source: libcfa/src/concurrency/kernel.cfa@ d34575b

//
// 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 : Peter A. Buhr
// Last Modified On : Thu Jul  9 06:22:54 2020
// Update Count     : 66
//

#define __cforall_thread__
// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__

//C Includes
#include <stddef.h>
#include <errno.h>
#include <string.h>
#include <stdio.h>
#include <fenv.h>
#include <signal.h>
#include <unistd.h>
#include <limits.h>                                                                             // PTHREAD_STACK_MIN
#include <sys/mman.h>                                                                   // mprotect
extern "C" {
#include <sys/resource.h>
}

//CFA Includes
#include "time.hfa"
#include "kernel_private.hfa"
#include "preemption.hfa"
#include "startup.hfa"

//Private includes
#define __CFA_INVOKE_PRIVATE__
#include "invoke.h"


//-----------------------------------------------------------------------------
// Some assembly required
#if defined( __i386 )
        #define CtxGet( ctx )        \
                __asm__ volatile (     \
                        "movl %%esp,%0\n"\
                        "movl %%ebp,%1\n"\
                        : "=rm" (ctx.SP),\
                                "=rm" (ctx.FP) \
                )

        // mxcr : SSE Status and Control bits (control bits are preserved across function calls)
        // fcw  : X87 FPU control word (preserved across function calls)
        #define __x87_store         \
                uint32_t __mxcr;      \
                uint16_t __fcw;       \
                __asm__ volatile (    \
                        "stmxcsr %0\n"  \
                        "fnstcw  %1\n"  \
                        : "=m" (__mxcr),\
                                "=m" (__fcw)  \
                )

        #define __x87_load         \
                __asm__ volatile (   \
                        "fldcw  %1\n"  \
                        "ldmxcsr %0\n" \
                        ::"m" (__mxcr),\
                                "m" (__fcw)  \
                )

#elif defined( __x86_64 )
        #define CtxGet( ctx )        \
                __asm__ volatile (     \
                        "movq %%rsp,%0\n"\
                        "movq %%rbp,%1\n"\
                        : "=rm" (ctx.SP),\
                                "=rm" (ctx.FP) \
                )

        #define __x87_store         \
                uint32_t __mxcr;      \
                uint16_t __fcw;       \
                __asm__ volatile (    \
                        "stmxcsr %0\n"  \
                        "fnstcw  %1\n"  \
                        : "=m" (__mxcr),\
                                "=m" (__fcw)  \
                )

        #define __x87_load          \
                __asm__ volatile (    \
                        "fldcw  %1\n"   \
                        "ldmxcsr %0\n"  \
                        :: "m" (__mxcr),\
                                "m" (__fcw)  \
                )


#elif defined( __ARM_ARCH )
#define CtxGet( ctx ) __asm__ ( \
                "mov %0,%%sp\n"   \
                "mov %1,%%r11\n"   \
        : "=rm" (ctx.SP), "=rm" (ctx.FP) )
#else
        #error unknown hardware architecture
#endif

//-----------------------------------------------------------------------------
//Start and stop routine for the kernel, declared first to make sure they run first
static void __kernel_startup (void) __attribute__(( constructor( STARTUP_PRIORITY_KERNEL ) ));
static void __kernel_shutdown(void) __attribute__(( destructor ( STARTUP_PRIORITY_KERNEL ) ));

//-----------------------------------------------------------------------------
// Kernel Scheduling logic
static $thread * __next_thread(cluster * this);
static bool __has_next_thread(cluster * this);
static void __run_thread(processor * this, $thread * dst);
static bool __wake_proc(processor *);
static bool __wake_one(struct __processor_id_t * id, cluster * cltr);
static void __halt(processor * this);

//-----------------------------------------------------------------------------
// Kernel storage
KERNEL_STORAGE(cluster,              mainCluster);
KERNEL_STORAGE(processor,            mainProcessor);
KERNEL_STORAGE($thread,              mainThread);
KERNEL_STORAGE(__stack_t,            mainThreadCtx);
KERNEL_STORAGE(__scheduler_RWLock_t, __scheduler_lock);
#if !defined(__CFA_NO_STATISTICS__)
KERNEL_STORAGE(__stats_t, mainProcStats);
#endif

cluster              * mainCluster;
processor            * mainProcessor;
$thread              * mainThread;
__scheduler_RWLock_t * __scheduler_lock;

extern "C" {
        struct { __dllist_t(cluster) list; __spinlock_t lock; } __cfa_dbg_global_clusters;
}

size_t __page_size = 0;

//-----------------------------------------------------------------------------
// Global state
thread_local struct KernelThreadData kernelTLS __attribute__ ((tls_model ( "initial-exec" ))) @= {
        NULL,                                                                                           // cannot use 0p
        NULL,
        NULL,
        { 1, false, false },
};

//-----------------------------------------------------------------------------
// Struct to steal stack
struct current_stack_info_t {
        __stack_t * storage;                                                            // pointer to stack object
        void * base;                                                                            // base of stack
        void * limit;                                                                           // stack grows towards stack limit
        void * context;                                                                         // address of cfa_context_t
};

void ?{}( current_stack_info_t & this ) {
        __stack_context_t ctx;
        CtxGet( ctx );
        this.base = ctx.FP;

        rlimit r;
        getrlimit( RLIMIT_STACK, &r);
        size_t size = r.rlim_cur;

        this.limit = (void *)(((intptr_t)this.base) - size);
        this.context = &storage_mainThreadCtx;
}

//-----------------------------------------------------------------------------
// Main thread construction

void ?{}( $coroutine & this, current_stack_info_t * info) with( this ) {
        stack.storage = info->storage;
        with(*stack.storage) {
                limit     = info->limit;
                base      = info->base;
        }
        __attribute__((may_alias)) intptr_t * istorage = (intptr_t*) &stack.storage;
        *istorage |= 0x1;
        name = "Main Thread";
        state = Start;
        starter = 0p;
        last = 0p;
        cancellation = 0p;
}

void ?{}( $thread & this, current_stack_info_t * info) with( this ) {
        ticket = 1;
        state = Start;
        self_cor{ info };
        curr_cor = &self_cor;
        curr_cluster = mainCluster;
        self_mon.owner = &this;
        self_mon.recursion = 1;
        self_mon_p = &self_mon;
        link.next = 0p;
        link.prev = 0p;

        node.next = 0p;
        node.prev = 0p;
        doregister(curr_cluster, this);

        monitors{ &self_mon_p, 1, (fptr_t)0 };
}

//-----------------------------------------------------------------------------
// Processor coroutine
void ?{}(processorCtx_t & this) {

}

// Construct the processor context of non-main processors
static void ?{}(processorCtx_t & this, processor * proc, current_stack_info_t * info) {
        (this.__cor){ info };
        this.proc = proc;
}

static void * __invoke_processor(void * arg);

void ?{}(processor & this, const char name[], cluster & _cltr) with( this ) {
        this.name = name;
        this.cltr = &_cltr;
        id = -1u;
        terminated{ 0 };
        destroyer = 0p;
        do_terminate = false;
        preemption_alarm = 0p;
        pending_preemption = false;
        runner.proc = &this;

        #if !defined(__CFA_NO_STATISTICS__)
                print_stats = 0;
                print_halts = false;
        #endif

        idle{};

        __cfadbg_print_safe(runtime_core, "Kernel : Starting core %p\n", &this);

        this.stack = __create_pthread( &this.kernel_thread, __invoke_processor, (void *)&this );
        __atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );

        __cfadbg_print_safe(runtime_core, "Kernel : core %p created\n", &this);
}

void ^?{}(processor & this) with( this ){
        if( ! __atomic_load_n(&do_terminate, __ATOMIC_ACQUIRE) ) {
                __cfadbg_print_safe(runtime_core, "Kernel : core %p signaling termination\n", &this);

                __atomic_store_n(&do_terminate, true, __ATOMIC_RELAXED);
                __wake_proc( &this );

                P( terminated );
                verify( kernelTLS.this_processor != &this);
        }

        int err = pthread_join( kernel_thread, 0p );
        if( err != 0 ) abort("KERNEL ERROR: joining processor %p caused error %s\n", &this, strerror(err));

        free( this.stack );

        __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
}

void ?{}(cluster & this, const char name[], Duration preemption_rate, unsigned io_flags) with( this ) {
        this.name = name;
        this.preemption_rate = preemption_rate;
        this.nprocessors = 0;
        ready_queue{};

        #if !defined(__CFA_NO_STATISTICS__)
                print_stats = 0;
                stats = alloc();
                __init_stats( stats );
        #endif

        threads{ __get };

        __kernel_io_startup( this, io_flags, &this == mainCluster );

        doregister(this);
}

void ^?{}(cluster & this) {
        __kernel_io_shutdown( this, &this == mainCluster );

        #if !defined(__CFA_NO_STATISTICS__)
                if( 0 != this.print_stats ) {
                        __print_stats( this.stats, this.print_stats, true, this.name, (void*)&this );
                }
                free( this.stats );
        #endif

        unregister(this);
}

//=============================================================================================
// Kernel Scheduling logic
//=============================================================================================
//Main of the processor contexts
void main(processorCtx_t & runner) {
        // Because of a bug, we couldn't initialized the seed on construction
        // Do it here
        kernelTLS.rand_seed ^= rdtscl();

        processor * this = runner.proc;
        verify(this);

        __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);

        // register the processor unless it's the main thread which is handled in the boot sequence
        if(this != mainProcessor) {
                this->id = doregister((__processor_id_t*)this);
                #if !defined(__CFA_NO_STATISTICS__)
                        if( this->print_halts ) {
                                __cfaabi_bits_print_safe( STDOUT_FILENO, "Processor : %d - %s (%p)\n", this->id, this->name, (void*)this);
                        }
                #endif

                // Lock the RWlock so no-one pushes/pops while we are changing the queue
                uint_fast32_t last_size = ready_mutate_lock();

                        // Adjust the ready queue size
                        ready_queue_grow( this->cltr );

                // Unlock the RWlock
                ready_mutate_unlock( last_size );
        }

        {
                // Setup preemption data
                preemption_scope scope = { this };

                __cfadbg_print_safe(runtime_core, "Kernel : core %p started\n", this);

                $thread * readyThread = 0p;
                for( unsigned int spin_count = 0;; spin_count++ ) {
                        // Try to get the next thread
                        readyThread = __next_thread( this->cltr );

                        // Check if we actually found a thread
                        if( readyThread ) {
                                /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
                                /* paranoid */ verifyf( readyThread->state == Ready || readyThread->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", readyThread->state, readyThread->preempted);
                                /* paranoid */ verifyf( readyThread->link.next == 0p, "Expected null got %p", readyThread->link.next );
                                __builtin_prefetch( readyThread->context.SP );

                                // We found a thread run it
                                __run_thread(this, readyThread);

                                /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
                        }

                        if(__atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST)) break;

                        if( !readyThread ) {
                                // Block until a thread is ready
                                __halt(this);
                        }
                }

                __cfadbg_print_safe(runtime_core, "Kernel : core %p stopping\n", this);
        }

        V( this->terminated );

        // unregister the processor unless it's the main thread which is handled in the boot sequence
        if(this != mainProcessor) {
                // Lock the RWlock so no-one pushes/pops while we are changing the queue
                uint_fast32_t last_size = ready_mutate_lock();

                        // Adjust the ready queue size
                        ready_queue_shrink( this->cltr );

                        // Make sure we aren't on the idle queue
                        #if !defined(__CFA_NO_STATISTICS__)
                                bool removed =
                        #endif
                        unsafe_remove( this->cltr->idles, this );

                        #if !defined(__CFA_NO_STATISTICS__)
                                if(removed) __tls_stats()->ready.sleep.exits++;
                        #endif

                // Unlock the RWlock
                ready_mutate_unlock( last_size );

                // Finally we don't need the read_lock any more
                unregister((__processor_id_t*)this);
        }
        else {
                // HACK : the coroutine context switch expects this_thread to be set
                // and it make sense for it to be set in all other cases except here
                // fake it
                kernelTLS.this_thread = mainThread;
        }

        __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
}

static int * __volatile_errno() __attribute__((noinline));
static int * __volatile_errno() { asm(""); return &errno; }

// KERNEL ONLY
// runThread runs a thread by context switching
// from the processor coroutine to the target thread
static void __run_thread(processor * this, $thread * thrd_dst) {
        $coroutine * proc_cor = get_coroutine(this->runner);

        // Update global state
        kernelTLS.this_thread = thrd_dst;

        // set state of processor coroutine to inactive
        verify(proc_cor->state == Active);
        proc_cor->state = Blocked;

        // Actually run the thread
        RUNNING:  while(true) {
                thrd_dst->preempted = __NO_PREEMPTION;
                thrd_dst->state = Active;

                __cfaabi_dbg_debug_do(
                        thrd_dst->park_stale   = true;
                        thrd_dst->unpark_stale = true;
                )

                /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
                /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
                /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst ); // add escape condition if we are setting up the processor
                /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->curr_cor == proc_cor, "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor

                // set context switch to the thread that the processor is executing
                verify( thrd_dst->context.SP );
                __cfactx_switch( &proc_cor->context, &thrd_dst->context );
                // when __cfactx_switch returns we are back in the processor coroutine

                /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too large.\n", thrd_dst );
                /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ), "ERROR : Destination $thread %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
                /* paranoid */ verify( kernelTLS.this_thread == thrd_dst );
                /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );


                // We just finished running a thread, there are a few things that could have happened.
                // 1 - Regular case : the thread has blocked and now one has scheduled it yet.
                // 2 - Racy case    : the thread has blocked but someone has already tried to schedule it.
                // 4 - Preempted
                // In case 1, we may have won a race so we can't write to the state again.
                // In case 2, we lost the race so we now own the thread.

                if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
                        // The thread was preempted, reschedule it and reset the flag
                        __schedule_thread( (__processor_id_t*)this, thrd_dst );
                        break RUNNING;
                }

                if(unlikely(thrd_dst->state == Halted)) {
                        // The thread has halted, it should never be scheduled/run again
                        // We may need to wake someone up here since
                        unpark( this->destroyer __cfaabi_dbg_ctx2 );
                        this->destroyer = 0p;
                        break RUNNING;
                }

                /* paranoid */ verify( thrd_dst->state == Active );
                thrd_dst->state = Blocked;

                // set state of processor coroutine to active and the thread to inactive
                int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
                __cfaabi_dbg_debug_do( thrd_dst->park_result = old_ticket; )
                switch(old_ticket) {
                        case 1:
                                // This is case 1, the regular case, nothing more is needed
                                break RUNNING;
                        case 2:
                                // This is case 2, the racy case, someone tried to run this thread before it finished blocking
                                // In this case, just run it again.
                                continue RUNNING;
                        default:
                                // This makes no sense, something is wrong abort
                                abort();
                }
        }

        // Just before returning to the processor, set the processor coroutine to active
        proc_cor->state = Active;
        kernelTLS.this_thread = 0p;
}

// KERNEL_ONLY
void returnToKernel() {
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        $coroutine * proc_cor = get_coroutine(kernelTLS.this_processor->runner);
        $thread * thrd_src = kernelTLS.this_thread;

        #if !defined(__CFA_NO_STATISTICS__)
                struct processor * last_proc = kernelTLS.this_processor;
        #endif

        // Run the thread on this processor
        {
                int local_errno = *__volatile_errno();
                #if defined( __i386 ) || defined( __x86_64 )
                        __x87_store;
                #endif
                verify( proc_cor->context.SP );
                __cfactx_switch( &thrd_src->context, &proc_cor->context );
                #if defined( __i386 ) || defined( __x86_64 )
                        __x87_load;
                #endif
                *__volatile_errno() = local_errno;
        }

        #if !defined(__CFA_NO_STATISTICS__)
                if(last_proc != kernelTLS.this_processor) {
                        __tls_stats()->ready.threads.migration++;
                }
        #endif

        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too small.\n", thrd_src );
        /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) > ((uintptr_t)__get_stack(thrd_src->curr_cor)->limit), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too large.\n", thrd_src );
}

// KERNEL_ONLY
// Context invoker for processors
// This is the entry point for processors (kernel threads)
// It effectively constructs a coroutine by stealing the pthread stack
static void * __invoke_processor(void * arg) {
        #if !defined( __CFA_NO_STATISTICS__ )
                __stats_t local_stats;
                __init_stats( &local_stats );
                kernelTLS.this_stats = &local_stats;
        #endif

        processor * proc = (processor *) arg;
        kernelTLS.this_processor = proc;
        kernelTLS.this_thread    = 0p;
        kernelTLS.preemption_state.[enabled, disable_count] = [false, 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;
        __stack_t ctx;
        info.storage = &ctx;
        (proc->runner){ proc, &info };

        __cfaabi_dbg_print_safe("Coroutine : created stack %p\n", get_coroutine(proc->runner)->stack.storage);

        //Set global state
        kernelTLS.this_thread = 0p;

        //We now have a proper context from which to schedule threads
        __cfadbg_print_safe(runtime_core, "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.
        get_coroutine(proc->runner)->state = Active;
        main( proc->runner );
        get_coroutine(proc->runner)->state = Halted;

        // Main routine of the core returned, the core is now fully terminated
        __cfadbg_print_safe(runtime_core, "Kernel : core %p main ended (%p)\n", proc, &proc->runner);

        #if !defined(__CFA_NO_STATISTICS__)
                __tally_stats(proc->cltr->stats, &local_stats);
                if( 0 != proc->print_stats ) {
                        __print_stats( &local_stats, proc->print_stats, true, proc->name, (void*)proc );
                }
        #endif

        return 0p;
}

static void Abort( int ret, const char func[] ) {
        if ( ret ) {                                                                            // pthread routines return errno values
                abort( "%s : internal error, error(%d) %s.", func, ret, strerror( ret ) );
        } // if
} // Abort

void * __create_pthread( pthread_t * pthread, void * (*start)(void *), void * arg ) {
        pthread_attr_t attr;

        Abort( pthread_attr_init( &attr ), "pthread_attr_init" ); // initialize attribute

        size_t stacksize;
        // default stack size, normally defined by shell limit
        Abort( pthread_attr_getstacksize( &attr, &stacksize ), "pthread_attr_getstacksize" );
        assert( stacksize >= PTHREAD_STACK_MIN );

        void * stack;
        __cfaabi_dbg_debug_do(
                stack = memalign( __page_size, stacksize + __page_size );
                // pthread has no mechanism to create the guard page in user supplied stack.
                if ( mprotect( stack, __page_size, PROT_NONE ) == -1 ) {
                        abort( "mprotect : internal error, mprotect failure, error(%d) %s.", errno, strerror( errno ) );
                } // if
        );
        __cfaabi_dbg_no_debug_do(
                stack = malloc( stacksize );
        );

        Abort( pthread_attr_setstack( &attr, stack, stacksize ), "pthread_attr_setstack" );

        Abort( pthread_create( pthread, &attr, start, arg ), "pthread_create" );
        return stack;
}

// KERNEL_ONLY
static void __kernel_first_resume( processor * this ) {
        $thread * src = mainThread;
        $coroutine * dst = get_coroutine(this->runner);

        verify( ! kernelTLS.preemption_state.enabled );

        kernelTLS.this_thread->curr_cor = dst;
        __stack_prepare( &dst->stack, 65000 );
        __cfactx_start(main, dst, this->runner, __cfactx_invoke_coroutine);

        verify( ! kernelTLS.preemption_state.enabled );

        dst->last = &src->self_cor;
        dst->starter = dst->starter ? dst->starter : &src->self_cor;

        // make sure the current state is still correct
        /* paranoid */ verify(src->state == Ready);

        // context switch to specified coroutine
        verify( dst->context.SP );
        __cfactx_switch( &src->context, &dst->context );
        // when __cfactx_switch returns we are back in the src coroutine

        mainThread->curr_cor = &mainThread->self_cor;

        // make sure the current state has been update
        /* paranoid */ verify(src->state == Active);

        verify( ! kernelTLS.preemption_state.enabled );
}

// KERNEL_ONLY
static void __kernel_last_resume( processor * this ) {
        $coroutine * src = &mainThread->self_cor;
        $coroutine * dst = get_coroutine(this->runner);

        verify( ! kernelTLS.preemption_state.enabled );
        verify( dst->starter == src );
        verify( dst->context.SP );

        // SKULLDUGGERY in debug the processors check that the
        // stack is still within the limit of the stack limits after running a thread.
        // that check doesn't make sense if we context switch to the processor using the
        // coroutine semantics. Since this is a special case, use the current context
        // info to populate these fields.
        __cfaabi_dbg_debug_do(
                __stack_context_t ctx;
                CtxGet( ctx );
                mainThread->context.SP = ctx.SP;
                mainThread->context.FP = ctx.FP;
        )

        // context switch to the processor
        __cfactx_switch( &src->context, &dst->context );
}

//-----------------------------------------------------------------------------
// Scheduler routines
// KERNEL ONLY
void __schedule_thread( struct __processor_id_t * id, $thread * thrd ) {
        /* paranoid */ verify( thrd );
        /* paranoid */ verify( thrd->state != Halted );
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
        /* paranoid */  if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
                                        "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
        /* paranoid */  if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
                                        "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
        /* paranoid */ #endif
        /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );

        if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;

        ready_schedule_lock  ( id );
                push( thrd->curr_cluster, thrd );

                #if !defined(__CFA_NO_STATISTICS__)
                        bool woke =
                #endif
                        __wake_one(id, thrd->curr_cluster);

                #if !defined(__CFA_NO_STATISTICS__)
                        if(woke) __tls_stats()->ready.sleep.wakes++;
                #endif
        ready_schedule_unlock( id );

        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
}

// KERNEL ONLY
static $thread * __next_thread(cluster * this) with( *this ) {
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );

        ready_schedule_lock  ( (__processor_id_t*)kernelTLS.this_processor );
                $thread * head = pop( this );
        ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );

        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        return head;
}

// KERNEL ONLY
static bool __has_next_thread(cluster * this) with( *this ) {
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );

        ready_schedule_lock  ( (__processor_id_t*)kernelTLS.this_processor );
                bool not_empty = query( this );
        ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );

        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        return not_empty;
}

// KERNEL ONLY unpark with out disabling interrupts
void __unpark(  struct __processor_id_t * id, $thread * thrd __cfaabi_dbg_ctx_param2 ) {
        // record activity
        __cfaabi_dbg_debug_do( char * old_caller = thrd->unpark_caller; )
        __cfaabi_dbg_record_thrd( *thrd, false, caller );

        int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
        __cfaabi_dbg_debug_do( thrd->unpark_result = old_ticket; thrd->unpark_state = thrd->state; )
        switch(old_ticket) {
                case 1:
                        // Wake won the race, the thread will reschedule/rerun itself
                        break;
                case 0:
                        /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
                        /* paranoid */ verify( thrd->state == Blocked );

                        // Wake lost the race,
                        __schedule_thread( id, thrd );
                        break;
                default:
                        // This makes no sense, something is wrong abort
                        abort();
        }
}

void unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
        if( !thrd ) return;

        disable_interrupts();
        __unpark( (__processor_id_t*)kernelTLS.this_processor, thrd __cfaabi_dbg_ctx_fwd2 );
        enable_interrupts( __cfaabi_dbg_ctx );
}

void park( __cfaabi_dbg_ctx_param ) {
        /* paranoid */ verify( kernelTLS.preemption_state.enabled );
        disable_interrupts();
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        /* paranoid */ verify( kernelTLS.this_thread->preempted == __NO_PREEMPTION );

        // record activity
        __cfaabi_dbg_record_thrd( *kernelTLS.this_thread, true, caller );

        returnToKernel();

        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        enable_interrupts( __cfaabi_dbg_ctx );
        /* paranoid */ verify( kernelTLS.preemption_state.enabled );

}

// KERNEL ONLY
void __leave_thread() {
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        returnToKernel();
        abort();
}

// KERNEL ONLY
bool force_yield( __Preemption_Reason reason ) {
        /* paranoid */ verify( kernelTLS.preemption_state.enabled );
        disable_interrupts();
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );

        $thread * thrd = kernelTLS.this_thread;
        /* paranoid */ verify(thrd->state == Active);

        // SKULLDUGGERY: It is possible that we are preempting this thread just before
        // it was going to park itself. If that is the case and it is already using the
        // intrusive fields then we can't use them to preempt the thread
        // If that is the case, abandon the preemption.
        bool preempted = false;
        if(thrd->link.next == 0p) {
                preempted = true;
                thrd->preempted = reason;
                returnToKernel();
        }

        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
        enable_interrupts_noPoll();
        /* paranoid */ verify( kernelTLS.preemption_state.enabled );

        return preempted;
}

//=============================================================================================
// Kernel Setup logic
//=============================================================================================
//-----------------------------------------------------------------------------
// Kernel boot procedures
static void __kernel_startup(void) {
        verify( ! kernelTLS.preemption_state.enabled );
        __cfadbg_print_safe(runtime_core, "Kernel : Starting\n");

        __page_size = sysconf( _SC_PAGESIZE );

        __cfa_dbg_global_clusters.list{ __get };
        __cfa_dbg_global_clusters.lock{};

        // Initialize the global scheduler lock
        __scheduler_lock = (__scheduler_RWLock_t*)&storage___scheduler_lock;
        (*__scheduler_lock){};

        // Initialize the main cluster
        mainCluster = (cluster *)&storage_mainCluster;
        (*mainCluster){"Main Cluster"};

        __cfadbg_print_safe(runtime_core, "Kernel : Main cluster ready\n");

        // Start by initializing the main thread
        // SKULLDUGGERY: the mainThread steals the process main thread
        // which will then be scheduled by the mainProcessor normally
        mainThread = ($thread *)&storage_mainThread;
        current_stack_info_t info;
        info.storage = (__stack_t*)&storage_mainThreadCtx;
        (*mainThread){ &info };

        __cfadbg_print_safe(runtime_core, "Kernel : Main thread ready\n");



        // Construct the processor context of the main processor
        void ?{}(processorCtx_t & this, processor * proc) {
                (this.__cor){ "Processor" };
                this.__cor.starter = 0p;
                this.proc = proc;
        }

        void ?{}(processor & this) with( this ) {
                name = "Main Processor";
                cltr = mainCluster;
                terminated{ 0 };
                do_terminate = false;
                preemption_alarm = 0p;
                pending_preemption = false;
                kernel_thread = pthread_self();
                id = -1u;

                #if !defined(__CFA_NO_STATISTICS__)
                        print_stats = false;
                        print_halts = false;
                #endif

                runner{ &this };
                __cfadbg_print_safe(runtime_core, "Kernel : constructed main processor context %p\n", &runner);

                __atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
        }

        // Initialize the main processor and the main processor ctx
        // (the coroutine that contains the processing control flow)
        mainProcessor = (processor *)&storage_mainProcessor;
        (*mainProcessor){};

        mainProcessor->id = doregister( (__processor_id_t*)mainProcessor);

        //initialize the global state variables
        kernelTLS.this_processor = mainProcessor;
        kernelTLS.this_thread    = mainThread;

        #if !defined( __CFA_NO_STATISTICS__ )
                kernelTLS.this_stats = (__stats_t *)& storage_mainProcStats;
                __init_stats( kernelTLS.this_stats );
        #endif

        // Enable preemption
        kernel_start_preemption();

        // Add the main thread to the ready queue
        // once resume is called on mainProcessor->runner the mainThread needs to be scheduled like any normal thread
        __schedule_thread((__processor_id_t *)mainProcessor, mainThread);

        // SKULLDUGGERY: Force a context switch to the main processor to set the main thread's context to the current UNIX
        // context. Hence, the main thread does not begin through __cfactx_invoke_thread, like all other threads. The trick here is that
        // mainThread is on the ready queue when this call is made.
        __kernel_first_resume( kernelTLS.this_processor );


        // THE SYSTEM IS NOW COMPLETELY RUNNING


        // Now that the system is up, finish creating systems that need threading
        __kernel_io_finish_start( *mainCluster );


        __cfadbg_print_safe(runtime_core, "Kernel : Started\n--------------------------------------------------\n\n");

        verify( ! kernelTLS.preemption_state.enabled );
        enable_interrupts( __cfaabi_dbg_ctx );
        verify( TL_GET( preemption_state.enabled ) );
}

static void __kernel_shutdown(void) {
        //Before we start shutting things down, wait for systems that need threading to shutdown
        __kernel_io_prepare_stop( *mainCluster );

        /* paranoid */ verify( TL_GET( preemption_state.enabled ) );
        disable_interrupts();
        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );

        __cfadbg_print_safe(runtime_core, "\n--------------------------------------------------\nKernel : Shutting down\n");

        // SKULLDUGGERY: Notify the mainProcessor it needs to terminates.
        // When its coroutine terminates, it return control to the mainThread
        // which is currently here
        __atomic_store_n(&mainProcessor->do_terminate, true, __ATOMIC_RELEASE);
        __kernel_last_resume( kernelTLS.this_processor );
        mainThread->self_cor.state = Halted;

        // THE SYSTEM IS NOW COMPLETELY STOPPED

        // Disable preemption
        kernel_stop_preemption();

        unregister((__processor_id_t*)mainProcessor);

        // Destroy the main processor and its context in reverse order of construction
        // These were manually constructed so we need manually destroy them
        void ^?{}(processor & this) with( this ){
                /* paranoid */ verify( this.do_terminate == true );
                __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
                __cfaabi_dbg_print_safe("Kernel : destroyed main processor context %p\n", &runner);
        }

        ^(*mainProcessor){};

        // 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
        /* paranoid */ verify(mainThread->self_cor.stack.storage == (__stack_t*)(((uintptr_t)&storage_mainThreadCtx)| 0x1));
        ^(*mainThread){};

        ^(*mainCluster){};

        ^(*__scheduler_lock){};

        ^(__cfa_dbg_global_clusters.list){};
        ^(__cfa_dbg_global_clusters.lock){};

        __cfadbg_print_safe(runtime_core, "Kernel : Shutdown complete\n");
}

//=============================================================================================
// Kernel Idle Sleep
//=============================================================================================
// Wake a thread from the front if there are any
static bool __wake_one(struct __processor_id_t * id, cluster * this) {
        /* paranoid */ verify( ready_schedule_islocked( id ) );

        // Check if there is a sleeping processor
        processor * p = pop(this->idles);

        // If no one is sleeping, we are done
        if( 0p == p ) return false;

        // We found a processor, wake it up
        post( p->idle );

        return true;
}

// Unconditionnaly wake a thread
static bool __wake_proc(processor * this) {
        __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);

        disable_interrupts();
                /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
                bool ret = post( this->idle );
        enable_interrupts( __cfaabi_dbg_ctx );

        return ret;
}

static void __halt(processor * this) with( *this ) {
        if( do_terminate ) return;

        #if !defined(__CFA_NO_STATISTICS__)
                __tls_stats()->ready.sleep.halts++;
        #endif
        // Push self to queue
        push(cltr->idles, *this);

        // Makre sure we don't miss a thread
        if( __has_next_thread(cltr) ) {
                // A thread was posted, make sure a processor is woken up
                struct __processor_id_t *id = (struct __processor_id_t *) this;
                ready_schedule_lock  ( id );
                        __wake_one( id, cltr );
                ready_schedule_unlock( id );
                #if !defined(__CFA_NO_STATISTICS__)
                        __tls_stats()->ready.sleep.cancels++;
                #endif
        }

        #if !defined(__CFA_NO_STATISTICS__)
                if(this->print_halts) {
                        __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->id, rdtscl());
                }
        #endif

        wait( idle );

        #if !defined(__CFA_NO_STATISTICS__)
                if(this->print_halts) {
                        __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->id, rdtscl());
                }
        #endif
}

//=============================================================================================
// Unexpected Terminating logic
//=============================================================================================
static __spinlock_t kernel_abort_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 __cfaabi_dbg_ctx2 );

        // first task to abort ?
        if ( kernel_abort_called ) {                    // not first task to abort ?
                unlock( kernel_abort_lock );

                sigset_t mask;
                sigemptyset( &mask );
                sigaddset( &mask, SIGALRM );            // block SIGALRM signals
                sigaddset( &mask, SIGUSR1 );            // block SIGALRM signals
                sigsuspend( &mask );                            // block the processor to prevent further damage during abort
                _exit( EXIT_FAILURE );                          // if processor unblocks before it is killed, terminate it
        }
        else {
                kernel_abort_called = true;
                unlock( kernel_abort_lock );
        }

        return kernelTLS.this_thread;
}

void kernel_abort_msg( void * kernel_data, char * abort_text, int abort_text_size ) {
        $thread * thrd = kernel_data;

        if(thrd) {
                int len = snprintf( abort_text, abort_text_size, "Error occurred while executing thread %.256s (%p)", thrd->self_cor.name, thrd );
                __cfaabi_bits_write( STDERR_FILENO, abort_text, len );

                if ( &thrd->self_cor != thrd->curr_cor ) {
                        len = snprintf( abort_text, abort_text_size, " in coroutine %.256s (%p).\n", thrd->curr_cor->name, thrd->curr_cor );
                        __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
                }
                else {
                        __cfaabi_bits_write( STDERR_FILENO, ".\n", 2 );
                }
        }
        else {
                int len = snprintf( abort_text, abort_text_size, "Error occurred outside of any thread.\n" );
                __cfaabi_bits_write( STDERR_FILENO, abort_text, len );
        }
}

int kernel_abort_lastframe( void ) __attribute__ ((__nothrow__)) {
        return get_coroutine(kernelTLS.this_thread) == get_coroutine(mainThread) ? 4 : 2;
}

static __spinlock_t kernel_debug_lock;

extern "C" {
        void __cfaabi_bits_acquire() {
                lock( kernel_debug_lock __cfaabi_dbg_ctx2 );
        }

        void __cfaabi_bits_release() {
                unlock( kernel_debug_lock );
        }
}

//=============================================================================================
// Kernel Utilities
//=============================================================================================
//-----------------------------------------------------------------------------
// Locks
void  ?{}( semaphore & this, int count = 1 ) {
        (this.lock){};
        this.count = count;
        (this.waiting){};
}
void ^?{}(semaphore & this) {}

bool P(semaphore & this) with( this ){
        lock( lock __cfaabi_dbg_ctx2 );
        count -= 1;
        if ( count < 0 ) {
                // queue current task
                append( waiting, kernelTLS.this_thread );

                // atomically release spin lock and block
                unlock( lock );
                park( __cfaabi_dbg_ctx );
                return true;
        }
        else {
            unlock( lock );
            return false;
        }
}

bool V(semaphore & this) with( this ) {
        $thread * thrd = 0p;
        lock( lock __cfaabi_dbg_ctx2 );
        count += 1;
        if ( count <= 0 ) {
                // remove task at head of waiting list
                thrd = pop_head( waiting );
        }

        unlock( lock );

        // make new owner
        unpark( thrd __cfaabi_dbg_ctx2 );

        return thrd != 0p;
}

bool V(semaphore & this, unsigned diff) with( this ) {
        $thread * thrd = 0p;
        lock( lock __cfaabi_dbg_ctx2 );
        int release = max(-count, (int)diff);
        count += diff;
        for(release) {
                unpark( pop_head( waiting ) __cfaabi_dbg_ctx2 );
        }

        unlock( lock );

        return thrd != 0p;
}

//-----------------------------------------------------------------------------
// Global Queues
void doregister( cluster     & cltr ) {
        lock      ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
        push_front( __cfa_dbg_global_clusters.list, cltr );
        unlock    ( __cfa_dbg_global_clusters.lock );
}

void unregister( cluster     & cltr ) {
        lock  ( __cfa_dbg_global_clusters.lock __cfaabi_dbg_ctx2);
        remove( __cfa_dbg_global_clusters.list, cltr );
        unlock( __cfa_dbg_global_clusters.lock );
}

void doregister( cluster * cltr, $thread & thrd ) {
        lock      (cltr->thread_list_lock __cfaabi_dbg_ctx2);
        cltr->nthreads += 1;
        push_front(cltr->threads, thrd);
        unlock    (cltr->thread_list_lock);
}

void unregister( cluster * cltr, $thread & thrd ) {
        lock  (cltr->thread_list_lock __cfaabi_dbg_ctx2);
        remove(cltr->threads, thrd );
        cltr->nthreads -= 1;
        unlock(cltr->thread_list_lock);
}

//-----------------------------------------------------------------------------
// Debug
__cfaabi_dbg_debug_do(
        extern "C" {
                void __cfaabi_dbg_record_lock(__spinlock_t & this, const char prev_name[]) {
                        this.prev_name = prev_name;
                        this.prev_thrd = kernelTLS.this_thread;
                }

                void __cfaabi_dbg_record_thrd($thread & this, bool park, const char prev_name[]) {
                        if(park) {
                                this.park_caller   = prev_name;
                                this.park_stale    = false;
                        }
                        else {
                                this.unpark_caller = prev_name;
                                this.unpark_stale  = false;
                        }
                }
        }
)

//-----------------------------------------------------------------------------
// Debug
bool threading_enabled(void) __attribute__((const)) {
        return true;
}

//-----------------------------------------------------------------------------
// Statistics
#if !defined(__CFA_NO_STATISTICS__)
        void print_halts( processor & this ) {
                this.print_halts = true;
        }
#endif
// Local Variables: //
// mode: c //
// tab-width: 4 //
// End: //