source: libcfa/src/concurrency/kernel/startup.cfa @ fe610ab

//
// Cforall Version 1.0.0 Copyright (C) 2020 University of Waterloo
//
// The contents of this file are covered under the licence agreement in the
// file "LICENCE" distributed with Cforall.
//
// kernel/startup.cfa --
//
// Author           : Thierry Delisle
// Created On       : Thu Jul 30 15:12:54 2020
// Last Modified By :
// Last Modified On :
// Update Count     :
//

#define __cforall_thread__
#define _GNU_SOURCE

// C Includes
#include <errno.h>                                      // errno
#include <signal.h>
#include <string.h>                                     // strerror
#include <unistd.h>                                     // sysconf

extern "C" {
        #include <limits.h>                             // PTHREAD_STACK_MIN
        #include <unistd.h>                             // syscall
        #include <sys/eventfd.h>                        // eventfd
        #include <sys/mman.h>                           // mprotect
        #include <sys/resource.h>                       // getrlimit
}

// CFA Includes
#include "kernel_private.hfa"
#include "startup.hfa"                                  // STARTUP_PRIORITY_XXX
#include "limits.hfa"
#include "math.hfa"

#define CFA_PROCESSOR_USE_MMAP 0

//-----------------------------------------------------------------------------
// 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) \
        )
#elif defined( __x86_64 )
        #define CtxGet( ctx ) __asm__ volatile ( \
                "movq %%rsp,%0\n" \
                "movq %%rbp,%1\n" \
                : "=rm" (ctx.SP), \
                  "=rm" (ctx.FP) \
        )
#elif defined( __aarch64__ )
        #define CtxGet( ctx ) __asm__ volatile ( \
                "mov %0, sp\n" \
                "mov %1, fp\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 ) ));

//-----------------------------------------------------------------------------
// Static Forward Declarations
struct current_stack_info_t;

static void * __invoke_processor(void * arg);
static void __kernel_first_resume( processor * this );
static void __kernel_last_resume ( processor * this );
static void init(processor & this, const char name[], cluster & _cltr, thread$ * initT);
static void deinit(processor & this);
static void doregister( struct cluster & cltr );
static void unregister( struct cluster & cltr );
static void register_tls( processor * this );
static void unregister_tls( processor * this );
static void ?{}( coroutine$ & this, current_stack_info_t * info);
static void ?{}( thread$ & this, current_stack_info_t * info);
static void ?{}(processorCtx_t & this) {}
static void ?{}(processorCtx_t & this, processor * proc, current_stack_info_t * info);

#if defined(__CFA_WITH_VERIFY__)
        static bool verify_fwd_bck_rng(void);
#endif

//-----------------------------------------------------------------------------
// Forward Declarations for other modules
extern void __kernel_alarm_startup(void);
extern void __kernel_alarm_shutdown(void);

//-----------------------------------------------------------------------------
// Other Forward Declarations
extern void __wake_proc(processor *);
extern int cfa_main_returned;                                                   // from interpose.cfa
uint32_t __global_random_prime = 4_294_967_291u, __global_random_mask = false;

//-----------------------------------------------------------------------------
// 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;
        extern size_t __cfa_page_size;
        extern int __map_prot;
}

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

#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
        // No data needed
#elif defined(CFA_HAVE_LINUX_RSEQ_H)
        extern "Cforall" {
                __attribute__((aligned(128))) thread_local volatile struct rseq __cfaabi_rseq @= {
                        .cpu_id : RSEQ_CPU_ID_UNINITIALIZED,
                };
        }
#else
        // No data needed
#endif

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


//=============================================================================================
// Kernel Setup logic
//=============================================================================================
//-----------------------------------------------------------------------------
// Kernel boot procedures
static void __kernel_startup(void) {
        /* paranoid */ verify( ! __preemption_enabled() );
        __cfadbg_print_safe(runtime_core, "Kernel : Starting\n");

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

        /* paranoid */ verify( verify_fwd_bck_rng() );

        // 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", 0};

        __cfadbg_print_safe(runtime_core, "Kernel : Main cluster 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 ) {
                ( this.terminated ){};
                ( this.runner ){};
                init( this, "Main Processor", *mainCluster, 0p );
                kernel_thread = pthread_self();

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

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

        register_tls( mainProcessor );

        // 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");

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

        #if !defined( __CFA_NO_STATISTICS__ )
                __cfaabi_tls.this_stats = (__stats_t *)& storage_mainProcStats;
                __init_stats( __cfaabi_tls.this_stats );
        #endif
        mainProcessor->local_data = &__cfaabi_tls;

        // Enable preemption
        __kernel_alarm_startup();

        // 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$(mainThread, UNPARK_LOCAL);

        // 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( __cfaabi_tls.this_processor );


        // THE SYSTEM IS NOW COMPLETELY RUNNING

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

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

}

static void __kernel_shutdown(void) {
        if(!cfa_main_returned) return;
        /* paranoid */ verify( __preemption_enabled() );
        disable_interrupts();
        /* paranoid */ verify( ! __preemption_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
        /* paranoid */ verify( !__atomic_load_n(&mainProcessor->do_terminate, __ATOMIC_ACQUIRE) );
        __atomic_store_n(&mainProcessor->do_terminate, true, __ATOMIC_RELEASE);
        __wake_proc( mainProcessor );
        __kernel_last_resume( __cfaabi_tls.this_processor );
        mainThread->self_cor.state = Halted;

        // THE SYSTEM IS NOW COMPLETELY STOPPED

        // Disable preemption
        __kernel_alarm_shutdown();

        #if !defined( __CFA_NO_STATISTICS__ )
                __stats_t * st = (__stats_t *)& storage_mainProcStats;
                __tally_stats(mainCluster->stats, st);
                if( 0 != mainProcessor->print_stats ) {
                        __print_stats( st, mainProcessor->print_stats, "Processor ", mainProcessor->name, (void*)mainProcessor );
                }
                #if defined(CFA_STATS_ARRAY)
                        __flush_stat( st, "Processor", mainProcessor );
                #endif
        #endif

        mainProcessor->local_data = 0p;

        unregister_tls( 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 ){
                deinit( this );

                /* paranoid */ verify( this.do_terminate == true );
                __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 Initial Scheduling logic
//=============================================================================================

// Context invoker for processors
// This is the entry point for processors (kernel threads) *except* for the main processor
// 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 );
                __cfaabi_tls.this_stats = &local_stats;
        #endif

        processor * proc = (processor *) arg;
        __cfaabi_tls.this_processor = proc;
        __cfaabi_tls.this_thread    = 0p;
        __cfaabi_tls.preemption_state.[enabled, disable_count] = [false, 1];
        proc->local_data = &__cfaabi_tls;

        register_tls( proc );

        // 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
        __cfaabi_tls.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, "Processor ", proc->name, (void*)proc );
                }
                #if defined(CFA_STATS_ARRAY)
                        __flush_stat( &local_stats, "Processor", proc );
                #endif
        #endif

        proc->local_data = 0p;

        unregister_tls( proc );

        return 0p;
}

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

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

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

        /* paranoid */ verify( ! __preemption_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);
        src->corctx_flag = true;

        // 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);

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

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

        /* paranoid */ verify( ! __preemption_enabled() );
        /* paranoid */ verify( dst->starter == src );
        /* paranoid */ 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 );
}


//=============================================================================================
// Kernel Object Constructors logic
//=============================================================================================
//-----------------------------------------------------------------------------
// Main thread construction
static 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;
}

static void ?{}( thread$ & this, current_stack_info_t * info) with( this ) {
        ticket = TICKET_RUNNING;
        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.ts   = -1llu;
        preferred = ready_queue_new_preferred();
        last_proc = 0p;
        random_state = __global_random_mask ? __global_random_prime : __global_random_prime ^ rdtscl();
        #if defined( __CFA_WITH_VERIFY__ )
                canary = 0x0D15EA5E0D15EA5Ep;
        #endif

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

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

//-----------------------------------------------------------------------------
// Processor
// 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 init(processor & this, const char name[], cluster & _cltr, thread$ * initT) with( this ) {
        this.name = name;
        this.cltr = &_cltr;
        this.rdq.its = 0;
        this.rdq.itr = 0;
        this.rdq.id  = MAX;
        this.rdq.target = MAX;
        this.rdq.last = MAX;
        this.rdq.cpu = 0;
        // this.rdq.cutoff = 0ull;
        do_terminate = false;
        preemption_alarm = 0p;
        pending_preemption = false;

        this.io.ctx = 0p;
        this.io.pending = false;
        this.io.dirty   = false;

        this.init.thrd = initT;

        this.local_data = 0p;

        this.idle_fd = eventfd(0, 0);
        if (idle_fd < 0) {
                abort("KERNEL ERROR: PROCESSOR EVENTFD - %s\n", strerror(errno));
        }

        this.idle_wctx.fd = 0;

        // I'm assuming these two are reserved for standard input and output
        // so I'm using them as sentinels with idle_wctx.
        /* paranoid */ verify( this.idle_fd != 0 );
        /* paranoid */ verify( this.idle_fd != 1 );

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

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

// Not a ctor, it just preps the destruction but should not destroy members
static void deinit(processor & this) {
        close(this.idle_fd);
}

void ?{}(processor & this, const char name[], cluster & _cltr, thread$ * initT) {
        ( this.terminated ){};
        ( this.runner ){};

        disable_interrupts();
                init( this, name, _cltr, initT );
        enable_interrupts();

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

        this.stack = __create_pthread( &this.kernel_thread, __invoke_processor, (void *)&this );
}

void ?{}(processor & this, const char name[], cluster & _cltr) {
        (this){name, _cltr, 0p};
}

void ^?{}(processor & this) with( this ){
        /* paranoid */ verify( !__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);
        __disable_interrupts_checked();
                __wake_proc( &this );
        __enable_interrupts_checked();

        wait( terminated );
        /* paranoid */ verify( active_processor() != &this);

        __destroy_pthread( kernel_thread, this.stack, 0p );

        disable_interrupts();
                deinit( this );
        enable_interrupts();
}

//-----------------------------------------------------------------------------
// Cluster
static void ?{}(__cluster_proc_list & this) {
        this.fdw   = 0p;
        this.idle  = 0;
        this.total = 0;
}

void ?{}(cluster & this, const char name[], Duration preemption_rate, unsigned num_io, const io_context_params & io_params) with( this ) {
        this.name = name;
        this.preemption_rate = preemption_rate;
        ready_queue{};

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

        threads{ __get };

        io.arbiter = create();
        io.params = io_params;

        doregister(this);

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

                // Adjust the ready queue size
                ready_queue_grow( &this );

        // Unlock the RWlock
        ready_mutate_unlock( last_size );
        enable_interrupts( false ); // Don't poll, could be in main cluster
}

void ^?{}(cluster & this) {
        destroy(this.io.arbiter);

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

                // Adjust the ready queue size
                ready_queue_shrink( &this );

        // Unlock the RWlock
        ready_mutate_unlock( last_size );
        enable_interrupts( false ); // Don't poll, could be in main cluster

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

        unregister(this);
}

//=============================================================================================
// Miscellaneous Initialization
//=============================================================================================
//-----------------------------------------------------------------------------
// Global Queues
static 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 );
}

static 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);
}

static void register_tls( processor * this ) {
        // Register and Lock the RWlock so no-one pushes/pops while we are changing the queue
        uint_fast32_t last_size;
        [this->unique_id, last_size] = ready_mutate_register();

                this->rdq.cpu = __kernel_getcpu();

                this->cltr->procs.total += 1u;
                insert_last(this->cltr->procs.actives, *this);

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

        // Unlock the RWlock
        ready_mutate_unlock( last_size );
}


static void unregister_tls( processor * this ) {
        // Lock the RWlock so no-one pushes/pops while we are changing the queue
        uint_fast32_t last_size = ready_mutate_lock();
                this->cltr->procs.total -= 1u;
                remove(*this);

                // clear the cluster so nothing gets pushed to local queues
                cluster * cltr = this->cltr;
                this->cltr = 0p;

                // Adjust the ready queue size
                ready_queue_shrink( cltr );

        // Unlock the RWlock and unregister: we don't need the read_lock any more
        ready_mutate_unregister( this->unique_id, last_size );
}

static void check( 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;

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

        size_t stacksize = max( PTHREAD_STACK_MIN, DEFAULT_STACK_SIZE );

        void * stack;
        #if CFA_PROCESSOR_USE_MMAP
                stacksize = ceiling( stacksize, __cfa_page_size ) + __cfa_page_size;
                stack = mmap(0p, stacksize, __map_prot, MAP_PRIVATE | MAP_ANONYMOUS, 0, 0);
                if(stack == ((void*)-1)) {
                        abort( "pthread stack creation : internal error, mmap failure, error(%d) %s.", errno, strerror( errno ) );
                }
                if ( mprotect( stack, __cfa_page_size, PROT_NONE ) == -1 ) {
                        abort( "pthread stack creation : internal error, mprotect failure, error(%d) %s.", errno, strerror( errno ) );
                } // if
        #else
                __cfaabi_dbg_debug_do(
                        stack = memalign( __cfa_page_size, stacksize + __cfa_page_size );
                        // pthread has no mechanism to create the guard page in user supplied stack.
                        if ( mprotect( stack, __cfa_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 );
                );
        #endif

        check( pthread_attr_setstack( &attr, stack, stacksize ), "pthread_attr_setstack" );
        check( pthread_create( pthread, &attr, start, arg ), "pthread_create" );
        return stack;
}

void __destroy_pthread( pthread_t pthread, void * stack, void ** retval ) {
        int err = pthread_join( pthread, retval );
        if( err != 0 ) abort("KERNEL ERROR: joining pthread %p caused error %s\n", (void*)pthread, strerror(err));

        #if CFA_PROCESSOR_USE_MMAP
                pthread_attr_t attr;

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

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

                if(munmap(stack, stacksize) == -1) {
                        abort( "pthread stack destruction : internal error, munmap failure, error(%d) %s.", errno, strerror( errno ) );
                }
        #else
                __cfaabi_dbg_debug_do(
                        // pthread has no mechanism to create the guard page in user supplied stack.
                        if ( mprotect( stack, __cfa_page_size, __map_prot ) == -1 ) {
                                abort( "mprotect : internal error, mprotect failure, error(%d) %s.", errno, strerror( errno ) );
                        } // if
                );
                free( stack );
        #endif
}

#if defined(__CFA_WITH_VERIFY__)
static bool verify_fwd_bck_rng(void) {
        __cfaabi_tls.ready_rng.fwd_seed = 25214903917_l64u * (rdtscl() ^ (uintptr_t)&verify_fwd_bck_rng);

        unsigned values[10];
        for(i; 10) {
                values[i] = __tls_rand_fwd();
        }

        __tls_rand_advance_bck();

        for ( i; 9 -~= 0 ) {
                if(values[i] != __tls_rand_bck()) {
                        return false;
                }
        }

        return true;
}
#endif