X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/316670eb35587141e969394ae8537d66b9211e80..c7d2c2c6ee645e10cbccdd01c6191873ec77239d:/osfmk/kern/syscall_subr.c diff --git a/osfmk/kern/syscall_subr.c b/osfmk/kern/syscall_subr.c index 89fc63b1b..f0c067b03 100644 --- a/osfmk/kern/syscall_subr.c +++ b/osfmk/kern/syscall_subr.c @@ -73,10 +73,11 @@ #include #include #include - +#include #ifdef MACH_BSD extern void workqueue_thread_yielded(void); +extern sched_call_t workqueue_get_sched_callback(void); #endif /* MACH_BSD */ @@ -97,14 +98,7 @@ __unused struct pfz_exit_args *args) * swtch and swtch_pri both attempt to context switch (logic in * thread_block no-ops the context switch if nothing would happen). * A boolean is returned that indicates whether there is anything - * else runnable. - * - * This boolean can be used by a thread waiting on a - * lock or condition: If FALSE is returned, the thread is justified - * in becoming a resource hog by continuing to spin because there's - * nothing else useful that the processor could do. If TRUE is - * returned, the thread should make one more check on the - * lock and then be a good citizen and really suspend. + * else runnable. That's no excuse to spin, though. */ static void @@ -199,15 +193,48 @@ __unused struct swtch_pri_args *args) return (result); } +static int +thread_switch_disable_workqueue_sched_callback(void) +{ + sched_call_t callback = workqueue_get_sched_callback(); + thread_t self = current_thread(); + if (!callback || self->sched_call != callback) { + return FALSE; + } + spl_t s = splsched(); + thread_lock(self); + thread_sched_call(self, NULL); + thread_unlock(self); + splx(s); + return TRUE; +} + +static void +thread_switch_enable_workqueue_sched_callback(void) +{ + sched_call_t callback = workqueue_get_sched_callback(); + thread_t self = current_thread(); + spl_t s = splsched(); + thread_lock(self); + thread_sched_call(self, callback); + thread_unlock(self); + splx(s); +} + static void thread_switch_continue(void) { register thread_t self = current_thread(); int option = self->saved.swtch.option; + boolean_t reenable_workq_callback = self->saved.swtch.reenable_workq_callback; + - if (option == SWITCH_OPTION_DEPRESS) + if (option == SWITCH_OPTION_DEPRESS || option == SWITCH_OPTION_OSLOCK_DEPRESS) thread_depress_abort_internal(self); + if (reenable_workq_callback) + thread_switch_enable_workqueue_sched_callback(); + thread_syscall_return(KERN_SUCCESS); /*NOTREACHED*/ } @@ -221,35 +248,60 @@ kern_return_t thread_switch( struct thread_switch_args *args) { - register thread_t thread, self = current_thread(); + thread_t thread = THREAD_NULL; + thread_t self = current_thread(); mach_port_name_t thread_name = args->thread_name; int option = args->option; mach_msg_timeout_t option_time = args->option_time; + uint32_t scale_factor = NSEC_PER_MSEC; + boolean_t reenable_workq_callback = FALSE; + boolean_t depress_option = FALSE; + boolean_t wait_option = FALSE; /* - * Process option. + * Validate and process option. */ switch (option) { case SWITCH_OPTION_NONE: - case SWITCH_OPTION_DEPRESS: + workqueue_thread_yielded(); + break; case SWITCH_OPTION_WAIT: - break; - + wait_option = TRUE; + workqueue_thread_yielded(); + break; + case SWITCH_OPTION_DEPRESS: + depress_option = TRUE; + workqueue_thread_yielded(); + break; + case SWITCH_OPTION_DISPATCH_CONTENTION: + scale_factor = NSEC_PER_USEC; + wait_option = TRUE; + if (thread_switch_disable_workqueue_sched_callback()) + reenable_workq_callback = TRUE; + break; + case SWITCH_OPTION_OSLOCK_DEPRESS: + depress_option = TRUE; + if (thread_switch_disable_workqueue_sched_callback()) + reenable_workq_callback = TRUE; + break; + case SWITCH_OPTION_OSLOCK_WAIT: + wait_option = TRUE; + if (thread_switch_disable_workqueue_sched_callback()) + reenable_workq_callback = TRUE; + break; default: return (KERN_INVALID_ARGUMENT); } - workqueue_thread_yielded(); - /* * Translate the port name if supplied. */ - if (thread_name != MACH_PORT_NULL) { - ipc_port_t port; + if (thread_name != MACH_PORT_NULL) { + ipc_port_t port; if (ipc_port_translate_send(self->task->itk_space, - thread_name, &port) == KERN_SUCCESS) { + thread_name, &port) == KERN_SUCCESS) { ip_reference(port); ip_unlock(port); @@ -257,79 +309,91 @@ thread_switch( ip_release(port); if (thread == self) { - (void)thread_deallocate_internal(thread); + thread_deallocate(thread); thread = THREAD_NULL; } } - else - thread = THREAD_NULL; } - else - thread = THREAD_NULL; + + if (option == SWITCH_OPTION_OSLOCK_DEPRESS || option == SWITCH_OPTION_OSLOCK_WAIT) { + if (thread != THREAD_NULL) { + + if (thread->task != self->task) { + /* + * OSLock boosting only applies to other threads + * in your same task (even if you have a port for + * a thread in another task) + */ + + thread_deallocate(thread); + thread = THREAD_NULL; + } else { + /* + * Attempt to kick the lock owner up to our same IO throttling tier. + * If the thread is currently blocked in throttle_lowpri_io(), + * it will immediately break out. + * + * TODO: SFI break out? + */ + int new_policy = proc_get_effective_thread_policy(self, TASK_POLICY_IO); + + set_thread_iotier_override(thread, new_policy); + } + } + } /* * Try to handoff if supplied. */ if (thread != THREAD_NULL) { - processor_t processor; - spl_t s; - - s = splsched(); - thread_lock(thread); - - /* - * Check that the thread is not bound - * to a different processor, and that realtime - * is not involved. - * - * Next, pull it off its run queue. If it - * doesn't come, it's not eligible. - */ - processor = current_processor(); - if (processor->current_pri < BASEPRI_RTQUEUES && - thread->sched_pri < BASEPRI_RTQUEUES && - (thread->bound_processor == PROCESSOR_NULL || - thread->bound_processor == processor) && - thread_run_queue_remove(thread) ) { - /* - * Hah, got it!! - */ - thread_unlock(thread); - - (void)thread_deallocate_internal(thread); - - if (option == SWITCH_OPTION_WAIT) + spl_t s = splsched(); + + /* This may return a different thread if the target is pushing on something */ + thread_t pulled_thread = thread_run_queue_remove_for_handoff(thread); + + KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED,MACH_SCHED_THREAD_SWITCH)|DBG_FUNC_NONE, + thread_tid(thread), thread->state, + pulled_thread ? TRUE : FALSE, 0, 0); + + if (pulled_thread != THREAD_NULL) { + /* We can't be dropping the last ref here */ + thread_deallocate_safe(thread); + + if (wait_option) assert_wait_timeout((event_t)assert_wait_timeout, THREAD_ABORTSAFE, - option_time, 1000*NSEC_PER_USEC); - else - if (option == SWITCH_OPTION_DEPRESS) + option_time, scale_factor); + else if (depress_option) thread_depress_ms(option_time); self->saved.swtch.option = option; + self->saved.swtch.reenable_workq_callback = reenable_workq_callback; - thread_run(self, (thread_continue_t)thread_switch_continue, NULL, thread); + thread_run(self, (thread_continue_t)thread_switch_continue, NULL, pulled_thread); /* NOTREACHED */ + panic("returned from thread_run!"); } - thread_unlock(thread); splx(s); thread_deallocate(thread); } - - if (option == SWITCH_OPTION_WAIT) - assert_wait_timeout((event_t)assert_wait_timeout, THREAD_ABORTSAFE, option_time, 1000*NSEC_PER_USEC); - else - if (option == SWITCH_OPTION_DEPRESS) + + if (wait_option) + assert_wait_timeout((event_t)assert_wait_timeout, THREAD_ABORTSAFE, option_time, scale_factor); + else if (depress_option) thread_depress_ms(option_time); - + self->saved.swtch.option = option; + self->saved.swtch.reenable_workq_callback = reenable_workq_callback; thread_block_reason((thread_continue_t)thread_switch_continue, NULL, AST_YIELD); - if (option == SWITCH_OPTION_DEPRESS) + if (depress_option) thread_depress_abort_internal(self); + if (reenable_workq_callback) + thread_switch_enable_workqueue_sched_callback(); + return (KERN_SUCCESS); } @@ -356,7 +420,7 @@ thread_depress_abstime( if (interval != 0) { clock_absolutetime_interval_to_deadline(interval, &deadline); - if (!timer_call_enter(&self->depress_timer, deadline, TIMER_CALL_CRITICAL)) + if (!timer_call_enter(&self->depress_timer, deadline, TIMER_CALL_USER_CRITICAL)) self->depress_timer_active++; } } @@ -371,7 +435,7 @@ thread_depress_ms( uint64_t abstime; clock_interval_to_absolutetime_interval( - interval, 1000*NSEC_PER_USEC, &abstime); + interval, NSEC_PER_MSEC, &abstime); thread_depress_abstime(abstime); } @@ -390,7 +454,7 @@ thread_depress_expire( thread_lock(thread); if (--thread->depress_timer_active == 0) { thread->sched_flags &= ~TH_SFLAG_DEPRESSED_MASK; - SCHED(compute_priority)(thread, FALSE); + thread_recompute_sched_pri(thread, FALSE); } thread_unlock(thread); splx(s); @@ -411,7 +475,7 @@ thread_depress_abort_internal( if (!(thread->sched_flags & TH_SFLAG_POLLDEPRESS)) { if (thread->sched_flags & TH_SFLAG_DEPRESSED_MASK) { thread->sched_flags &= ~TH_SFLAG_DEPRESSED_MASK; - SCHED(compute_priority)(thread, FALSE); + thread_recompute_sched_pri(thread, FALSE); result = KERN_SUCCESS; } @@ -453,12 +517,13 @@ thread_poll_yield( self->sched_flags |= TH_SFLAG_POLLDEPRESS; abstime += (total_computation >> sched_poll_yield_shift); - if (!timer_call_enter(&self->depress_timer, abstime, TIMER_CALL_CRITICAL)) + if (!timer_call_enter(&self->depress_timer, abstime, TIMER_CALL_USER_CRITICAL)) self->depress_timer_active++; - thread_unlock(self); - if ((preempt = csw_check(myprocessor)) != AST_NONE) + if ((preempt = csw_check(myprocessor, AST_NONE)) != AST_NONE) ast_on(preempt); + + thread_unlock(self); } } splx(s);