source: trunk/kernel/libk/remote_barrier.c @ 645

/*
 * remote_barrier.c -  POSIX barrier implementation.
 *
 * Author   Alain Greiner (2016,2017,2018,2019)
 *
 * Copyright (c) UPMC Sorbonne Universites
 *
 * This file is part of ALMOS-MKH.
 *
 * ALMOS-MKH is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License as published by
 * the Free Software Foundation; version 2.0 of the License.
 *
 * ALMOS-MKH is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with ALMOS-MKH; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
 */

#include <hal_kernel_types.h>
#include <hal_macros.h>
#include <hal_remote.h>
#include <hal_irqmask.h>
#include <remote_busylock.h>
#include <thread.h>
#include <kmem.h>
#include <printk.h>
#include <process.h>
#include <vmm.h>
#include <remote_barrier.h>

////////////////////////////////////////////////////
//  generic (implementation independant) functions
////////////////////////////////////////////////////

///////////////////////////////////////////////////
xptr_t generic_barrier_from_ident( intptr_t  ident )
{
    // get pointer on local process_descriptor
    process_t * process = CURRENT_THREAD->process;

    // get pointers on reference process
    xptr_t         ref_xp  = process->ref_xp;
    cxy_t          ref_cxy = GET_CXY( ref_xp );
    process_t    * ref_ptr = (process_t *)GET_PTR( ref_xp );

    // get extended pointer on root of barriers list
    xptr_t root_xp = XPTR( ref_cxy , &ref_ptr->barrier_root );

    // scan reference process barriers list
    xptr_t              iter_xp;
    xptr_t              barrier_xp;
    cxy_t               barrier_cxy;
    generic_barrier_t * barrier_ptr;
    intptr_t            current;
    bool_t              found = false;

    XLIST_FOREACH( root_xp , iter_xp )
    {
        barrier_xp  = XLIST_ELEMENT( iter_xp , generic_barrier_t , list );
        barrier_cxy = GET_CXY( barrier_xp );
        barrier_ptr = (generic_barrier_t *)GET_PTR( barrier_xp );
        current     = (intptr_t)hal_remote_lpt( XPTR( barrier_cxy , &barrier_ptr->ident ) );
        if( ident == current )
        {
            found = true;
            break;
        }
    }

    if( found == false )  return XPTR_NULL;
    else                  return barrier_xp;

} // end generic_barrier_from_ident()

//////////////////////////////////////////////////////////////
error_t generic_barrier_create( intptr_t                ident,
                                uint32_t                count,
                                pthread_barrierattr_t * attr )
{
    generic_barrier_t * gen_barrier_ptr;  // local pointer on generic barrier descriptor
    void              * barrier;          // local pointer on implementation barrier descriptor      
    kmem_req_t          req;              // kmem request

    // get pointer on local process_descriptor
    process_t * process = CURRENT_THREAD->process;

    // get pointers on reference process
    xptr_t         ref_xp  = process->ref_xp;
    cxy_t          ref_cxy = GET_CXY( ref_xp );
    process_t    * ref_ptr = (process_t *)GET_PTR( ref_xp );

    // allocate memory for generic barrier descriptor
    req.type   = KMEM_KCM;
    req.order  = bits_log2( sizeof(generic_barrier_t) );
    req.flags  = AF_ZERO | AF_KERNEL;
    gen_barrier_ptr = kmem_remote_alloc( ref_cxy , &req );

    if( gen_barrier_ptr == NULL )
    {
        printk("\n[ERROR] in %s : cannot create generic barrier\n", __FUNCTION__ );
        return -1;
    }

    // create implementation specific barrier descriptor
    if( attr == NULL )                                    // simple barrier implementation
    {
        // create simple barrier descriptor
         barrier = simple_barrier_create( count );

        if( barrier == NULL ) return -1;
    }
    else                                                  // QDT barrier implementation
    {
        uint32_t x_size   = attr->x_size;
        uint32_t y_size   = attr->y_size;
        uint32_t nthreads = attr->nthreads;

        // check attributes / count
        if( (x_size * y_size * nthreads) != count )
        {
            printk("\n[ERROR] in %s : count(%d) != x_size(%d) * y_size(%d) * nthreads(%d)\n",
            __FUNCTION__, count, x_size, y_size, nthreads );
            return -1;
        }

        // create DQT barrier descriptor
        barrier = dqt_barrier_create( x_size , y_size , nthreads );

        if( barrier == NULL ) return -1;
    }

    // initialize the generic barrier descriptor
    hal_remote_spt( XPTR( ref_cxy , &gen_barrier_ptr->ident  ) , (void*)ident );
    hal_remote_s32( XPTR( ref_cxy , &gen_barrier_ptr->is_dqt ) , (attr != NULL) );
    hal_remote_spt( XPTR( ref_cxy , &gen_barrier_ptr->extend ) , barrier );

    // build extended pointers on lock, root and entry for reference process xlist
    xptr_t root_xp  = XPTR( ref_cxy , &ref_ptr->barrier_root );
    xptr_t lock_xp  = XPTR( ref_cxy , &ref_ptr->sync_lock );
    xptr_t entry_xp = XPTR( ref_cxy , &gen_barrier_ptr->list );

    // register barrier in reference process xlist of barriers
    remote_busylock_acquire( lock_xp );
    xlist_add_first( root_xp , entry_xp );
    remote_busylock_release( lock_xp );

    return 0;

}  // en generic_barrier_create() 

/////////////////////////////////////////////////////
void generic_barrier_destroy( xptr_t gen_barrier_xp )
{
    kmem_req_t  req;              // kmem request

    // get pointer on local process_descriptor
    process_t * process = CURRENT_THREAD->process;

    // get pointers on reference process
    xptr_t      ref_xp  = process->ref_xp;
    cxy_t       ref_cxy = GET_CXY( ref_xp );
    process_t * ref_ptr = GET_PTR( ref_xp );

    // get cluster and local pointer on generic barrier descriptor
    generic_barrier_t * gen_barrier_ptr = GET_PTR( gen_barrier_xp );
    cxy_t               gen_barrier_cxy = GET_CXY( gen_barrier_xp );

    // get barrier type and extension pointer
    bool_t  is_dqt = hal_remote_l32( XPTR( gen_barrier_cxy , &gen_barrier_ptr->is_dqt ) );
    void  * extend = hal_remote_lpt( XPTR( gen_barrier_cxy , &gen_barrier_ptr->extend ) );

    // build extended pointer on implementation dependant barrier descriptor
    xptr_t barrier_xp = XPTR( gen_barrier_cxy , extend );

    // delete the implementation specific barrier
    if( is_dqt ) dqt_barrier_destroy( barrier_xp );
    else         simple_barrier_destroy( barrier_xp );

    // build extended pointers on lock and entry for reference process xlist
    xptr_t  lock_xp  = XPTR( ref_cxy , &ref_ptr->sync_lock );
    xptr_t  entry_xp = XPTR( gen_barrier_cxy , &gen_barrier_ptr->list );

    // remove barrier from reference process xlist
    remote_busylock_acquire( lock_xp );
    xlist_unlink( entry_xp );
    remote_busylock_release( lock_xp );

    // release memory allocated to barrier descriptor
    req.type          = KMEM_KCM;
    req.ptr           = gen_barrier_ptr;
    kmem_remote_free( ref_cxy , &req );

}  // end generic_barrier_destroy()

//////////////////////////////////////////////////
void generic_barrier_wait( xptr_t gen_barrier_xp )
{
    // get generic barrier descriptor cluster and pointer
    cxy_t               gen_barrier_cxy = GET_CXY( gen_barrier_xp );
    generic_barrier_t * gen_barrier_ptr = GET_PTR( gen_barrier_xp );

    // get implementation type and extend local pointer
    bool_t  is_dqt = hal_remote_l32( XPTR( gen_barrier_cxy , &gen_barrier_ptr->is_dqt ) );
    void  * extend = hal_remote_lpt( XPTR( gen_barrier_cxy , &gen_barrier_ptr->extend ) );

    // build extended pointer on implementation specific barrier descriptor
    xptr_t barrier_xp = XPTR( gen_barrier_cxy , extend );

    // call the relevant wait function
    if( is_dqt ) dqt_barrier_wait( barrier_xp );
    else         simple_barrier_wait( barrier_xp );
    
}  // end generic_barrier_wait()

/////////////////////////////////////////////////////
void generic_barrier_display( xptr_t gen_barrier_xp )
{
    // get cluster and local pointer
    generic_barrier_t * gen_barrier_ptr = GET_PTR( gen_barrier_xp );
    cxy_t               gen_barrier_cxy = GET_CXY( gen_barrier_xp );

    // get barrier type and extend pointer
    bool_t  is_dqt = hal_remote_l32( XPTR( gen_barrier_cxy , &gen_barrier_ptr->is_dqt ) );
    void  * extend = hal_remote_lpt( XPTR( gen_barrier_cxy , &gen_barrier_ptr->extend ) );

    // buil extended pointer on the implementation specific barrier descriptor
    xptr_t barrier_xp = XPTR( gen_barrier_cxy , extend );

    // display barrier state
    if( is_dqt ) dqt_barrier_display( barrier_xp );
    else         simple_barrier_display( barrier_xp );
}



/////////////////////////////////////////////////////////////
//      simple barrier functions
/////////////////////////////////////////////////////////////

///////////////////////////////////////////////////////////
simple_barrier_t * simple_barrier_create( uint32_t  count )
{
    kmem_req_t         req;
    simple_barrier_t * barrier;

    // get pointer on local client process descriptor
    thread_t  * this    = CURRENT_THREAD;
    process_t * process = this->process;

    // get reference process cluster
    xptr_t         ref_xp  = process->ref_xp;
    cxy_t          ref_cxy = GET_CXY( ref_xp );

    // allocate memory for simple barrier descriptor
    req.type   = KMEM_KCM;
    req.order  = bits_log2( sizeof(simple_barrier_t) );
    req.flags  = AF_ZERO | AF_KERNEL;
    barrier    = kmem_remote_alloc( ref_cxy , &req );

    if( barrier == NULL )
    {
        printk("\n[ERROR] in %s : cannot create simple barrier\n", __FUNCTION__ );
        return NULL;
    }

    // initialise simple barrier descriptor
    hal_remote_s32      ( XPTR( ref_cxy , &barrier->arity )      , count );
    hal_remote_s32      ( XPTR( ref_cxy , &barrier->current    ) , 0 );
    hal_remote_s32      ( XPTR( ref_cxy , &barrier->sense      ) , 0 );

    xlist_root_init     ( XPTR( ref_cxy , &barrier->root ) );
    remote_busylock_init( XPTR( ref_cxy , &barrier->lock ) , LOCK_BARRIER_STATE );

#if DEBUG_BARRIER_CREATE
uint32_t cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_CREATE )
printk("\n[%s] thread[%x,%x] created barrier (%x,%x) / count %d / cycle %d\n",
__FUNCTION__, process->pid, this->trdid, ref_cxy, barrier, count, cycle );
#endif

    return barrier;

}  // end simple_barrier_create()

////////////////////////////////////////////////
void simple_barrier_destroy( xptr_t barrier_xp )
{
    kmem_req_t  req;

    // get barrier cluster and local pointer
    cxy_t              barrier_cxy = GET_CXY( barrier_xp );
    simple_barrier_t * barrier_ptr = GET_PTR( barrier_xp );

    // release memory allocated for barrier descriptor
    req.type = KMEM_KCM;
    req.ptr  = barrier_ptr;
    kmem_remote_free( barrier_cxy , &req );

#if DEBUG_BARRIER_DESTROY
uint32_t    cycle   = (uint32_t)hal_get_cycles();
thread_t  * this    = CURRENT_THREAD;
process_t * process = this->process;
if( cycle > DEBUG_BARRIER_DESTROY )
printk("\n[%s] thread[%x,%x] deleted barrier (%x,%x) / cycle %d\n",
__FUNCTION__, process->pid, this->trdid, barrier_ptr, barrier_cxy, cycle );
#endif

}  // end simple_barrier_destroy()

/////////////////////////////////////////////
void simple_barrier_wait( xptr_t barrier_xp )
{
    uint32_t  expected;
    uint32_t  sense;
    uint32_t  current;
    uint32_t  arity;
    xptr_t    root_xp;
    xptr_t    lock_xp;
    xptr_t    current_xp;
    xptr_t    sense_xp;
    xptr_t    arity_xp;

    // get pointer on calling thread
    thread_t * this = CURRENT_THREAD;

    // check calling thread can yield
    thread_assert_can_yield( this , __FUNCTION__ );

    // get cluster and local pointer on remote barrier
    simple_barrier_t * barrier_ptr = GET_PTR( barrier_xp );
    cxy_t              barrier_cxy = GET_CXY( barrier_xp );

#if DEBUG_BARRIER_WAIT
uint32_t cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] enter / barrier (%x,%x) / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, barrier_cxy, barrier_ptr, cycle );
#endif

    // build extended pointers on various barrier descriptor fields
    lock_xp    = XPTR( barrier_cxy , &barrier_ptr->lock );
    root_xp    = XPTR( barrier_cxy , &barrier_ptr->root );
    current_xp = XPTR( barrier_cxy , &barrier_ptr->current );
    sense_xp   = XPTR( barrier_cxy , &barrier_ptr->sense );
    arity_xp   = XPTR( barrier_cxy , &barrier_ptr->arity );

    // take busylock protecting the barrier state
    remote_busylock_acquire( lock_xp );

    // get sense and threads values from barrier descriptor
    sense = hal_remote_l32( sense_xp );
    arity = hal_remote_l32( arity_xp );

    // compute expected value
    if ( sense == 0 ) expected = 1;
    else              expected = 0;

    // increment current number of arrived threads / get value before increment
    current = hal_remote_atomic_add( current_xp , 1 );

    // last thread reset current, toggle sense, and activate all waiting threads
    // other threads block, register in queue, and deschedule

    if( current == (arity - 1) )                       // last thread
    {
        hal_remote_s32( current_xp , 0 );
        hal_remote_s32( sense_xp , expected );

        // unblock all waiting threads
        while( xlist_is_empty( root_xp ) == false )
        {
            // get pointers on first waiting thread
            xptr_t     thread_xp  = XLIST_FIRST( root_xp , thread_t , wait_list );
            cxy_t      thread_cxy = GET_CXY( thread_xp );
            thread_t * thread_ptr = GET_PTR( thread_xp );

#if (DEBUG_BARRIER_WAIT & 1)
trdid_t     trdid   = hal_remote_l32( XPTR( thread_cxy , &thread_ptr->trdid ) );
process_t * process = hal_remote_lpt( XPTR( thread_cxy , &thread_ptr->process ) );
pid_t       pid     = hal_remote_l32( XPTR( thread_cxy , &process->pid ) ); 
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] unblocks thread[%x,%x]\n",
__FUNCTION__, this->process->pid, this->trdid, pid, trdid );
#endif

            // remove waiting thread from queue
            xlist_unlink( XPTR( thread_cxy , &thread_ptr->wait_list ) );

            // unblock waiting thread
            thread_unblock( thread_xp , THREAD_BLOCKED_USERSYNC );
        }

        // release busylock protecting the barrier
        remote_busylock_release( lock_xp );
    }
    else                                             // not the last thread
    {

#if (DEBUG_BARRIER_WAIT & 1)
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] blocks\n",
__FUNCTION__, this->process->pid, this->trdid );
#endif

        // register calling thread in barrier waiting queue
        xlist_add_last( root_xp , XPTR( local_cxy , &this->wait_list ) );

        // block calling thread
        thread_block( XPTR( local_cxy , this ) , THREAD_BLOCKED_USERSYNC );

        // release busylock protecting the remote_barrier
        remote_busylock_release( lock_xp );

        // deschedule
        sched_yield("blocked on barrier");
    }

#if DEBUG_BARRIER_WAIT
cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] exit / barrier (%x,%x) / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, barrier_cxy, barrier_ptr, cycle );
#endif

}  // end simple_barrier_wait()

/////////////////////////////////////////////////
void simple_barrier_display( xptr_t  barrier_xp )
{
    // get cluster and local pointer on simple barrier 
    simple_barrier_t * barrier_ptr = GET_PTR( barrier_xp );
    cxy_t              barrier_cxy = GET_CXY( barrier_xp );

    // get barrier global parameters
    uint32_t current  = hal_remote_l32( XPTR( barrier_cxy , &barrier_ptr->current ) );
    uint32_t arity    = hal_remote_l32( XPTR( barrier_cxy , &barrier_ptr->arity   ) );

    printk("\n***** simple barrier : %d arrived threads on %d *****\n",
    current, arity );

}   // end simple_barrier_display()




/////////////////////////////////////////////////////////////
//      DQT barrier functions
/////////////////////////////////////////////////////////////

static void dqt_barrier_increment( xptr_t node_xp );

#if DEBUG_BARRIER_CREATE
void dqt_barrier_display( xptr_t  barrier_xp );
#endif

///////////////////////////////////////////////////////
dqt_barrier_t * dqt_barrier_create( uint32_t    x_size,
                                    uint32_t    y_size,
                                    uint32_t    nthreads )
{
    dqt_barrier_t * barrier;       // local pointer on DQT barrier descriptor
    xptr_t          barrier_xp;    // extended pointer on DQT barrier descriptor
    uint32_t        z;             // actual DQT size == max(x_size,y_size)
    uint32_t        levels;        // actual number of DQT levels
    uint32_t        x;             // X coordinate in QDT mesh
    uint32_t        y;             // Y coordinate in QDT mesh
    uint32_t        l;             // level coordinate
    kmem_req_t      req;           // kmem request

    // compute number of DQT levels, depending on the mesh size
    z      = (x_size > y_size) ? x_size : y_size;
    levels = (z < 2) ? 1 : (z < 3) ? 2 : (z < 5) ? 3 : (z < 9) ? 4 : 5;

// check x_size and y_size arguments
assert( (z <= 16) , "DQT mesh size larger than (16*16)\n");

// check size of an array of 5 DQT nodes
assert( (sizeof(dqt_node_t) * 5 <= 512 ), "array of DQT nodes larger than 512 bytes\n");

// check size of DQT barrier descriptor
assert( (sizeof(dqt_barrier_t) <= 0x4000 ), "DQT barrier descriptor larger than 4 pages\n");

    // get pointer on client thread and process descriptors
    thread_t  * this    = CURRENT_THREAD;
    process_t * process = this->process;

#if DEBUG_BARRIER_CREATE
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_CREATE ) 
printk("\n[%s] thread[%x,%x] enter : x_size %d / y_size %d / levels %d / cycle %d\n",
__FUNCTION__, process->pid, this->trdid, x_size, y_size, levels, cycle );
#endif

    // get reference process cluster
    xptr_t         ref_xp  = process->ref_xp;
    cxy_t          ref_cxy = GET_CXY( ref_xp );

    // 1. allocate 4 small pages for the DQT barrier descriptor in reference cluster
    req.type   = KMEM_PPM;
    req.order  = 2;                     // 4 small pages == 16 Kbytes                     
    req.flags  = AF_ZERO | AF_KERNEL;
    barrier    = kmem_remote_alloc( ref_cxy , &req );

    if( barrier == NULL )
    {
        printk("\n[ERROR] in %s : cannot create DQT barrier\n", __FUNCTION__ );
        return NULL;
    }

    // get pointers on DQT barrier descriptor in reference cluster
    barrier_xp = XPTR( ref_cxy , barrier );

    // initialize global parameters in DQT barrier descriptor
    hal_remote_s32( XPTR( ref_cxy , &barrier->x_size   ) , x_size );
    hal_remote_s32( XPTR( ref_cxy , &barrier->y_size   ) , x_size );
    hal_remote_s32( XPTR( ref_cxy , &barrier->nthreads ) , nthreads );

#if DEBUG_BARRIER_CREATE
if( cycle > DEBUG_BARRIER_CREATE ) 
printk("\n[%s] thread[%x,%x] created DQT barrier descriptor(%x,%x)\n",
__FUNCTION__, process->pid, this->trdid, ref_cxy, barrier );
#endif

    // 2. allocate memory for an array of 5 DQT nodes  
    //    in all existing clusters covered by the DQDT
    //    (5 nodes per cluster <= 512 bytes per cluster)
    //    and complete barrier descriptor initialisation. 
    for ( x = 0 ; x < x_size ; x++ )
    {
        for ( y = 0 ; y < y_size ; y++ )
        {
            cxy_t  cxy = HAL_CXY_FROM_XY( x , y );   // target cluster identifier
            xptr_t local_array_xp;                   // xptr of nodes array in cluster cxy

            // allocate memory in existing clusters only
            if( LOCAL_CLUSTER->cluster_info[x][y] )
            {
                req.type  = KMEM_KCM;
                req.order = 9;                    // 512 bytes
                req.flags = AF_ZERO | AF_KERNEL;

                void * ptr = kmem_remote_alloc( cxy , &req );

                if( ptr == NULL )
                {
                    printk("\n[ERROR] in %s : cannot allocate DQT in cluster %x\n",
                    __FUNCTION__, cxy );
                    return NULL;
                }
       
                // build extended pointer on local node array in cluster cxy         
                            local_array_xp = XPTR( cxy , ptr );

                // initialize the node_xp[x][y][l] array in barrier descriptor
                for ( l = 0 ; l < levels ; l++ )
                {
                    xptr_t  node_xp = local_array_xp + ( l * sizeof(dqt_node_t) );
                    hal_remote_s64( XPTR( ref_cxy , &barrier->node_xp[x][y][l] ), node_xp );

#if (DEBUG_BARRIER_CREATE & 1)
if( cycle > DEBUG_BARRIER_CREATE )
printk(" - dqt_node_xp[%d,%d,%d] = (%x,%x) / &dqt_node_xp = %x\n",
x , y , l , GET_CXY( node_xp ), GET_PTR( node_xp ), &barrier->node_xp[x][y][l] );
#endif
                }  
            }
            else   // register XPTR_NULL for all non-existing entries
            {
                for ( l = 0 ; l < levels ; l++ )
                {
                    hal_remote_s64( XPTR( ref_cxy , &barrier->node_xp[x][y][l] ), XPTR_NULL );
                }
            }
        }  // end for y 
    }  // end for x

#if DEBUG_BARRIER_CREATE
if( cycle > DEBUG_BARRIER_CREATE ) 
printk("\n[%s] thread[%x,%x] initialized array of pointers in DQT barrier\n",
__FUNCTION__, process->pid, this->trdid );
#endif

    // 3. initialise all distributed DQT nodes using remote accesses
    //    and the pointers stored in the node_xp[x][y][l] array
    for ( x = 0 ; x < x_size ; x++ )
    {
        for ( y = 0 ; y < y_size ; y++ )
        {
            // initialize existing clusters only
            if( LOCAL_CLUSTER->cluster_info[x][y] )
            {
                for ( l = 0 ; l < levels ; l++ )
                {
                                    xptr_t    parent_xp;
                    xptr_t    child_xp[4];
                    uint32_t  arity = 0;

                    // get DQT node pointers
                    xptr_t       node_xp  = hal_remote_l64( XPTR( ref_cxy,
                                            &barrier->node_xp[x][y][l] ) );
                    cxy_t        node_cxy = GET_CXY( node_xp );
                    dqt_node_t * node_ptr = GET_PTR( node_xp ); 

                    // compute arity and child_xp[i] 
                    if (l == 0 )                            // bottom DQT node
                    {
                        arity       = nthreads;

                        child_xp[0] = XPTR_NULL;
                        child_xp[1] = XPTR_NULL;
                        child_xp[2] = XPTR_NULL;
                        child_xp[3] = XPTR_NULL;
                    }
                    else                                    // not a bottom DQT node
                    {
                        arity = 0;

                        // only few non-bottom nodes must be initialised
                        if( ((x & ((1<<l)-1)) == 0) && ((y & ((1<<l)-1)) == 0) )
                        {
                            uint32_t cx[4];       // x coordinate for children
                            uint32_t cy[4];       // y coordinate for children
                            uint32_t i;

                            // the child0 coordinates are equal to the parent coordinates
                            // other children coordinates depend on the level value
                            cx[0] = x;
                            cy[0] = y;

                            cx[1] = x;
                            cy[1] = y + (1 << (l-1));

                            cx[2] = x + (1 << (l-1));
                            cy[2] = y;

                            cx[3] = x + (1 << (l-1));
                            cy[3] = y + (1 << (l-1));

                            for ( i = 0 ; i < 4 ; i++ )
                            {
                                // child pointer is NULL if  outside the mesh
                                if ( (cx[i] < x_size) && (cy[i] < y_size) ) 
                                {
                                    // get child_xp[i] 
                                    child_xp[i] = hal_remote_l64( XPTR( ref_cxy,
                                                  &barrier->node_xp[cx[i]][cy[i]][l-1] ) );

                                    // increment arity
                                    arity++;
                                }
                                else
                                {
                                    child_xp[i] = XPTR_NULL;
                                }
                            }
                        }
                    }

                    // compute parent_xp
                    if( l == (levels - 1) )                      // root DQT node
                    {
                        parent_xp = XPTR_NULL;
                    }
                    else                                          // not the root
                    {
                        uint32_t px = 0;           // parent X coordinate
                        uint32_t py = 0;           // parent Y coordinate
                        bool_t   found = false;

                        // compute macro_cluster x_min, x_max, y_min, y_max                
                        uint32_t x_min = x & ~((1<<(l+1))-1);
                        uint32_t x_max = x_min + (1<<(l+1));
                        uint32_t y_min = y & ~((1<<(l+1))-1);
                        uint32_t y_max = y_min + (1<<(l+1));

                        // scan all clusters in macro-cluster[x][y][l] / take first active
                        for( px = x_min ; px < x_max ; px++ )
                        {
                            for( py = y_min ; py < y_max ; py++ )
                            {
                                if( LOCAL_CLUSTER->cluster_info[px][py] ) found = true;
                                if( found ) break;
                            }
                            if( found ) break;
                        }

                        parent_xp = hal_remote_l64( XPTR( ref_cxy ,
                                    &barrier->node_xp[px][py][l+1] ) );
                    }

                    // initializes  the DQT node
                    hal_remote_s32( XPTR( node_cxy , &node_ptr->arity )       , arity );   
                    hal_remote_s32( XPTR( node_cxy , &node_ptr->current )     , 0 );   
                    hal_remote_s32( XPTR( node_cxy , &node_ptr->sense )       , 0 );   
                    hal_remote_s32( XPTR( node_cxy , &node_ptr->level )       , l );   
                    hal_remote_s64( XPTR( node_cxy , &node_ptr->parent_xp )   , parent_xp );
                    hal_remote_s64( XPTR( node_cxy , &node_ptr->child_xp[0] ) , child_xp[0] );
                    hal_remote_s64( XPTR( node_cxy , &node_ptr->child_xp[1] ) , child_xp[1] );
                    hal_remote_s64( XPTR( node_cxy , &node_ptr->child_xp[2] ) , child_xp[2] );
                    hal_remote_s64( XPTR( node_cxy , &node_ptr->child_xp[3] ) , child_xp[3] );

                    xlist_root_init( XPTR( node_cxy , &node_ptr->root ) );

                    remote_busylock_init( XPTR( node_cxy , &node_ptr->lock ),
                                          LOCK_BARRIER_STATE );
                }
            }
        }
    }

#if DEBUG_BARRIER_CREATE 
cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_CREATE ) 
printk("\n[%s] thread[%x,%x] completed DQT barrier initialisation / cycle %d\n",
__FUNCTION__, process->pid, this->trdid, cycle );
dqt_barrier_display( barrier_xp );
#endif

    return barrier;

}  // end dqt_barrier_create()

///////////////////////////////////////////////
void dqt_barrier_destroy( xptr_t   barrier_xp )
{
    kmem_req_t   req;                      // kmem request
    uint32_t     x;
    uint32_t     y;


    // get DQT barrier descriptor cluster and local pointer
    dqt_barrier_t * barrier_ptr = GET_PTR( barrier_xp );
    cxy_t           barrier_cxy = GET_CXY( barrier_xp );

#if DEBUG_BARRIER_DESTROY
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_DESTROY ) 
printk("\n[%s] thread[%x,%x] enter for barrier (%x,%x) / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, barrier_cxy, barrier_ptr, cycle );
#endif

    // get x_size and y_size global parameters
    uint32_t x_size = hal_remote_l32( XPTR( barrier_cxy , &barrier_ptr->x_size ) );
    uint32_t y_size = hal_remote_l32( XPTR( barrier_cxy , &barrier_ptr->y_size ) );

    // 1. release memory allocated for the DQT nodes
    //    in all clusters covered by the QDT mesh
    for ( x = 0 ; x < x_size ; x++ )
    {
        for ( y = 0 ; y < y_size ; y++ )
        {
            // compute target cluster identifier
            cxy_t   cxy = HAL_CXY_FROM_XY( x , y ); 

            // existing cluster only
            if( LOCAL_CLUSTER->cluster_info[x][y] )
            {
                // get local pointer on dqt_nodes array in target cluster  
                xptr_t  buf_xp_xp = XPTR( barrier_cxy , &barrier_ptr->node_xp[x][y][0] );
                xptr_t  buf_xp    = hal_remote_l64( buf_xp_xp );
                void  * buf       = GET_PTR( buf_xp );

assert( (cxy == GET_CXY(buf_xp)) , "bad extended pointer on dqt_nodes array\n" );

                req.type  = KMEM_KCM;
                req.ptr   = buf;
                kmem_remote_free( cxy , &req );

#if DEBUG_BARRIER_DESTROY
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_DESTROY ) 
printk("\n[%s] thread[%x,%x] released node array %x in cluster %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, buf, cxy, cycle );
#endif
            }
        }
    }

    // 2. release memory allocated for barrier descriptor in ref cluster
    req.type = KMEM_PPM;
    req.ptr  = barrier_ptr;
    kmem_remote_free( barrier_cxy , &req );

#if DEBUG_BARRIER_DESTROY
cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_DESTROY ) 
printk("\n[%s] thread[%x,%x] release barrier descriptor (%x,%x) / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, barrier_cxy, barrier_ptr, cycle );
#endif

}  // end dqt_barrier_destroy()

////////////////////////////////////////////
void dqt_barrier_wait( xptr_t   barrier_xp )
{
    thread_t * this = CURRENT_THREAD;

    // check calling thread can yield
    thread_assert_can_yield( this , __FUNCTION__ );

    // get cluster and local pointer on DQT barrier descriptor
    dqt_barrier_t * barrier_ptr = GET_PTR( barrier_xp );
    cxy_t           barrier_cxy = GET_CXY( barrier_xp );

#if DEBUG_BARRIER_WAIT
uint32_t cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] enter / barrier (%x,%x) / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, barrier_cxy, barrier_ptr, cycle );
#endif

    // get extended pointer on local bottom DQT node
    uint32_t x       = HAL_X_FROM_CXY( local_cxy );
    uint32_t y       = HAL_Y_FROM_CXY( local_cxy );
    xptr_t   node_xp = hal_remote_l64( XPTR( barrier_cxy , &barrier_ptr->node_xp[x][y][0] ) );

    // call recursive function to traverse DQT from bottom to root
    dqt_barrier_increment( node_xp );

#if DEBUG_BARRIER_WAIT
cycle = (uint32_t)hal_get_cycles();
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] exit / barrier (%x,%x) / cycle %d\n",
__FUNCTION__, this->trdid, this->process->pid, barrier_cxy, barrier_ptr, cycle );
#endif

}  // end dqt_barrier_wait()

//////////////////////////////////////////////
void dqt_barrier_display( xptr_t  barrier_xp )
{
    // get cluster and local pointer on DQT barrier 
    dqt_barrier_t * barrier_ptr = GET_PTR( barrier_xp );
    cxy_t           barrier_cxy = GET_CXY( barrier_xp );

    // get barrier global parameters
    uint32_t x_size   = hal_remote_l32( XPTR( barrier_cxy , &barrier_ptr->x_size ) );
    uint32_t y_size   = hal_remote_l32( XPTR( barrier_cxy , &barrier_ptr->y_size ) );
    uint32_t nthreads = hal_remote_l32( XPTR( barrier_cxy , &barrier_ptr->nthreads ) );

    // compute size and number of DQT levels
    uint32_t z      = (x_size > y_size) ? x_size : y_size;
    uint32_t levels = (z < 2) ? 1 : (z < 3) ? 2 : (z < 5) ? 3 : (z < 9) ? 4 : 5;

    printk("\n***** DQT barrier : x_size %d / y_size %d / nthreads %d / levels %d *****\n",
    x_size, y_size, nthreads, levels );

    uint32_t x , y , l;

    for ( x = 0 ; x < x_size ; x++ )
    {
        for ( y = 0 ; y < y_size ; y++ )
        {
            printk(" - cluster[%d,%d]\n", x , y );

            for ( l = 0 ; l < levels ; l++ )
            {
                // get pointers on target node
                xptr_t       node_xp  = hal_remote_l64( XPTR( barrier_cxy ,
                                        &barrier_ptr->node_xp[x][y][l] ) );
                dqt_node_t * node_ptr = GET_PTR( node_xp );
                cxy_t        node_cxy = GET_CXY( node_xp );

                if( node_xp != XPTR_NULL )
                {
                     uint32_t level = hal_remote_l32( XPTR( node_cxy , &node_ptr->level       ));
                     xptr_t   pa_xp = hal_remote_l32( XPTR( node_cxy , &node_ptr->parent_xp   ));
                     xptr_t   c0_xp = hal_remote_l32( XPTR( node_cxy , &node_ptr->child_xp[0] ));
                     xptr_t   c1_xp = hal_remote_l32( XPTR( node_cxy , &node_ptr->child_xp[1] ));
                     xptr_t   c2_xp = hal_remote_l32( XPTR( node_cxy , &node_ptr->child_xp[2] ));
                     xptr_t   c3_xp = hal_remote_l32( XPTR( node_cxy , &node_ptr->child_xp[3] ));

                     printk("   . level %d : (%x,%x) / P(%x,%x) / C0(%x,%x)"
                            " C1(%x,%x) / C2(%x,%x) / C3(%x,%x)\n",
                     level, node_cxy, node_ptr, 
                     GET_CXY(pa_xp), GET_PTR(pa_xp),
                     GET_CXY(c0_xp), GET_PTR(c0_xp),
                     GET_CXY(c1_xp), GET_PTR(c1_xp),
                     GET_CXY(c2_xp), GET_PTR(c2_xp),
                     GET_CXY(c3_xp), GET_PTR(c3_xp) );
                }
            }
        }
    }
}   // end dqt_barrier_display()


//////////////////////////////////////////////////////////////////////////////////////////
// This static (recursive) function is called by the dqt_barrier_wait() function.
// It traverses the DQT from bottom to root, and decrements the "current" variables.
// For each traversed node, it blocks and deschedules if it is not the last expected
//  thread. The last arrived thread reset the local node before returning.
//////////////////////////////////////////////////////////////////////////////////////////
static void dqt_barrier_increment( xptr_t  node_xp )
{
    uint32_t   expected;
    uint32_t   sense;
    uint32_t   arity;

    thread_t * this = CURRENT_THREAD;

    // get node cluster and local pointer
    dqt_node_t * node_ptr = GET_PTR( node_xp );
    cxy_t        node_cxy = GET_CXY( node_xp );

    // build relevant extended pointers
    xptr_t  arity_xp   = XPTR( node_cxy , &node_ptr->arity );
    xptr_t  sense_xp   = XPTR( node_cxy , &node_ptr->sense );
    xptr_t  current_xp = XPTR( node_cxy , &node_ptr->current );
    xptr_t  lock_xp    = XPTR( node_cxy , &node_ptr->lock );
    xptr_t  root_xp    = XPTR( node_cxy , &node_ptr->root );

#if DEBUG_BARRIER_WAIT
uint32_t   cycle = (uint32_t)hal_get_cycles();
uint32_t   level = hal_remote_l32( XPTR( node_cxy, &node_ptr->level ) );
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] increments DQT node(%d,%d,%d) / cycle %d\n",
__FUNCTION__ , this->process->pid, this->trdid, 
HAL_X_FROM_CXY(node_cxy), HAL_Y_FROM_CXY(node_cxy), level );
#endif

    // get extended pointer on parent node
    xptr_t  parent_xp  = hal_remote_l64( XPTR( node_cxy , &node_ptr->parent_xp ) );

    // take busylock
    remote_busylock_acquire( lock_xp );
    
    // get sense and arity values from barrier descriptor
    sense = hal_remote_l32( sense_xp );
    arity = hal_remote_l32( arity_xp );

    // compute expected value 
    expected = (sense == 0) ? 1 : 0;

    // increment current number of arrived threads / get value before increment
    uint32_t current = hal_remote_atomic_add( current_xp , 1 );

    // last arrived thread reset the local node, makes the recursive call 
    // on parent node, and reactivates all waiting thread when returning.
    // other threads block, register in queue, and deschedule.

    if ( current == (arity - 1) )                        // last thread  
    {

#if DEBUG_BARRIER_WAIT
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] reset DQT node(%d,%d,%d)\n",
__FUNCTION__ , this->process->pid, this->trdid,
HAL_X_FROM_CXY(node_cxy), HAL_Y_FROM_CXY(node_cxy), level );
#endif
        // reset the current node
        hal_remote_s32( sense_xp   , expected );
        hal_remote_s32( current_xp , 0 );

        // release busylock protecting the current node
        remote_busylock_release( lock_xp );

        // recursive call on parent node when current node is not the root
        if( parent_xp != XPTR_NULL) dqt_barrier_increment( parent_xp );

        // unblock all waiting threads on this node
        while( xlist_is_empty( root_xp ) == false )
        {
            // get pointers on first waiting thread
            xptr_t     thread_xp  = XLIST_FIRST( root_xp , thread_t , wait_list );
            cxy_t      thread_cxy = GET_CXY( thread_xp );
            thread_t * thread_ptr = GET_PTR( thread_xp );

#if (DEBUG_BARRIER_WAIT & 1)
trdid_t     trdid   = hal_remote_l32( XPTR( thread_cxy , &thread_ptr->trdid ) );
process_t * process = hal_remote_lpt( XPTR( thread_cxy , &thread_ptr->process ) );
pid_t       pid     = hal_remote_l32( XPTR( thread_cxy , &process->pid ) ); 
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] unblock thread[%x,%x]\n",
__FUNCTION__, this->process->pid, this->trdid, pid, trdid );
#endif
            // remove waiting thread from queue
            xlist_unlink( XPTR( thread_cxy , &thread_ptr->wait_list ) );

            // unblock waiting thread
            thread_unblock( thread_xp , THREAD_BLOCKED_USERSYNC );
        }
    }
    else                                               // not the last thread
    {
        // get extended pointer on xlist entry from thread
        xptr_t  entry_xp = XPTR( local_cxy , &this->wait_list );
        
        // register calling thread in barrier waiting queue
        xlist_add_last( root_xp , entry_xp );

        // block calling thread
        thread_block( XPTR( local_cxy , this ) , THREAD_BLOCKED_USERSYNC );

        // release busylock protecting the remote_barrier
        remote_busylock_release( lock_xp );

#if DEBUG_BARRIER_WAIT
if( cycle > DEBUG_BARRIER_WAIT )
printk("\n[%s] thread[%x,%x] blocks on node(%d,%d,%d)\n",
__FUNCTION__ , this->process->pid, this->trdid,
HAL_X_FROM_CXY(node_cxy), HAL_Y_FROM_CXY(node_cxy), level );
#endif
        // deschedule
        sched_yield("blocked on barrier");
    }

    return;

} // end dqt_barrier_decrement()