// -*- Mode: CFA -*- // // Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo // // The contents of this file are covered under the licence agreement in the // file "LICENCE" distributed with Cforall. // // kernel.c -- // // Author : Thierry Delisle // Created On : Tue Jan 17 12:27:26 2017 // Last Modified By : Thierry Delisle // Last Modified On : -- // Update Count : 0 // #include "startup.h" //Start and stop routine for the kernel, declared first to make sure they run first void kernel_startup(void) __attribute__(( constructor( STARTUP_PRIORITY_KERNEL ) )); void kernel_shutdown(void) __attribute__(( destructor ( STARTUP_PRIORITY_KERNEL ) )); //Header #include "kernel_private.h" //C Includes #include extern "C" { #include #include #include #include #include } //CFA Includes #include "libhdr.h" #include "preemption.h" //Private includes #define __CFA_INVOKE_PRIVATE__ #include "invoke.h" //----------------------------------------------------------------------------- // Kernel storage #define KERNEL_STORAGE(T,X) static char X##_storage[sizeof(T)] KERNEL_STORAGE(processorCtx_t, systemProcessorCtx); KERNEL_STORAGE(cluster, systemCluster); KERNEL_STORAGE(system_proc_t, systemProcessor); KERNEL_STORAGE(thread_desc, mainThread); KERNEL_STORAGE(machine_context_t, mainThread_context); cluster * systemCluster; system_proc_t * systemProcessor; thread_desc * mainThread; //----------------------------------------------------------------------------- // Global state volatile thread_local processor * this_processor; volatile thread_local coroutine_desc * this_coroutine; volatile thread_local thread_desc * this_thread; volatile thread_local unsigned short disable_preempt_count = 1; //----------------------------------------------------------------------------- // Main thread construction struct current_stack_info_t { machine_context_t ctx; unsigned int size; // size of stack void *base; // base of stack void *storage; // pointer to stack void *limit; // stack grows towards stack limit void *context; // address of cfa_context_t void *top; // address of top of storage }; void ?{}( current_stack_info_t * this ) { CtxGet( &this->ctx ); this->base = this->ctx.FP; this->storage = this->ctx.SP; rlimit r; getrlimit( RLIMIT_STACK, &r); this->size = r.rlim_cur; this->limit = (void *)(((intptr_t)this->base) - this->size); this->context = &mainThread_context_storage; this->top = this->base; } void ?{}( coStack_t * this, current_stack_info_t * info) { this->size = info->size; this->storage = info->storage; this->limit = info->limit; this->base = info->base; this->context = info->context; this->top = info->top; this->userStack = true; } void ?{}( coroutine_desc * this, current_stack_info_t * info) { (&this->stack){ info }; this->name = "Main Thread"; this->errno_ = 0; this->state = Start; } void ?{}( thread_desc * this, current_stack_info_t * info) { (&this->cor){ info }; } //----------------------------------------------------------------------------- // Processor coroutine void ?{}(processorCtx_t * this, processor * proc) { (&this->__cor){ "Processor" }; this->proc = proc; proc->runner = this; } void ?{}(processorCtx_t * this, processor * proc, current_stack_info_t * info) { (&this->__cor){ info }; this->proc = proc; proc->runner = this; } void ?{}(processor * this) { this{ systemCluster }; } void ?{}(processor * this, cluster * cltr) { this->cltr = cltr; (&this->terminated){}; this->is_terminated = false; this->preemption_alarm = NULL; this->preemption = default_preemption(); this->pending_preemption = false; start( this ); } void ?{}(processor * this, cluster * cltr, processorCtx_t * runner) { this->cltr = cltr; (&this->terminated){}; this->is_terminated = false; this->preemption_alarm = NULL; this->preemption = default_preemption(); this->pending_preemption = false; this->kernel_thread = pthread_self(); this->runner = runner; LIB_DEBUG_PRINT_SAFE("Kernel : constructing system processor context %p\n", runner); runner{ this }; } LIB_DEBUG_DO( bool validate( alarm_list_t * this ); ) void ?{}(system_proc_t * this, cluster * cltr, processorCtx_t * runner) { (&this->alarms){}; (&this->alarm_lock){}; this->pending_alarm = false; (&this->proc){ cltr, runner }; verify( validate( &this->alarms ) ); } void ^?{}(processor * this) { if( ! this->is_terminated ) { LIB_DEBUG_PRINT_SAFE("Kernel : core %p signaling termination\n", this); this->is_terminated = true; wait( &this->terminated ); } } void ?{}(cluster * this) { ( &this->ready_queue ){}; ( &this->lock ){}; } void ^?{}(cluster * this) { } //============================================================================================= // Kernel Scheduling logic //============================================================================================= //Main of the processor contexts void main(processorCtx_t * runner) { processor * this = runner->proc; LIB_DEBUG_PRINT_SAFE("Kernel : core %p starting\n", this); { // Setup preemption data preemption_scope scope = { this }; LIB_DEBUG_PRINT_SAFE("Kernel : core %p started\n", this); thread_desc * readyThread = NULL; for( unsigned int spin_count = 0; ! this->is_terminated; spin_count++ ) { readyThread = nextThread( this->cltr ); if(readyThread) { verify( disable_preempt_count > 0 ); runThread(this, readyThread); verify( disable_preempt_count > 0 ); //Some actions need to be taken from the kernel finishRunning(this); spin_count = 0; } else { spin(this, &spin_count); } } LIB_DEBUG_PRINT_SAFE("Kernel : core %p stopping\n", this); } signal( &this->terminated ); LIB_DEBUG_PRINT_SAFE("Kernel : core %p terminated\n", this); } // runThread runs a thread by context switching // from the processor coroutine to the target thread void runThread(processor * this, thread_desc * dst) { coroutine_desc * proc_cor = get_coroutine(this->runner); coroutine_desc * thrd_cor = get_coroutine(dst); //Reset the terminating actions here this->finish.action_code = No_Action; //Update global state this_thread = dst; // Context Switch to the thread ThreadCtxSwitch(proc_cor, thrd_cor); // when ThreadCtxSwitch returns we are back in the processor coroutine } // Once a thread has finished running, some of // its final actions must be executed from the kernel void finishRunning(processor * this) { if( this->finish.action_code == Release ) { unlock( this->finish.lock ); } else if( this->finish.action_code == Schedule ) { ScheduleThread( this->finish.thrd ); } else if( this->finish.action_code == Release_Schedule ) { unlock( this->finish.lock ); ScheduleThread( this->finish.thrd ); } else if( this->finish.action_code == Release_Multi ) { for(int i = 0; i < this->finish.lock_count; i++) { unlock( this->finish.locks[i] ); } } else if( this->finish.action_code == Release_Multi_Schedule ) { for(int i = 0; i < this->finish.lock_count; i++) { unlock( this->finish.locks[i] ); } for(int i = 0; i < this->finish.thrd_count; i++) { ScheduleThread( this->finish.thrds[i] ); } } else { assert(this->finish.action_code == No_Action); } } // Handles spinning logic // TODO : find some strategy to put cores to sleep after some time void spin(processor * this, unsigned int * spin_count) { (*spin_count)++; } // Context invoker for processors // This is the entry point for processors (kernel threads) // It effectively constructs a coroutine by stealing the pthread stack void * CtxInvokeProcessor(void * arg) { processor * proc = (processor *) arg; this_processor = proc; this_coroutine = NULL; this_thread = NULL; disable_preempt_count = 1; // SKULLDUGGERY: We want to create a context for the processor coroutine // which is needed for the 2-step context switch. However, there is no reason // to waste the perfectly valid stack create by pthread. current_stack_info_t info; machine_context_t ctx; info.context = &ctx; processorCtx_t proc_cor_storage = { proc, &info }; LIB_DEBUG_PRINT_SAFE("Coroutine : created stack %p\n", proc_cor_storage.__cor.stack.base); //Set global state this_coroutine = &proc->runner->__cor; this_thread = NULL; //We now have a proper context from which to schedule threads LIB_DEBUG_PRINT_SAFE("Kernel : core %p created (%p, %p)\n", proc, proc->runner, &ctx); // SKULLDUGGERY: Since the coroutine doesn't have its own stack, we can't // resume it to start it like it normally would, it will just context switch // back to here. Instead directly call the main since we already are on the // appropriate stack. proc_cor_storage.__cor.state = Active; main( &proc_cor_storage ); proc_cor_storage.__cor.state = Halted; // Main routine of the core returned, the core is now fully terminated LIB_DEBUG_PRINT_SAFE("Kernel : core %p main ended (%p)\n", proc, proc->runner); return NULL; } void start(processor * this) { LIB_DEBUG_PRINT_SAFE("Kernel : Starting core %p\n", this); // SIGALRM must only be caught by the system processor sigset_t old_mask; bool is_system_proc = this_processor == &systemProcessor->proc; if ( is_system_proc ) { // Child kernel-thread inherits the signal mask from the parent kernel-thread. So one special case for the // system processor creating the user processor => toggle the blocking SIGALRM on system processor, create user // processor, and toggle back (below) previous signal mask of the system processor. sigset_t new_mask; sigemptyset( &new_mask ); sigemptyset( &old_mask ); sigaddset( &new_mask, SIGALRM ); if ( sigprocmask( SIG_BLOCK, &new_mask, &old_mask ) == -1 ) { abortf( "internal error, sigprocmask" ); } assert( ! sigismember( &old_mask, SIGALRM ) ); } pthread_create( &this->kernel_thread, NULL, CtxInvokeProcessor, (void*)this ); // Toggle back previous signal mask of system processor. if ( is_system_proc ) { if ( sigprocmask( SIG_SETMASK, &old_mask, NULL ) == -1 ) { abortf( "internal error, sigprocmask" ); } // if } // if LIB_DEBUG_PRINT_SAFE("Kernel : core %p started\n", this); } //----------------------------------------------------------------------------- // Scheduler routines void ScheduleThread( thread_desc * thrd ) { // if( !thrd ) return; assert( thrd ); assert( thrd->cor.state != Halted ); verify( disable_preempt_count > 0 ); verifyf( thrd->next == NULL, "Expected null got %p", thrd->next ); lock( &systemProcessor->proc.cltr->lock, __PRETTY_FUNCTION__ ); append( &systemProcessor->proc.cltr->ready_queue, thrd ); unlock( &systemProcessor->proc.cltr->lock ); verify( disable_preempt_count > 0 ); } thread_desc * nextThread(cluster * this) { verify( disable_preempt_count > 0 ); lock( &this->lock, __PRETTY_FUNCTION__ ); thread_desc * head = pop_head( &this->ready_queue ); unlock( &this->lock ); verify( disable_preempt_count > 0 ); return head; } void BlockInternal() { disable_interrupts(); verify( disable_preempt_count > 0 ); suspend(); verify( disable_preempt_count > 0 ); enable_interrupts( __PRETTY_FUNCTION__ ); } void BlockInternal( spinlock * lock ) { disable_interrupts(); this_processor->finish.action_code = Release; this_processor->finish.lock = lock; verify( disable_preempt_count > 0 ); suspend(); verify( disable_preempt_count > 0 ); enable_interrupts( __PRETTY_FUNCTION__ ); } void BlockInternal( thread_desc * thrd ) { disable_interrupts(); assert( thrd->cor.state != Halted ); this_processor->finish.action_code = Schedule; this_processor->finish.thrd = thrd; verify( disable_preempt_count > 0 ); suspend(); verify( disable_preempt_count > 0 ); enable_interrupts( __PRETTY_FUNCTION__ ); } void BlockInternal( spinlock * lock, thread_desc * thrd ) { disable_interrupts(); this_processor->finish.action_code = Release_Schedule; this_processor->finish.lock = lock; this_processor->finish.thrd = thrd; verify( disable_preempt_count > 0 ); suspend(); verify( disable_preempt_count > 0 ); enable_interrupts( __PRETTY_FUNCTION__ ); } void BlockInternal(spinlock ** locks, unsigned short count) { disable_interrupts(); this_processor->finish.action_code = Release_Multi; this_processor->finish.locks = locks; this_processor->finish.lock_count = count; verify( disable_preempt_count > 0 ); suspend(); verify( disable_preempt_count > 0 ); enable_interrupts( __PRETTY_FUNCTION__ ); } void BlockInternal(spinlock ** locks, unsigned short lock_count, thread_desc ** thrds, unsigned short thrd_count) { disable_interrupts(); this_processor->finish.action_code = Release_Multi_Schedule; this_processor->finish.locks = locks; this_processor->finish.lock_count = lock_count; this_processor->finish.thrds = thrds; this_processor->finish.thrd_count = thrd_count; verify( disable_preempt_count > 0 ); suspend(); verify( disable_preempt_count > 0 ); enable_interrupts( __PRETTY_FUNCTION__ ); } //============================================================================================= // Kernel Setup logic //============================================================================================= //----------------------------------------------------------------------------- // Kernel boot procedures void kernel_startup(void) { LIB_DEBUG_PRINT_SAFE("Kernel : Starting\n"); // Start by initializing the main thread // SKULLDUGGERY: the mainThread steals the process main thread // which will then be scheduled by the systemProcessor normally mainThread = (thread_desc *)&mainThread_storage; current_stack_info_t info; mainThread{ &info }; LIB_DEBUG_PRINT_SAFE("Kernel : Main thread ready\n"); // Initialize the system cluster systemCluster = (cluster *)&systemCluster_storage; systemCluster{}; LIB_DEBUG_PRINT_SAFE("Kernel : System cluster ready\n"); // Initialize the system processor and the system processor ctx // (the coroutine that contains the processing control flow) systemProcessor = (system_proc_t *)&systemProcessor_storage; systemProcessor{ systemCluster, (processorCtx_t *)&systemProcessorCtx_storage }; // Add the main thread to the ready queue // once resume is called on systemProcessor->runner the mainThread needs to be scheduled like any normal thread ScheduleThread(mainThread); //initialize the global state variables this_processor = &systemProcessor->proc; this_thread = mainThread; this_coroutine = &mainThread->cor; disable_preempt_count = 1; // Enable preemption kernel_start_preemption(); // SKULLDUGGERY: Force a context switch to the system processor to set the main thread's context to the current UNIX // context. Hence, the main thread does not begin through CtxInvokeThread, like all other threads. The trick here is that // mainThread is on the ready queue when this call is made. resume( systemProcessor->proc.runner ); // THE SYSTEM IS NOW COMPLETELY RUNNING LIB_DEBUG_PRINT_SAFE("Kernel : Started\n--------------------------------------------------\n\n"); enable_interrupts( __PRETTY_FUNCTION__ ); } void kernel_shutdown(void) { LIB_DEBUG_PRINT_SAFE("\n--------------------------------------------------\nKernel : Shutting down\n"); disable_interrupts(); // SKULLDUGGERY: Notify the systemProcessor it needs to terminates. // When its coroutine terminates, it return control to the mainThread // which is currently here systemProcessor->proc.is_terminated = true; suspend(); // THE SYSTEM IS NOW COMPLETELY STOPPED // Disable preemption kernel_stop_preemption(); // Destroy the system processor and its context in reverse order of construction // These were manually constructed so we need manually destroy them ^(systemProcessor->proc.runner){}; ^(systemProcessor){}; // Final step, destroy the main thread since it is no longer needed // Since we provided a stack to this taxk it will not destroy anything ^(mainThread){}; LIB_DEBUG_PRINT_SAFE("Kernel : Shutdown complete\n"); } static spinlock kernel_abort_lock; static spinlock kernel_debug_lock; static bool kernel_abort_called = false; void * kernel_abort (void) __attribute__ ((__nothrow__)) { // abort cannot be recursively entered by the same or different processors because all signal handlers return when // the globalAbort flag is true. lock( &kernel_abort_lock, __PRETTY_FUNCTION__ ); // first task to abort ? if ( !kernel_abort_called ) { // not first task to abort ? kernel_abort_called = true; unlock( &kernel_abort_lock ); } else { unlock( &kernel_abort_lock ); sigset_t mask; sigemptyset( &mask ); sigaddset( &mask, SIGALRM ); // block SIGALRM signals sigaddset( &mask, SIGUSR1 ); // block SIGUSR1 signals sigsuspend( &mask ); // block the processor to prevent further damage during abort _exit( EXIT_FAILURE ); // if processor unblocks before it is killed, terminate it } return this_thread; } void kernel_abort_msg( void * kernel_data, char * abort_text, int abort_text_size ) { thread_desc * thrd = kernel_data; int len = snprintf( abort_text, abort_text_size, "Error occurred while executing task %.256s (%p)", thrd->cor.name, thrd ); __lib_debug_write( STDERR_FILENO, abort_text, len ); if ( thrd != this_coroutine ) { len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", this_coroutine->name, this_coroutine ); __lib_debug_write( STDERR_FILENO, abort_text, len ); } else { __lib_debug_write( STDERR_FILENO, ".\n", 2 ); } } extern "C" { void __lib_debug_acquire() { lock(&kernel_debug_lock, __PRETTY_FUNCTION__); } void __lib_debug_release() { unlock(&kernel_debug_lock); } } //============================================================================================= // Kernel Utilities //============================================================================================= //----------------------------------------------------------------------------- // Locks void ?{}( spinlock * this ) { this->lock = 0; } void ^?{}( spinlock * this ) { } bool try_lock( spinlock * this, const char * caller ) { bool ret = this->lock == 0 && __sync_lock_test_and_set_4( &this->lock, 1 ) == 0; this->prev = caller; return ret; } void lock( spinlock * this, const char * caller ) { for ( unsigned int i = 1;; i += 1 ) { if ( this->lock == 0 && __sync_lock_test_and_set_4( &this->lock, 1 ) == 0 ) break; } this->prev = caller; } void unlock( spinlock * this ) { __sync_lock_release_4( &this->lock ); } void ?{}( signal_once * this ) { this->cond = false; } void ^?{}( signal_once * this ) { } void wait( signal_once * this ) { lock( &this->lock, __PRETTY_FUNCTION__ ); if( !this->cond ) { append( &this->blocked, (thread_desc*)this_thread ); BlockInternal( &this->lock ); } else { unlock( &this->lock ); } } void signal( signal_once * this ) { lock( &this->lock, __PRETTY_FUNCTION__ ); { this->cond = true; disable_interrupts(); thread_desc * it; while( it = pop_head( &this->blocked) ) { ScheduleThread( it ); } enable_interrupts( __PRETTY_FUNCTION__ ); } unlock( &this->lock ); } //----------------------------------------------------------------------------- // Queues void ?{}( __thread_queue_t * this ) { this->head = NULL; this->tail = &this->head; } void append( __thread_queue_t * this, thread_desc * t ) { verify(this->tail != NULL); *this->tail = t; this->tail = &t->next; } thread_desc * pop_head( __thread_queue_t * this ) { thread_desc * head = this->head; if( head ) { this->head = head->next; if( !head->next ) { this->tail = &this->head; } head->next = NULL; } return head; } void ?{}( __condition_stack_t * this ) { this->top = NULL; } void push( __condition_stack_t * this, __condition_criterion_t * t ) { verify( !t->next ); t->next = this->top; this->top = t; } __condition_criterion_t * pop( __condition_stack_t * this ) { __condition_criterion_t * top = this->top; if( top ) { this->top = top->next; top->next = NULL; } return top; } // Local Variables: // // mode: c // // tab-width: 4 // // End: //