source: libcfa/src/concurrency/kernel.cfa@ bbfe226

//
// 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 : Mon Aug 31 07:08:20 2020
// Update Count     : 71
//

#define __cforall_thread__
#define _GNU_SOURCE

// #define __CFA_DEBUG_PRINT_RUNTIME_CORE__

//C Includes
#include <errno.h>
#include <stdio.h>
#include <string.h>
#include <signal.h>
#include <unistd.h>
extern "C" {
        #include <sys/eventfd.h>
        #include <sys/uio.h>
}

//CFA Includes
#include "kernel_private.hfa"
#include "preemption.hfa"
#include "strstream.hfa"
#include "device/cpu.hfa"
#include "io/types.hfa"

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

#if !defined(__CFA_NO_STATISTICS__)
        #define __STATS( ...) __VA_ARGS__
#else
        #define __STATS( ...)
#endif

//-----------------------------------------------------------------------------
// 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__ )
        #define __x87_store
        #define __x87_load

#elif defined( __aarch64__ )
        #define __x87_store              \
                uint32_t __fpcntl[2];    \
                __asm__ volatile (    \
                        "mrs x9, FPCR\n" \
                        "mrs x10, FPSR\n"  \
                        "stp x9, x10, %0\n"  \
                        : "=m" (__fpcntl) : : "x9", "x10" \
                )

        #define __x87_load         \
                __asm__ volatile (    \
                        "ldp x9, x10, %0\n"  \
                        "msr FPSR, x10\n"  \
                        "msr FPCR, x9\n" \
                : "=m" (__fpcntl) : : "x9", "x10" \
                )

#else
        #error unsupported hardware architecture
#endif

extern thread$ * mainThread;
extern processor * mainProcessor;

//-----------------------------------------------------------------------------
// Kernel Scheduling logic
static thread$ * __next_thread(cluster * this);
static thread$ * __next_thread_slow(cluster * this);
static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
static void __run_thread(processor * this, thread$ * dst);
static void __wake_one(cluster * cltr);

static void idle_sleep(processor * proc, io_future_t & future, iovec & iov);
static bool mark_idle (__cluster_proc_list & idles, processor & proc);
static void mark_awake(__cluster_proc_list & idles, processor & proc);

extern void __cfa_io_start( processor * );
extern bool __cfa_io_drain( processor * );
extern bool __cfa_io_flush( processor *, int min_comp );
extern void __cfa_io_stop ( processor * );
static inline bool __maybe_io_drain( processor * );

#if defined(CFA_WITH_IO_URING_IDLE)
        extern bool __kernel_read(processor * proc, io_future_t & future, iovec &, int fd);
#endif

extern void __disable_interrupts_hard();
extern void __enable_interrupts_hard();


//=============================================================================================
// 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
        __cfaabi_tls.rand_seed ^= rdtscl();
        __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&runner);
        __tls_rand_advance_bck();

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

        io_future_t future; // used for idle sleep when io_uring is present
        future.self.ptr = 1p;  // mark it as already fulfilled so we know if there is a pending request or not
        eventfd_t idle_val;
        iovec idle_iovec = { &idle_val, sizeof(idle_val) };

        __cfa_io_start( 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->unique_id, this->name, (void*)this);
                }
        #endif

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

                // if we need to run some special setup, now is the time to do it.
                if(this->init.thrd) {
                        this->init.thrd->curr_cluster = this->cltr;
                        __run_thread(this, this->init.thrd);
                }

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

                thread$ * readyThread = 0p;
                MAIN_LOOP:
                for() {
                        #define OLD_MAIN 1
                        #if OLD_MAIN
                        // Check if there is pending io
                        __maybe_io_drain( this );

                        // Try to get the next thread
                        readyThread = __next_thread( this->cltr );

                        if( !readyThread ) {
                                __tls_stats()->io.flush.idle++;
                                __cfa_io_flush( this, 0 );

                                readyThread = __next_thread_slow( this->cltr );
                        }

                        HALT:
                        if( !readyThread ) {
                                // Don't block if we are done
                                if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;

                                // Push self to idle stack
                                if(!mark_idle(this->cltr->procs, * this)) continue MAIN_LOOP;

                                // Confirm the ready-queue is empty
                                readyThread = __next_thread_slow( this->cltr );
                                if( readyThread ) {
                                        // A thread was found, cancel the halt
                                        mark_awake(this->cltr->procs, * this);

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

                                        // continue the mai loop
                                        break HALT;
                                }

                                idle_sleep( this, future, idle_iovec );

                                // We were woken up, remove self from idle
                                mark_awake(this->cltr->procs, * this);

                                // DON'T just proceed, start looking again
                                continue MAIN_LOOP;
                        }

                        /* paranoid */ verify( readyThread );

                        // Reset io dirty bit
                        this->io.dirty = false;

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

                        // Are we done?
                        if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;

                        if(this->io.pending && !this->io.dirty) {
                                __tls_stats()->io.flush.dirty++;
                                __cfa_io_flush( this, 0 );
                        }

                        #else
                                #warning new kernel loop
                        SEARCH: {
                                /* paranoid */ verify( ! __preemption_enabled() );

                                // First, lock the scheduler since we are searching for a thread
                                ready_schedule_lock();

                                // Try to get the next thread
                                readyThread = pop_fast( this->cltr );
                                if(readyThread) { ready_schedule_unlock(); break SEARCH; }

                                // If we can't find a thread, might as well flush any outstanding I/O
                                if(this->io.pending) { __cfa_io_flush( this, 0 ); }

                                // Spin a little on I/O, just in case
                                for(5) {
                                        __maybe_io_drain( this );
                                        readyThread = pop_fast( this->cltr );
                                        if(readyThread) { ready_schedule_unlock(); break SEARCH; }
                                }

                                // no luck, try stealing a few times
                                for(5) {
                                        if( __maybe_io_drain( this ) ) {
                                                readyThread = pop_fast( this->cltr );
                                        } else {
                                                readyThread = pop_slow( this->cltr );
                                        }
                                        if(readyThread) { ready_schedule_unlock(); break SEARCH; }
                                }

                                // still no luck, search for a thread
                                readyThread = pop_search( this->cltr );
                                if(readyThread) { ready_schedule_unlock(); break SEARCH; }

                                // Don't block if we are done
                                if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
                                        ready_schedule_unlock();
                                        break MAIN_LOOP;
                                }

                                __STATS( __tls_stats()->ready.sleep.halts++; )

                                // Push self to idle stack
                                ready_schedule_unlock();
                                if(!mark_idle(this->cltr->procs, * this)) goto SEARCH;
                                ready_schedule_lock();

                                // Confirm the ready-queue is empty
                                __maybe_io_drain( this );
                                readyThread = pop_search( this->cltr );
                                ready_schedule_unlock();

                                if( readyThread ) {
                                        // A thread was found, cancel the halt
                                        mark_awake(this->cltr->procs, * this);

                                        __STATS( __tls_stats()->ready.sleep.cancels++; )

                                        // continue the main loop
                                        break SEARCH;
                                }

                                __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
                                __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);

                                {
                                        eventfd_t val;
                                        ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
                                        if(ret < 0) {
                                                switch((int)errno) {
                                                case EAGAIN:
                                                #if EAGAIN != EWOULDBLOCK
                                                        case EWOULDBLOCK:
                                                #endif
                                                case EINTR:
                                                        // No need to do anything special here, just assume it's a legitimate wake-up
                                                        break;
                                                default:
                                                        abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
                                                }
                                        }
                                }

                                        __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )

                                // We were woken up, remove self from idle
                                mark_awake(this->cltr->procs, * this);

                                // DON'T just proceed, start looking again
                                continue MAIN_LOOP;
                        }

                RUN_THREAD:
                        /* paranoid */ verify( ! __preemption_enabled() );
                        /* paranoid */ verify( readyThread );

                        // Reset io dirty bit
                        this->io.dirty = false;

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

                        // Are we done?
                        if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;

                        if(this->io.pending && !this->io.dirty) {
                                __cfa_io_flush( this, 0 );
                        }

                        ready_schedule_lock();
                        __maybe_io_drain( this );
                        ready_schedule_unlock();
                        #endif
                }

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

        for(int i = 0; !available(future); i++) {
                if(i > 1000) __cfaabi_dbg_write( "ERROR: kernel has bin spinning on a flush after exit loop.\n", 60);
                __cfa_io_flush( this, 1 );
        }

        __cfa_io_stop( this );

        post( 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
                __cfaabi_tls.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) {
        /* paranoid */ verify( ! __preemption_enabled() );
        /* paranoid */ verifyf( thrd_dst->state == Ready || thrd_dst->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", thrd_dst->state, thrd_dst->preempted);
        /* paranoid */ verifyf( thrd_dst->link.next == 0p, "Expected null got %p", thrd_dst->link.next );
        __builtin_prefetch( thrd_dst->context.SP );

        __cfadbg_print_safe(runtime_core, "Kernel : core %p running thread %p (%s)\n", this, thrd_dst, thrd_dst->self_cor.name);

        coroutine$ * proc_cor = get_coroutine(this->runner);

        // 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;

                // Update global state
                kernelTLS().this_thread = thrd_dst;

                /* paranoid */ verify( ! __preemption_enabled() );
                /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
                /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
                /* paranoid */ verify( thrd_dst->context.SP );
                /* paranoid */ verify( thrd_dst->state != Halted );
                /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) < ((uintptr_t)__get_stack(thrd_dst->curr_cor)->base ) || thrd_dst->curr_cor == proc_cor || thrd_dst->corctx_flag, "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 || thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too large.\n", thrd_dst ); // add escape condition if we are setting up the processor
                /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );



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



                /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_dst->canary );
                /* paranoid */ verifyf( ((uintptr_t)thrd_dst->context.SP) > ((uintptr_t)__get_stack(thrd_dst->curr_cor)->limit) || thrd_dst->corctx_flag, "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 ) || thrd_dst->corctx_flag, "ERROR : Destination thread$ %p has been corrupted.\n StackPointer too small.\n", thrd_dst );
                /* paranoid */ verify( thrd_dst->context.SP );
                /* paranoid */ verify( thrd_dst->curr_cluster == this->cltr );
                /* paranoid */ verify( kernelTLS().this_thread == thrd_dst );
                /* paranoid */ verify( ! __preemption_enabled() );

                // Reset global state
                kernelTLS().this_thread = 0p;

                // 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$( thrd_dst, UNPARK_LOCAL );
                        break RUNNING;
                }

                if(unlikely(thrd_dst->state == Halting)) {
                        // The thread has halted, it should never be scheduled/run again
                        // finish the thread
                        __thread_finish( thrd_dst );
                        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);
                switch(old_ticket) {
                        case TICKET_RUNNING:
                                // This is case 1, the regular case, nothing more is needed
                                break RUNNING;
                        case TICKET_UNBLOCK:
                                #if !defined(__CFA_NO_STATISTICS__)
                                        __tls_stats()->ready.threads.threads++;
                                #endif
                                // 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;

        __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);

        #if !defined(__CFA_NO_STATISTICS__)
                __tls_stats()->ready.threads.threads--;
        #endif

        /* paranoid */ verify( ! __preemption_enabled() );
}

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

        __STATS( thrd_src->last_proc = kernelTLS().this_processor; )

        // Run the thread on this processor
        {
                int local_errno = *__volatile_errno();
                #if defined( __i386 ) || defined( __x86_64 )
                        __x87_store;
                #endif
                /* paranoid */ verify( proc_cor->context.SP );
                /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
                __cfactx_switch( &thrd_src->context, &proc_cor->context );
                /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd_src->canary );
                #if defined( __i386 ) || defined( __x86_64 )
                        __x87_load;
                #endif
                *__volatile_errno() = local_errno;
        }

        #if !defined(__CFA_NO_STATISTICS__)
                /* paranoid */ verify( thrd_src->last_proc != 0p );
                if(thrd_src->last_proc != kernelTLS().this_processor) {
                        __tls_stats()->ready.threads.migration++;
                }
        #endif

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

//-----------------------------------------------------------------------------
// Scheduler routines
// KERNEL ONLY
static void __schedule_thread( thread$ * thrd, unpark_hint hint ) {
        /* paranoid */ verify( ! __preemption_enabled() );
        /* paranoid */ verify( ready_schedule_islocked());
        /* paranoid */ verify( thrd );
        /* paranoid */ verify( thrd->state != Halted );
        /* paranoid */ verify( thrd->curr_cluster );
        /* 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 );
        /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );

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

        // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
        struct cluster * cl = thrd->curr_cluster;
        __STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )

        // push the thread to the cluster ready-queue
        push( cl, thrd, hint );

        // variable thrd is no longer safe to use
        thrd = 0xdeaddeaddeaddeadp;

        // wake the cluster using the save variable.
        __wake_one( cl );

        #if !defined(__CFA_NO_STATISTICS__)
                if( kernelTLS().this_stats ) {
                        __tls_stats()->ready.threads.threads++;
                        if(outside) {
                                __tls_stats()->ready.threads.extunpark++;
                        }
                }
                else {
                        __atomic_fetch_add(&cl->stats->ready.threads.threads, 1, __ATOMIC_RELAXED);
                        __atomic_fetch_add(&cl->stats->ready.threads.extunpark, 1, __ATOMIC_RELAXED);
                }
        #endif

        /* paranoid */ verify( ready_schedule_islocked());
        /* paranoid */ verify( ! __preemption_enabled() );
}

void schedule_thread$( thread$ * thrd, unpark_hint hint ) {
        ready_schedule_lock();
                __schedule_thread( thrd, hint );
        ready_schedule_unlock();
}

// KERNEL ONLY
static inline thread$ * __next_thread(cluster * this) with( *this ) {
        /* paranoid */ verify( ! __preemption_enabled() );

        ready_schedule_lock();
                thread$ * thrd = pop_fast( this );
        ready_schedule_unlock();

        /* paranoid */ verify( ! __preemption_enabled() );
        return thrd;
}

// KERNEL ONLY
static inline thread$ * __next_thread_slow(cluster * this) with( *this ) {
        /* paranoid */ verify( ! __preemption_enabled() );

        ready_schedule_lock();
                thread$ * thrd;
                for(25) {
                        thrd = pop_slow( this );
                        if(thrd) goto RET;
                }
                thrd = pop_search( this );

                RET:
        ready_schedule_unlock();

        /* paranoid */ verify( ! __preemption_enabled() );
        return thrd;
}

static inline bool __must_unpark( thread$ * thrd ) {
        int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
        switch(old_ticket) {
                case TICKET_RUNNING:
                        // Wake won the race, the thread will reschedule/rerun itself
                        return false;
                case TICKET_BLOCKED:
                        /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
                        /* paranoid */ verify( thrd->state == Blocked );
                        return true;
                default:
                        // This makes no sense, something is wrong abort
                        abort("Thread %p (%s) has mismatch park/unpark\n", thrd, thrd->self_cor.name);
        }
}

void __kernel_unpark( thread$ * thrd, unpark_hint hint ) {
        /* paranoid */ verify( ! __preemption_enabled() );
        /* paranoid */ verify( ready_schedule_islocked());

        if( !thrd ) return;

        if(__must_unpark(thrd)) {
                // Wake lost the race,
                __schedule_thread( thrd, hint );
        }

        /* paranoid */ verify( ready_schedule_islocked());
        /* paranoid */ verify( ! __preemption_enabled() );
}

void unpark( thread$ * thrd, unpark_hint hint ) {
        if( !thrd ) return;

        if(__must_unpark(thrd)) {
                disable_interrupts();
                        // Wake lost the race,
                        schedule_thread$( thrd, hint );
                enable_interrupts(false);
        }
}

void park( void ) {
        __disable_interrupts_checked();
                /* paranoid */ verify( kernelTLS().this_thread->preempted == __NO_PREEMPTION );
                returnToKernel();
        __enable_interrupts_checked();

}

extern "C" {
        // Leave the thread monitor
        // last routine called by a thread.
        // Should never return
        void __cfactx_thrd_leave() {
                thread$ * thrd = active_thread();
                monitor$ * this = &thrd->self_mon;

                // Lock the monitor now
                lock( this->lock __cfaabi_dbg_ctx2 );

                disable_interrupts();

                /* paranoid */ verify( ! __preemption_enabled() );
                /* paranoid */ verify( thrd->state == Active );
                /* paranoid */ verify( 0x0D15EA5E0D15EA5Ep == thrd->canary );
                /* paranoid */ verify( kernelTLS().this_thread == thrd );
                /* paranoid */ verify( thrd->context.SP );
                /* paranoid */ verifyf( ((uintptr_t)thrd->context.SP) > ((uintptr_t)__get_stack(thrd->curr_cor)->limit), "ERROR : thread$ %p has been corrupted.\n StackPointer too large.\n", thrd );
                /* paranoid */ verifyf( ((uintptr_t)thrd->context.SP) < ((uintptr_t)__get_stack(thrd->curr_cor)->base ), "ERROR : thread$ %p has been corrupted.\n StackPointer too small.\n", thrd );

                thrd->state = Halting;
                if( TICKET_RUNNING != thrd->ticket ) { abort( "Thread terminated with pending unpark" ); }
                if( thrd != this->owner ) { abort( "Thread internal monitor has incorrect owner" ); }
                if( this->recursion != 1) { abort( "Thread internal monitor has unbalanced recursion" ); }

                // Leave the thread
                returnToKernel();

                // Control flow should never reach here!
                abort();
        }
}

// KERNEL ONLY
bool force_yield( __Preemption_Reason reason ) {
        __disable_interrupts_checked();
                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();
                }
        __enable_interrupts_checked( false );
        return preempted;
}

//=============================================================================================
// Kernel Idle Sleep
//=============================================================================================
// Wake a thread from the front if there are any
static void __wake_one(cluster * this) {
        eventfd_t val;

        /* paranoid */ verify( ! __preemption_enabled() );
        /* paranoid */ verify( ready_schedule_islocked() );

        // Check if there is a sleeping processor
        struct __fd_waitctx * fdp = __atomic_load_n(&this->procs.fdw, __ATOMIC_SEQ_CST);

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

        int fd = 1;
        if( __atomic_load_n(&fdp->fd, __ATOMIC_SEQ_CST) != 1 ) {
                fd = __atomic_exchange_n(&fdp->fd, 1, __ATOMIC_RELAXED);
        }

        switch(fd) {
        case 0:
                // If the processor isn't ready to sleep then the exchange will already wake it up
                #if !defined(__CFA_NO_STATISTICS__)
                        if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.early++;
                        } else { __atomic_fetch_add(&this->stats->ready.sleep.early, 1, __ATOMIC_RELAXED); }
                #endif
                break;
        case 1:
                // If someone else already said they will wake them: we are done
                #if !defined(__CFA_NO_STATISTICS__)
                        if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.seen++;
                        } else { __atomic_fetch_add(&this->stats->ready.sleep.seen, 1, __ATOMIC_RELAXED); }
                #endif
                break;
        default:
                // If the processor was ready to sleep, we need to wake it up with an actual write
                val = 1;
                eventfd_write( fd, val );

                #if !defined(__CFA_NO_STATISTICS__)
                        if( kernelTLS().this_stats ) { __tls_stats()->ready.sleep.wakes++;
                        } else { __atomic_fetch_add(&this->stats->ready.sleep.wakes, 1, __ATOMIC_RELAXED); }
                #endif
                break;
        }

        /* paranoid */ verify( ready_schedule_islocked() );
        /* paranoid */ verify( ! __preemption_enabled() );

        return;
}

// Unconditionnaly wake a thread
void __wake_proc(processor * this) {
        /* paranoid */ verify( ! __preemption_enabled() );

        __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this);

        this->idle_wctx.fd = 1;

        eventfd_t val;
        val = 1;
        eventfd_write( this->idle_fd, val );

        /* paranoid */ verify( ! __preemption_enabled() );
}

static void idle_sleep(processor * this, io_future_t & future, iovec & iov) {
        // Tell everyone we are ready to go do sleep
        for() {
                int expected = this->idle_wctx.fd;

                // Someone already told us to wake-up! No time for a nap.
                if(expected == 1) { return; }

                // Try to mark that we are going to sleep
                if(__atomic_compare_exchange_n(&this->idle_wctx.fd, &expected, this->idle_fd, false,  __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) ) {
                        // Every one agreed, taking a nap
                        break;
                }
        }


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

                __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);

                {
                        eventfd_t val;
                        ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
                        if(ret < 0) {
                                switch((int)errno) {
                                case EAGAIN:
                                #if EAGAIN != EWOULDBLOCK
                                        case EWOULDBLOCK:
                                #endif
                                case EINTR:
                                        // No need to do anything special here, just assume it's a legitimate wake-up
                                        break;
                                default:
                                        abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
                                }
                        }
                }

                #if !defined(__CFA_NO_STATISTICS__)
                        if(this->print_halts) {
                                __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl());
                        }
                #endif
        #else
                // Do we already have a pending read
                if(available(future)) {
                        // There is no pending read, we need to add one
                        reset(future);

                        __kernel_read(this, future, iov, this->idle_fd );
                }

                __cfa_io_flush( this, 1 );
        #endif
}

static bool mark_idle(__cluster_proc_list & this, processor & proc) {
        #if !defined(__CFA_NO_STATISTICS__)
                __tls_stats()->ready.sleep.halts++;
        #endif

        proc.idle_wctx.fd = 0;

        /* paranoid */ verify( ! __preemption_enabled() );
        if(!try_lock( this )) return false;
                this.idle++;
                /* paranoid */ verify( this.idle <= this.total );
                remove(proc);
                insert_first(this.idles, proc);

                __atomic_store_n(&this.fdw, &proc.idle_wctx, __ATOMIC_SEQ_CST);
        unlock( this );
        /* paranoid */ verify( ! __preemption_enabled() );

        return true;
}

static void mark_awake(__cluster_proc_list & this, processor & proc) {
        /* paranoid */ verify( ! __preemption_enabled() );
        lock( this );
                this.idle--;
                /* paranoid */ verify( this.idle >= 0 );
                remove(proc);
                insert_last(this.actives, proc);

                {
                        struct __fd_waitctx * wctx = 0;
                        if(!this.idles`isEmpty) wctx = &this.idles`first.idle_wctx;
                        __atomic_store_n(&this.fdw, wctx, __ATOMIC_SEQ_CST);
                }

        unlock( this );
        /* paranoid */ verify( ! __preemption_enabled() );
}

//=============================================================================================
// Unexpected Terminating logic
//=============================================================================================
void __kernel_abort_msg( char * abort_text, int abort_text_size ) {
        thread$ * thrd = __cfaabi_tls.this_thread;

        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(__cfaabi_tls.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
//=============================================================================================
#if defined(CFA_HAVE_LINUX_IO_URING_H)
#include "io/types.hfa"
#endif

static inline bool __maybe_io_drain( processor * proc ) {
        bool ret = false;
        #if defined(CFA_HAVE_LINUX_IO_URING_H)
                __cfadbg_print_safe(runtime_core, "Kernel : core %p checking io for ring %d\n", proc, proc->io.ctx->fd);

                // Check if we should drain the queue
                $io_context * ctx = proc->io.ctx;
                unsigned head = *ctx->cq.head;
                unsigned tail = *ctx->cq.tail;
                if(head == tail) return false;
                #if OLD_MAIN
                        ready_schedule_lock();
                        ret = __cfa_io_drain( proc );
                        ready_schedule_unlock();
                #else
                        ret = __cfa_io_drain( proc );
                #endif
        #endif
        return ret;
}

//-----------------------------------------------------------------------------
// 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;
                }
        }
)

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

//-----------------------------------------------------------------------------
// Statistics
#if !defined(__CFA_NO_STATISTICS__)
        void print_halts( processor & this ) {
                this.print_halts = true;
        }

        static void crawl_list( cluster * cltr, dlist(processor) & list, unsigned count ) {
                /* paranoid */ verify( cltr->stats );

                processor * it = &list`first;
                for(unsigned i = 0; i < count; i++) {
                        /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
                        /* paranoid */ verify( it->local_data->this_stats );
                        // __print_stats( it->local_data->this_stats, cltr->print_stats, "Processor", it->name, (void*)it );
                        __tally_stats( cltr->stats, it->local_data->this_stats );
                        it = &(*it)`next;
                }
        }

        void crawl_cluster_stats( cluster & this ) {
                // Stop the world, otherwise stats could get really messed-up
                // this doesn't solve all problems but does solve many
                // so it's probably good enough
                disable_interrupts();
                uint_fast32_t last_size = ready_mutate_lock();

                        crawl_list(&this, this.procs.actives, this.procs.total - this.procs.idle);
                        crawl_list(&this, this.procs.idles  , this.procs.idle );

                // Unlock the RWlock
                ready_mutate_unlock( last_size );
                enable_interrupts();
        }


        void print_stats_now( cluster & this, int flags ) {
                crawl_cluster_stats( this );
                __print_stats( this.stats, this.print_stats, "Cluster", this.name, (void*)&this );
        }
#endif
// Local Variables: //
// mode: c //
// tab-width: 4 //
// End: //