary_map); cpuid >= 0; cpuid = lsb_next(idle_secondary_map, cpuid)) {
+ processor = processor_array[cpuid];
- /*
- * Make sure that we pick a running processor,
- * and that the correct processor set is locked.
- */
- do {
- if (pmeta != PROCESSOR_META_NULL) {
- if (cset != pmeta->primary->processor_set) {
- pset_unlock(cset);
+ processor_t cprimary = processor->processor_primary;
+
+ integer_t primary_pri = cprimary->current_pri;
+
+ /*
+ * TODO: This should also make the same decisions
+ * as secondary_can_run_realtime_thread
+ *
+ * TODO: Keep track of the pending preemption priority
+ * of the primary to make this more accurate.
+ */
+
+ /* If the primary is running a no-smt thread, then don't choose its secondary */
+ if (cprimary->state == PROCESSOR_RUNNING &&
+ processor_active_thread_no_smt(cprimary)) {
+ continue;
+ }
- cset = pmeta->primary->processor_set;
- pset_lock(cset);
+ /*
+ * Find the idle secondary processor with the lowest priority primary
+ *
+ * We will choose this processor as a fallback if we find no better
+ * primary to preempt.
+ */
+ if (primary_pri < lowest_idle_secondary_priority) {
+ lp_idle_secondary_processor = processor;
+ lowest_idle_secondary_priority = primary_pri;
}
- if (!queue_empty(&pmeta->idle_queue))
- return ((processor_t)queue_first(&pmeta->idle_queue));
+ /* Find the the lowest priority active primary with idle secondary */
+ if (primary_pri < lowest_unpaired_primary_priority) {
+ /* If the primary processor is offline or starting up, it's not a candidate for this path */
+ if (cprimary->state != PROCESSOR_RUNNING &&
+ cprimary->state != PROCESSOR_DISPATCHING) {
+ continue;
+ }
+
+ if (!cprimary->is_recommended) {
+ continue;
+ }
+
+ /* if the primary is pending preemption, don't try to re-preempt it */
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, cprimary->cpu_id)) {
+ continue;
+ }
+
+ if (SCHED(priority_is_urgent)(thread->sched_pri) == FALSE &&
+ bit_test(pset->pending_AST_PREEMPT_cpu_mask, cprimary->cpu_id)) {
+ continue;
+ }
- pmeta = PROCESSOR_META_NULL;
+ lowest_unpaired_primary_priority = primary_pri;
+ lp_unpaired_primary_processor = cprimary;
+ }
}
/*
- * If we haven't been able to choose a processor,
- * pick the boot processor and return it.
+ * We prefer preempting a primary processor over waking up its secondary.
+ * The secondary will then be woken up by the preempted thread.
*/
- if (processor == PROCESSOR_NULL) {
- processor = master_processor;
+ if (thread->sched_pri > lowest_unpaired_primary_priority) {
+ pset->last_chosen = lp_unpaired_primary_processor->cpu_id;
+ return lp_unpaired_primary_processor;
+ }
+
+ /*
+ * We prefer preempting a lower priority active processor over directly
+ * waking up an idle secondary.
+ * The preempted thread will then find the idle secondary.
+ */
+ if (thread->sched_pri > lowest_priority) {
+ pset->last_chosen = lp_processor->cpu_id;
+ return lp_processor;
+ }
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
/*
- * Check that the correct processor set is
- * returned locked.
+ * For realtime threads, the most important aspect is
+ * scheduling latency, so we will pick an active
+ * secondary processor in this pset, or preempt
+ * another RT thread with a further deadline before
+ * going to the next pset.
*/
- if (cset != processor->processor_set) {
- pset_unlock(cset);
- cset = processor->processor_set;
- pset_lock(cset);
+ if (sched_allow_rt_smt && (thread->sched_pri > lowest_secondary_priority)) {
+ pset->last_chosen = lp_paired_secondary_processor->cpu_id;
+ return lp_paired_secondary_processor;
+ }
+
+ if (thread->realtime.deadline < furthest_deadline) {
+ return fd_processor;
}
+ }
- return (processor);
+ /*
+ * lc_processor is used to indicate the best processor set run queue
+ * on which to enqueue a thread when all available CPUs are busy with
+ * higher priority threads, so try to make sure it is initialized.
+ */
+ if (lc_processor == PROCESSOR_NULL) {
+ cpumap_t available_map = ((pset->cpu_state_map[PROCESSOR_IDLE] |
+ pset->cpu_state_map[PROCESSOR_RUNNING] |
+ pset->cpu_state_map[PROCESSOR_DISPATCHING]) &
+ pset->recommended_bitmask);
+ cpuid = lsb_first(available_map);
+ if (cpuid >= 0) {
+ lc_processor = processor_array[cpuid];
+ lowest_count = SCHED(processor_runq_count)(lc_processor);
+ }
}
/*
- * Check that the processor set for the chosen
- * processor is locked.
+ * Move onto the next processor set.
+ *
+ * If all primary processors in this pset are running a higher
+ * priority thread, move on to next pset. Only when we have
+ * exhausted the search for primary processors do we
+ * fall back to secondaries.
*/
- if (cset != processor->processor_set) {
- pset_unlock(cset);
+ nset = next_pset(pset);
+
+ if (nset != starting_pset) {
+ pset = change_locked_pset(pset, nset);
+ }
+ } while (nset != starting_pset);
- cset = processor->processor_set;
- pset_lock(cset);
+ /*
+ * Make sure that we pick a running processor,
+ * and that the correct processor set is locked.
+ * Since we may have unlocked the candidate processor's
+ * pset, it may have changed state.
+ *
+ * All primary processors are running a higher priority
+ * thread, so the only options left are enqueuing on
+ * the secondary processor that would perturb the least priority
+ * primary, or the least busy primary.
+ */
+ boolean_t fallback_processor = false;
+ do {
+ /* lowest_priority is evaluated in the main loops above */
+ if (lp_idle_secondary_processor != PROCESSOR_NULL) {
+ processor = lp_idle_secondary_processor;
+ lp_idle_secondary_processor = PROCESSOR_NULL;
+ } else if (lp_paired_secondary_processor != PROCESSOR_NULL) {
+ processor = lp_paired_secondary_processor;
+ lp_paired_secondary_processor = PROCESSOR_NULL;
+ } else if (lc_processor != PROCESSOR_NULL) {
+ processor = lc_processor;
+ lc_processor = PROCESSOR_NULL;
+ } else {
+ /*
+ * All processors are executing higher priority threads, and
+ * the lowest_count candidate was not usable.
+ *
+ * For AMP platforms running the clutch scheduler always
+ * return a processor from the requested pset to allow the
+ * thread to be enqueued in the correct runq. For non-AMP
+ * platforms, simply return the master_processor.
+ */
+ fallback_processor = true;
+#if CONFIG_SCHED_EDGE
+ processor = processor_array[lsb_first(starting_pset->primary_map)];
+#else /* CONFIG_SCHED_EDGE */
+ processor = master_processor;
+#endif /* CONFIG_SCHED_EDGE */
}
/*
- * We must verify that the chosen processor is still available.
+ * Check that the correct processor set is
+ * returned locked.
+ */
+ pset = change_locked_pset(pset, processor->processor_set);
+
+ /*
+ * We must verify that the chosen processor is still available.
+ * The cases where we pick the master_processor or the fallback
+ * processor are execptions, since we may need enqueue a thread
+ * on its runqueue if this is the last remaining processor
+ * during pset shutdown.
+ *
+ * <rdar://problem/47559304> would really help here since it
+ * gets rid of the weird last processor SHUTDOWN case where
+ * the pset is still schedulable.
*/
- if (processor->state == PROCESSOR_INACTIVE ||
- processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE)
+ if (processor != master_processor && (fallback_processor == false) && (processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE)) {
processor = PROCESSOR_NULL;
+ }
} while (processor == PROCESSOR_NULL);
- return (processor);
+ pset->last_chosen = processor->cpu_id;
+ return processor;
+}
+
+/*
+ * Default implementation of SCHED(choose_node)()
+ * for single node systems
+ */
+pset_node_t
+sched_choose_node(__unused thread_t thread)
+{
+ return &pset_node0;
+}
+
+/*
+ * choose_starting_pset:
+ *
+ * Choose a starting processor set for the thread.
+ * May return a processor hint within the pset.
+ *
+ * Returns a starting processor set, to be used by
+ * choose_processor.
+ *
+ * The thread must be locked. The resulting pset is unlocked on return,
+ * and is chosen without taking any pset locks.
+ */
+processor_set_t
+choose_starting_pset(pset_node_t node, thread_t thread, processor_t *processor_hint)
+{
+ processor_set_t pset;
+ processor_t processor = PROCESSOR_NULL;
+
+ if (thread->affinity_set != AFFINITY_SET_NULL) {
+ /*
+ * Use affinity set policy hint.
+ */
+ pset = thread->affinity_set->aset_pset;
+ } else if (thread->last_processor != PROCESSOR_NULL) {
+ /*
+ * Simple (last processor) affinity case.
+ */
+ processor = thread->last_processor;
+ pset = processor->processor_set;
+ } else {
+ /*
+ * No Affinity case:
+ *
+ * Utilitize a per task hint to spread threads
+ * among the available processor sets.
+ * NRG this seems like the wrong thing to do.
+ * See also task->pset_hint = pset in thread_setrun()
+ */
+ task_t task = thread->task;
+
+ pset = task->pset_hint;
+ if (pset == PROCESSOR_SET_NULL) {
+ pset = current_processor()->processor_set;
+ }
+
+ pset = choose_next_pset(pset);
+ }
+
+ if (!bit_test(node->pset_map, pset->pset_id)) {
+ /* pset is not from this node so choose one that is */
+ int id = lsb_first(node->pset_map);
+ assert(id >= 0);
+ pset = pset_array[id];
+ }
+
+ if (bit_count(node->pset_map) == 1) {
+ /* Only a single pset in this node */
+ goto out;
+ }
+
+ bool avoid_cpu0 = false;
+
+#if defined(__x86_64__)
+ if ((thread->sched_pri >= BASEPRI_RTQUEUES) && sched_avoid_cpu0) {
+ /* Avoid the pset containing cpu0 */
+ avoid_cpu0 = true;
+ /* Assert that cpu0 is in pset0. I expect this to be true on __x86_64__ */
+ assert(bit_test(pset_array[0]->cpu_bitmask, 0));
+ }
+#endif
+
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+ pset_map_t rt_target_map = atomic_load(&node->pset_non_rt_primary_map);
+ if ((avoid_cpu0 && pset->pset_id == 0) || !bit_test(rt_target_map, pset->pset_id)) {
+ if (avoid_cpu0) {
+ rt_target_map = bit_ror64(rt_target_map, 1);
+ }
+ int rotid = lsb_first(rt_target_map);
+ if (rotid >= 0) {
+ int id = avoid_cpu0 ? ((rotid + 1) & 63) : rotid;
+ pset = pset_array[id];
+ goto out;
+ }
+ }
+ if (!pset->is_SMT || !sched_allow_rt_smt) {
+ /* All psets are full of RT threads - fall back to choose processor to find the furthest deadline RT thread */
+ goto out;
+ }
+ rt_target_map = atomic_load(&node->pset_non_rt_map);
+ if ((avoid_cpu0 && pset->pset_id == 0) || !bit_test(rt_target_map, pset->pset_id)) {
+ if (avoid_cpu0) {
+ rt_target_map = bit_ror64(rt_target_map, 1);
+ }
+ int rotid = lsb_first(rt_target_map);
+ if (rotid >= 0) {
+ int id = avoid_cpu0 ? ((rotid + 1) & 63) : rotid;
+ pset = pset_array[id];
+ goto out;
+ }
+ }
+ /* All psets are full of RT threads - fall back to choose processor to find the furthest deadline RT thread */
+ } else {
+ pset_map_t idle_map = atomic_load(&node->pset_idle_map);
+ if (!bit_test(idle_map, pset->pset_id)) {
+ int next_idle_pset_id = lsb_first(idle_map);
+ if (next_idle_pset_id >= 0) {
+ pset = pset_array[next_idle_pset_id];
+ }
+ }
+ }
+
+out:
+ if ((processor != PROCESSOR_NULL) && (processor->processor_set != pset)) {
+ processor = PROCESSOR_NULL;
+ }
+ if (processor != PROCESSOR_NULL) {
+ *processor_hint = processor;
+ }
+
+ return pset;
}
/*
*/
void
thread_setrun(
- thread_t thread,
- integer_t options)
+ thread_t thread,
+ sched_options_t options)
{
- processor_t processor;
- processor_set_t pset;
+ processor_t processor;
+ processor_set_t pset;
+
+ assert((thread->state & (TH_RUN | TH_WAIT | TH_UNINT | TH_TERMINATE | TH_TERMINATE2)) == TH_RUN);
+ assert(thread->runq == PROCESSOR_NULL);
-#if DEBUG
- assert(thread_runnable(thread));
-#endif
-
/*
* Update priority if needed.
*/
- if (SCHED(can_update_priority)(thread))
+ if (SCHED(can_update_priority)(thread)) {
SCHED(update_priority)(thread);
+ }
+
+ thread->sfi_class = sfi_thread_classify(thread);
assert(thread->runq == PROCESSOR_NULL);
/*
* Unbound case.
*/
- if (thread->affinity_set != AFFINITY_SET_NULL) {
- /*
- * Use affinity set policy hint.
- */
- pset = thread->affinity_set->aset_pset;
- pset_lock(pset);
-
- processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread);
- }
- 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 ((thread->last_processor != processor) && (thread->last_processor != PROCESSOR_NULL)) {
- KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_LPA_BROKEN)|DBG_FUNC_NONE,
- (uintptr_t)thread_tid(thread), (uintptr_t)thread->last_processor->cpu_id, (uintptr_t)processor->cpu_id, thread->last_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;
+ processor_t processor_hint = PROCESSOR_NULL;
+ pset_node_t node = SCHED(choose_node)(thread);
+ processor_set_t starting_pset = choose_starting_pset(node, thread, &processor_hint);
- pset = task->pset_hint;
- if (pset == PROCESSOR_SET_NULL)
- pset = current_processor()->processor_set;
+ pset_lock(starting_pset);
- pset = choose_next_pset(pset);
- pset_lock(pset);
+ processor = SCHED(choose_processor)(starting_pset, processor_hint, thread);
+ pset = processor->processor_set;
+ task_t task = thread->task;
+ task->pset_hint = pset; /* NRG this is done without holding the task lock */
- processor = SCHED(choose_processor)(pset, PROCESSOR_NULL, thread);
- task->pset_hint = processor->processor_set;
- }
- }
- else {
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-1, processor->cpu_id, processor->state, 0);
+ } else {
/*
* Bound case:
*
processor = thread->bound_processor;
pset = processor->processor_set;
pset_lock(pset);
+
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-2, processor->cpu_id, processor->state, 0);
}
/*
- * Dispatch the thread on the choosen processor.
+ * Dispatch the thread on the chosen processor.
+ * TODO: This should be based on sched_mode, not sched_pri
*/
- if (thread->sched_pri >= BASEPRI_RTQUEUES)
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
realtime_setrun(processor, thread);
- else if (thread->sched_mode == TH_MODE_FAIRSHARE)
- fairshare_setrun(processor, thread);
- else
+ } else {
processor_setrun(processor, thread, options);
+ }
+ /* pset is now unlocked */
+ if (thread->bound_processor == PROCESSOR_NULL) {
+ SCHED(check_spill)(pset, thread);
+ }
}
processor_set_t
task_choose_pset(
- task_t task)
+ task_t task)
{
- processor_set_t pset = task->pset_hint;
+ processor_set_t pset = task->pset_hint;
- if (pset != PROCESSOR_SET_NULL)
+ if (pset != PROCESSOR_SET_NULL) {
pset = choose_next_pset(pset);
+ }
- return (pset);
+ return pset;
}
-#if defined(CONFIG_SCHED_TRADITIONAL)
-
/*
- * processor_queue_shutdown:
- *
- * Shutdown a processor run queue by
- * re-dispatching non-bound threads.
+ * Check for a preemption point in
+ * the current context.
*
- * Associated pset must be locked, and is
- * returned unlocked.
+ * Called at splsched with thread locked.
*/
-void
-processor_queue_shutdown(
- processor_t processor)
+ast_t
+csw_check(
+ thread_t thread,
+ processor_t processor,
+ ast_t check_reason)
{
- processor_set_t pset = processor->processor_set;
- run_queue_t rq = runq_for_processor(processor);
- queue_t queue = rq->queues + rq->highq;
- int pri = rq->highq, count = rq->count;
- thread_t next, thread;
- queue_head_t tqueue;
-
- queue_init(&tqueue);
-
- while (count > 0) {
- thread = (thread_t)queue_first(queue);
- while (!queue_end(queue, (queue_entry_t)thread)) {
- next = (thread_t)queue_next((queue_entry_t)thread);
-
- if (thread->bound_processor == PROCESSOR_NULL) {
- remqueue((queue_entry_t)thread);
-
- thread->runq = PROCESSOR_NULL;
- SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
- runq_consider_decr_bound_count(processor, thread);
- rq->count--;
- if (SCHED(priority_is_urgent)(pri)) {
- rq->urgency--; assert(rq->urgency >= 0);
- }
- if (queue_empty(queue)) {
- if (pri != IDLEPRI)
- clrbit(MAXPRI - pri, rq->bitmap);
- rq->highq = MAXPRI - ffsbit(rq->bitmap);
- }
+ processor_set_t pset = processor->processor_set;
- enqueue_tail(&tqueue, (queue_entry_t)thread);
- }
- count--;
+ assert(thread == processor->active_thread);
- thread = next;
- }
+ pset_lock(pset);
- queue--; pri--;
- }
+ processor_state_update_from_thread(processor, thread);
- pset_unlock(pset);
+ ast_t preempt = csw_check_locked(thread, processor, pset, check_reason);
- while ((thread = (thread_t)dequeue_head(&tqueue)) != THREAD_NULL) {
- thread_lock(thread);
+ /* Acknowledge the IPI if we decided not to preempt */
- thread_setrun(thread, SCHED_TAILQ);
+ if ((preempt & AST_URGENT) == 0) {
+ bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
- thread_unlock(thread);
+ if ((preempt & AST_PREEMPT) == 0) {
+ bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
}
-}
-#endif /* CONFIG_SCHED_TRADITIONAL */
+ pset_unlock(pset);
+
+ return preempt;
+}
/*
- * Check for a preemption point in
- * the current context.
- *
- * Called at splsched.
+ * Check for preemption at splsched with
+ * pset and thread locked
*/
ast_t
-csw_check(
- processor_t processor)
+csw_check_locked(
+ thread_t thread,
+ processor_t processor,
+ processor_set_t pset,
+ ast_t check_reason)
{
- ast_t result = AST_NONE;
- thread_t thread = processor->active_thread;
+ ast_t result;
- if (first_timeslice(processor)) {
- if (rt_runq.count > 0)
- return (AST_PREEMPT | AST_URGENT);
+ if (processor->first_timeslice) {
+ if (rt_runq_count(pset) > 0) {
+ return check_reason | AST_PREEMPT | AST_URGENT;
+ }
+ } else {
+ if (rt_runq_count(pset) > 0) {
+ if (BASEPRI_RTQUEUES > processor->current_pri) {
+ return check_reason | AST_PREEMPT | AST_URGENT;
+ } else {
+ return check_reason | AST_PREEMPT;
+ }
+ }
}
- else {
- if (rt_runq.count > 0 && BASEPRI_RTQUEUES >= processor->current_pri)
- return (AST_PREEMPT | AST_URGENT);
+
+ /*
+ * If the current thread is running on a processor that is no longer recommended,
+ * urgently preempt it, at which point thread_select() should
+ * try to idle the processor and re-dispatch the thread to a recommended processor.
+ */
+ if (!processor->is_recommended) {
+ return check_reason | AST_PREEMPT | AST_URGENT;
}
result = SCHED(processor_csw_check)(processor);
- if (result != AST_NONE)
- return (result);
+ if (result != AST_NONE) {
+ return check_reason | result | (thread_is_eager_preempt(thread) ? AST_URGENT : AST_NONE);
+ }
- if (SCHED(should_current_thread_rechoose_processor)(processor))
- return (AST_PREEMPT);
-
- if (machine_processor_is_inactive(processor))
- return (AST_PREEMPT);
+ /*
+ * Same for avoid-processor
+ *
+ * TODO: Should these set AST_REBALANCE?
+ */
+ if (SCHED(avoid_processor_enabled) && SCHED(thread_avoid_processor)(processor, thread)) {
+ return check_reason | AST_PREEMPT;
+ }
+
+ /*
+ * 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
+ */
- if (thread->state & TH_SUSP)
- return (AST_PREEMPT);
+ if (processor->current_pri < BASEPRI_RTQUEUES &&
+ processor->processor_primary != processor) {
+ return check_reason | AST_PREEMPT;
+ }
+
+ if (thread->state & TH_SUSP) {
+ return check_reason | AST_PREEMPT;
+ }
+
+#if CONFIG_SCHED_SFI
+ /*
+ * Current thread may not need to be preempted, but maybe needs
+ * an SFI wait?
+ */
+ result = sfi_thread_needs_ast(thread, NULL);
+ if (result != AST_NONE) {
+ return check_reason | result;
+ }
+#endif
- return (AST_NONE);
+ return AST_NONE;
}
/*
- * set_sched_pri:
- *
- * Set the scheduled priority of the specified thread.
- *
- * This may cause the thread to change queues.
- *
- * Thread must be locked.
+ * Handle preemption IPI or IPI in response to setting an AST flag
+ * Triggered by cause_ast_check
+ * Called at splsched
*/
void
-set_sched_pri(
- thread_t thread,
- int priority)
+ast_check(processor_t processor)
{
- boolean_t removed = thread_run_queue_remove(thread);
+ if (processor->state != PROCESSOR_RUNNING &&
+ processor->state != PROCESSOR_SHUTDOWN) {
+ return;
+ }
- thread->sched_pri = priority;
- if (removed)
- thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
- else
- if (thread->state & TH_RUN) {
- processor_t processor = thread->last_processor;
+ thread_t thread = processor->active_thread;
- if (thread == current_thread()) {
- ast_t preempt;
+ assert(thread == current_thread());
- processor->current_pri = priority;
- processor->current_thmode = thread->sched_mode;
- if ((preempt = csw_check(processor)) != AST_NONE)
- ast_on(preempt);
- }
- else
- if ( processor != PROCESSOR_NULL &&
- processor->active_thread == thread )
- cause_ast_check(processor);
- }
-}
+ thread_lock(thread);
-#if 0
+ /*
+ * Propagate thread ast to processor.
+ * (handles IPI in response to setting AST flag)
+ */
+ ast_propagate(thread);
-static void
-run_queue_check(
- run_queue_t rq,
- thread_t thread)
-{
- queue_t q;
- queue_entry_t qe;
+ /*
+ * Stash the old urgency and perfctl values to find out if
+ * csw_check updates them.
+ */
+ thread_urgency_t old_urgency = processor->current_urgency;
+ perfcontrol_class_t old_perfctl_class = processor->current_perfctl_class;
- if (rq != thread->runq)
- panic("run_queue_check: thread runq");
+ ast_t preempt;
- if (thread->sched_pri > MAXPRI || thread->sched_pri < MINPRI)
- panic("run_queue_check: thread sched_pri");
+ if ((preempt = csw_check(thread, processor, AST_NONE)) != AST_NONE) {
+ ast_on(preempt);
+ }
- q = &rq->queues[thread->sched_pri];
- qe = queue_first(q);
- while (!queue_end(q, qe)) {
- if (qe == (queue_entry_t)thread)
- return;
+ if (old_urgency != processor->current_urgency) {
+ /*
+ * Urgency updates happen with the thread lock held (ugh).
+ * TODO: This doesn't notice QoS changes...
+ */
+ uint64_t urgency_param1, urgency_param2;
- qe = queue_next(qe);
+ thread_urgency_t urgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2);
+ thread_tell_urgency(urgency, urgency_param1, urgency_param2, 0, thread);
}
- panic("run_queue_check: end");
-}
-
-#endif /* DEBUG */
+ thread_unlock(thread);
-#if defined(CONFIG_SCHED_TRADITIONAL)
-
-/* locks the runqueue itself */
-
-static boolean_t
-processor_queue_remove(
- processor_t processor,
- thread_t thread)
-{
- void * rqlock;
- run_queue_t rq;
-
- rqlock = &processor->processor_set->sched_lock;
- rq = runq_for_processor(processor);
-
- simple_lock(rqlock);
- if (processor == thread->runq) {
- /*
- * Thread is on a run queue and we have a lock on
- * that run queue.
- */
- runq_consider_decr_bound_count(processor, thread);
- run_queue_remove(rq, thread);
- }
- else {
+ if (old_perfctl_class != processor->current_perfctl_class) {
/*
- * The thread left the run queue before we could
- * lock the run queue.
+ * We updated the perfctl class of this thread from another core.
+ * Let CLPC know that the currently running thread has a new
+ * class.
*/
- assert(thread->runq == PROCESSOR_NULL);
- processor = PROCESSOR_NULL;
- }
-
- simple_unlock(rqlock);
- return (processor != PROCESSOR_NULL);
+ machine_switch_perfcontrol_state_update(PERFCONTROL_ATTR_UPDATE,
+ mach_approximate_time(), 0, thread);
+ }
}
-#endif /* CONFIG_SCHED_TRADITIONAL */
/*
- * thread_run_queue_remove:
+ * set_sched_pri:
*
- * Remove a thread from a current run queue and
- * return TRUE if successful.
+ * Set the scheduled priority of the specified thread.
+ *
+ * This may cause the thread to change queues.
*
* Thread must be locked.
*/
-boolean_t
-thread_run_queue_remove(
- thread_t thread)
+void
+set_sched_pri(
+ thread_t thread,
+ int16_t new_priority,
+ set_sched_pri_options_t options)
{
- processor_t processor = thread->runq;
+ bool is_current_thread = (thread == current_thread());
+ bool removed_from_runq = false;
+ bool lazy_update = ((options & SETPRI_LAZY) == SETPRI_LAZY);
+
+ int16_t old_priority = thread->sched_pri;
+
+ /* If we're already at this priority, no need to mess with the runqueue */
+ if (new_priority == old_priority) {
+#if CONFIG_SCHED_CLUTCH
+ /* For the first thread in the system, the priority is correct but
+ * th_sched_bucket is still TH_BUCKET_RUN. Since the clutch
+ * scheduler relies on the bucket being set for all threads, update
+ * its bucket here.
+ */
+ if (thread->th_sched_bucket == TH_BUCKET_RUN) {
+ assert(is_current_thread);
+ SCHED(update_thread_bucket)(thread);
+ }
+#endif /* CONFIG_SCHED_CLUTCH */
+
+ return;
+ }
+
+ if (is_current_thread) {
+ assert(thread->state & TH_RUN);
+ assert(thread->runq == PROCESSOR_NULL);
+ } else {
+ removed_from_runq = thread_run_queue_remove(thread);
+ }
+ thread->sched_pri = new_priority;
+
+#if CONFIG_SCHED_CLUTCH
/*
- * If processor is PROCESSOR_NULL, the thread will stay out of the
- * run queues because the caller locked the thread. Otherwise
- * the thread is on a run queue, but could be chosen for dispatch
- * and removed.
+ * Since for the clutch scheduler, the thread's bucket determines its runq
+ * in the hierarchy it is important to update the bucket when the thread
+ * lock is held and the thread has been removed from the runq hierarchy.
*/
- if (processor != PROCESSOR_NULL) {
- queue_t q;
+ SCHED(update_thread_bucket)(thread);
+
+#endif /* CONFIG_SCHED_CLUTCH */
+
+ KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY),
+ (uintptr_t)thread_tid(thread),
+ thread->base_pri,
+ thread->sched_pri,
+ thread->sched_usage,
+ 0);
+
+ if (removed_from_runq) {
+ thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ);
+ } else if (is_current_thread) {
+ processor_t processor = thread->last_processor;
+ assert(processor == current_processor());
+
+ thread_urgency_t old_urgency = processor->current_urgency;
/*
- * The processor run queues are locked by the
- * processor set. Real-time priorities use a
- * global queue with a dedicated lock.
+ * When dropping in priority, check if the thread no longer belongs on core.
+ * If a thread raises its own priority, don't aggressively rebalance it.
+ * <rdar://problem/31699165>
+ *
+ * csw_check does a processor_state_update_from_thread, but
+ * we should do our own if we're being lazy.
*/
- if (thread->sched_mode == TH_MODE_FAIRSHARE) {
- return SCHED(fairshare_queue_remove)(thread);
- }
-
- if (thread->sched_pri < BASEPRI_RTQUEUES) {
- return SCHED(processor_queue_remove)(processor, thread);
+ if (!lazy_update && new_priority < old_priority) {
+ ast_t preempt;
+
+ if ((preempt = csw_check(thread, processor, AST_NONE)) != AST_NONE) {
+ ast_on(preempt);
+ }
+ } else {
+ processor_state_update_from_thread(processor, thread);
}
- simple_lock(&rt_lock);
- q = &rt_runq.queue;
+ /*
+ * set_sched_pri doesn't alter RT params. We expect direct base priority/QoS
+ * class alterations from user space to occur relatively infrequently, hence
+ * those are lazily handled. QoS classes have distinct priority bands, and QoS
+ * inheritance is expected to involve priority changes.
+ */
+ if (processor->current_urgency != old_urgency) {
+ uint64_t urgency_param1, urgency_param2;
- if (processor == thread->runq) {
- /*
- * Thread is on a run queue and we have a lock on
- * that run queue.
- */
- remqueue((queue_entry_t)thread);
- SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count);
- rt_runq.count--;
+ thread_urgency_t new_urgency = thread_get_urgency(thread,
+ &urgency_param1, &urgency_param2);
- thread->runq = PROCESSOR_NULL;
+ thread_tell_urgency(new_urgency, urgency_param1,
+ urgency_param2, 0, thread);
}
- else {
- /*
- * The thread left the run queue before we could
- * lock the run queue.
- */
- assert(thread->runq == PROCESSOR_NULL);
- processor = PROCESSOR_NULL;
+
+ /* TODO: only call this if current_perfctl_class changed */
+ uint64_t ctime = mach_approximate_time();
+ machine_thread_going_on_core(thread, processor->current_urgency, 0, 0, ctime);
+ } else if (thread->state & TH_RUN) {
+ processor_t processor = thread->last_processor;
+
+ if (!lazy_update &&
+ processor != PROCESSOR_NULL &&
+ processor != current_processor() &&
+ processor->active_thread == thread) {
+ cause_ast_check(processor);
}
+ }
+}
+
+/*
+ * thread_run_queue_remove_for_handoff
+ *
+ * Pull a thread or its (recursive) push target out of the runqueue
+ * so that it is ready for thread_run()
+ *
+ * Called at splsched
+ *
+ * Returns the thread that was pulled or THREAD_NULL if no thread could be pulled.
+ * This may be different than the thread that was passed in.
+ */
+thread_t
+thread_run_queue_remove_for_handoff(thread_t thread)
+{
+ thread_t pulled_thread = THREAD_NULL;
+
+ thread_lock(thread);
- simple_unlock(&rt_lock);
+ /*
+ * Check that the thread is not bound to a different processor,
+ * NO_SMT flag is not set on the thread, cluster type of
+ * processor matches with thread if the thread is pinned to a
+ * particular cluster 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 ((thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == processor)
+ && (!thread_no_smt(thread))
+ && (processor->current_pri < BASEPRI_RTQUEUES)
+ && (thread->sched_pri < BASEPRI_RTQUEUES)
+#if __AMP__
+ && ((!(thread->sched_flags & TH_SFLAG_PCORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_P)
+ && ((!(thread->sched_flags & TH_SFLAG_ECORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_E)
+#endif /* __AMP__ */
+ ) {
+ if (thread_run_queue_remove(thread)) {
+ pulled_thread = thread;
+ }
}
- return (processor != PROCESSOR_NULL);
-}
+ thread_unlock(thread);
-#if defined(CONFIG_SCHED_TRADITIONAL)
+ return pulled_thread;
+}
/*
- * steal_processor_thread:
+ * thread_prepare_for_handoff
+ *
+ * Make the thread ready for handoff.
+ * If the thread was runnable then pull it off the runq, if the thread could
+ * not be pulled, return NULL.
*
- * Locate a thread to steal from the processor and
- * return it.
+ * If the thread was woken up from wait for handoff, make sure it is not bound to
+ * different processor.
*
- * Associated pset must be locked. Returns THREAD_NULL
- * on failure.
+ * 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.
*/
-static thread_t
-steal_processor_thread(
- processor_t processor)
-{
- run_queue_t rq = runq_for_processor(processor);
- queue_t queue = rq->queues + rq->highq;
- int pri = rq->highq, count = rq->count;
- thread_t thread;
-
- while (count > 0) {
- thread = (thread_t)queue_first(queue);
- while (!queue_end(queue, (queue_entry_t)thread)) {
- if (thread->bound_processor == PROCESSOR_NULL) {
- remqueue((queue_entry_t)thread);
-
- thread->runq = PROCESSOR_NULL;
- SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
- runq_consider_decr_bound_count(processor, thread);
- rq->count--;
- if (SCHED(priority_is_urgent)(pri)) {
- rq->urgency--; assert(rq->urgency >= 0);
- }
- if (queue_empty(queue)) {
- if (pri != IDLEPRI)
- clrbit(MAXPRI - pri, rq->bitmap);
- rq->highq = MAXPRI - ffsbit(rq->bitmap);
- }
+thread_t
+thread_prepare_for_handoff(thread_t thread, thread_handoff_option_t option)
+{
+ thread_t pulled_thread = THREAD_NULL;
- return (thread);
- }
- count--;
+ if (option & THREAD_HANDOFF_SETRUN_NEEDED) {
+ processor_t processor = current_processor();
+ thread_lock(thread);
- thread = (thread_t)queue_next((queue_entry_t)thread);
+ /*
+ * Check that the thread is not bound to a different processor,
+ * NO_SMT flag is not set on the thread and cluster type of
+ * processor matches with thread if the thread is pinned to a
+ * particular cluster. Call setrun instead if above conditions
+ * are not satisfied.
+ */
+ if ((thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == processor)
+ && (!thread_no_smt(thread))
+#if __AMP__
+ && ((!(thread->sched_flags & TH_SFLAG_PCORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_P)
+ && ((!(thread->sched_flags & TH_SFLAG_ECORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_E)
+#endif /* __AMP__ */
+ ) {
+ pulled_thread = thread;
+ } else {
+ thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
}
-
- queue--; pri--;
+ thread_unlock(thread);
+ } else {
+ pulled_thread = thread_run_queue_remove_for_handoff(thread);
}
- return (THREAD_NULL);
+ return pulled_thread;
}
/*
- * Locate and steal a thread, beginning
- * at the pset.
+ * thread_run_queue_remove:
+ *
+ * Remove a thread from its current run queue and
+ * return TRUE if successful.
+ *
+ * Thread must be locked.
*
- * The pset must be locked, and is returned
- * unlocked.
+ * 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.
*
- * Returns the stolen thread, or THREAD_NULL on
- * failure.
+ * 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.
*/
-static thread_t
-steal_thread(
- processor_set_t pset)
+boolean_t
+thread_run_queue_remove(
+ thread_t thread)
{
- processor_set_t nset, cset = pset;
- processor_t processor;
- thread_t thread;
+ boolean_t removed = FALSE;
+ processor_t processor = thread->runq;
- do {
- processor = (processor_t)queue_first(&cset->active_queue);
- while (!queue_end(&cset->active_queue, (queue_entry_t)processor)) {
- if (runq_for_processor(processor)->count > 0) {
- thread = steal_processor_thread(processor);
- if (thread != THREAD_NULL) {
- remqueue((queue_entry_t)processor);
- enqueue_tail(&cset->active_queue, (queue_entry_t)processor);
+ if ((thread->state & (TH_RUN | TH_WAIT)) == TH_WAIT) {
+ /* Thread isn't runnable */
+ assert(thread->runq == PROCESSOR_NULL);
+ return FALSE;
+ }
- pset_unlock(cset);
+ 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.
+ */
- return (thread);
- }
- }
+ return FALSE;
+ }
- processor = (processor_t)queue_next((queue_entry_t)processor);
- }
+ if (thread->sched_pri < BASEPRI_RTQUEUES) {
+ return SCHED(processor_queue_remove)(processor, thread);
+ }
- nset = next_pset(cset);
+ processor_set_t pset = processor->processor_set;
- if (nset != pset) {
- pset_unlock(cset);
+ pset_lock(pset);
- cset = nset;
- pset_lock(cset);
- }
- } while (nset != pset);
+ if (thread->runq != PROCESSOR_NULL) {
+ /*
+ * Thread is on the RT run queue and we have a lock on
+ * that run queue.
+ */
- pset_unlock(cset);
+ remqueue(&thread->runq_links);
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_decr(pset);
- return (THREAD_NULL);
-}
+ thread->runq = PROCESSOR_NULL;
+
+ removed = TRUE;
+ }
-static thread_t steal_thread_disabled(
- processor_set_t pset)
-{
pset_unlock(pset);
- return (THREAD_NULL);
+ return removed;
}
-#endif /* CONFIG_SCHED_TRADITIONAL */
+/*
+ * Put the thread back where it goes after a thread_run_queue_remove
+ *
+ * Thread must have been removed under the same thread lock hold
+ *
+ * thread locked, at splsched
+ */
+void
+thread_run_queue_reinsert(thread_t thread, sched_options_t options)
+{
+ assert(thread->runq == PROCESSOR_NULL);
+ assert(thread->state & (TH_RUN));
+ thread_setrun(thread, options);
+}
-int
-thread_get_urgency(uint64_t *rt_period, uint64_t *rt_deadline)
+void
+sys_override_cpu_throttle(boolean_t enable_override)
{
- processor_t processor;
- thread_t thread;
-
- processor = current_processor();
+ if (enable_override) {
+ cpu_throttle_enabled = 0;
+ } else {
+ cpu_throttle_enabled = 1;
+ }
+}
+
+thread_urgency_t
+thread_get_urgency(thread_t thread, uint64_t *arg1, uint64_t *arg2)
+{
+ uint64_t urgency_param1 = 0, urgency_param2 = 0;
+
+ thread_urgency_t urgency;
- thread = processor->next_thread;
+ if (thread == NULL || (thread->state & TH_IDLE)) {
+ urgency_param1 = 0;
+ urgency_param2 = 0;
- if (thread != NULL) {
- if (thread->sched_mode == TH_MODE_REALTIME) {
+ urgency = THREAD_URGENCY_NONE;
+ } else if (thread->sched_mode == TH_MODE_REALTIME) {
+ urgency_param1 = thread->realtime.period;
+ urgency_param2 = thread->realtime.deadline;
- if (rt_period != NULL)
- *rt_period = thread->realtime.period;
- if (rt_deadline != NULL)
- *rt_deadline = thread->realtime.deadline;
+ urgency = THREAD_URGENCY_REAL_TIME;
+ } else if (cpu_throttle_enabled &&
+ (thread->sched_pri <= MAXPRI_THROTTLE) &&
+ (thread->base_pri <= MAXPRI_THROTTLE)) {
+ /*
+ * Threads that are running at low priority but are not
+ * tagged with a specific QoS are separated out from
+ * the "background" urgency. Performance management
+ * subsystem can decide to either treat these threads
+ * as normal threads or look at other signals like thermal
+ * levels for optimal power/perf tradeoffs for a platform.
+ */
+ boolean_t thread_lacks_qos = (proc_get_effective_thread_policy(thread, TASK_POLICY_QOS) == THREAD_QOS_UNSPECIFIED); //thread_has_qos_policy(thread);
+ boolean_t task_is_suppressed = (proc_get_effective_task_policy(thread->task, TASK_POLICY_SUP_ACTIVE) == 0x1);
- KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_GET_URGENCY), THREAD_URGENCY_REAL_TIME, thread->realtime.period,
- (thread->realtime.deadline >> 32), thread->realtime.deadline, 0);
+ /*
+ * Background urgency applied when thread priority is
+ * MAXPRI_THROTTLE or lower and thread is not promoted
+ * and thread has a QoS specified
+ */
+ urgency_param1 = thread->sched_pri;
+ urgency_param2 = thread->base_pri;
- return (THREAD_URGENCY_REAL_TIME);
- } else if ((thread->sched_pri <= MAXPRI_THROTTLE) &&
- (thread->priority <= MAXPRI_THROTTLE)) {
- KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_GET_URGENCY), THREAD_URGENCY_BACKGROUND, thread->sched_pri, thread->priority, 0, 0);
- return (THREAD_URGENCY_BACKGROUND);
+ if (thread_lacks_qos && !task_is_suppressed) {
+ urgency = THREAD_URGENCY_LOWPRI;
+ } else {
+ urgency = THREAD_URGENCY_BACKGROUND;
}
- else
- KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_GET_URGENCY), THREAD_URGENCY_NORMAL, 0, 0, 0, 0);
+ } else {
+ /* For otherwise unclassified threads, report throughput QoS parameters */
+ urgency_param1 = proc_get_effective_thread_policy(thread, TASK_POLICY_THROUGH_QOS);
+ urgency_param2 = proc_get_effective_task_policy(thread->task, TASK_POLICY_THROUGH_QOS);
+ urgency = THREAD_URGENCY_NORMAL;
+ }
- return (THREAD_URGENCY_NORMAL);
+ if (arg1 != NULL) {
+ *arg1 = urgency_param1;
+ }
+ if (arg2 != NULL) {
+ *arg2 = urgency_param2;
}
- else
- KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_GET_URGENCY), THREAD_URGENCY_NONE, 0, 0, 0, 0);
- return (THREAD_URGENCY_NONE);
+
+ return urgency;
}
+perfcontrol_class_t
+thread_get_perfcontrol_class(thread_t thread)
+{
+ /* Special case handling */
+ if (thread->state & TH_IDLE) {
+ return PERFCONTROL_CLASS_IDLE;
+ }
+ if (thread->task == kernel_task) {
+ return PERFCONTROL_CLASS_KERNEL;
+ }
+ if (thread->sched_mode == TH_MODE_REALTIME) {
+ return PERFCONTROL_CLASS_REALTIME;
+ }
+
+ /* perfcontrol_class based on base_pri */
+ if (thread->base_pri <= MAXPRI_THROTTLE) {
+ return PERFCONTROL_CLASS_BACKGROUND;
+ } else if (thread->base_pri <= BASEPRI_UTILITY) {
+ return PERFCONTROL_CLASS_UTILITY;
+ } else if (thread->base_pri <= BASEPRI_DEFAULT) {
+ return PERFCONTROL_CLASS_NONUI;
+ } else if (thread->base_pri <= BASEPRI_FOREGROUND) {
+ return PERFCONTROL_CLASS_UI;
+ } else {
+ return PERFCONTROL_CLASS_ABOVEUI;
+ }
+}
/*
* This is the processor idle loop, which just looks for other threads
thread_t
processor_idle(
- thread_t thread,
- processor_t processor)
+ thread_t thread,
+ processor_t processor)
{
- processor_set_t pset = processor->processor_set;
- thread_t new_thread;
- int state;
+ processor_set_t pset = processor->processor_set;
+
(void)splsched();
KERNEL_DEBUG_CONSTANT_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);
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_START,
+ (uintptr_t)thread_tid(thread), 0, 0, 0, 0);
- timer_switch(&PROCESSOR_DATA(processor, system_state),
- mach_absolute_time(), &PROCESSOR_DATA(processor, idle_state));
- PROCESSOR_DATA(processor, current_state) = &PROCESSOR_DATA(processor, idle_state);
+ SCHED_STATS_INC(idle_transitions);
+ assert(processor->running_timers_active == false);
- while (processor->next_thread == THREAD_NULL && SCHED(processor_queue_empty)(processor) && rt_runq.count == 0 && SCHED(fairshare_runq_count)() == 0 &&
- (thread == THREAD_NULL || ((thread->state & (TH_WAIT|TH_SUSP)) == TH_WAIT && !thread->wake_active))) {
- 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);
+ uint64_t ctime = mach_absolute_time();
- machine_idle();
+ timer_switch(&processor->system_state, ctime, &processor->idle_state);
+ processor->current_state = &processor->idle_state;
- (void)splsched();
+ cpu_quiescent_counter_leave(ctime);
- 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);
+ while (1) {
+ /*
+ * Ensure that updates to my processor and pset state,
+ * made by the IPI source processor before sending the IPI,
+ * are visible on this processor now (even though we don't
+ * take the pset lock yet).
+ */
+ atomic_thread_fence(memory_order_acquire);
- if (processor->state == PROCESSOR_INACTIVE && !machine_processor_is_inactive(processor))
+ if (processor->state != PROCESSOR_IDLE) {
break;
- }
+ }
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id)) {
+ break;
+ }
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ if (bit_test(pset->pending_deferred_AST_cpu_mask, processor->cpu_id)) {
+ break;
+ }
+#endif
+ if (processor->is_recommended && (processor->processor_primary == processor)) {
+ if (rt_runq_count(pset)) {
+ break;
+ }
+ } else {
+ if (SCHED(processor_bound_count)(processor)) {
+ break;
+ }
+ }
- 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);
+ IDLE_KERNEL_DEBUG_CONSTANT(
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_NONE, (uintptr_t)thread_tid(thread), rt_runq_count(pset), SCHED(processor_runq_count)(processor), -1, 0);
- pset_lock(pset);
+ machine_track_platform_idle(TRUE);
- state = processor->state;
- if (state == PROCESSOR_DISPATCHING) {
- /*
- * Commmon case -- cpu dispatched.
- */
- new_thread = processor->next_thread;
- processor->next_thread = THREAD_NULL;
- processor->state = PROCESSOR_RUNNING;
+ machine_idle();
+ /* returns with interrupts enabled */
- if (SCHED(processor_queue_has_priority)(processor, new_thread->sched_pri, FALSE) ||
- (rt_runq.count > 0 && BASEPRI_RTQUEUES >= new_thread->sched_pri) ) {
- processor->deadline = UINT64_MAX;
+ machine_track_platform_idle(FALSE);
- pset_unlock(pset);
+ (void)splsched();
- 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);
+ /*
+ * Check if we should call sched_timeshare_consider_maintenance() here.
+ * The CPU was woken out of idle due to an interrupt and we should do the
+ * call only if the processor is still idle. If the processor is non-idle,
+ * the threads running on the processor would do the call as part of
+ * context swithing.
+ */
+ if (processor->state == PROCESSOR_IDLE) {
+ sched_timeshare_consider_maintenance(mach_absolute_time());
+ }
- 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);
+ IDLE_KERNEL_DEBUG_CONSTANT(
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_NONE, (uintptr_t)thread_tid(thread), rt_runq_count(pset), SCHED(processor_runq_count)(processor), -2, 0);
- return (THREAD_NULL);
+ 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;
+ }
}
+ }
- pset_unlock(pset);
+ ctime = mach_absolute_time();
- 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);
-
- return (new_thread);
- }
- else
- if (state == PROCESSOR_IDLE) {
- remqueue((queue_entry_t)processor);
+ timer_switch(&processor->idle_state, ctime, &processor->system_state);
+ processor->current_state = &processor->system_state;
- processor->state = PROCESSOR_RUNNING;
- enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
- }
- else
- if (state == PROCESSOR_INACTIVE) {
- processor->state = PROCESSOR_RUNNING;
- enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
- }
- else
- if (state == PROCESSOR_SHUTDOWN) {
- /*
- * Going off-line. Force a
- * reschedule.
- */
- if ((new_thread = processor->next_thread) != THREAD_NULL) {
- processor->next_thread = THREAD_NULL;
- processor->deadline = UINT64_MAX;
+ cpu_quiescent_counter_join(ctime);
- pset_unlock(pset);
+ ast_t reason = AST_NONE;
- thread_lock(new_thread);
- thread_setrun(new_thread, SCHED_HEADQ);
- thread_unlock(new_thread);
+ /* We're handling all scheduling AST's */
+ ast_off(AST_SCHEDULING);
- 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);
- }
- }
+ /*
+ * thread_select will move the processor from dispatching to running,
+ * or put it in idle if there's nothing to do.
+ */
+ thread_t current_thread = current_thread();
- pset_unlock(pset);
+ thread_lock(current_thread);
+ thread_t new_thread = thread_select(current_thread, processor, &reason);
+ thread_unlock(current_thread);
+
+ assert(processor->running_timers_active == false);
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);
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_END,
+ (uintptr_t)thread_tid(thread), processor->state, (uintptr_t)thread_tid(new_thread), reason, 0);
+
+ return new_thread;
}
/*
* Each processor has a dedicated thread which
* executes the idle loop when there is no suitable
* previous context.
+ *
+ * This continuation is entered with interrupts disabled.
*/
void
-idle_thread(void)
+idle_thread(__assert_only void* parameter,
+ __unused wait_result_t result)
{
- processor_t processor = current_processor();
- thread_t new_thread;
+ assert(ml_get_interrupts_enabled() == FALSE);
+ assert(parameter == NULL);
+
+ processor_t processor = current_processor();
+
+ /*
+ * Ensure that anything running in idle context triggers
+ * preemption-disabled checks.
+ */
+ disable_preemption();
+
+ /*
+ * Enable interrupts temporarily to handle any pending interrupts
+ * or IPIs before deciding to sleep
+ */
+ spllo();
+
+ thread_t new_thread = processor_idle(THREAD_NULL, processor);
+ /* returns with interrupts disabled */
+
+ enable_preemption();
- new_thread = processor_idle(THREAD_NULL, processor);
if (new_thread != THREAD_NULL) {
- thread_run(processor->idle_thread, (thread_continue_t)idle_thread, NULL, new_thread);
+ thread_run(processor->idle_thread,
+ idle_thread, NULL, new_thread);
/*NOTREACHED*/
}
- thread_block((thread_continue_t)idle_thread);
+ thread_block(idle_thread);
/*NOTREACHED*/
}
kern_return_t
idle_thread_create(
- processor_t processor)
+ processor_t processor)
{
- kern_return_t result;
- thread_t thread;
- spl_t s;
+ kern_return_t result;
+ thread_t thread;
+ spl_t s;
+ char name[MAXTHREADNAMESIZE];
+
+ result = kernel_thread_create(idle_thread, NULL, MAXPRI_KERNEL, &thread);
+ if (result != KERN_SUCCESS) {
+ return result;
+ }
- result = kernel_thread_create((thread_continue_t)idle_thread, NULL, MAXPRI_KERNEL, &thread);
- if (result != KERN_SUCCESS)
- return (result);
+ snprintf(name, sizeof(name), "idle #%d", processor->cpu_id);
+ thread_set_thread_name(thread, name);
s = splsched();
thread_lock(thread);
thread->bound_processor = processor;
processor->idle_thread = thread;
- thread->sched_pri = thread->priority = IDLEPRI;
+ thread->sched_pri = thread->base_pri = IDLEPRI;
thread->state = (TH_RUN | TH_IDLE);
+ thread->options |= TH_OPT_IDLE_THREAD;
thread_unlock(thread);
splx(s);
thread_deallocate(thread);
- return (KERN_SUCCESS);
+ return KERN_SUCCESS;
}
/*
void
sched_startup(void)
{
- kern_return_t result;
- thread_t thread;
+ kern_return_t result;
+ thread_t thread;
+
+ simple_lock_init(&sched_vm_group_list_lock, 0);
+
+#if __arm__ || __arm64__
+ simple_lock_init(&sched_recommended_cores_lock, 0);
+#endif /* __arm__ || __arm64__ */
result = kernel_thread_start_priority((thread_continue_t)sched_init_thread,
- (void *)SCHED(maintenance_continuation),
- MAXPRI_KERNEL, &thread);
- if (result != KERN_SUCCESS)
+ NULL, 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
thread_block(THREAD_CONTINUE_NULL);
}
-#if defined(CONFIG_SCHED_TRADITIONAL)
+#if __arm64__
+static _Atomic uint64_t sched_perfcontrol_callback_deadline;
+#endif /* __arm64__ */
+
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+
+static volatile uint64_t sched_maintenance_deadline;
+static uint64_t sched_tick_last_abstime;
+static uint64_t sched_tick_delta;
+uint64_t sched_tick_max_delta;
-static uint64_t sched_tick_deadline = 0;
/*
* sched_init_thread:
* Perform periodic bookkeeping functions about ten
* times per second.
*/
-static void
-sched_traditional_tick_continue(void)
+void
+sched_timeshare_maintenance_continue(void)
{
- uint64_t abstime = mach_absolute_time();
+ uint64_t sched_tick_ctime, late_time;
+
+ struct sched_update_scan_context scan_context = {
+ .earliest_bg_make_runnable_time = UINT64_MAX,
+ .earliest_normal_make_runnable_time = UINT64_MAX,
+ .earliest_rt_make_runnable_time = UINT64_MAX
+ };
+
+ sched_tick_ctime = mach_absolute_time();
+
+ if (__improbable(sched_tick_last_abstime == 0)) {
+ sched_tick_last_abstime = sched_tick_ctime;
+ late_time = 0;
+ sched_tick_delta = 1;
+ } else {
+ late_time = sched_tick_ctime - sched_tick_last_abstime;
+ sched_tick_delta = late_time / sched_tick_interval;
+ /* Ensure a delta of 1, since the interval could be slightly
+ * smaller than the sched_tick_interval due to dispatch
+ * latencies.
+ */
+ sched_tick_delta = MAX(sched_tick_delta, 1);
+
+ /* 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);
+ }
- sched_tick++;
+ scan_context.sched_tick_last_abstime = sched_tick_last_abstime;
+ KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_MAINTENANCE) | DBG_FUNC_START,
+ sched_tick_delta, late_time, 0, 0, 0);
+
+ /* Add a number of pseudo-ticks corresponding to the elapsed interval
+ * This could be greater than 1 if substantial intervals where
+ * all processors are idle occur, which rarely occurs in practice.
+ */
+
+ sched_tick += sched_tick_delta;
+
+ update_vm_info();
/*
* Compute various averages.
*/
- compute_averages();
+ compute_averages(sched_tick_delta);
/*
* Scan the run queues for threads which
- * may need to be updated.
+ * may need to be updated, and find the earliest runnable thread on the runqueue
+ * to report its latency.
*/
- thread_update_scan();
+ SCHED(thread_update_scan)(&scan_context);
- if (sched_tick_deadline == 0)
- sched_tick_deadline = abstime;
-
- clock_deadline_for_periodic_event(sched_tick_interval, abstime,
- &sched_tick_deadline);
+ SCHED(rt_runq_scan)(&scan_context);
- assert_wait_deadline((event_t)sched_traditional_tick_continue, THREAD_UNINT, sched_tick_deadline);
- thread_block((thread_continue_t)sched_traditional_tick_continue);
- /*NOTREACHED*/
-}
+ uint64_t ctime = mach_absolute_time();
-#endif /* CONFIG_SCHED_TRADITIONAL */
+ uint64_t bg_max_latency = (ctime > scan_context.earliest_bg_make_runnable_time) ?
+ ctime - scan_context.earliest_bg_make_runnable_time : 0;
-void
-sched_init_thread(void (*continuation)(void))
-{
- thread_block(THREAD_CONTINUE_NULL);
+ uint64_t default_max_latency = (ctime > scan_context.earliest_normal_make_runnable_time) ?
+ ctime - scan_context.earliest_normal_make_runnable_time : 0;
- continuation();
+ uint64_t realtime_max_latency = (ctime > scan_context.earliest_rt_make_runnable_time) ?
+ ctime - scan_context.earliest_rt_make_runnable_time : 0;
- /*NOTREACHED*/
-}
+ machine_max_runnable_latency(bg_max_latency, default_max_latency, realtime_max_latency);
-#if defined(CONFIG_SCHED_TRADITIONAL)
+ /*
+ * Check to see if the special sched VM group needs attention.
+ */
+ sched_vm_group_maintenance();
-/*
- * thread_update_scan / runq_scan:
- *
- * Scan the run queues to account for timesharing threads
+#if __arm__ || __arm64__
+ /* Check to see if the recommended cores failsafe is active */
+ sched_recommended_cores_maintenance();
+#endif /* __arm__ || __arm64__ */
+
+
+#if DEBUG || DEVELOPMENT
+#if __x86_64__
+#include <i386/misc_protos.h>
+ /* Check for long-duration interrupts */
+ mp_interrupt_watchdog();
+#endif /* __x86_64__ */
+#endif /* DEBUG || DEVELOPMENT */
+
+ KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_MAINTENANCE) | DBG_FUNC_END,
+ sched_pri_shifts[TH_BUCKET_SHARE_FG], sched_pri_shifts[TH_BUCKET_SHARE_BG],
+ sched_pri_shifts[TH_BUCKET_SHARE_UT], sched_pri_shifts[TH_BUCKET_SHARE_DF], 0);
+
+ assert_wait((event_t)sched_timeshare_maintenance_continue, THREAD_UNINT);
+ thread_block((thread_continue_t)sched_timeshare_maintenance_continue);
+ /*NOTREACHED*/
+}
+
+static uint64_t sched_maintenance_wakeups;
+
+/*
+ * 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_timeshare_consider_maintenance(uint64_t ctime)
+{
+ cpu_quiescent_counter_checkin(ctime);
+
+ uint64_t deadline = sched_maintenance_deadline;
+
+ if (__improbable(ctime >= deadline)) {
+ if (__improbable(current_thread() == sched_maintenance_thread)) {
+ return;
+ }
+ OSMemoryBarrier();
+
+ uint64_t ndeadline = ctime + sched_tick_interval;
+
+ if (__probable(os_atomic_cmpxchg(&sched_maintenance_deadline, deadline, ndeadline, seq_cst))) {
+ thread_wakeup((event_t)sched_timeshare_maintenance_continue);
+ sched_maintenance_wakeups++;
+ }
+ }
+
+#if !CONFIG_SCHED_CLUTCH
+ /*
+ * Only non-clutch schedulers use the global load calculation EWMA algorithm. For clutch
+ * scheduler, the load is maintained at the thread group and bucket level.
+ */
+ uint64_t load_compute_deadline = os_atomic_load_wide(&sched_load_compute_deadline, relaxed);
+
+ if (__improbable(load_compute_deadline && ctime >= load_compute_deadline)) {
+ uint64_t new_deadline = 0;
+ if (os_atomic_cmpxchg(&sched_load_compute_deadline, load_compute_deadline, new_deadline, relaxed)) {
+ compute_sched_load();
+ new_deadline = ctime + sched_load_compute_interval_abs;
+ os_atomic_store_wide(&sched_load_compute_deadline, new_deadline, relaxed);
+ }
+ }
+#endif /* CONFIG_SCHED_CLUTCH */
+
+#if __arm64__
+ uint64_t perf_deadline = os_atomic_load(&sched_perfcontrol_callback_deadline, relaxed);
+
+ if (__improbable(perf_deadline && ctime >= perf_deadline)) {
+ /* CAS in 0, if success, make callback. Otherwise let the next context switch check again. */
+ if (os_atomic_cmpxchg(&sched_perfcontrol_callback_deadline, perf_deadline, 0, relaxed)) {
+ machine_perfcontrol_deadline_passed(perf_deadline);
+ }
+ }
+#endif /* __arm64__ */
+}
+
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+
+void
+sched_init_thread(void)
+{
+ thread_block(THREAD_CONTINUE_NULL);
+
+ thread_t thread = current_thread();
+
+ thread_set_thread_name(thread, "sched_maintenance_thread");
+
+ sched_maintenance_thread = thread;
+
+ SCHED(maintenance_continuation)();
+
+ /*NOTREACHED*/
+}
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+
+/*
+ * 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
* disabling preemption for long periods.
*/
-#define THREAD_UPDATE_SIZE 128
+#define THREAD_UPDATE_SIZE 128
-static thread_t thread_update_array[THREAD_UPDATE_SIZE];
-static int thread_update_count = 0;
+static thread_t thread_update_array[THREAD_UPDATE_SIZE];
+static uint32_t thread_update_count = 0;
-/*
- * Scan a runq for candidate threads.
- *
- * Returns TRUE if retry is needed.
- */
-static boolean_t
-runq_scan(
- run_queue_t runq)
-{
- register int count;
- register queue_t q;
- register thread_t thread;
-
- if ((count = runq->count) > 0) {
- q = runq->queues + runq->highq;
- while (count > 0) {
- queue_iterate(q, thread, thread_t, links) {
- if ( thread->sched_stamp != sched_tick &&
- (thread->sched_mode == TH_MODE_TIMESHARE) ) {
- if (thread_update_count == THREAD_UPDATE_SIZE)
- return (TRUE);
-
- thread_update_array[thread_update_count++] = thread;
- thread_reference_internal(thread);
- }
+/* Returns TRUE if thread was added, FALSE if thread_update_array is full */
+boolean_t
+thread_update_add_thread(thread_t thread)
+{
+ if (thread_update_count == THREAD_UPDATE_SIZE) {
+ return FALSE;
+ }
- count--;
- }
+ thread_update_array[thread_update_count++] = thread;
+ thread_reference_internal(thread);
+ return TRUE;
+}
+
+void
+thread_update_process_threads(void)
+{
+ assert(thread_update_count <= THREAD_UPDATE_SIZE);
+
+ for (uint32_t i = 0; i < thread_update_count; i++) {
+ thread_t thread = thread_update_array[i];
+ assert_thread_magic(thread);
+ thread_update_array[i] = THREAD_NULL;
- 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);
}
- return (FALSE);
+ thread_update_count = 0;
}
-static void
-thread_update_scan(void)
+static boolean_t
+runq_scan_thread(
+ thread_t thread,
+ sched_update_scan_context_t scan_context)
{
- boolean_t restart_needed = FALSE;
- processor_t processor = processor_list;
- processor_set_t pset;
- thread_t thread;
- spl_t s;
-
- do {
- do {
- pset = processor->processor_set;
+ assert_thread_magic(thread);
- s = splsched();
- pset_lock(pset);
+ if (thread->sched_stamp != sched_tick &&
+ thread->sched_mode == TH_MODE_TIMESHARE) {
+ if (thread_update_add_thread(thread) == FALSE) {
+ return TRUE;
+ }
+ }
- restart_needed = runq_scan(runq_for_processor(processor));
+ 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;
+ }
+ }
- pset_unlock(pset);
- splx(s);
+ return FALSE;
+}
- if (restart_needed)
- break;
+/*
+ * 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)
+{
+ int count = runq->count;
+ int queue_index;
- thread = processor->idle_thread;
- if (thread != THREAD_NULL && thread->sched_stamp != sched_tick) {
- if (thread_update_count == THREAD_UPDATE_SIZE) {
- restart_needed = TRUE;
- break;
- }
+ assert(count >= 0);
- thread_update_array[thread_update_count++] = thread;
- thread_reference_internal(thread);
- }
- } while ((processor = processor->processor_list) != NULL);
+ if (count == 0) {
+ return FALSE;
+ }
- /*
- * Ok, we now have a collection of candidates -- fix them.
- */
- while (thread_update_count > 0) {
- thread = thread_update_array[--thread_update_count];
- thread_update_array[thread_update_count] = THREAD_NULL;
+ for (queue_index = bitmap_first(runq->bitmap, NRQS);
+ queue_index >= 0;
+ queue_index = bitmap_next(runq->bitmap, queue_index)) {
+ thread_t thread;
+ circle_queue_t queue = &runq->queues[queue_index];
- s = splsched();
- thread_lock(thread);
- if ( !(thread->state & (TH_WAIT)) ) {
- if (SCHED(can_update_priority)(thread))
- SCHED(update_priority)(thread);
+ cqe_foreach_element(thread, queue, runq_links) {
+ assert(count > 0);
+ if (runq_scan_thread(thread, scan_context) == TRUE) {
+ return TRUE;
}
- thread_unlock(thread);
- splx(s);
+ count--;
+ }
+ }
- thread_deallocate(thread);
- }
- } while (restart_needed);
+ return FALSE;
}
-#endif /* CONFIG_SCHED_TRADITIONAL */
+#if CONFIG_SCHED_CLUTCH
boolean_t
-thread_eager_preemption(thread_t thread)
+sched_clutch_timeshare_scan(
+ queue_t thread_queue,
+ uint16_t thread_count,
+ sched_update_scan_context_t scan_context)
+{
+ if (thread_count == 0) {
+ return FALSE;
+ }
+
+ thread_t thread;
+ qe_foreach_element_safe(thread, thread_queue, th_clutch_timeshare_link) {
+ if (runq_scan_thread(thread, scan_context) == TRUE) {
+ return TRUE;
+ }
+ thread_count--;
+ }
+
+ assert(thread_count == 0);
+ return FALSE;
+}
+
+
+#endif /* CONFIG_SCHED_CLUTCH */
+
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+
+bool
+thread_is_eager_preempt(thread_t thread)
{
- return ((thread->sched_flags & TH_SFLAG_EAGERPREEMPT) != 0);
+ return thread->sched_flags & TH_SFLAG_EAGERPREEMPT;
}
void
-thread_set_eager_preempt(thread_t thread)
+thread_set_eager_preempt(thread_t thread)
{
- spl_t x;
- processor_t p;
- ast_t ast = AST_NONE;
+ spl_t s = splsched();
+ thread_lock(thread);
- x = splsched();
- p = current_processor();
+ assert(!thread_is_eager_preempt(thread));
- thread_lock(thread);
thread->sched_flags |= TH_SFLAG_EAGERPREEMPT;
if (thread == current_thread()) {
+ /* csw_check updates current_is_eagerpreempt on the processor */
+ ast_t ast = csw_check(thread, current_processor(), AST_NONE);
+
thread_unlock(thread);
- ast = csw_check(p);
if (ast != AST_NONE) {
- (void) thread_block_reason(THREAD_CONTINUE_NULL, NULL, ast);
+ thread_block_reason(THREAD_CONTINUE_NULL, NULL, ast);
}
} else {
- p = thread->last_processor;
+ processor_t last_processor = thread->last_processor;
- if (p != PROCESSOR_NULL && p->state == PROCESSOR_RUNNING &&
- p->active_thread == thread) {
- cause_ast_check(p);
+ if (last_processor != PROCESSOR_NULL &&
+ last_processor->state == PROCESSOR_RUNNING &&
+ last_processor->active_thread == thread) {
+ cause_ast_check(last_processor);
}
-
+
thread_unlock(thread);
}
- splx(x);
+ splx(s);
}
void
-thread_clear_eager_preempt(thread_t thread)
+thread_clear_eager_preempt(thread_t thread)
{
- spl_t x;
-
- x = splsched();
+ spl_t s = splsched();
thread_lock(thread);
+ assert(thread_is_eager_preempt(thread));
+
thread->sched_flags &= ~TH_SFLAG_EAGERPREEMPT;
-
+
+ if (thread == current_thread()) {
+ current_processor()->current_is_eagerpreempt = false;
+ }
+
thread_unlock(thread);
- splx(x);
+ splx(s);
}
+
/*
* Scheduling statistics
*/
void
sched_stats_handle_csw(processor_t processor, int reasons, int selfpri, int otherpri)
{
- struct processor_sched_statistics *stats;
+ struct sched_statistics *stats;
boolean_t to_realtime = FALSE;
-
- stats = &processor->processor_data.sched_stats;
+
+ stats = PERCPU_GET_RELATIVE(sched_stats, processor, processor);
stats->csw_count++;
if (otherpri >= BASEPRI_REALTIME) {
if (selfpri >= BASEPRI_REALTIME) {
stats->preempted_rt_count++;
- }
+ }
if (to_realtime) {
stats->preempted_by_rt_count++;
}
-
}
}
void
-sched_stats_handle_runq_change(struct runq_stats *stats, int old_count)
+sched_stats_handle_runq_change(struct runq_stats *stats, int old_count)
{
uint64_t timestamp = mach_absolute_time();
#undef thread_wakeup
void
thread_wakeup(
- event_t x);
+ event_t x);
void
thread_wakeup(
- event_t x)
+ event_t x)
{
- thread_wakeup_with_result(x, THREAD_AWAKENED);
+ thread_wakeup_with_result(x, THREAD_AWAKENED);
}
boolean_t
preemption_enabled(void)
{
- return (get_preemption_level() == 0 && ml_get_interrupts_enabled());
+ return get_preemption_level() == 0 && ml_get_interrupts_enabled();
}
-#if DEBUG
-static boolean_t
-thread_runnable(
- thread_t thread)
+static void
+sched_timer_deadline_tracking_init(void)
+{
+ nanoseconds_to_absolutetime(TIMER_DEADLINE_TRACKING_BIN_1_DEFAULT, &timer_deadline_tracking_bin_1);
+ nanoseconds_to_absolutetime(TIMER_DEADLINE_TRACKING_BIN_2_DEFAULT, &timer_deadline_tracking_bin_2);
+}
+
+#if __arm__ || __arm64__
+
+uint32_t perfcontrol_requested_recommended_cores = ALL_CORES_RECOMMENDED;
+uint32_t perfcontrol_requested_recommended_core_count = MAX_CPUS;
+bool perfcontrol_failsafe_active = false;
+bool perfcontrol_sleep_override = false;
+
+uint64_t perfcontrol_failsafe_maintenance_runnable_time;
+uint64_t perfcontrol_failsafe_activation_time;
+uint64_t perfcontrol_failsafe_deactivation_time;
+
+/* data covering who likely caused it and how long they ran */
+#define FAILSAFE_NAME_LEN 33 /* (2*MAXCOMLEN)+1 from size of p_name */
+char perfcontrol_failsafe_name[FAILSAFE_NAME_LEN];
+int perfcontrol_failsafe_pid;
+uint64_t perfcontrol_failsafe_tid;
+uint64_t perfcontrol_failsafe_thread_timer_at_start;
+uint64_t perfcontrol_failsafe_thread_timer_last_seen;
+uint32_t perfcontrol_failsafe_recommended_at_trigger;
+
+/*
+ * Perf controller calls here to update the recommended core bitmask.
+ * If the failsafe is active, we don't immediately apply the new value.
+ * Instead, we store the new request and use it after the failsafe deactivates.
+ *
+ * If the failsafe is not active, immediately apply the update.
+ *
+ * No scheduler locks are held, no other locks are held that scheduler might depend on,
+ * interrupts are enabled
+ *
+ * currently prototype is in osfmk/arm/machine_routines.h
+ */
+void
+sched_perfcontrol_update_recommended_cores(uint32_t recommended_cores)
+{
+ assert(preemption_enabled());
+
+ spl_t s = splsched();
+ simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL);
+
+ perfcontrol_requested_recommended_cores = recommended_cores;
+ perfcontrol_requested_recommended_core_count = __builtin_popcountll(recommended_cores);
+
+ if ((perfcontrol_failsafe_active == false) && (perfcontrol_sleep_override == false)) {
+ sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores);
+ } else {
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_NONE,
+ perfcontrol_requested_recommended_cores,
+ sched_maintenance_thread->last_made_runnable_time, 0, 0, 0);
+ }
+
+ simple_unlock(&sched_recommended_cores_lock);
+ splx(s);
+}
+
+void
+sched_override_recommended_cores_for_sleep(void)
+{
+ spl_t s = splsched();
+ simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL);
+
+ if (perfcontrol_sleep_override == false) {
+ perfcontrol_sleep_override = true;
+ sched_update_recommended_cores(ALL_CORES_RECOMMENDED);
+ }
+
+ simple_unlock(&sched_recommended_cores_lock);
+ splx(s);
+}
+
+void
+sched_restore_recommended_cores_after_sleep(void)
+{
+ spl_t s = splsched();
+ simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL);
+
+ if (perfcontrol_sleep_override == true) {
+ perfcontrol_sleep_override = false;
+ sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores);
+ }
+
+ simple_unlock(&sched_recommended_cores_lock);
+ splx(s);
+}
+
+/*
+ * Consider whether we need to activate the recommended cores failsafe
+ *
+ * Called from quantum timer interrupt context of a realtime thread
+ * No scheduler locks are held, interrupts are disabled
+ */
+void
+sched_consider_recommended_cores(uint64_t ctime, thread_t cur_thread)
+{
+ /*
+ * Check if a realtime thread is starving the system
+ * and bringing up non-recommended cores would help
+ *
+ * TODO: Is this the correct check for recommended == possible cores?
+ * TODO: Validate the checks without the relevant lock are OK.
+ */
+
+ if (__improbable(perfcontrol_failsafe_active == TRUE)) {
+ /* keep track of how long the responsible thread runs */
+
+ simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL);
+
+ if (perfcontrol_failsafe_active == TRUE &&
+ cur_thread->thread_id == perfcontrol_failsafe_tid) {
+ perfcontrol_failsafe_thread_timer_last_seen = timer_grab(&cur_thread->user_timer) +
+ timer_grab(&cur_thread->system_timer);
+ }
+
+ simple_unlock(&sched_recommended_cores_lock);
+
+ /* we're already trying to solve the problem, so bail */
+ return;
+ }
+
+ /* The failsafe won't help if there are no more processors to enable */
+ if (__probable(perfcontrol_requested_recommended_core_count >= processor_count)) {
+ return;
+ }
+
+ uint64_t too_long_ago = ctime - perfcontrol_failsafe_starvation_threshold;
+
+ /* Use the maintenance thread as our canary in the coal mine */
+ thread_t m_thread = sched_maintenance_thread;
+
+ /* If it doesn't look bad, nothing to see here */
+ if (__probable(m_thread->last_made_runnable_time >= too_long_ago)) {
+ return;
+ }
+
+ /* It looks bad, take the lock to be sure */
+ thread_lock(m_thread);
+
+ if (m_thread->runq == PROCESSOR_NULL ||
+ (m_thread->state & (TH_RUN | TH_WAIT)) != TH_RUN ||
+ m_thread->last_made_runnable_time >= too_long_ago) {
+ /*
+ * Maintenance thread is either on cpu or blocked, and
+ * therefore wouldn't benefit from more cores
+ */
+ thread_unlock(m_thread);
+ return;
+ }
+
+ uint64_t maintenance_runnable_time = m_thread->last_made_runnable_time;
+
+ thread_unlock(m_thread);
+
+ /*
+ * There are cores disabled at perfcontrol's recommendation, but the
+ * system is so overloaded that the maintenance thread can't run.
+ * That likely means that perfcontrol can't run either, so it can't fix
+ * the recommendation. We have to kick in a failsafe to keep from starving.
+ *
+ * When the maintenance thread has been starved for too long,
+ * ignore the recommendation from perfcontrol and light up all the cores.
+ *
+ * TODO: Consider weird states like boot, sleep, or debugger
+ */
+
+ simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL);
+
+ if (perfcontrol_failsafe_active == TRUE) {
+ simple_unlock(&sched_recommended_cores_lock);
+ return;
+ }
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_START,
+ perfcontrol_requested_recommended_cores, maintenance_runnable_time, 0, 0, 0);
+
+ perfcontrol_failsafe_active = TRUE;
+ perfcontrol_failsafe_activation_time = mach_absolute_time();
+ perfcontrol_failsafe_maintenance_runnable_time = maintenance_runnable_time;
+ perfcontrol_failsafe_recommended_at_trigger = perfcontrol_requested_recommended_cores;
+
+ /* Capture some data about who screwed up (assuming that the thread on core is at fault) */
+ task_t task = cur_thread->task;
+ perfcontrol_failsafe_pid = task_pid(task);
+ strlcpy(perfcontrol_failsafe_name, proc_name_address(task->bsd_info), sizeof(perfcontrol_failsafe_name));
+
+ perfcontrol_failsafe_tid = cur_thread->thread_id;
+
+ /* Blame the thread for time it has run recently */
+ uint64_t recent_computation = (ctime - cur_thread->computation_epoch) + cur_thread->computation_metered;
+
+ uint64_t last_seen = timer_grab(&cur_thread->user_timer) + timer_grab(&cur_thread->system_timer);
+
+ /* Compute the start time of the bad behavior in terms of the thread's on core time */
+ perfcontrol_failsafe_thread_timer_at_start = last_seen - recent_computation;
+ perfcontrol_failsafe_thread_timer_last_seen = last_seen;
+
+ /* Ignore the previously recommended core configuration */
+ sched_update_recommended_cores(ALL_CORES_RECOMMENDED);
+
+ simple_unlock(&sched_recommended_cores_lock);
+}
+
+/*
+ * Now that our bacon has been saved by the failsafe, consider whether to turn it off
+ *
+ * Runs in the context of the maintenance thread, no locks held
+ */
+static void
+sched_recommended_cores_maintenance(void)
{
- return ((thread->state & (TH_RUN|TH_WAIT)) == TH_RUN);
+ /* Common case - no failsafe, nothing to be done here */
+ if (__probable(perfcontrol_failsafe_active == FALSE)) {
+ return;
+ }
+
+ uint64_t ctime = mach_absolute_time();
+
+ boolean_t print_diagnostic = FALSE;
+ char p_name[FAILSAFE_NAME_LEN] = "";
+
+ spl_t s = splsched();
+ simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL);
+
+ /* Check again, under the lock, to avoid races */
+ if (perfcontrol_failsafe_active == FALSE) {
+ goto out;
+ }
+
+ /*
+ * Ensure that the other cores get another few ticks to run some threads
+ * If we don't have this hysteresis, the maintenance thread is the first
+ * to run, and then it immediately kills the other cores
+ */
+ if ((ctime - perfcontrol_failsafe_activation_time) < perfcontrol_failsafe_starvation_threshold) {
+ goto out;
+ }
+
+ /* Capture some diagnostic state under the lock so we can print it out later */
+
+ int pid = perfcontrol_failsafe_pid;
+ uint64_t tid = perfcontrol_failsafe_tid;
+
+ uint64_t thread_usage = perfcontrol_failsafe_thread_timer_last_seen -
+ perfcontrol_failsafe_thread_timer_at_start;
+ uint32_t rec_cores_before = perfcontrol_failsafe_recommended_at_trigger;
+ uint32_t rec_cores_after = perfcontrol_requested_recommended_cores;
+ uint64_t failsafe_duration = ctime - perfcontrol_failsafe_activation_time;
+ strlcpy(p_name, perfcontrol_failsafe_name, sizeof(p_name));
+
+ print_diagnostic = TRUE;
+
+ /* Deactivate the failsafe and reinstate the requested recommendation settings */
+
+ perfcontrol_failsafe_deactivation_time = ctime;
+ perfcontrol_failsafe_active = FALSE;
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_END,
+ perfcontrol_requested_recommended_cores, failsafe_duration, 0, 0, 0);
+
+ sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores);
+
+out:
+ simple_unlock(&sched_recommended_cores_lock);
+ splx(s);
+
+ if (print_diagnostic) {
+ uint64_t failsafe_duration_ms = 0, thread_usage_ms = 0;
+
+ absolutetime_to_nanoseconds(failsafe_duration, &failsafe_duration_ms);
+ failsafe_duration_ms = failsafe_duration_ms / NSEC_PER_MSEC;
+
+ absolutetime_to_nanoseconds(thread_usage, &thread_usage_ms);
+ thread_usage_ms = thread_usage_ms / NSEC_PER_MSEC;
+
+ printf("recommended core failsafe kicked in for %lld ms "
+ "likely due to %s[%d] thread 0x%llx spending "
+ "%lld ms on cpu at realtime priority - "
+ "new recommendation: 0x%x -> 0x%x\n",
+ failsafe_duration_ms, p_name, pid, tid, thread_usage_ms,
+ rec_cores_before, rec_cores_after);
+ }
+}
+
+#endif /* __arm__ || __arm64__ */
+
+kern_return_t
+sched_processor_enable(processor_t processor, boolean_t enable)
+{
+ assert(preemption_enabled());
+
+ spl_t s = splsched();
+ simple_lock(&sched_recommended_cores_lock, LCK_GRP_NULL);
+
+ if (enable) {
+ bit_set(usercontrol_requested_recommended_cores, processor->cpu_id);
+ } else {
+ bit_clear(usercontrol_requested_recommended_cores, processor->cpu_id);
+ }
+
+#if __arm__ || __arm64__
+ if ((perfcontrol_failsafe_active == false) && (perfcontrol_sleep_override == false)) {
+ sched_update_recommended_cores(perfcontrol_requested_recommended_cores & usercontrol_requested_recommended_cores);
+ } else {
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_REC_CORES_FAILSAFE) | DBG_FUNC_NONE,
+ perfcontrol_requested_recommended_cores,
+ sched_maintenance_thread->last_made_runnable_time, 0, 0, 0);
+ }
+#else /* __arm__ || __arm64__ */
+ sched_update_recommended_cores(usercontrol_requested_recommended_cores);
+#endif /* !__arm__ || __arm64__ */
+
+ simple_unlock(&sched_recommended_cores_lock);
+ splx(s);
+
+ return KERN_SUCCESS;
+}
+
+
+/*
+ * Apply a new recommended cores mask to the processors it affects
+ * Runs after considering failsafes and such
+ *
+ * Iterate over processors and update their ->is_recommended field.
+ * If a processor is running, we let it drain out at its next
+ * quantum expiration or blocking point. If a processor is idle, there
+ * may be more work for it to do, so IPI it.
+ *
+ * interrupts disabled, sched_recommended_cores_lock is held
+ */
+static void
+sched_update_recommended_cores(uint64_t recommended_cores)
+{
+ processor_set_t pset, nset;
+ processor_t processor;
+ uint64_t needs_exit_idle_mask = 0x0;
+ uint32_t avail_count;
+
+ processor = processor_list;
+ pset = processor->processor_set;
+
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_UPDATE_REC_CORES) | DBG_FUNC_START,
+ recommended_cores,
+#if __arm__ || __arm64__
+ perfcontrol_failsafe_active, 0, 0);
+#else /* __arm__ || __arm64__ */
+ 0, 0, 0);
+#endif /* ! __arm__ || __arm64__ */
+
+ if (__builtin_popcountll(recommended_cores) == 0) {
+ bit_set(recommended_cores, master_processor->cpu_id); /* add boot processor or we hang */
+ }
+
+ /* First set recommended cores */
+ pset_lock(pset);
+ avail_count = 0;
+ do {
+ nset = processor->processor_set;
+ if (nset != pset) {
+ pset_unlock(pset);
+ pset = nset;
+ pset_lock(pset);
+ }
+
+ if (bit_test(recommended_cores, processor->cpu_id)) {
+ processor->is_recommended = TRUE;
+ bit_set(pset->recommended_bitmask, processor->cpu_id);
+
+ if (processor->state == PROCESSOR_IDLE) {
+ if (processor != current_processor()) {
+ bit_set(needs_exit_idle_mask, processor->cpu_id);
+ }
+ }
+ if (processor->state != PROCESSOR_OFF_LINE) {
+ avail_count++;
+ SCHED(pset_made_schedulable)(processor, pset, false);
+ }
+ }
+ } while ((processor = processor->processor_list) != NULL);
+ pset_unlock(pset);
+
+ /* Now shutdown not recommended cores */
+ processor = processor_list;
+ pset = processor->processor_set;
+
+ pset_lock(pset);
+ do {
+ nset = processor->processor_set;
+ if (nset != pset) {
+ pset_unlock(pset);
+ pset = nset;
+ pset_lock(pset);
+ }
+
+ if (!bit_test(recommended_cores, processor->cpu_id)) {
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+
+ processor->is_recommended = FALSE;
+ bit_clear(pset->recommended_bitmask, processor->cpu_id);
+
+ if ((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_DISPATCHING)) {
+ ipi_type = SCHED_IPI_IMMEDIATE;
+ }
+ SCHED(processor_queue_shutdown)(processor);
+ /* pset unlocked */
+
+ SCHED(rt_queue_shutdown)(processor);
+
+ if (ipi_type != SCHED_IPI_NONE) {
+ if (processor == current_processor()) {
+ ast_on(AST_PREEMPT);
+ } else {
+ sched_ipi_perform(processor, ipi_type);
+ }
+ }
+
+ pset_lock(pset);
+ }
+ } while ((processor = processor->processor_list) != NULL);
+
+ processor_avail_count_user = avail_count;
+#if defined(__x86_64__)
+ commpage_update_active_cpus();
+#endif
+
+ pset_unlock(pset);
+
+ /* Issue all pending IPIs now that the pset lock has been dropped */
+ for (int cpuid = lsb_first(needs_exit_idle_mask); cpuid >= 0; cpuid = lsb_next(needs_exit_idle_mask, cpuid)) {
+ processor = processor_array[cpuid];
+ machine_signal_idle(processor);
+ }
+
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_UPDATE_REC_CORES) | DBG_FUNC_END,
+ needs_exit_idle_mask, 0, 0, 0);
+}
+
+void
+thread_set_options(uint32_t thopt)
+{
+ spl_t x;
+ thread_t t = current_thread();
+
+ x = splsched();
+ thread_lock(t);
+
+ t->options |= thopt;
+
+ thread_unlock(t);
+ splx(x);
+}
+
+void
+thread_set_pending_block_hint(thread_t thread, block_hint_t block_hint)
+{
+ thread->pending_block_hint = block_hint;
+}
+
+uint32_t
+qos_max_parallelism(int qos, uint64_t options)
+{
+ return SCHED(qos_max_parallelism)(qos, options);
+}
+
+uint32_t
+sched_qos_max_parallelism(__unused int qos, uint64_t options)
+{
+ host_basic_info_data_t hinfo;
+ mach_msg_type_number_t count = HOST_BASIC_INFO_COUNT;
+ /* Query the machine layer for core information */
+ __assert_only kern_return_t kret = host_info(host_self(), HOST_BASIC_INFO,
+ (host_info_t)&hinfo, &count);
+ assert(kret == KERN_SUCCESS);
+
+ if (options & QOS_PARALLELISM_COUNT_LOGICAL) {
+ return hinfo.logical_cpu;
+ } else {
+ return hinfo.physical_cpu;
+ }
+}
+
+int sched_allow_NO_SMT_threads = 1;
+bool
+thread_no_smt(thread_t thread)
+{
+ return sched_allow_NO_SMT_threads && (thread->bound_processor == PROCESSOR_NULL) && ((thread->sched_flags & TH_SFLAG_NO_SMT) || (thread->task->t_flags & TF_NO_SMT));
+}
+
+bool
+processor_active_thread_no_smt(processor_t processor)
+{
+ return sched_allow_NO_SMT_threads && !processor->current_is_bound && processor->current_is_NO_SMT;
+}
+
+#if __arm64__
+
+/*
+ * Set up or replace old timer with new timer
+ *
+ * Returns true if canceled old timer, false if it did not
+ */
+boolean_t
+sched_perfcontrol_update_callback_deadline(uint64_t new_deadline)
+{
+ /*
+ * Exchange deadline for new deadline, if old deadline was nonzero,
+ * then I cancelled the callback, otherwise I didn't
+ */
+
+ return os_atomic_xchg(&sched_perfcontrol_callback_deadline, new_deadline,
+ relaxed) != 0;
+}
+
+#endif /* __arm64__ */
+
+#if CONFIG_SCHED_EDGE
+
+#define SCHED_PSET_LOAD_EWMA_TC_NSECS 10000000u
+
+/*
+ * sched_edge_pset_running_higher_bucket()
+ *
+ * Routine to calculate cumulative running counts for each scheduling
+ * bucket. This effectively lets the load calculation calculate if a
+ * cluster is running any threads at a QoS lower than the thread being
+ * migrated etc.
+ */
+
+static void
+sched_edge_pset_running_higher_bucket(processor_set_t pset, uint32_t *running_higher)
+{
+ bitmap_t *active_map = &pset->cpu_state_map[PROCESSOR_RUNNING];
+
+ /* Edge Scheduler Optimization */
+ for (int cpu = bitmap_first(active_map, MAX_CPUS); cpu >= 0; cpu = bitmap_next(active_map, cpu)) {
+ sched_bucket_t cpu_bucket = os_atomic_load(&pset->cpu_running_buckets[cpu], relaxed);
+ for (sched_bucket_t bucket = cpu_bucket; bucket < TH_BUCKET_SCHED_MAX; bucket++) {
+ running_higher[bucket]++;
+ }
+ }
+}
+
+/*
+ * sched_update_pset_load_average()
+ *
+ * Updates the load average for each sched bucket for a cluster.
+ * This routine must be called with the pset lock held.
+ */
+void
+sched_update_pset_load_average(processor_set_t pset, uint64_t curtime)
+{
+ if (pset->online_processor_count == 0) {
+ /* Looks like the pset is not runnable any more; nothing to do here */
+ return;
+ }
+
+ /*
+ * Edge Scheduler Optimization
+ *
+ * See if more callers of this routine can pass in timestamps to avoid the
+ * mach_absolute_time() call here.
+ */
+
+ if (!curtime) {
+ curtime = mach_absolute_time();
+ }
+ uint64_t last_update = os_atomic_load(&pset->pset_load_last_update, relaxed);
+ int64_t delta_ticks = curtime - last_update;
+ if (delta_ticks < 0) {
+ return;
+ }
+
+ uint64_t delta_nsecs = 0;
+ absolutetime_to_nanoseconds(delta_ticks, &delta_nsecs);
+
+ if (__improbable(delta_nsecs > UINT32_MAX)) {
+ delta_nsecs = UINT32_MAX;
+ }
+
+ uint32_t running_higher[TH_BUCKET_SCHED_MAX] = {0};
+ sched_edge_pset_running_higher_bucket(pset, running_higher);
+
+ for (sched_bucket_t sched_bucket = TH_BUCKET_FIXPRI; sched_bucket < TH_BUCKET_SCHED_MAX; sched_bucket++) {
+ uint64_t old_load_average = os_atomic_load(&pset->pset_load_average[sched_bucket], relaxed);
+ uint64_t old_load_average_factor = old_load_average * SCHED_PSET_LOAD_EWMA_TC_NSECS;
+ uint32_t current_runq_depth = (sched_edge_cluster_cumulative_count(&pset->pset_clutch_root, sched_bucket) + rt_runq_count(pset) + running_higher[sched_bucket]) / pset->online_processor_count;
+
+ /*
+ * For the new load average multiply current_runq_depth by delta_nsecs (which resuts in a 32.0 value).
+ * Since we want to maintain the load average as a 24.8 fixed arithmetic value for precision, the
+ * new load averga needs to be shifted before it can be added to the old load average.
+ */
+ uint64_t new_load_average_factor = (current_runq_depth * delta_nsecs) << SCHED_PSET_LOAD_EWMA_FRACTION_BITS;
+
+ /*
+ * For extremely parallel workloads, it is important that the load average on a cluster moves zero to non-zero
+ * instantly to allow threads to be migrated to other (potentially idle) clusters quickly. Hence use the EWMA
+ * when the system is already loaded; otherwise for an idle system use the latest load average immediately.
+ */
+ int old_load_shifted = (int)((old_load_average + SCHED_PSET_LOAD_EWMA_ROUND_BIT) >> SCHED_PSET_LOAD_EWMA_FRACTION_BITS);
+ boolean_t load_uptick = (old_load_shifted == 0) && (current_runq_depth != 0);
+ boolean_t load_downtick = (old_load_shifted != 0) && (current_runq_depth == 0);
+ uint64_t load_average;
+ if (load_uptick || load_downtick) {
+ load_average = (current_runq_depth << SCHED_PSET_LOAD_EWMA_FRACTION_BITS);
+ } else {
+ /* Indicates a loaded system; use EWMA for load average calculation */
+ load_average = (old_load_average_factor + new_load_average_factor) / (delta_nsecs + SCHED_PSET_LOAD_EWMA_TC_NSECS);
+ }
+ os_atomic_store(&pset->pset_load_average[sched_bucket], load_average, relaxed);
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED_CLUTCH, MACH_SCHED_EDGE_LOAD_AVG) | DBG_FUNC_NONE, pset->pset_cluster_id, (load_average >> SCHED_PSET_LOAD_EWMA_FRACTION_BITS), load_average & SCHED_PSET_LOAD_EWMA_FRACTION_MASK, sched_bucket);
+ }
+ os_atomic_store(&pset->pset_load_last_update, curtime, relaxed);
+}
+
+void
+sched_update_pset_avg_execution_time(processor_set_t pset, uint64_t execution_time, uint64_t curtime, sched_bucket_t sched_bucket)
+{
+ pset_execution_time_t old_execution_time_packed, new_execution_time_packed;
+ uint64_t avg_thread_execution_time = 0;
+
+ os_atomic_rmw_loop(&pset->pset_execution_time[sched_bucket].pset_execution_time_packed,
+ old_execution_time_packed.pset_execution_time_packed,
+ new_execution_time_packed.pset_execution_time_packed, relaxed, {
+ uint64_t last_update = old_execution_time_packed.pset_execution_time_last_update;
+ int64_t delta_ticks = curtime - last_update;
+ if (delta_ticks < 0) {
+ /*
+ * Its possible that another CPU came in and updated the pset_execution_time
+ * before this CPU could do it. Since the average execution time is meant to
+ * be an approximate measure per cluster, ignore the older update.
+ */
+ os_atomic_rmw_loop_give_up(return );
+ }
+ uint64_t delta_nsecs = 0;
+ absolutetime_to_nanoseconds(delta_ticks, &delta_nsecs);
+
+ uint64_t nanotime = 0;
+ absolutetime_to_nanoseconds(execution_time, &nanotime);
+ uint64_t execution_time_us = nanotime / NSEC_PER_USEC;
+
+ uint64_t old_execution_time = (old_execution_time_packed.pset_avg_thread_execution_time * SCHED_PSET_LOAD_EWMA_TC_NSECS);
+ uint64_t new_execution_time = (execution_time_us * delta_nsecs);
+
+ avg_thread_execution_time = (old_execution_time + new_execution_time) / (delta_nsecs + SCHED_PSET_LOAD_EWMA_TC_NSECS);
+ new_execution_time_packed.pset_avg_thread_execution_time = avg_thread_execution_time;
+ new_execution_time_packed.pset_execution_time_last_update = curtime;
+ });
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_PSET_AVG_EXEC_TIME) | DBG_FUNC_NONE, pset->pset_cluster_id, avg_thread_execution_time, sched_bucket);
+}
+
+#else /* CONFIG_SCHED_EDGE */
+
+void
+sched_update_pset_load_average(processor_set_t pset, __unused uint64_t curtime)
+{
+ int non_rt_load = pset->pset_runq.count;
+ int load = ((bit_count(pset->cpu_state_map[PROCESSOR_RUNNING]) + non_rt_load + rt_runq_count(pset)) << PSET_LOAD_NUMERATOR_SHIFT);
+ int new_load_average = ((int)pset->load_average + load) >> 1;
+
+ pset->load_average = new_load_average;
+#if (DEVELOPMENT || DEBUG)
+#if __AMP__
+ if (pset->pset_cluster_type == PSET_AMP_P) {
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_PSET_LOAD_AVERAGE) | DBG_FUNC_NONE, sched_get_pset_load_average(pset, 0), (bit_count(pset->cpu_state_map[PROCESSOR_RUNNING]) + pset->pset_runq.count + rt_runq_count(pset)));
+ }
+#endif
+#endif
+}
+
+void
+sched_update_pset_avg_execution_time(__unused processor_set_t pset, __unused uint64_t execution_time, __unused uint64_t curtime, __unused sched_bucket_t sched_bucket)
+{
+}
+#endif /* CONFIG_SCHED_EDGE */
+
+/* pset is locked */
+static bool
+processor_is_fast_track_candidate_for_realtime_thread(processor_set_t pset, processor_t processor)
+{
+ int cpuid = processor->cpu_id;
+#if defined(__x86_64__)
+ if (sched_avoid_cpu0 && (cpuid == 0)) {
+ return false;
+ }
+#endif
+
+ cpumap_t fasttrack_map = pset_available_cpumap(pset) & ~pset->pending_AST_URGENT_cpu_mask & ~pset->realtime_map;
+
+ return bit_test(fasttrack_map, cpuid);
+}
+
+/* pset is locked */
+static processor_t
+choose_processor_for_realtime_thread(processor_set_t pset, processor_t skip_processor, bool consider_secondaries)
+{
+#if defined(__x86_64__)
+ bool avoid_cpu0 = sched_avoid_cpu0 && bit_test(pset->cpu_bitmask, 0);
+#else
+ const bool avoid_cpu0 = false;
+#endif
+
+ cpumap_t cpu_map = pset_available_cpumap(pset) & ~pset->pending_AST_URGENT_cpu_mask & ~pset->realtime_map;
+ if (skip_processor) {
+ bit_clear(cpu_map, skip_processor->cpu_id);
+ }
+
+ cpumap_t primary_map = cpu_map & pset->primary_map;
+ if (avoid_cpu0) {
+ primary_map = bit_ror64(primary_map, 1);
+ }
+
+ int rotid = lsb_first(primary_map);
+ if (rotid >= 0) {
+ int cpuid = avoid_cpu0 ? ((rotid + 1) & 63) : rotid;
+
+ processor_t processor = processor_array[cpuid];
+
+ return processor;
+ }
+
+ if (!pset->is_SMT || !sched_allow_rt_smt || !consider_secondaries) {
+ goto out;
+ }
+
+ /* Consider secondary processors */
+ cpumap_t secondary_map = cpu_map & ~pset->primary_map;
+ if (avoid_cpu0) {
+ /* Also avoid cpu1 */
+ secondary_map = bit_ror64(secondary_map, 2);
+ }
+ rotid = lsb_first(secondary_map);
+ if (rotid >= 0) {
+ int cpuid = avoid_cpu0 ? ((rotid + 2) & 63) : rotid;
+
+ processor_t processor = processor_array[cpuid];
+
+ return processor;
+ }
+
+out:
+ if (skip_processor) {
+ return PROCESSOR_NULL;
+ }
+
+ /*
+ * If we didn't find an obvious processor to choose, but there are still more CPUs
+ * not already running realtime threads than realtime threads in the realtime run queue,
+ * this thread belongs in this pset, so choose some other processor in this pset
+ * to ensure the thread is enqueued here.
+ */
+ cpumap_t non_realtime_map = pset_available_cpumap(pset) & pset->primary_map & ~pset->realtime_map;
+ if (bit_count(non_realtime_map) > rt_runq_count(pset)) {
+ cpu_map = non_realtime_map;
+ assert(cpu_map != 0);
+ int cpuid = bit_first(cpu_map);
+ assert(cpuid >= 0);
+ return processor_array[cpuid];
+ }
+
+ if (!pset->is_SMT || !sched_allow_rt_smt || !consider_secondaries) {
+ goto skip_secondaries;
+ }
+
+ non_realtime_map = pset_available_cpumap(pset) & ~pset->realtime_map;
+ if (bit_count(non_realtime_map) > rt_runq_count(pset)) {
+ cpu_map = non_realtime_map;
+ assert(cpu_map != 0);
+ int cpuid = bit_first(cpu_map);
+ assert(cpuid >= 0);
+ return processor_array[cpuid];
+ }
+
+skip_secondaries:
+ return PROCESSOR_NULL;
+}
+
+/* pset is locked */
+static bool
+all_available_primaries_are_running_realtime_threads(processor_set_t pset)
+{
+ cpumap_t cpu_map = pset_available_cpumap(pset) & pset->primary_map & ~pset->realtime_map;
+ return rt_runq_count(pset) > bit_count(cpu_map);
+}
+
+#if defined(__x86_64__)
+/* pset is locked */
+static bool
+these_processors_are_running_realtime_threads(processor_set_t pset, uint64_t these_map)
+{
+ cpumap_t cpu_map = pset_available_cpumap(pset) & these_map & ~pset->realtime_map;
+ return rt_runq_count(pset) > bit_count(cpu_map);
+}
+#endif
+
+static bool
+sched_ok_to_run_realtime_thread(processor_set_t pset, processor_t processor)
+{
+ bool ok_to_run_realtime_thread = true;
+#if defined(__x86_64__)
+ if (sched_avoid_cpu0 && processor->cpu_id == 0) {
+ ok_to_run_realtime_thread = these_processors_are_running_realtime_threads(pset, pset->primary_map & ~0x1);
+ } else if (sched_avoid_cpu0 && (processor->cpu_id == 1) && processor->is_SMT) {
+ ok_to_run_realtime_thread = sched_allow_rt_smt && these_processors_are_running_realtime_threads(pset, ~0x2);
+ } else if (processor->processor_primary != processor) {
+ ok_to_run_realtime_thread = (sched_allow_rt_smt && all_available_primaries_are_running_realtime_threads(pset));
+ }
+#else
+ (void)pset;
+ (void)processor;
+#endif
+ return ok_to_run_realtime_thread;
+}
+
+void
+sched_pset_made_schedulable(__unused processor_t processor, processor_set_t pset, boolean_t drop_lock)
+{
+ if (drop_lock) {
+ pset_unlock(pset);
+ }
+}
+
+void
+thread_set_no_smt(bool set)
+{
+ if (!system_is_SMT) {
+ /* Not a machine that supports SMT */
+ return;
+ }
+
+ thread_t thread = current_thread();
+
+ spl_t s = splsched();
+ thread_lock(thread);
+ if (set) {
+ thread->sched_flags |= TH_SFLAG_NO_SMT;
+ }
+ thread_unlock(thread);
+ splx(s);
+}
+
+bool
+thread_get_no_smt(void)
+{
+ return current_thread()->sched_flags & TH_SFLAG_NO_SMT;
+}
+
+extern void task_set_no_smt(task_t);
+void
+task_set_no_smt(task_t task)
+{
+ if (!system_is_SMT) {
+ /* Not a machine that supports SMT */
+ return;
+ }
+
+ if (task == TASK_NULL) {
+ task = current_task();
+ }
+
+ task_lock(task);
+ task->t_flags |= TF_NO_SMT;
+ task_unlock(task);
+}
+
+#if DEBUG || DEVELOPMENT
+extern void sysctl_task_set_no_smt(char no_smt);
+void
+sysctl_task_set_no_smt(char no_smt)
+{
+ if (!system_is_SMT) {
+ /* Not a machine that supports SMT */
+ return;
+ }
+
+ task_t task = current_task();
+
+ task_lock(task);
+ if (no_smt == '1') {
+ task->t_flags |= TF_NO_SMT;
+ }
+ task_unlock(task);
+}
+
+extern char sysctl_task_get_no_smt(void);
+char
+sysctl_task_get_no_smt(void)
+{
+ task_t task = current_task();
+
+ if (task->t_flags & TF_NO_SMT) {
+ return '1';
+ }
+ return '0';
+}
+#endif /* DEVELOPMENT || DEBUG */
+
+
+__private_extern__ void
+thread_bind_cluster_type(thread_t thread, char cluster_type, bool soft_bound)
+{
+#if __AMP__
+ spl_t s = splsched();
+ thread_lock(thread);
+ thread->sched_flags &= ~(TH_SFLAG_ECORE_ONLY | TH_SFLAG_PCORE_ONLY | TH_SFLAG_BOUND_SOFT);
+ if (soft_bound) {
+ thread->sched_flags |= TH_SFLAG_BOUND_SOFT;
+ }
+ switch (cluster_type) {
+ case 'e':
+ case 'E':
+ thread->sched_flags |= TH_SFLAG_ECORE_ONLY;
+ break;
+ case 'p':
+ case 'P':
+ thread->sched_flags |= TH_SFLAG_PCORE_ONLY;
+ break;
+ default:
+ break;
+ }
+ thread_unlock(thread);
+ splx(s);
+
+ if (thread == current_thread()) {
+ thread_block(THREAD_CONTINUE_NULL);
+ }
+#else /* __AMP__ */
+ (void)thread;
+ (void)cluster_type;
+ (void)soft_bound;
+#endif /* __AMP__ */
+}
+
+#if DEVELOPMENT || DEBUG
+extern int32_t sysctl_get_bound_cpuid(void);
+int32_t
+sysctl_get_bound_cpuid(void)
+{
+ int32_t cpuid = -1;
+ thread_t self = current_thread();
+
+ processor_t processor = self->bound_processor;
+ if (processor == NULL) {
+ cpuid = -1;
+ } else {
+ cpuid = processor->cpu_id;
+ }
+
+ return cpuid;
+}
+
+extern kern_return_t sysctl_thread_bind_cpuid(int32_t cpuid);
+kern_return_t
+sysctl_thread_bind_cpuid(int32_t cpuid)
+{
+ processor_t processor = PROCESSOR_NULL;
+
+ if (cpuid == -1) {
+ goto unbind;
+ }
+
+ if (cpuid < 0 || cpuid >= MAX_SCHED_CPUS) {
+ return KERN_INVALID_VALUE;
+ }
+
+ processor = processor_array[cpuid];
+ if (processor == PROCESSOR_NULL) {
+ return KERN_INVALID_VALUE;
+ }
+
+#if __AMP__
+
+ thread_t thread = current_thread();
+
+ if (thread->sched_flags & (TH_SFLAG_ECORE_ONLY | TH_SFLAG_PCORE_ONLY)) {
+ if ((thread->sched_flags & TH_SFLAG_BOUND_SOFT) == 0) {
+ /* Cannot hard-bind an already hard-cluster-bound thread */
+ return KERN_NOT_SUPPORTED;
+ }
+ }
+
+#endif /* __AMP__ */
+
+unbind:
+ thread_bind(processor);
+
+ thread_block(THREAD_CONTINUE_NULL);
+ return KERN_SUCCESS;
}
-#endif /* DEBUG */
+#endif /* DEVELOPMENT || DEBUG */
projects / apple / xnu.git / blobdiff