source: libcfa/src/concurrency/kernel.cfa@ d5d3a90

//
// 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 <errno.h>
#include <stdio.h>
#include <signal.h>
#include <unistd.h>

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

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


//-----------------------------------------------------------------------------
// Some assembly required
#if defined( __i386 )
        // 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 __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 )
#else
        #error unknown hardware architecture
#endif

extern $thread * mainThread;
extern processor * mainProcessor;

//-----------------------------------------------------------------------------
// 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_one(struct __processor_id_t * id, cluster * cltr);
static void __halt(processor * this);
bool __wake_proc(processor *);

//=============================================================================================
// 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);
        #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

        {
                // 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 );

        if(this == mainProcessor) {
                // 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 );
}

//-----------------------------------------------------------------------------
// 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_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 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
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 = ( $thread * ) 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;
}

//-----------------------------------------------------------------------------
// 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: //