source: trunk/kernel/mm/vmm.c @ 654

/*
 * vmm.c - virtual memory manager related operations definition.
 *
 * Authors   Ghassan Almaless (2008,2009,2010,2011, 2012)
 *           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 <kernel_config.h>
#include <hal_kernel_types.h>
#include <hal_special.h>
#include <hal_gpt.h>
#include <hal_vmm.h>
#include <hal_irqmask.h>
#include <hal_macros.h>
#include <printk.h>
#include <memcpy.h>
#include <remote_queuelock.h>
#include <list.h>
#include <xlist.h>
#include <bits.h>
#include <process.h>
#include <thread.h>
#include <vseg.h>
#include <cluster.h>
#include <scheduler.h>
#include <vfs.h>
#include <mapper.h>
#include <page.h>
#include <kmem.h>
#include <vmm.h>
#include <hal_exception.h>

////////////////////////////////////////////////////////////////////////////////////////////
//   Extern global variables
////////////////////////////////////////////////////////////////////////////////////////////

extern  process_t  process_zero;      // allocated in cluster.c 

////////////////////////////////////////////////////////////////////////////////////////////
// This static function is called by the vmm_user_init() function.
// It initialises the free lists of vsegs used by the VMM MMAP allocator.
// It makes the assumption that HEAP_BASE == 1 Gbytes and HEAP_SIZE == 2 Gbytes.
////////////////////////////////////////////////////////////////////////////////////////////
static void vmm_stack_init( vmm_t * vmm )
{

// check STACK zone
assert( ((CONFIG_VMM_STACK_SIZE * CONFIG_THREADS_MAX_PER_CLUSTER) <=
(CONFIG_VMM_VSPACE_SIZE - CONFIG_VMM_STACK_BASE)) , "STACK zone too small\n");

    // get pointer on STACK allocator 
    stack_mgr_t * mgr = &vmm->stack_mgr;

    mgr->bitmap   = 0;
    mgr->vpn_base = CONFIG_VMM_STACK_BASE;
    busylock_init( &mgr->lock , LOCK_VMM_STACK );

}

////////////////////////////////////////////////////////////////////////////////////////////
// This static function is called by the vmm_create_vseg() function, and implements
// the VMM STACK specific allocator. Depending on the local thread index <ltid>,
// it ckeks availability of the corresponding slot in the process STACKS region,
// allocates a vseg descriptor, and initializes the "vpn_base" and "vpn_size" fields.
////////////////////////////////////////////////////////////////////////////////////////////
// @ vmm      : [in]  pointer on VMM.
// @ ltid     : [in]  requested slot == local user thread identifier.
////////////////////////////////////////////////////////////////////////////////////////////
static vseg_t * vmm_stack_alloc( vmm_t  * vmm,
                                 ltid_t   ltid )
{

// check ltid argument
assert( (ltid <= ((CONFIG_VMM_VSPACE_SIZE - CONFIG_VMM_STACK_BASE) / CONFIG_VMM_STACK_SIZE)),
"slot index %d too large for an user stack vseg", ltid );

    // get stack allocator pointer
    stack_mgr_t * mgr = &vmm->stack_mgr;

    // get lock protecting stack allocator
    busylock_acquire( &mgr->lock );

// check requested slot is available
assert( (bitmap_state( &mgr->bitmap , ltid ) == false),
"slot index %d already allocated", ltid );

    // allocate a vseg descriptor
    vseg_t * vseg = vseg_alloc();

    if( vseg == NULL )
        {
        // release lock protecting free lists
        busylock_release( &mgr->lock );

        printk("\n[ERROR] %s cannot allocate memory for vseg in cluster %x\n",
        __FUNCTION__ , local_cxy );

        return NULL;
    }

    // update bitmap
    bitmap_set( &mgr->bitmap , ltid );

    // release lock on stack allocator
    busylock_release( &mgr->lock );

    // set "vpn_base" & "vpn_size" fields (first page non allocated)
    vseg->vpn_base = mgr->vpn_base + (ltid * CONFIG_VMM_STACK_SIZE) + 1;
    vseg->vpn_size = CONFIG_VMM_STACK_SIZE - 1;

    return vseg;

} // end vmm_stack_alloc()

////////////////////////////////////////////////////////////////////////////////////////////
// This static function is called by the vmm_remove_vseg() function, and implements
// the VMM STACK specific desallocator.
// It updates the bitmap to release the corresponding slot in the process STACKS region,
// and releases memory allocated to vseg descriptor.
////////////////////////////////////////////////////////////////////////////////////////////
// @ vmm      : [in] pointer on VMM.
// @ vseg     : [in] pointer on released vseg.
////////////////////////////////////////////////////////////////////////////////////////////
static void vmm_stack_free( vmm_t  * vmm,
                            vseg_t * vseg )
{
    // get stack allocator pointer
    stack_mgr_t * mgr = &vmm->stack_mgr;

    // compute slot index
    uint32_t index = (vseg->vpn_base - 1 - mgr->vpn_base) / CONFIG_VMM_STACK_SIZE;

// check index
assert( (index <= ((CONFIG_VMM_VSPACE_SIZE - CONFIG_VMM_STACK_BASE) / CONFIG_VMM_STACK_SIZE)),
"slot index %d too large for an user stack vseg", index );

// check released slot is allocated
assert( (bitmap_state( &mgr->bitmap , index ) == true),
"released slot index %d non allocated", index );

    // get lock on stack allocator
    busylock_acquire( &mgr->lock );

    // update stacks_bitmap
    bitmap_clear( &mgr->bitmap , index );

    // release lock on stack allocator
    busylock_release( &mgr->lock );

    // release memory allocated to vseg descriptor
    vseg_free( vseg );

}  // end vmm_stack_free()



////////////////////////////////////////////////////////////////////////////////////////////
// This function display the current state of the VMM MMAP allocator of a process VMM
// identified by the <vmm> argument.
////////////////////////////////////////////////////////////////////////////////////////////
void vmm_mmap_display( vmm_t * vmm )
{
    uint32_t  order;
    xptr_t    root_xp;
    xptr_t    iter_xp;

    // get pointer on process
    process_t * process = (process_t *)(((char*)vmm) - OFFSETOF( process_t , vmm ));

    // get process PID
    pid_t pid = process->pid;

    // get pointer on VMM MMAP allocator
    mmap_mgr_t * mgr = &vmm->mmap_mgr;

    // display header
    printk("***** VMM MMAP allocator / process %x *****\n", pid );

    // scan the array of free lists of vsegs
    for( order = 0 ; order <= CONFIG_VMM_HEAP_MAX_ORDER ; order++ )
    {
        root_xp = XPTR( local_cxy , &mgr->free_list_root[order] );

        if( !xlist_is_empty( root_xp ) )
        {
            printk(" - %d (%x pages) : ", order , 1<<order );

            XLIST_FOREACH( root_xp , iter_xp )
            {
                xptr_t   vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist );
                vseg_t * vseg    = GET_PTR( vseg_xp );

                printk("%x | ", vseg->vpn_base );
            }

            printk("\n");
        }
    }
}  // end vmm_mmap_display()

////////////////////////////////////////////////////////////////////////////////////////////
// This static function is called by the vmm_user_init() function.
// It initialises the free lists of vsegs used by the VMM MMAP allocator.
// TODO this function is only valid for 32 bits cores, and makes three assumptions:
// HEAP_BASE == 1 Gbytes / HEAP_SIZE == 2 Gbytes / MMAP_MAX_SIZE == 1 Gbytes
////////////////////////////////////////////////////////////////////////////////////////////
void vmm_mmap_init( vmm_t * vmm )
{

// check HEAP base and size 
assert( (CONFIG_VMM_HEAP_BASE == 0x40000) & (CONFIG_VMM_STACK_BASE == 0xc0000),
"CONFIG_VMM_HEAP_BASE != 0x40000 or CONFIG_VMM_STACK_BASE != 0xc0000" );

// check  MMAP vseg max order 
assert( (CONFIG_VMM_HEAP_MAX_ORDER == 18), "max mmap vseg size is 256K pages" );

    // get pointer on MMAP allocator 
    mmap_mgr_t * mgr = &vmm->mmap_mgr;

    // initialize HEAP base and size
    mgr->vpn_base        = CONFIG_VMM_HEAP_BASE;
    mgr->vpn_size        = CONFIG_VMM_STACK_BASE - CONFIG_VMM_HEAP_BASE;

    // initialize lock
    busylock_init( &mgr->lock , LOCK_VMM_MMAP );

    // initialize free lists
    uint32_t   i;
    for( i = 0 ; i <= CONFIG_VMM_HEAP_MAX_ORDER ; i++ )
    {
        xlist_root_init( XPTR( local_cxy , &mgr->free_list_root[i] ) );
    }

    // allocate and register first 1 Gbytes vseg
    vseg_t * vseg0 = vseg_alloc();

assert( (vseg0 != NULL) , "cannot allocate vseg" );

    vseg0->vpn_base = CONFIG_VMM_HEAP_BASE;
    vseg0->vpn_size = CONFIG_VMM_HEAP_BASE;

    xlist_add_first( XPTR( local_cxy , &mgr->free_list_root[CONFIG_VMM_HEAP_MAX_ORDER] ),
                     XPTR( local_cxy , &vseg0->xlist ) );

    // allocate and register second 1 Gbytes vseg
    vseg_t * vseg1 = vseg_alloc();

assert( (vseg1 != NULL) , "cannot allocate vseg" );

    vseg1->vpn_base = CONFIG_VMM_HEAP_BASE << 1;
    vseg1->vpn_size = CONFIG_VMM_HEAP_BASE;

    xlist_add_first( XPTR( local_cxy , &mgr->free_list_root[CONFIG_VMM_HEAP_MAX_ORDER] ),
                     XPTR( local_cxy , &vseg1->xlist ) );

#if DEBUG_VMM_MMAP
thread_t * this = CURRENT_THREAD;
uint32_t cycle = (uint32_t)hal_get_cycles();
printk("\n[%s] thread[%x,%x] / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, cycle );
vmm_mmap_display( vmm );
#endif

}  // end vmm_mmap_init()

////////////////////////////////////////////////////////////////////////////////////////////
// This static function is called by the vmm_create_vseg() function, and implements
// the VMM MMAP specific allocator.  Depending on the requested number of pages <npages>,
// it get a free vseg from the relevant free_list, and initializes the "vpn_base" and
// "vpn_size" fields.
////////////////////////////////////////////////////////////////////////////////////////////
// @ vmm      : [in] pointer on VMM.
// @ npages   : [in] requested number of pages.
// @ returns local pointer on vseg if success / returns NULL if failure.
////////////////////////////////////////////////////////////////////////////////////////////
static vseg_t * vmm_mmap_alloc( vmm_t * vmm,
                                vpn_t   npages )
{

#if DEBUG_VMM_MMAP
thread_t * this = CURRENT_THREAD;
uint32_t cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_MMAP < cycle )
printk("\n[%s] thread[%x,%x] for %x pages / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, npages, cycle );
#endif

    // number of allocated pages must be power of 2
    // compute actual size and order 
    vpn_t    required_vpn_size = POW2_ROUNDUP( npages );
    uint32_t required_order    = bits_log2( required_vpn_size );

    // get mmap allocator pointer
    mmap_mgr_t * mgr = &vmm->mmap_mgr;

    // take lock protecting free lists in MMAP allocator
    busylock_acquire( &mgr->lock );

    // initialises the while loop variables
    uint32_t   current_order = required_order;
    vseg_t   * current_vseg  = NULL;

    // search a free vseg equal or larger than requested size
        while( current_order <= CONFIG_VMM_HEAP_MAX_ORDER )
        {
        // build extended pointer on free_pages_root[current_order]
        xptr_t root_xp = XPTR( local_cxy , &mgr->free_list_root[current_order] );

                if( !xlist_is_empty( root_xp ) )
                {
            // get extended pointer on first vseg in this free_list
                        xptr_t current_vseg_xp = XLIST_FIRST( root_xp , vseg_t , xlist );
            current_vseg = GET_PTR( current_vseg_xp );

            // build extended pointer on xlist field in vseg descriptor
            xptr_t list_entry_xp = XPTR( local_cxy , &current_vseg->xlist );

            // remove this vseg from the free_list
                        xlist_unlink( list_entry_xp );

                        break;  
                }

        // increment loop index
        current_order++;

    }  // end while loop 

    if( current_vseg == NULL )  // return failure 
    {
        // release lock protecting free lists
        busylock_release( &mgr->lock );

        printk("\n[ERROR] %s cannot allocate ) %d page(s) in cluster %x\n",
        __FUNCTION__, npages , local_cxy );

        return NULL;
    }

        // split recursively the found vseg in smaller vsegs
    // if required, and update the free-lists accordingly
        while( current_order > required_order )
        {
        // get found vseg base and size
        vpn_t  vpn_base = current_vseg->vpn_base;
        vpn_t  vpn_size = current_vseg->vpn_size;
        
        // allocate a new vseg for the upper half of current vseg
            vseg_t * new_vseg = vseg_alloc();

            if( new_vseg == NULL )
        {
                // release lock protecting free lists
            busylock_release( &mgr->lock );

            printk("\n[ERROR] %s cannot allocate memory for vseg in cluster %x\n",
            __FUNCTION__ , local_cxy );

            return NULL;
            }

        // initialise new vseg (upper half of found vseg)
        new_vseg->vmm      = vmm;
        new_vseg->vpn_base = vpn_base + (vpn_size >> 1);
        new_vseg->vpn_size = vpn_size >> 1;

        // insert new vseg in relevant free_list
                xlist_add_first( XPTR( local_cxy , &mgr->free_list_root[current_order-1] ),
                         XPTR( local_cxy , &new_vseg->xlist ) );

        // update found vseg
        current_vseg->vpn_size = vpn_size>>1; 

        // update order
                current_order --;
        }

        // release lock protecting free lists
        busylock_release( &mgr->lock );

#if DEBUG_VMM_MMAP
vmm_mmap_display( vmm );
#endif

    return current_vseg;

}  // end vmm_mmap_alloc()

////////////////////////////////////////////////////////////////////////////////////////////
// This static function implements the VMM MMAP specific desallocator.
// It is called by the vmm_remove_vseg() function.
// It releases the vseg to the relevant free_list, after trying (recursively) to
// merge it to the buddy vseg.
////////////////////////////////////////////////////////////////////////////////////////////
// @ vmm      : [in] pointer on VMM.
// @ vseg     : [in] pointer on released vseg.
////////////////////////////////////////////////////////////////////////////////////////////
static void vmm_mmap_free( vmm_t  * vmm,
                           vseg_t * vseg )
{

#if DEBUG_VMM_MMAP
thread_t * this = CURRENT_THREAD;
uint32_t cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_MMAP < cycle )
printk("\n[%s] thread[%x,%x] for vpn_base %x / vpn_size %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, vseg->vpn_base, vseg->vpn_size, cycle );
#endif

    vseg_t * buddy_vseg;

    // get mmap allocator pointer
    mmap_mgr_t * mgr = &vmm->mmap_mgr;

    // take lock protecting free lists
    busylock_acquire( &mgr->lock );

    // initialise loop variables 
    // released_vseg is the currently released vseg
    vseg_t * released_vseg     = vseg;
    uint32_t released_order    = bits_log2( vseg->vpn_size );

        // iteratively merge the released vseg to the buddy vseg
        // release the current page and exit when buddy not found
    while( released_order <= CONFIG_VMM_HEAP_MAX_ORDER )
    {
        // compute buddy_vseg vpn_base 
                vpn_t buddy_vpn_base = released_vseg->vpn_base ^ (1 << released_order);
        
        // build extended pointer on free_pages_root[current_order]
        xptr_t root_xp = XPTR( local_cxy , &mgr->free_list_root[released_order] );

        // scan this free list to find the buddy vseg
        xptr_t   iter_xp;
        buddy_vseg = NULL;
        XLIST_FOREACH( root_xp , iter_xp )
        {
            xptr_t   current_vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist );
            vseg_t * current_vseg    = GET_PTR( current_vseg_xp );

            if( current_vseg->vpn_base == buddy_vpn_base )
            {
                buddy_vseg = current_vseg;
                break;
            }
        }
        
        if( buddy_vseg != NULL )     // buddy found => merge released & buddy
        {
            // update released vseg fields
            released_vseg->vpn_size = buddy_vseg->vpn_size<<1;
            if( released_vseg->vpn_base > buddy_vseg->vpn_base) 
                released_vseg->vpn_base = buddy_vseg->vpn_base;

            // remove buddy vseg from free_list
            xlist_unlink( XPTR( local_cxy , &buddy_vseg->xlist ) );

            // release memory allocated to buddy descriptor
            vseg_free( buddy_vseg );
        }
        else                         // buddy not found => register & exit
        {
            // register released vseg in free list
            xlist_add_first( root_xp , XPTR( local_cxy , &released_vseg->xlist ) );

            // exit while loop
            break;
        }

        // increment released_order
        released_order++;
    }

    // release lock
    busylock_release( &mgr->lock );

#if DEBUG_VMM_MMAP
vmm_mmap_display( vmm );
#endif

}  // end vmm_mmap_free()

////////////////////////////////////////////////////////////////////////////////////////////
// This static function registers one vseg in the VSL of a local process descriptor.
////////////////////////////////////////////////////////////////////////////////////////////
// vmm       : [in] pointer on VMM.
// vseg      : [in] pointer on vseg.
////////////////////////////////////////////////////////////////////////////////////////////
void vmm_attach_vseg_to_vsl( vmm_t  * vmm,
                             vseg_t * vseg )
{
    // update vseg descriptor
    vseg->vmm = vmm;

    // increment vsegs number
    vmm->vsegs_nr++;

    // add vseg in vmm list
    xlist_add_last( XPTR( local_cxy , &vmm->vsegs_root ),
                    XPTR( local_cxy , &vseg->xlist ) );

}  // end vmm_attach_vseg_from_vsl()

////////////////////////////////////////////////////////////////////////////////////////////
// This static function removes one vseg from the VSL of a local process descriptor.
////////////////////////////////////////////////////////////////////////////////////////////
// vmm       : [in] pointer on VMM.
// vseg      : [in] pointer on vseg.
////////////////////////////////////////////////////////////////////////////////////////////
void vmm_detach_vseg_from_vsl( vmm_t  * vmm,
                               vseg_t * vseg )
{
    // update vseg descriptor
    vseg->vmm = NULL;

    // decrement vsegs number
    vmm->vsegs_nr--;

    // remove vseg from VSL
    xlist_unlink( XPTR( local_cxy , &vseg->xlist ) );

}  // end vmm_detach_from_vsl()

////////////////////////////////////////////
error_t vmm_user_init( process_t * process )
{

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

    // get pointer on VMM
    vmm_t   * vmm = &process->vmm;

// check UTILS zone 
assert( ((CONFIG_VMM_ARGS_SIZE + CONFIG_VMM_ENVS_SIZE) <= 
         (CONFIG_VMM_ELF_BASE - CONFIG_VMM_UTILS_BASE)) ,
         "UTILS zone too small\n" );

    // initialize lock protecting the VSL 
        remote_queuelock_init( XPTR( local_cxy , &vmm->vsl_lock ) , LOCK_VMM_VSL );


    // initialize STACK allocator
    vmm_stack_init( vmm );

    // initialize MMAP allocator
    vmm_mmap_init( vmm );

    // initialize instrumentation counters
        vmm->false_pgfault_nr    = 0;
        vmm->local_pgfault_nr    = 0;
        vmm->global_pgfault_nr   = 0;
        vmm->false_pgfault_cost  = 0;
        vmm->local_pgfault_cost  = 0;
        vmm->global_pgfault_cost = 0;

/*
    // register "args" vseg in VSL
    base = CONFIG_VMM_UTILS_BASE << CONFIG_PPM_PAGE_SHIFT;
    size = CONFIG_VMM_ARGS_SIZE << CONFIG_PPM_PAGE_SHIFT;

    vseg_args = vmm_create_vseg( process,
                                 VSEG_TYPE_DATA,
                                 base,
                                 size,
                                 0,             // file_offset unused
                                 0,             // file_size unused
                                 XPTR_NULL,     // mapper_xp unused
                                 local_cxy );
    if( vseg_args == NULL )
    {
        printk("\n[ERROR] in %s : cannot register args vseg\n", __FUNCTION__ );
        return -1;
    }

    vmm->args_vpn_base = base;

    // register "envs" vseg in VSL
    base = (CONFIG_VMM_UTILS_BASE + CONFIG_VMM_ARGS_SIZE) << CONFIG_PPM_PAGE_SHIFT;
    size = CONFIG_VMM_ENVS_SIZE << CONFIG_PPM_PAGE_SHIFT;

    vseg_envs = vmm_create_vseg( process,
                                 VSEG_TYPE_DATA,
                                 base,
                                 size,
                                 0,             // file_offset unused
                                 0,             // file_size unused
                                 XPTR_NULL,     // mapper_xp unused
                                 local_cxy );
    if( vseg_envs == NULL )
    {
        printk("\n[ERROR] in %s : cannot register envs vseg\n", __FUNCTION__ );
        return -1;
    }

    vmm->envs_vpn_base = base;
*/
    hal_fence();

#if DEBUG_VMM_USER_INIT
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_USER_INIT )
printk("\n[%s] thread[%x,%x] exit for process %x in cluster %x / cycle %d\n", 
__FUNCTION__, this->process->pid, this->trdid, process->pid, local_cxy, cycle );
#endif

    return 0;

}  // end vmm_user_init()

//////////////////////////////////////////
void vmm_user_reset( process_t * process )
{
    xptr_t       vseg_xp;
        vseg_t     * vseg;
    vseg_type_t  vseg_type;

#if DEBUG_VMM_USER_RESET
uint32_t   cycle;
thread_t * this = CURRENT_THREAD;
#endif

#if (DEBUG_VMM_USER_RESET & 1 )
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_USER_RESET < cycle )
printk("\n[%s] thread[%x,%x] enter for process %x in cluster %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, local_cxy, cycle );
#endif

#if (DEBUG_VMM_USER_RESET & 1 )
if( DEBUG_VMM_USER_RESET < cycle )
hal_vmm_display( XPTR( local_cxy , process ) , true );
#endif

    // get pointer on local VMM
    vmm_t * vmm = &process->vmm;

    // build extended pointer on VSL root and VSL lock
    xptr_t   root_xp = XPTR( local_cxy , &vmm->vsegs_root );
    xptr_t   lock_xp = XPTR( local_cxy , &vmm->vsl_lock );

    // take the VSL lock
        remote_queuelock_acquire( lock_xp );

    // scan the VSL to delete all non kernel vsegs
    // (we don't use a FOREACH in case of item deletion)
    xptr_t   iter_xp;
    xptr_t   next_xp;
        for( iter_xp = hal_remote_l64( root_xp ) ; 
         iter_xp != root_xp ;
         iter_xp = next_xp )
        {
        // save extended pointer on next item in xlist
        next_xp = hal_remote_l64( iter_xp );

        // get pointers on current vseg in VSL
        vseg_xp   = XLIST_ELEMENT( iter_xp , vseg_t , xlist );
        vseg      = GET_PTR( vseg_xp );
        vseg_type = vseg->type;

#if( DEBUG_VMM_USER_RESET & 1 )
if( DEBUG_VMM_USER_RESET < cycle )
printk("\n[%s] found %s vseg / vpn_base %x / vpn_size %d\n",
__FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size );
#endif
        // delete non kernel vseg  
        if( (vseg_type != VSEG_TYPE_KCODE) && 
            (vseg_type != VSEG_TYPE_KDATA) && 
            (vseg_type != VSEG_TYPE_KDEV ) )
        {
            // remove vseg from VSL
            vmm_remove_vseg( process , vseg );

#if( DEBUG_VMM_USER_RESET & 1 )
if( DEBUG_VMM_USER_RESET < cycle )
printk("\n[%s] %s vseg deleted / vpn_base %x / vpn_size %d\n",
__FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size );
#endif
        }
        else
        {

#if( DEBUG_VMM_USER_RESET & 1 )
if( DEBUG_VMM_USER_RESET < cycle )
printk("\n[%s] keep %s vseg / vpn_base %x / vpn_size %d\n",
__FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size );
#endif
        }
        }  // end loop on vsegs in VSL

    // release the VSL lock
        remote_queuelock_release( lock_xp );

// FIXME il faut gérer les process copies...

#if DEBUG_VMM_USER_RESET
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_USER_RESET < cycle )
printk("\n[%s] thread[%x,%x] exit for process %x in cluster %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, local_cxy , cycle );
#endif

#if (DEBUG_VMM_USER_RESET & 1 )
if( DEBUG_VMM_USER_RESET < cycle )
hal_vmm_display( XPTR( local_cxy , process ) , true );
#endif

}  // end vmm_user_reset()

/////////////////////////////////////////////////
void vmm_global_delete_vseg( process_t * process,
                             intptr_t    base )
{
    cxy_t           owner_cxy;
    lpid_t          owner_lpid;
    reg_t           save_sr;

    xptr_t          process_lock_xp;
    xptr_t          process_root_xp;
    xptr_t          process_iter_xp;

    xptr_t          remote_process_xp;
    cxy_t           remote_process_cxy;
    process_t     * remote_process_ptr;

    xptr_t          vsl_root_xp;
    xptr_t          vsl_lock_xp;
    xptr_t          vsl_iter_xp;

    rpc_desc_t      rpc;                  // shared rpc descriptor for parallel RPCs
    uint32_t        responses;            // RPC responses counter

    thread_t      * this    = CURRENT_THREAD;
    pid_t           pid     = process->pid;
    cluster_t     * cluster = LOCAL_CLUSTER;

#if DEBUG_VMM_GLOBAL_DELETE_VSEG
uint32_t cycle = (uint32_t)hal_get_cycles();
#endif

#if (DEBUG_VMM_GLOBAL_DELETE_VSEG & 1)
if( DEBUG_VMM_GLOBAL_DELETE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] enters / process %x / base %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, base, cycle );
#endif

    // initialize a shared RPC descriptor
    rpc.rsp       = &responses;
    rpc.blocking  = false;                  // non blocking behaviour for rpc_send()
    rpc.index     = RPC_VMM_REMOVE_VSEG;
    rpc.thread    = this;
    rpc.lid       = this->core->lid;
    rpc.args[0]   = this->process->pid;
    rpc.args[1]   = base;

    // get owner process cluster and local index 
    owner_cxy        = CXY_FROM_PID( pid );
    owner_lpid       = LPID_FROM_PID( pid );

    // get extended pointer on root and lock of process copies xlist in owner cluster
    process_root_xp  = XPTR( owner_cxy , &cluster->pmgr.copies_root[owner_lpid] );
    process_lock_xp  = XPTR( owner_cxy , &cluster->pmgr.copies_lock[owner_lpid] );

    // mask IRQs
    hal_disable_irq( &save_sr );

    // client thread blocks itself
    thread_block( XPTR( local_cxy , this ) , THREAD_BLOCKED_RPC );

    // take the lock protecting process copies
    remote_queuelock_acquire( process_lock_xp );

    // initialize responses counter
    responses = 0;

    // loop on process copies
    XLIST_FOREACH( process_root_xp , process_iter_xp )
    {
        // get cluster and local pointer on remote process
        remote_process_xp  = XLIST_ELEMENT( process_iter_xp , process_t , copies_list );
        remote_process_ptr = GET_PTR( remote_process_xp );
        remote_process_cxy = GET_CXY( remote_process_xp );

        // build extended pointers on remote VSL root and lock
        vsl_root_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.vsegs_root );
        vsl_lock_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.vsl_lock );

        // get lock on remote VSL
        remote_queuelock_acquire( vsl_lock_xp );

        // loop on vsegs in remote process VSL
        XLIST_FOREACH( vsl_root_xp , vsl_iter_xp )
        {
            // get pointers on current vseg 
            xptr_t   vseg_xp  = XLIST_ELEMENT( vsl_iter_xp , vseg_t , xlist );
            vseg_t * vseg_ptr = GET_PTR( vseg_xp );

            // get current vseg base address
            intptr_t vseg_base = (intptr_t)hal_remote_lpt( XPTR( remote_process_cxy,
                                                                 &vseg_ptr->min ) );

            if( vseg_base == base )   // found searched vseg
            {
                // atomically increment responses counter
                hal_atomic_add( &responses , 1 );

#if (DEBUG_VMM_GLOBAL_DELETE_VSEG & 1)
if( DEBUG_VMM_GLOBAL_DELETE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] register RPC request in cluster %x\n",
__FUNCTION__, this->process->pid, this->trdid, remote_process_cxy );
#endif
                // send RPC to remote cluster
                rpc_send( remote_process_cxy , &rpc );

                // exit loop on vsegs
                break;
            }
        }  // end of loop on vsegs

        // release lock on remote VSL
        remote_queuelock_release( vsl_lock_xp );

    }  // end of loop on process copies

    // release the lock protecting process copies
    remote_queuelock_release( process_lock_xp );

#if (DEBUG_VMM_GLOBAL_DELETE_VSEG & 1)
if( DEBUG_VMM_GLOBAL_DELETE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] deschedule / process %x / base %x\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, base );
#endif

    // client thread deschedule
    sched_yield("blocked on rpc_vmm_delete_vseg");
 
    // restore IRQs
    hal_restore_irq( save_sr );

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

}  // end vmm_global_delete_vseg()

////////////////////////////////////////////////
void vmm_global_resize_vseg( process_t * process,
                             intptr_t    base,
                             intptr_t    new_base,
                             intptr_t    new_size )
{
    cxy_t           owner_cxy;
    lpid_t          owner_lpid;
    reg_t           save_sr;

    xptr_t          process_lock_xp;
    xptr_t          process_root_xp;
    xptr_t          process_iter_xp;

    xptr_t          remote_process_xp;
    cxy_t           remote_process_cxy;
    process_t     * remote_process_ptr;

    xptr_t          vsl_root_xp;
    xptr_t          vsl_lock_xp;
    xptr_t          vsl_iter_xp;

    rpc_desc_t      rpc;                  // shared rpc descriptor for parallel RPCs
    uint32_t        responses;            // RPC responses counter

    thread_t      * this    = CURRENT_THREAD; 
    pid_t           pid     = process->pid;
    cluster_t     * cluster = LOCAL_CLUSTER;

#if DEBUG_VMM_GLOBAL_RESIZE_VSEG
uint32_t cycle = (uint32_t)hal_get_cycles();
#endif

#if (DEBUG_VMM_GLOBAL_RESIZE_VSEG & 1)
if( DEBUG_VMM_GLOBAL_RESIZE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] : process %x / base %x / new_base %x / new_size %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, base, new_base, new_size, cycle );
#endif

    // initialize a shared RPC descriptor
    rpc.rsp       = &responses;
    rpc.blocking  = false;                  // non blocking behaviour for rpc_send()
    rpc.index     = RPC_VMM_REMOVE_VSEG;
    rpc.thread    = this;
    rpc.lid       = this->core->lid;
    rpc.args[0]   = this->process->pid;
    rpc.args[1]   = base;
    rpc.args[2]   = new_base;
    rpc.args[3]   = new_size;

    // get owner process cluster and local index
    owner_cxy        = CXY_FROM_PID( pid );
    owner_lpid       = LPID_FROM_PID( pid );

    // get extended pointer on root and lock of process copies xlist in owner cluster
    process_root_xp  = XPTR( owner_cxy , &cluster->pmgr.copies_root[owner_lpid] );
    process_lock_xp  = XPTR( owner_cxy , &cluster->pmgr.copies_lock[owner_lpid] );

    // mask IRQs
    hal_disable_irq( &save_sr );

    // client thread blocks itself
    thread_block( XPTR( local_cxy , this ) , THREAD_BLOCKED_RPC );

    // take the lock protecting process copies
    remote_queuelock_acquire( process_lock_xp );

    // initialize responses counter
    responses = 0;

    // loop on process copies
    XLIST_FOREACH( process_root_xp , process_iter_xp )
    {
        // get cluster and local pointer on remote process
        remote_process_xp  = XLIST_ELEMENT( process_iter_xp , process_t , copies_list );
        remote_process_ptr = GET_PTR( remote_process_xp );
        remote_process_cxy = GET_CXY( remote_process_xp );

        // build extended pointers on remote VSL root and lock
        vsl_root_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.vsegs_root );
        vsl_lock_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.vsl_lock );

        // get lock on remote VSL
        remote_queuelock_acquire( vsl_lock_xp );

        // loop on vsegs in remote process VSL
        XLIST_FOREACH( vsl_root_xp , vsl_iter_xp )
        {
            // get pointers on current vseg 
            xptr_t   vseg_xp  = XLIST_ELEMENT( vsl_iter_xp , vseg_t , xlist );
            vseg_t * vseg_ptr = GET_PTR( vseg_xp );

            // get current vseg base address
            intptr_t vseg_base = (intptr_t)hal_remote_lpt( XPTR( remote_process_cxy,
                                                                 &vseg_ptr->min ) );

            if( vseg_base == base )   // found searched vseg
            {
                // atomically increment responses counter
                hal_atomic_add( &responses , 1 );

#if (DEBUG_VMM_GLOBAL_RESIZE_VSEG & 1)
if( DEBUG_VMM_GLOBAL_RESIZE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] register RPC request in cluster %x\n",
__FUNCTION__, this->process->pid, this->trdid, remote_process_cxy );
#endif
                // send RPC to remote cluster
                rpc_send( remote_process_cxy , & rpc );

                // exit loop on vsegs
                break;
            }

        }  // end of loop on vsegs

#if (DEBUG_VMM_GLOBAL_RESIZE_VSEG & 1)
if( DEBUG_VMM_GLOBAL_RESIZE_VSEG < cycle )
hal_vmm_display( remote_process_xp , false );
#endif

        // release lock on remote VSL
        remote_queuelock_release( vsl_lock_xp );

    }  // end of loop on process copies

    // release the lock protecting process copies
    remote_queuelock_release( process_lock_xp );

#if (DEBUG_VMM_GLOBAL_RESIZE_VSEG & 1)
if( DEBUG_VMM_GLOBAL_RESIZE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] deschedule / process %x / base %x\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, base );
#endif

    // client thread deschedule
    sched_yield("blocked on rpc_vmm_delete_vseg");

    // restore IRQs
    hal_restore_irq( save_sr );

#if DEBUG_VMM_GLOBAL_RESIZE_VSEG
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_GLOBAL_RESIZE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] exit for process %x / base %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid , base, cycle );
#endif

}  // end vmm_global_resize_vseg()

////////////////////////////////////////////////
void vmm_global_update_pte( process_t * process,
                            vpn_t       vpn,
                            uint32_t    attr,
                            ppn_t       ppn )
{
    pid_t           pid;
    cxy_t           owner_cxy;
    lpid_t          owner_lpid;

    xlist_entry_t * process_root_ptr;
    xptr_t          process_root_xp;
    xptr_t          process_iter_xp;

    xptr_t          remote_process_xp;
    cxy_t           remote_process_cxy;
    process_t     * remote_process_ptr;
    xptr_t          remote_gpt_xp;

#if DEBUG_VMM_GLOBAL_UPDATE_PTE
uint32_t cycle = (uint32_t)hal_get_cycles();
thread_t * this = CURRENT_THREAD;
#endif


#if (DEBUG_VMM_GLOBAL_UPDATE_PTE & 1)
if( DEBUG_VMM_GLOBAL_UPDATE_PTE < cycle )
printk("\n[%s] thread[%x,%x] enter for process %x / vpn %x / attr %x / ppn %x / ycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, vpn, attr, ppn, cycle );
#endif

    // get owner process cluster and local index
    pid              = process->pid;
    owner_cxy        = CXY_FROM_PID( pid );
    owner_lpid       = LPID_FROM_PID( pid );

    // get extended pointer on root of process copies xlist in owner cluster
    process_root_ptr = &LOCAL_CLUSTER->pmgr.copies_root[owner_lpid];
    process_root_xp  = XPTR( owner_cxy , process_root_ptr );

    // loop on process copies
    XLIST_FOREACH( process_root_xp , process_iter_xp )
    {
        // get cluster and local pointer on remote process
        remote_process_xp  = XLIST_ELEMENT( process_iter_xp , process_t , copies_list );
        remote_process_ptr = GET_PTR( remote_process_xp );
        remote_process_cxy = GET_CXY( remote_process_xp );

#if (DEBUG_VMM_GLOBAL_UPDATE_PTE & 1)
if( DEBUG_VMM_GLOBAL_UPDATE_PTE < cycle )
printk("\n[%s] thread[%x,%x] handling vpn %x for process %x in cluster %x\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, process->pid, remote_process_cxy );
#endif

        // get extended pointer on remote gpt
        remote_gpt_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.gpt );

        // update remote GPT
        hal_gpt_update_pte( remote_gpt_xp, vpn, attr, ppn );
    }  

#if DEBUG_VMM_GLOBAL_UPDATE_PTE
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_GLOBAL_UPDATE_PTE < cycle )
printk("\n[%s] thread[%x,%x] exit for process %x / vpn %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid , vpn , cycle );
#endif

#if (DEBUG_VMM_GLOBAL_UPDATE_PTE & 1)
hal_vmm_display( process , true );
#endif

}  // end vmm_global_update_pte()

///////////////////////////////////////
void vmm_set_cow( process_t * process )
{
    vmm_t         * vmm;

    xlist_entry_t * process_root_ptr;
    xptr_t          process_root_xp;
    xptr_t          process_iter_xp;

    xptr_t          remote_process_xp;
    cxy_t           remote_process_cxy;
    process_t     * remote_process_ptr;
    xptr_t          remote_gpt_xp;

    xptr_t          vseg_root_xp;
    xptr_t          vseg_iter_xp;

    xptr_t          vseg_xp;
    vseg_t        * vseg;

    pid_t           pid;
    cxy_t           owner_cxy;
    lpid_t          owner_lpid;

    // get target process PID
    pid = process->pid;

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

#if (DEBUG_VMM_SET_COW & 1)
if( DEBUG_VMM_SET_COW < cycle )
hal_vmm_display( process , true );
#endif

// check cluster is reference
assert( (XPTR( local_cxy , process ) == process->ref_xp),
"local cluster must be process reference cluster\n");

    // get pointer on reference VMM
    vmm = &process->vmm;

    // get extended pointer on root of process copies xlist in owner cluster
    owner_cxy        = CXY_FROM_PID( pid );
    owner_lpid       = LPID_FROM_PID( pid );
    process_root_ptr = &LOCAL_CLUSTER->pmgr.copies_root[owner_lpid];
    process_root_xp  = XPTR( owner_cxy , process_root_ptr );

    // get extended pointer on root of vsegs xlist from reference VMM
    vseg_root_xp  = XPTR( local_cxy , &vmm->vsegs_root ); 

    // loop on target process copies
    XLIST_FOREACH( process_root_xp , process_iter_xp )
    {
        // get cluster and local pointer on remote process copy
        remote_process_xp  = XLIST_ELEMENT( process_iter_xp , process_t , copies_list );
        remote_process_ptr = GET_PTR( remote_process_xp );
        remote_process_cxy = GET_CXY( remote_process_xp );

#if (DEBUG_VMM_SET_COW & 1)
if( DEBUG_VMM_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] (%x) handles process %x in cluster %x\n",
__FUNCTION__, this->process->pid, this->trdid, this, pid, remote_process_cxy );
#endif

        // get extended pointer on remote gpt
        remote_gpt_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.gpt );

        // loop on vsegs in (local) reference process VSL
        XLIST_FOREACH( vseg_root_xp , vseg_iter_xp )
        {
            // get pointer on vseg
            vseg_xp  = XLIST_ELEMENT( vseg_iter_xp , vseg_t , xlist );
            vseg     = GET_PTR( vseg_xp );

            // get vseg type, base and size
            uint32_t type     = vseg->type;
            vpn_t    vpn_base = vseg->vpn_base;
            vpn_t    vpn_size = vseg->vpn_size;

#if (DEBUG_VMM_SET_COW & 1)
if( DEBUG_VMM_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] found vseg %s / vpn_base = %x / vpn_size = %x\n",
__FUNCTION__, this->process->pid, this->trdid, vseg_type_str(type), vpn_base, vpn_size );
#endif
            // only DATA, ANON and REMOTE vsegs
            if( (type == VSEG_TYPE_DATA)  ||
                (type == VSEG_TYPE_ANON)  ||
                (type == VSEG_TYPE_REMOTE) )
            {
                vpn_t      vpn;
                uint32_t   attr;
                ppn_t      ppn;
                xptr_t     page_xp;
                cxy_t      page_cxy;
                page_t   * page_ptr;
                xptr_t     forks_xp;
                xptr_t     lock_xp;

                // update flags in remote GPT 
                hal_gpt_set_cow( remote_gpt_xp,
                                 vpn_base,
                                 vpn_size ); 

                // atomically increment pending forks counter in physical pages,
                // this is only done once, when handling the reference copy 
                if( remote_process_cxy == local_cxy )
                {

#if (DEBUG_VMM_SET_COW & 1)
if( DEBUG_VMM_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] handles vseg %s / vpn_base = %x / vpn_size = %x\n",
__FUNCTION__, this->process->pid, this->trdid, vseg_type_str(type), vpn_base, vpn_size );
#endif
                    // scan all pages in vseg 
                    for( vpn = vpn_base ; vpn < (vpn_base + vpn_size) ; vpn++ )
                    {
                        // get page attributes and PPN from reference GPT
                        hal_gpt_get_pte( remote_gpt_xp , vpn , &attr , &ppn ); 

                        // atomically update pending forks counter if page is mapped
                        if( attr & GPT_MAPPED )
                        {
                            // get pointers and cluster on page descriptor
                            page_xp  = ppm_ppn2page( ppn );
                            page_cxy = GET_CXY( page_xp );
                            page_ptr = GET_PTR( page_xp );

                            // get extended pointers on "forks" and "lock"
                            forks_xp = XPTR( page_cxy , &page_ptr->forks );
                            lock_xp  = XPTR( page_cxy , &page_ptr->lock );

                            // take lock protecting "forks" counter
                            remote_busylock_acquire( lock_xp );

                            // increment "forks"
                            hal_remote_atomic_add( forks_xp , 1 );

                            // release lock protecting "forks" counter
                            remote_busylock_release( lock_xp );
                        }
                    }   // end loop on vpn

#if (DEBUG_VMM_SET_COW & 1)
if( DEBUG_VMM_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] completes vseg %s / vpn_base = %x / vpn_size = %x\n",
__FUNCTION__, this->process->pid, this->trdid, vseg_type_str(type), vpn_base, vpn_size );
#endif
                }   // end if local
            }   // end if vseg type
        }   // end loop on vsegs
    }   // end loop on process copies
 
#if DEBUG_VMM_SET_COW
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] exit for process %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid , cycle );
#endif

}  // end vmm_set-cow()

/////////////////////////////////////////////////
error_t vmm_fork_copy( process_t * child_process,
                       xptr_t      parent_process_xp )
{
    error_t     error;
    cxy_t       parent_cxy;
    process_t * parent_process;
    vmm_t     * parent_vmm;
    xptr_t      parent_lock_xp;
    vmm_t     * child_vmm;
    xptr_t      iter_xp;
    xptr_t      parent_vseg_xp;
    vseg_t    * parent_vseg;
    vseg_t    * child_vseg;
    uint32_t    type;
    vpn_t       vpn;            
    vpn_t       vpn_base;
    vpn_t       vpn_size;
    xptr_t      parent_root_xp;
    bool_t      mapped; 
    ppn_t       ppn;

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

    // get parent process cluster and local pointer
    parent_cxy     = GET_CXY( parent_process_xp );
    parent_process = GET_PTR( parent_process_xp );

    // get local pointers on parent and child VMM
    parent_vmm = &parent_process->vmm; 
    child_vmm  = &child_process->vmm;

    // build extended pointer on parent VSL root and lock
    parent_root_xp = XPTR( parent_cxy , &parent_vmm->vsegs_root );
    parent_lock_xp = XPTR( parent_cxy , &parent_vmm->vsl_lock );

    // take the lock protecting the parent VSL 
    remote_queuelock_acquire( parent_lock_xp );

    // loop on parent VSL xlist
    XLIST_FOREACH( parent_root_xp , iter_xp )
    {
        // get pointers on current parent vseg
        parent_vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist );
        parent_vseg    = GET_PTR( parent_vseg_xp );

        // get vseg type
        type = hal_remote_l32( XPTR( parent_cxy , &parent_vseg->type ) );
        
#if DEBUG_VMM_FORK_COPY
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_FORK_COPY < cycle )
printk("\n[%s] thread[%x,%x] found parent vseg %s / vpn_base = %x / cycle %d\n",
__FUNCTION__ , this->process->pid, this->trdid, vseg_type_str(type),
hal_remote_l32( XPTR( parent_cxy , &parent_vseg->vpn_base ) ) , cycle );
#endif

        // all parent vsegs - but STACK and kernel vsegs - must be copied in child VSL
        if( (type != VSEG_TYPE_STACK) && (type != VSEG_TYPE_KCODE) &&
            (type != VSEG_TYPE_KDATA) && (type != VSEG_TYPE_KDEV) )
        {
            // allocate memory for a new child vseg
            child_vseg = vseg_alloc();
            if( child_vseg == NULL )   // release all allocated vsegs
            {
                vmm_destroy( child_process );
                printk("\n[ERROR] in %s : cannot create vseg for child\n", __FUNCTION__ );
                return -1;
            }

            // copy parent vseg to child vseg
            vseg_init_from_ref( child_vseg , parent_vseg_xp );

            // build extended pointer on child VSL lock
            xptr_t child_lock_xp = XPTR( local_cxy , &child_vmm->vsl_lock );
 
            // take the child VSL lock 
            remote_queuelock_acquire( child_lock_xp );

            // register child vseg in child VSL
            vmm_attach_vseg_to_vsl( child_vmm , child_vseg );

            // release the child VSL lock
            remote_queuelock_release( child_lock_xp );

#if DEBUG_VMM_FORK_COPY
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_FORK_COPY < cycle )
printk("\n[%s] thread[%x,%x] copied vseg %s / vpn_base = %x to child VSL / cycle %d\n",
__FUNCTION__ , this->process->pid, this->trdid, vseg_type_str(type),
hal_remote_l32( XPTR( parent_cxy , &parent_vseg->vpn_base ) ) , cycle );
#endif
            // copy DATA, ANON, REMOTE, FILE parent GPT entries to child GPT
            if( type != VSEG_TYPE_CODE )
            {
                // activate the COW for DATA, ANON, REMOTE vsegs only
                // cow = ( type != VSEG_TYPE_FILE );

                vpn_base = child_vseg->vpn_base;
                vpn_size = child_vseg->vpn_size;

                // scan pages in parent vseg
                for( vpn = vpn_base ; vpn < (vpn_base + vpn_size) ; vpn++ )
                {
                    error = hal_gpt_pte_copy( &child_vmm->gpt,
                                              vpn,
                                              XPTR( parent_cxy , &parent_vmm->gpt ),
                                              vpn,
                                              false,      // does not handle COW flag
                                              &ppn,       // unused
                                              &mapped );  // unused
                    if( error )
                    {
                        vmm_destroy( child_process );
                        printk("\n[ERROR] in %s : cannot copy GPT\n", __FUNCTION__ );
                        return -1;
                    }

#if DEBUG_VMM_FORK_COPY
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_FORK_COPY < cycle )
printk("\n[%s] thread[%x,%x] copied vpn %x to child GPT / cycle %d\n",
__FUNCTION__ , this->process->pid, this->trdid , vpn , cycle );
#endif
                }
            }   // end if no code & no stack
        }   // end if no stack
    }   // end loop on vsegs

    // release the parent VSL lock in read mode
    remote_queuelock_release( parent_lock_xp );

/* deprecated [AG] : this is already done by the vmm_user_init() funcfion

    // initialize the child VMM STACK allocator
    vmm_stack_init( child_vmm );

    // initialize the child VMM MMAP allocator
    vmm_mmap_init( child_vmm );

    // initialize instrumentation counters
        child_vmm->false_pgfault_nr    = 0;
        child_vmm->local_pgfault_nr    = 0;
        child_vmm->global_pgfault_nr   = 0;
        child_vmm->false_pgfault_cost  = 0;
        child_vmm->local_pgfault_cost  = 0;
        child_vmm->global_pgfault_cost = 0;
*/
    // copy base addresses from parent VMM to child VMM
    child_vmm->args_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->args_vpn_base));
    child_vmm->envs_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->envs_vpn_base));
    child_vmm->heap_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->heap_vpn_base));
    child_vmm->code_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->code_vpn_base));
    child_vmm->data_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->data_vpn_base));

    child_vmm->entry_point = (intptr_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->entry_point));

    hal_fence();

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

    return 0;

}  // vmm_fork_copy()

///////////////////////////////////////
void vmm_destroy( process_t * process )
{
    xptr_t   vseg_xp;
        vseg_t * vseg;

#if DEBUG_VMM_DESTROY
uint32_t   cycle = (uint32_t)hal_get_cycles();
thread_t * this  = CURRENT_THREAD;
if( DEBUG_VMM_DESTROY < cycle )
printk("\n[%s] thread[%x,%x] enter for process %x in cluster %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, local_cxy, cycle );
#endif

#if (DEBUG_VMM_DESTROY & 1 )
if( DEBUG_VMM_DESTROY < cycle )
hal_vmm_display( XPTR( local_cxy, process ) , true );
#endif

    // get pointer on local VMM
    vmm_t  * vmm = &process->vmm;

    // build extended pointer on VSL root, VSL lock and GPT lock
    xptr_t   vsl_root_xp = XPTR( local_cxy , &vmm->vsegs_root );
    xptr_t   vsl_lock_xp = XPTR( local_cxy , &vmm->vsl_lock );

    // take the VSL lock
    remote_queuelock_acquire( vsl_lock_xp );

    // scan the VSL to delete all registered vsegs
    // (we don't use a FOREACH in case of item deletion)
    xptr_t  iter_xp;
    xptr_t  next_xp;
        for( iter_xp = hal_remote_l64( vsl_root_xp ) ; 
         iter_xp != vsl_root_xp ;
         iter_xp = next_xp )
        {
        // save extended pointer on next item in xlist
        next_xp = hal_remote_l64( iter_xp );

        // get pointers on current vseg in VSL
        vseg_xp   = XLIST_ELEMENT( iter_xp , vseg_t , xlist );
        vseg      = GET_PTR( vseg_xp );

        // delete vseg and release physical pages 
        vmm_remove_vseg( process , vseg );

#if( DEBUG_VMM_DESTROY & 1 )
if( DEBUG_VMM_DESTROY < cycle )
printk("\n[%s] %s vseg deleted / vpn_base %x / vpn_size %d\n",
__FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size );
#endif

        }

    // release the VSL lock
    remote_queuelock_release( vsl_lock_xp );

    // remove all registered MMAP vsegs from free_lists in MMAP allocator
    uint32_t i;
    for( i = 0 ; i <= CONFIG_VMM_HEAP_MAX_ORDER ; i++ )
    {
        // build extended pointer on free list root
        xptr_t root_xp = XPTR( local_cxy , &vmm->mmap_mgr.free_list_root[i] );
  
        // scan zombi_list[i]
            while( !xlist_is_empty( root_xp ) )
            {
                    vseg_xp = XLIST_FIRST( root_xp , vseg_t , xlist );
            vseg    = GET_PTR( vseg_xp );

#if( DEBUG_VMM_DESTROY & 1 )
if( DEBUG_VMM_DESTROY < cycle )
printk("\n[%s] found zombi vseg / vpn_base %x / vpn_size %d\n",
__FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size );
#endif
            // clean vseg descriptor
            vseg->vmm = NULL;

            // remove vseg from  zombi_list
            xlist_unlink( XPTR( local_cxy , &vseg->xlist ) );

                    // release vseg descriptor
            vseg_free( vseg );

#if( DEBUG_VMM_DESTROY & 1 )
if( DEBUG_VMM_DESTROY < cycle )
printk("\n[%s] zombi vseg released / vpn_base %x / vpn_size %d\n",
__FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size );
#endif
            }
    }

    // release memory allocated to the GPT itself
    hal_gpt_destroy( &vmm->gpt );

#if DEBUG_VMM_DESTROY
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_DESTROY < cycle )
printk("\n[%s] thread[%x,%x] exit for process %x in cluster %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, process->pid, local_cxy , cycle );
#endif

}  // end vmm_destroy()

/////////////////////////////////////////////////
vseg_t * vmm_check_conflict( process_t * process,
                             vpn_t       vpn_base,
                             vpn_t       vpn_size )
{
    vmm_t        * vmm = &process->vmm;

    // scan the VSL
        vseg_t       * vseg;
    xptr_t         iter_xp;
    xptr_t         vseg_xp;
    xptr_t         root_xp = XPTR( local_cxy , &vmm->vsegs_root );

        XLIST_FOREACH( root_xp , iter_xp )
        {
                vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist );
        vseg    = GET_PTR( vseg_xp );

                if( ((vpn_base + vpn_size) > vseg->vpn_base) &&
             (vpn_base < (vseg->vpn_base + vseg->vpn_size)) ) return vseg;
        }
    return NULL;

}  // end vmm_check_conflict()

////////////////////////////////////////////////
vseg_t * vmm_create_vseg( process_t   * process,
                              vseg_type_t   type,
                          intptr_t      base,         // ltid for VSEG_TYPE_STACK
                              uint32_t      size,
                          uint32_t      file_offset,
                          uint32_t      file_size,
                          xptr_t        mapper_xp,
                          cxy_t         cxy )
{
    vseg_t     * vseg;          // pointer on allocated vseg descriptor 

#if DEBUG_VMM_CREATE_VSEG
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle;
#endif

#if (DEBUG_VMM_CREATE_VSEG & 1)
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_CREATE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] enter / process %x / %s / base %x / cxy %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid,
process->pid, vseg_type_str(type), base, cxy, cycle );
#endif

    // get pointer on VMM 
        vmm_t * vmm    = &process->vmm;

    // allocate a vseg descriptor and initialize it, depending on type
    // we use specific allocators for "stack" and "mmap" types

    /////////////////////////////
    if( type == VSEG_TYPE_STACK )
    {
        // get vseg from STACK allocator
        vseg = vmm_stack_alloc( vmm , base );    // base == ltid
       
        if( vseg == NULL )
        {
            printk("\n[ERROR] %s cannot create %s vseg for process %x in cluster %x\n",
            __FUNCTION__ , vseg_type_str( type ) , process->pid , local_cxy );
            return NULL;
        }

        // initialize vseg
        vseg->type = type;
        vseg->vmm  = vmm;
        vseg->min  = vseg->vpn_base << CONFIG_PPM_PAGE_SHIFT;
        vseg->max  = vseg->min + (vseg->vpn_size << CONFIG_PPM_PAGE_SHIFT);
        vseg->cxy  = cxy;

        vseg_init_flags( vseg , type );
    }
    /////////////////////////////////
    else if( type == VSEG_TYPE_FILE )
    {
        // compute page index (in mapper) for first and last byte
        vpn_t    vpn_min    = file_offset >> CONFIG_PPM_PAGE_SHIFT;
        vpn_t    vpn_max    = (file_offset + size - 1) >> CONFIG_PPM_PAGE_SHIFT;

        // compute offset in first page and number of pages
        uint32_t offset = file_offset & CONFIG_PPM_PAGE_MASK;
        vpn_t    npages      = vpn_max - vpn_min + 1;

        // get vseg from MMAP allocator
        vseg = vmm_mmap_alloc( vmm , npages );

        if( vseg == NULL )
        {
            printk("\n[ERROR] %s cannot create %s vseg for process %x in cluster %x\n",
            __FUNCTION__ , vseg_type_str( type ) , process->pid , local_cxy );
            return NULL;
        }

        // initialize vseg 
        vseg->type        = type;
        vseg->vmm         = vmm;
        vseg->min         = (vseg->vpn_base << CONFIG_PPM_PAGE_SHIFT) + offset; 
        vseg->max         = vseg->min + size;
        vseg->file_offset = file_offset;
        vseg->file_size   = file_size;
        vseg->mapper_xp   = mapper_xp;
        vseg->cxy         = cxy;

        vseg_init_flags( vseg , type );
    }
    /////////////////////////////////////////////////////////////////
    else if( (type == VSEG_TYPE_ANON) || (type == VSEG_TYPE_REMOTE) )
    {
        // compute number of required pages in virtual space
        vpn_t npages = size >> CONFIG_PPM_PAGE_SHIFT;
        if( size & CONFIG_PPM_PAGE_MASK) npages++;
        
        // allocate vseg from MMAP allocator
        vseg = vmm_mmap_alloc( vmm , npages );

        if( vseg == NULL )
        {
            printk("\n[ERROR] %s cannot create %s vseg for process %x in cluster %x\n",
            __FUNCTION__ , vseg_type_str( type ) , process->pid , local_cxy );
            return NULL;
        }

        // initialize vseg
        vseg->type = type;
        vseg->vmm  = vmm;
        vseg->min  = vseg->vpn_base << CONFIG_PPM_PAGE_SHIFT;
        vseg->max  = vseg->min + (vseg->vpn_size << CONFIG_PPM_PAGE_SHIFT);
        vseg->cxy  = cxy;

        vseg_init_flags( vseg , type );
    }
    /////////////////////////////////////////////////////////////////
    else    // VSEG_TYPE_DATA, VSEG_TYPE_CODE or KERNEL vseg
    {
        uint32_t vpn_min = base >> CONFIG_PPM_PAGE_SHIFT;
        uint32_t vpn_max = (base + size - 1) >> CONFIG_PPM_PAGE_SHIFT;

        // allocate vseg descriptor
            vseg = vseg_alloc();

            if( vseg == NULL )
            {
            printk("\n[ERROR] %s cannot create %s vseg for process %x in cluster %x\n",
            __FUNCTION__ , vseg_type_str( type ) , process->pid , local_cxy );
            return NULL;
            }
        // initialize vseg
        vseg->type        = type;
        vseg->vmm         = vmm;
        vseg->min         = base;
        vseg->max         = base + size;
        vseg->vpn_base    = base >> CONFIG_PPM_PAGE_SHIFT;
        vseg->vpn_size    = vpn_max - vpn_min + 1;
        vseg->file_offset = file_offset;
        vseg->file_size   = file_size;
        vseg->mapper_xp   = mapper_xp;
        vseg->cxy         = cxy;

        vseg_init_flags( vseg , type );
    }

    // check collisions
    vseg_t * existing_vseg = vmm_check_conflict( process , vseg->vpn_base , vseg->vpn_size );

    if( existing_vseg != NULL )
    {
        printk("\n[ERROR] in %s for process %x : new vseg %s [vpn_base %x / vpn_size %x]\n"
               "        overlap existing vseg %s [vpn_base %x / vpn_size %x]\n",
        __FUNCTION__ , process->pid, vseg_type_str(vseg->type), vseg->vpn_base, vseg->vpn_size, 
        vseg_type_str(existing_vseg->type), existing_vseg->vpn_base, existing_vseg->vpn_size );
        vseg_free( vseg );
        return NULL;
    }

    // build extended pointer on VSL lock
    xptr_t lock_xp = XPTR( local_cxy , &vmm->vsl_lock );
 
    // take the VSL lock in write mode
    remote_queuelock_acquire( lock_xp );

    // attach vseg to VSL
        vmm_attach_vseg_to_vsl( vmm , vseg );

    // release the VSL lock
    remote_queuelock_release( lock_xp );

#if DEBUG_VMM_CREATE_VSEG 
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_CREATE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] exit / %s / vpn_base %x / vpn_size %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid,
vseg_type_str(type), vseg->vpn_base, vseg->vpn_size, cycle );
#endif

        return vseg;

}  // vmm_create_vseg()

////////////////////////////////////////////////////////////////////////////////////////////
// This static function is called by the vmm_remove_vseg() and vmm_resize_vseg() functions.
// Depending on the vseg <type>, it decrements the physical page refcount, and
// conditionnally release to the relevant kmem the physical page identified by <ppn>.
////////////////////////////////////////////////////////////////////////////////////////////
// @ process  : local pointer on process.
// @ vseg     : local pointer on vseg.
// @ ppn      : released pysical page index. 
////////////////////////////////////////////////////////////////////////////////////////////
static void vmm_ppn_release( process_t * process,
                             vseg_t    * vseg,
                             ppn_t       ppn )
{
    bool_t do_release;

    // get vseg type
    vseg_type_t type = vseg->type;

    // compute is_ref
    bool_t is_ref = (GET_CXY( process->ref_xp ) == local_cxy);

    // get pointers on physical page descriptor
    xptr_t   page_xp  = ppm_ppn2page( ppn );
    cxy_t    page_cxy = GET_CXY( page_xp );
    page_t * page_ptr = GET_PTR( page_xp );

    // decrement page refcount
    xptr_t count_xp = XPTR( page_cxy , &page_ptr->refcount );
    hal_remote_atomic_add( count_xp , -1 );

    // compute the do_release condition depending on vseg type
    if( (type == VSEG_TYPE_FILE)  ||
        (type == VSEG_TYPE_KCODE) || 
        (type == VSEG_TYPE_KDATA) || 
        (type == VSEG_TYPE_KDEV) )            
    {
        // no physical page release for FILE and KERNEL 
        do_release = false;
    }
    else if( (type == VSEG_TYPE_CODE)  ||
             (type == VSEG_TYPE_STACK) ) 
    {
        // always release physical page for private vsegs
        do_release = true;
    }
    else if( (type == VSEG_TYPE_ANON)  ||
             (type == VSEG_TYPE_REMOTE) )
    {
        // release physical page if reference cluster
        do_release = is_ref;
    }
    else if( is_ref )  // vseg_type == DATA in reference cluster
    {
        // get extended pointers on forks and lock field in page descriptor
        xptr_t forks_xp = XPTR( page_cxy , &page_ptr->forks );
        xptr_t lock_xp  = XPTR( page_cxy , &page_ptr->lock );

        // take lock protecting "forks" counter
        remote_busylock_acquire( lock_xp );

        // get number of pending forks from page descriptor
        uint32_t forks = hal_remote_l32( forks_xp );

        // decrement pending forks counter if required
        if( forks )  hal_remote_atomic_add( forks_xp , -1 );

        // release lock protecting "forks" counter
        remote_busylock_release( lock_xp );

        // release physical page if forks == 0
        do_release = (forks == 0); 
    }
    else              // vseg_type == DATA not in reference cluster
    {
        // no physical page release if not in reference cluster 
        do_release = false;
    }

    // release physical page to relevant kmem when required
    if( do_release )
    {
        ppm_remote_free_pages( page_cxy , page_ptr );

#if DEBUG_VMM_PPN_RELEASE
thread_t * this = CURRENT_THREAD;
if( DEBUG_VMM_PPN_RELEASE < cycle )
printk("\n[%s] thread[%x,%x] released ppn %x to kmem\n",
__FUNCTION__, this->process->pid, this->trdid, ppn );
#endif

    }
} // end vmm_ppn_release()

//////////////////////////////////////////
void vmm_remove_vseg( process_t * process,
                      vseg_t    * vseg )
{
    uint32_t    vseg_type;  // vseg type
    vpn_t       vpn;        // VPN of current PTE
    vpn_t       vpn_min;    // VPN of first PTE
    vpn_t       vpn_max;    // VPN of last PTE (excluded)
    ppn_t       ppn;        // current PTE ppn value
    uint32_t    attr;       // current PTE attributes

// check arguments
assert( (process != NULL), "process argument is NULL" );
assert( (vseg    != NULL), "vseg argument is NULL" );

    // get pointers on local process VMM 
    vmm_t * vmm = &process->vmm;

    // build extended pointer on GPT
    xptr_t gpt_xp = XPTR( local_cxy , &vmm->gpt );

    // get relevant vseg infos
    vseg_type = vseg->type;
    vpn_min   = vseg->vpn_base;
    vpn_max   = vpn_min + vseg->vpn_size;

#if DEBUG_VMM_REMOVE_VSEG
uint32_t   cycle = (uint32_t)hal_get_cycles();
thread_t * this  = CURRENT_THREAD;
#endif

#if (DEBUG_VMM_REMOVE_VSEG & 1 )
if( DEBUG_VMM_REMOVE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] enters / process %x / type %s / base %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, 
process->pid, vseg_type_str(vseg->type), vseg->min, cycle );
#endif

    // loop on PTEs in GPT to unmap all mapped PTE
        for( vpn = vpn_min ; vpn < vpn_max ; vpn++ )
    {
        // get ppn and attr
        hal_gpt_get_pte( gpt_xp , vpn , &attr , &ppn );

        if( attr & GPT_MAPPED )  // PTE is mapped
        { 

#if( DEBUG_VMM_REMOVE_VSEG & 1 )
if( DEBUG_VMM_REMOVE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] unmap vpn %x / ppn %x / type %s\n",
__FUNCTION__, this->process->pid, this->trdid, vpn , ppn, vseg_type_str(vseg_type) );
#endif
            // unmap GPT entry in local GPT
            hal_gpt_reset_pte( gpt_xp , vpn );

            // release physical page when required
            vmm_ppn_release( process , vseg , ppn );
        }
    }

    // remove vseg from VSL 
    vmm_detach_vseg_from_vsl( vmm , vseg );

    // release vseg descriptor depending on vseg type
    if( vseg_type == VSEG_TYPE_STACK )
    {
        // release slot to local stack allocator
        vmm_stack_free( vmm , vseg );
    }
    else if( (vseg_type == VSEG_TYPE_ANON) || 
             (vseg_type == VSEG_TYPE_FILE) || 
             (vseg_type == VSEG_TYPE_REMOTE) )  
    {
        // release vseg to local mmap allocator
        vmm_mmap_free( vmm , vseg );
    }
    else
    {
        // release vseg descriptor to local kmem
        vseg_free( vseg );
    }

#if DEBUG_VMM_REMOVE_VSEG
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_REMOVE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] exit / process %x / type %s / base %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, 
process->pid, vseg_type_str(vseg->type), vseg->min, cycle );
#endif

}  // end vmm_remove_vseg()

/////////////////////////////////////////////
void vmm_resize_vseg( process_t * process,
                      vseg_t    * vseg,
                      intptr_t    new_base,
                      intptr_t    new_size )
{
    vpn_t     vpn;
    ppn_t     ppn;
    uint32_t  attr;

// check arguments
assert( (process != NULL), "process argument is NULL" );
assert( (vseg    != NULL), "vseg argument is NULL" );

#if DEBUG_VMM_RESIZE_VSEG
uint32_t   cycle = (uint32_t)hal_get_cycles();
thread_t * this  = CURRENT_THREAD;
#endif

#if (DEBUG_VMM_RESIZE_VSEG & 1)
if( DEBUG_VMM_RESIZE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] enter / process %x / %s / base %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, 
process->pid, vseg_type_str(vseg->type), old_base, cycle );
#endif

    // get existing vseg vpn_min and vpn_max
    vpn_t     old_vpn_min = vseg->vpn_base;
    vpn_t     old_vpn_max = old_vpn_min + vseg->vpn_size - 1;

    // compute new vseg vpn_min & vpn_max  
    intptr_t min          = new_base;
    intptr_t max          = new_base + new_size;
    vpn_t    new_vpn_min  = min >> CONFIG_PPM_PAGE_SHIFT;
    vpn_t    new_vpn_max  = (max - 1) >> CONFIG_PPM_PAGE_SHIFT;

    // build extended pointer on GPT
    xptr_t gpt_xp = XPTR( local_cxy , &process->vmm.gpt );

    // loop on PTEs in GPT to unmap PTE if (oldd_vpn_min <= vpn < new_vpn_min)
        for( vpn = old_vpn_min ; vpn < new_vpn_min ; vpn++ )
    {
        // get ppn and attr
        hal_gpt_get_pte( gpt_xp , vpn , &attr , &ppn );

        if( attr & GPT_MAPPED )  // PTE is mapped
        { 

#if( DEBUG_VMM_RESIZE_VSEG & 1 )
if( DEBUG_VMM_RESIZE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] unmap vpn %x / ppn %x / %s",
__FUNCTION__, this->process->pid, this->trdid, vpn , ppn, vseg_type_str(vseg_type) );
#endif
            // unmap GPT entry 
            hal_gpt_reset_pte( gpt_xp , vpn );

            // release physical page when required
            vmm_ppn_release( process , vseg , ppn );
        }
    }

    // loop on PTEs in GPT to unmap PTE if (new vpn_max <= vpn < old_vpn_max)
        for( vpn = new_vpn_max ; vpn < old_vpn_max ; vpn++ )
    {
        // get ppn and attr
        hal_gpt_get_pte( gpt_xp , vpn , &attr , &ppn );

        if( attr & GPT_MAPPED )  // PTE is mapped
        { 

#if( DEBUG_VMM_RESIZE_VSEG & 1 )
if( DEBUG_VMM_RESIZE_VSEG < cycle )
printk("\n[%s] thread[%x,%x] unmap vpn %x / ppn %x / %s",
__FUNCTION__, this->process->pid, this->trdid, vpn , ppn, vseg_type_str(vseg_type) );
#endif
            // unmap GPT entry in local GPT
            hal_gpt_reset_pte( gpt_xp , vpn );

            // release physical page when required
            vmm_ppn_release( process , vseg , ppn );
        }
    }

    // resize vseg in VSL
    vseg->min      = min;
    vseg->max      = max;
    vseg->vpn_base = new_vpn_min;
    vseg->vpn_size = new_vpn_max - new_vpn_min + 1;

#if DEBUG_VMM_RESIZE_VSEG
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_RESIZE_VSEG < cycle )
printk("[%s] thread[%x,%x] exit / process %x / %s / base %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, 
process->pid, vseg_type_str(vseg->type), vseg->min, cycle );
#endif

}  // end vmm_resize_vseg

/////////////////////////////////////////////////////////////////////////////////////////////
// This static function is called twice by the vmm_get_vseg() function.
// It scan the - possibly remote - VSL defined by the <vmm_xp> argument to find the vseg
// containing a given virtual address <vaddr>. It uses remote accesses to access the remote
// VSL if required. The VSL lock protecting the VSL must be taken by the caller.
/////////////////////////////////////////////////////////////////////////////////////////////
// @ vmm_xp  : extended pointer on the process VMM.
// @ vaddr   : virtual address.
// @ return local pointer on remote vseg if success / return NULL if not found.
/////////////////////////////////////////////////////////////////////////////////////////////
static vseg_t * vmm_vseg_from_vaddr( xptr_t     vmm_xp,
                                     intptr_t   vaddr )
{
    xptr_t   iter_xp;
    xptr_t   vseg_xp;
    vseg_t * vseg;
    intptr_t min;
    intptr_t max;

    // get cluster and local pointer on target VMM
    vmm_t * vmm_ptr = GET_PTR( vmm_xp );
    cxy_t   vmm_cxy = GET_CXY( vmm_xp );

    // build extended pointer on VSL root
    xptr_t root_xp = XPTR( vmm_cxy , &vmm_ptr->vsegs_root );

    // scan the list of vsegs in VSL
    XLIST_FOREACH( root_xp , iter_xp )
    {
        vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist );
        vseg    = GET_PTR( vseg_xp );

        min = hal_remote_l32( XPTR( vmm_cxy , &vseg->min ) );
        max = hal_remote_l32( XPTR( vmm_cxy , &vseg->max ) );

        // return success when match
        if( (vaddr >= min) && (vaddr < max) ) return vseg;
    }

    // return failure
    return NULL;

}  // end vmm_vseg_from_vaddr()

///////////////////////////////////////////
error_t  vmm_get_vseg( process_t * process,
                       intptr_t    vaddr,
                       vseg_t   ** found_vseg )
{
    xptr_t    loc_lock_xp;     // extended pointer on local VSL lock
    xptr_t    ref_lock_xp;     // extended pointer on reference VSL lock
    vseg_t  * loc_vseg;        // local pointer on local vseg
    vseg_t  * ref_vseg;        // local pointer on reference vseg

    // build extended pointer on local VSL lock
    loc_lock_xp = XPTR( local_cxy , &process->vmm.vsl_lock );
     
    // get local VSL lock 
    remote_queuelock_acquire( loc_lock_xp );

    // try to get vseg from local VMM
    loc_vseg = vmm_vseg_from_vaddr( XPTR( local_cxy, &process->vmm ) , vaddr );

    if (loc_vseg == NULL)   // vseg not found => access reference VSL
    {
        // get extended pointer on reference process
        xptr_t ref_xp = process->ref_xp;

        // get cluster and local pointer on reference process
        cxy_t       ref_cxy = GET_CXY( ref_xp );
        process_t * ref_ptr = GET_PTR( ref_xp );

        // build extended pointer on reference VSL lock
        ref_lock_xp = XPTR( ref_cxy , &ref_ptr->vmm.vsl_lock );
     
        // get reference VSL lock 
        remote_queuelock_acquire( ref_lock_xp );

        // try to get vseg from reference VMM
        ref_vseg = vmm_vseg_from_vaddr( XPTR( ref_cxy , &ref_ptr->vmm ) , vaddr );

        if( ref_vseg == NULL )  // vseg not found => return error
        {
            printk("\n[ERROR] in %s : vaddr %x in process %x out of segment\n",
            __FUNCTION__, vaddr, process->pid );

            // release reference VSL lock 
            remote_queuelock_release( ref_lock_xp );

            return -1;
        }
        else                    // vseg found => try to update local VSL 
        {
            // allocate a local vseg descriptor
            loc_vseg = vseg_alloc();

            if( loc_vseg == NULL )   // no memory => return error
            {
                printk("\n[ERROR] in %s : vaddr %x in process %x / no memory for local vseg\n",
                __FUNCTION__, vaddr, process->pid );

                // release reference VSL & local VSL locks
                remote_queuelock_release( ref_lock_xp );
                remote_queuelock_release( loc_lock_xp );

                return -1;
            }
            else                     // update local VSL and return success
            {
                // initialize local vseg
                vseg_init_from_ref( loc_vseg , XPTR( ref_cxy , ref_vseg ) );

                // register local vseg in local VSL
                vmm_attach_vseg_to_vsl( &process->vmm , loc_vseg );

                // release reference VSL & local VSL locks
                remote_queuelock_release( ref_lock_xp );
                remote_queuelock_release( loc_lock_xp );

                *found_vseg = loc_vseg;
                return 0;
            }
        }
    }
    else                        // vseg found in local VSL => return success
    {
        // release local VSL lock 
        remote_queuelock_release( loc_lock_xp );

        *found_vseg = loc_vseg;
        return 0;
    }
}  // end vmm_get_vseg()

//////////////////////////////////////////////////////////////////////////////////////
// This static function compute the target cluster to allocate a physical page
// for a given <vpn> in a given <vseg>, allocates the page and returns an extended 
// pointer on the allocated page descriptor.
// The vseg cannot have the FILE type.
//////////////////////////////////////////////////////////////////////////////////////
// @ vseg   : local pointer on vseg.
// @ vpn    : unmapped vpn.
// @ return an extended pointer on the allocated page
//////////////////////////////////////////////////////////////////////////////////////
static xptr_t vmm_page_allocate( vseg_t * vseg,
                                 vpn_t    vpn )
{

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

    xptr_t       page_xp;
    cxy_t        page_cxy;
    page_t     * page_ptr;
    uint32_t     index;

    uint32_t     type   = vseg->type;
    uint32_t     flags  = vseg->flags;
    uint32_t     x_size = LOCAL_CLUSTER->x_size;
    uint32_t     y_size = LOCAL_CLUSTER->y_size;

// check vseg type
assert( ( type != VSEG_TYPE_FILE ) , "illegal vseg type\n" );

    if( flags & VSEG_DISTRIB )    // distributed => cxy depends on vpn LSB
    {
        index    = vpn & ((x_size * y_size) - 1);
        page_cxy = HAL_CXY_FROM_XY( (index / y_size) , (index % y_size) );

        // If the cluster selected from VPN's LSBs is empty, we select one randomly
        if ( cluster_is_active( page_cxy ) == false )
        {
            page_cxy = cluster_random_select();
        }
    }
    else                          // other cases => cxy specified in vseg
    {
        page_cxy = vseg->cxy;
    }

    // allocate one small physical page from target cluster
    page_ptr = ppm_remote_alloc_pages( page_cxy , 0 );

    page_xp = XPTR( page_cxy , page_ptr );

#if DEBUG_VMM_PAGE_ALLOCATE
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_VMM_PAGE_ALLOCATE < cycle )
printk("\n[%s] thread[%x,%x] exit for vpn %x / ppn %x / cycle %d\n",
__FUNCTION__ , this->process->pid, this->trdid, vpn, ppm_page2ppn(page_xp), cycle );
#endif

    return page_xp;

}  // end vmm_page_allocate()  

////////////////////////////////////////
error_t vmm_get_one_ppn( vseg_t * vseg,
                         vpn_t    vpn,
                         ppn_t  * ppn )
{
    error_t    error;
    xptr_t     page_xp;           // extended pointer on physical page descriptor
    uint32_t   page_id;           // missing page index in vseg mapper
    uint32_t   type;              // vseg type;

    type      = vseg->type;
    page_id   = vpn - vseg->vpn_base;

#if DEBUG_VMM_GET_ONE_PPN 
uint32_t   cycle = (uint32_t)hal_get_cycles();
thread_t * this  = CURRENT_THREAD;
if( (DEBUG_VMM_GET_ONE_PPN < cycle) && (vpn == 0x40b) )
printk("\n[%s] thread[%x,%x] enter for vpn %x / type %s / page_id  %d / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, vseg_type_str(type), page_id, cycle );
#endif

    // FILE type : get the physical page from the file mapper
    if( type == VSEG_TYPE_FILE )
    {
        // get extended pointer on mapper
        xptr_t mapper_xp = vseg->mapper_xp;

assert( (mapper_xp != XPTR_NULL),
"mapper not defined for a FILE vseg\n" );
        
        // get extended pointer on page descriptor
        page_xp = mapper_remote_get_page( mapper_xp , page_id );

        if ( page_xp == XPTR_NULL ) return EINVAL;
    }

    // Other types : allocate a physical page from target cluster,
    // as defined by vseg type and vpn value
    else
    {
        // allocate one physical page 
        page_xp = vmm_page_allocate( vseg , vpn );

        if( page_xp == XPTR_NULL ) return -1;

        // initialise missing page from .elf file mapper for DATA and CODE types
        // the vseg->mapper_xp field is an extended pointer on the .elf file mapper
        if( (type == VSEG_TYPE_CODE) || (type == VSEG_TYPE_DATA) )
        {
            // get extended pointer on mapper
            xptr_t     mapper_xp = vseg->mapper_xp;

assert( (mapper_xp != XPTR_NULL),
"mapper not defined for a CODE or DATA vseg\n" );
        
            // compute missing page offset in vseg
            uint32_t offset = page_id << CONFIG_PPM_PAGE_SHIFT;

            // compute missing page offset in .elf file
            uint32_t elf_offset = vseg->file_offset + offset;

#if (DEBUG_VMM_GET_ONE_PPN & 0x1)
if( (DEBUG_VMM_GET_ONE_PPN < cycle) && (vpn == 0x40b) )
printk("\n[%s] thread[%x,%x] for vpn = %x / elf_offset = %x\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, elf_offset );
#endif
            // compute extended pointer on page base 
            xptr_t base_xp  = ppm_page2base( page_xp );

            // file_size (in .elf mapper) can be smaller than vseg_size (BSS)
            uint32_t file_size = vseg->file_size;

            if( file_size < offset )                 // missing page fully in  BSS 
            {

#if (DEBUG_VMM_GET_ONE_PPN & 0x1)
if( (DEBUG_VMM_GET_ONE_PPN < cycle) && (vpn == 0x40b) )
printk("\n[%s] thread[%x,%x] for vpn  %x / fully in BSS\n",
__FUNCTION__, this->process->pid, this->trdid, vpn );
#endif
                if( GET_CXY( page_xp ) == local_cxy )
                {
                    memset( GET_PTR( base_xp ) , 0 , CONFIG_PPM_PAGE_SIZE );
                }
                else
                {
                   hal_remote_memset( base_xp , 0 , CONFIG_PPM_PAGE_SIZE );       
                }
            }
            else if( file_size >= (offset + CONFIG_PPM_PAGE_SIZE) )  // fully in  mapper
            {

#if (DEBUG_VMM_GET_ONE_PPN & 0x1)
if( (DEBUG_VMM_GET_ONE_PPN < cycle) && (vpn == 0x40b) )
printk("\n[%s] thread[%x,%x] for vpn  %x / fully in mapper\n",
__FUNCTION__, this->process->pid, this->trdid, vpn );
#endif
                error = mapper_move_kernel( mapper_xp,
                                            true,             // to_buffer
                                            elf_offset,
                                            base_xp,
                                            CONFIG_PPM_PAGE_SIZE ); 
                if( error ) return EINVAL;
            }
            else  // both in mapper and in BSS :
                  // - (file_size - offset)             bytes from mapper
                  // - (page_size + offset - file_size) bytes from BSS
            {

#if (DEBUG_VMM_GET_ONE_PPN & 0x1)
if( (DEBUG_VMM_GET_ONE_PPN < cycle) && (vpn == 0x40b) )
printk("\n[%s] thread[%x,%x] for vpn  %x / both mapper & BSS\n"
"      %d bytes from mapper / %d bytes from BSS\n",
__FUNCTION__, this->process->pid, this->trdid, vpn,
file_size - offset , offset + CONFIG_PPM_PAGE_SIZE - file_size  );
#endif
                // initialize mapper part
                error = mapper_move_kernel( mapper_xp,
                                            true,         // to buffer
                                            elf_offset,
                                            base_xp,
                                            file_size - offset ); 
                if( error ) return EINVAL;

                // initialize BSS part
                if( GET_CXY( page_xp ) == local_cxy )
                {
                    memset( GET_PTR( base_xp ) + file_size - offset , 0 , 
                            offset + CONFIG_PPM_PAGE_SIZE - file_size );
                }
                else
                {
                   hal_remote_memset( base_xp + file_size - offset , 0 , 
                                      offset + CONFIG_PPM_PAGE_SIZE - file_size );
                }
            }    
        }  // end initialisation for CODE or DATA types   
    } 

    // return ppn
    *ppn = ppm_page2ppn( page_xp );

#if DEBUG_VMM_GET_ONE_PPN
cycle = (uint32_t)hal_get_cycles();
if( (DEBUG_VMM_GET_ONE_PPN < cycle) && (vpn == 0x40b) )
printk("\n[%s] thread[%x,%x] exit for vpn %x / ppn %x / cycle %d\n",
__FUNCTION__ , this->process->pid, this->trdid , vpn , *ppn, cycle );
#endif

    return 0;

}  // end vmm_get_one_ppn()

///////////////////////////////////////////////////
error_t vmm_handle_page_fault( process_t * process,
                               vpn_t       vpn )
{
    vseg_t         * vseg;            // vseg containing vpn
    uint32_t         attr;            // PTE_ATTR value
    ppn_t            ppn;             // PTE_PPN value
    uint32_t         ref_attr;        // PTE_ATTR value in reference GPT
    ppn_t            ref_ppn;         // PTE_PPN value in reference GPT
    cxy_t            ref_cxy;         // reference cluster for missing vpn
    process_t      * ref_ptr;         // reference process for missing vpn
    xptr_t           local_gpt_xp;    // extended pointer on local GPT
    xptr_t           ref_gpt_xp;      // extended pointer on reference GPT
    error_t          error;           // value returned by called functions

    thread_t * this  = CURRENT_THREAD;

#if (CONFIG_INSTRUMENTATION_PGFAULTS || DEBUG_VMM_HANDLE_PAGE_FAULT)
uint32_t start_cycle = (uint32_t)hal_get_cycles();
#endif

#if DEBUG_VMM_HANDLE_PAGE_FAULT
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] enter for vpn %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, start_cycle );
#endif

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
hal_vmm_display( this->process , true );
#endif

    // get local vseg (access to reference VSL can be required)
    error = vmm_get_vseg( process, 
                          (intptr_t)vpn<<CONFIG_PPM_PAGE_SHIFT,
                          &vseg );
    if( error )
    {
        printk("\n[ERROR] in %s : vpn %x in thread[%x,%x] not in registered vseg\n",
        __FUNCTION__ , vpn , process->pid, this->trdid );
        
        return EXCP_USER_ERROR;
    }

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] found vseg %s\n",
__FUNCTION__, this->process->pid, this->trdid, vseg_type_str(vseg->type) );
#endif

    // build extended pointer on local GPT
    local_gpt_xp  = XPTR( local_cxy , &process->vmm.gpt );

    // lock PTE in local GPT and get current PPN and attributes
    error = hal_gpt_lock_pte( local_gpt_xp,
                              vpn,
                              &attr,
                              &ppn );
    if( error )
    {
        printk("\n[PANIC] in %s : cannot lock PTE in local GPT / vpn %x / process %x\n",
        __FUNCTION__ , vpn , process->pid );
        
        return EXCP_KERNEL_PANIC;
    }

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] locked vpn %x in cluster %x\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, local_cxy );
#endif

    // handle page fault only if local PTE still unmapped after lock
    if( (attr & GPT_MAPPED) == 0 )
    {
        // get reference process cluster and local pointer
        ref_cxy = GET_CXY( process->ref_xp );
        ref_ptr = GET_PTR( process->ref_xp );

        /////////////// private vseg or (local == reference)
        /////////////// => access only the local GPT 
        if( (vseg->type == VSEG_TYPE_STACK) ||
            (vseg->type == VSEG_TYPE_CODE)  ||
            (ref_cxy    == local_cxy ) )
        {

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] access local gpt : cxy %x / ref_cxy %x / type %s / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid,
local_cxy, ref_cxy, vseg_type_str(vseg->type), (uint32_t)hal_get_cycles() );
#endif
            // allocate and initialise a physical page 
            error = vmm_get_one_ppn( vseg , vpn , &ppn );

            if( error )
            {
                printk("\n[ERROR] in %s : no physical page / process = %x / vpn = %x\n",
                __FUNCTION__ , process->pid , vpn );

                // unlock PTE in local GPT 
                hal_gpt_unlock_pte( local_gpt_xp , vpn );

                return EXCP_KERNEL_PANIC;
            }

            // define attr from vseg flags 
            attr = GPT_MAPPED | GPT_SMALL | GPT_READABLE;
            if( vseg->flags & VSEG_USER  ) attr |= GPT_USER;
            if( vseg->flags & VSEG_WRITE ) attr |= GPT_WRITABLE;
            if( vseg->flags & VSEG_EXEC  ) attr |= GPT_EXECUTABLE;
            if( vseg->flags & VSEG_CACHE ) attr |= GPT_CACHABLE;

            // set PTE to local GPT 
            // it unlocks this PTE
            hal_gpt_set_pte( local_gpt_xp,
                             vpn,
                             attr,
                             ppn );

#if (CONFIG_INSTRUMENTATION_PGFAULTS || DEBUG_VMM_HANDLE_PAGE_FAULT)
uint32_t end_cycle = (uint32_t)hal_get_cycles();
uint32_t cost      = end_cycle - start_cycle;
#endif

#if DEBUG_VMM_HANDLE_PAGE_FAULT
if( (end_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] handled local pgfault / ppn %x / attr %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, ppn, attr, end_cycle );
#endif

#if CONFIG_INSTRUMENTATION_PGFAULTS 
this->info.local_pgfault_nr++;
this->info.local_pgfault_cost += cost;
if( cost > this->info.local_pgfault_max ) this->info.local_pgfault_max = cost;
#endif
            return EXCP_NON_FATAL;

        }   // end local GPT access 

        /////////////////// public vseg and (local != reference) 
        /////////////////// => access ref GPT to update local GPT
        else                                
        {

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] access ref gpt : cxy %x / ref_cxy %x / type %s / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, 
local_cxy, ref_cxy, vseg_type_str(vseg->type), (uint32_t)hal_get_cycles() );
#endif
            // build extended pointer on reference GPT
            ref_gpt_xp = XPTR( ref_cxy , &ref_ptr->vmm.gpt );

            // lock PTE in reference GPT and get current PPN and attributes
            error = hal_gpt_lock_pte( ref_gpt_xp,
                                      vpn,
                                      &ref_attr,
                                      &ref_ppn );
            if( error )
            {
                printk("\n[PANIC] in %s : cannot lock PTE in ref GPT / vpn %x / process %x\n",
                __FUNCTION__ , vpn , process->pid );
        
                // unlock PTE in local GPT
                hal_gpt_unlock_pte( local_gpt_xp , vpn );
                    
                return EXCP_KERNEL_PANIC;
            }

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] get pte from ref gpt / attr %x / ppn %x\n",
__FUNCTION__, this->process->pid, this->trdid, ref_attr, ref_ppn );
#endif

            if( ref_attr & GPT_MAPPED )        // false page fault 
            {
                // update local GPT from reference GPT values
                // this unlocks the PTE in local GPT
                hal_gpt_set_pte( local_gpt_xp,
                                 vpn,
                                 ref_attr,
                                 ref_ppn );

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] updated local gpt for a false pgfault\n",
__FUNCTION__, this->process->pid, this->trdid );
#endif

                // unlock the PTE in reference GPT
                hal_gpt_unlock_pte( ref_gpt_xp, vpn );
                             
#if (DEBUG_VMM_HANDLE_PAGE_FAULT &1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] unlock the ref gpt after a false pgfault\n",
__FUNCTION__, this->process->pid, this->trdid );
#endif

#if (CONFIG_INSTRUMENTATION_PGFAULTS || DEBUG_VMM_HANDLE_PAGE_FAULT)
uint32_t end_cycle = (uint32_t)hal_get_cycles();
uint32_t cost      = end_cycle - start_cycle;
#endif

#if DEBUG_VMM_HANDLE_PAGE_FAULT
if( (end_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] handled false pgfault / ppn %x / attr %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, ref_ppn, ref_attr, end_cycle );
#endif

#if CONFIG_INSTRUMENTATION_PGFAULTS 
this->info.false_pgfault_nr++;
this->info.false_pgfault_cost += cost;
if( cost > this->info.false_pgfault_max ) this->info.false_pgfault_max = cost;
#endif
                return EXCP_NON_FATAL;
            }
            else                            // true page fault 
            {
                // allocate and initialise a physical page depending on the vseg type
                error = vmm_get_one_ppn( vseg , vpn , &ppn );

                if( error )
                {
                    printk("\n[ERROR] in %s : no memory / process = %x / vpn = %x\n",
                    __FUNCTION__ , process->pid , vpn );

                    // unlock PTE in local GPT and in reference GPT
                    hal_gpt_unlock_pte( local_gpt_xp , vpn );
                    hal_gpt_unlock_pte( ref_gpt_xp   , vpn );
                    
                    return EXCP_KERNEL_PANIC;
                }

                // define attr from vseg flags 
                attr = GPT_MAPPED | GPT_SMALL | GPT_READABLE;
                if( vseg->flags & VSEG_USER  ) attr |= GPT_USER;
                if( vseg->flags & VSEG_WRITE ) attr |= GPT_WRITABLE;
                if( vseg->flags & VSEG_EXEC  ) attr |= GPT_EXECUTABLE;
                if( vseg->flags & VSEG_CACHE ) attr |= GPT_CACHABLE;

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] build a new PTE for a true pgfault\n",
__FUNCTION__, this->process->pid, this->trdid );
#endif
                // set PTE in reference GPT
                // this unlock the PTE 
                hal_gpt_set_pte( ref_gpt_xp,
                                 vpn,
                                 attr,
                                 ppn );

#if (DEBUG_VMM_HANDLE_PAGE_FAULT & 1)
if( (start_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] set new PTE in ref gpt for a true page fault\n",
__FUNCTION__, this->process->pid, this->trdid );
#endif

                // set PTE in local GPT 
                // this unlock the PTE 
                hal_gpt_set_pte( local_gpt_xp,
                                 vpn,
                                 attr,
                                 ppn );

#if (CONFIG_INSTRUMENTATION_PGFAULTS || DEBUG_VMM_HANDLE_PAGE_FAULT)
uint32_t end_cycle = (uint32_t)hal_get_cycles();
uint32_t cost      = end_cycle - start_cycle;
#endif

#if DEBUG_VMM_HANDLE_PAGE_FAULT
if( (end_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] handled global pgfault / ppn %x / attr %x / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, ppn, attr, end_cycle );
#endif

#if CONFIG_INSTRUMENTATION_PGFAULTS 
this->info.global_pgfault_nr++;
this->info.global_pgfault_cost += cost;
if( cost > this->info.global_pgfault_max ) this->info.global_pgfault_max = cost;
#endif
                return EXCP_NON_FATAL;
            }
        }
    }
    else   // page has been locally mapped by another concurrent thread
    {
        // unlock the PTE in local GPT
        hal_gpt_unlock_pte( local_gpt_xp , vpn );

#if (CONFIG_INSTRUMENTATION_PGFAULTS || DEBUG_VMM_HANDLE_PAGE_FAULT)
uint32_t end_cycle = (uint32_t)hal_get_cycles();
uint32_t cost      = end_cycle - start_cycle;
#endif

#if DEBUG_VMM_HANDLE_PAGE_FAULT
if( (end_cycle > DEBUG_VMM_HANDLE_PAGE_FAULT) && (vpn > 0) )
printk("\n[%s] handled by another thread / vpn %x / ppn %x / attr %x / cycle %d\n",
__FUNCTION__, vpn, ppn, attr, end_cycle );
#endif

#if CONFIG_INSTRUMENTATION_PGFAULTS 
this->info.false_pgfault_nr++;
this->info.false_pgfault_cost += cost;
if( cost > this->info.false_pgfault_max ) this->info.false_pgfault_max = cost;
#endif
        return EXCP_NON_FATAL;
    }

}   // end vmm_handle_page_fault()

////////////////////////////////////////////
error_t vmm_handle_cow( process_t * process,
                        vpn_t       vpn )
{
    vseg_t         * vseg;            // vseg containing vpn
    xptr_t           gpt_xp;          // extended pointer on GPT (local or reference)
    gpt_t          * gpt_ptr;         // local pointer on GPT (local or reference)
    cxy_t            gpt_cxy;         // GPT cluster identifier
    uint32_t         old_attr;        // current PTE_ATTR value
    ppn_t            old_ppn;         // current PTE_PPN value
    uint32_t         new_attr;        // new PTE_ATTR value
    ppn_t            new_ppn;         // new PTE_PPN value
    cxy_t            ref_cxy;         // reference process cluster
    process_t      * ref_ptr;         // local pointer on reference process
    error_t          error;

    thread_t * this  = CURRENT_THREAD;

#if DEBUG_VMM_HANDLE_COW
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] enter for vpn %x / core[%x,%d] / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, local_cxy, this->core->lid, cycle );
#endif

#if ((DEBUG_VMM_HANDLE_COW & 3) == 3 )
hal_vmm_display( XPTR( local_cxy , process ) , true );
#endif

    // get local vseg
    error = vmm_get_vseg( process, 
                          (intptr_t)vpn<<CONFIG_PPM_PAGE_SHIFT,
                          &vseg );
    if( error )
    {
        printk("\n[ERROR] in %s : vpn %x in thread[%x,%x] not in a registered vseg\n",
        __FUNCTION__, vpn, process->pid, this->trdid );

        return EXCP_USER_ERROR;
    }

#if DEBUG_VMM_HANDLE_COW 
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] get vseg %s\n",
__FUNCTION__, this->process->pid, this->trdid, vseg_type_str(vseg->type) );
#endif

    // get reference process cluster and local pointer
    ref_cxy = GET_CXY( process->ref_xp );
    ref_ptr = GET_PTR( process->ref_xp );

    // build pointers on relevant GPT
    // - access only local GPT for a private vseg  
    // - access reference GPT and all copies for a public vseg
    if( (vseg->type == VSEG_TYPE_STACK) || (vseg->type == VSEG_TYPE_CODE) )
    {
        gpt_cxy = local_cxy;
        gpt_ptr = &process->vmm.gpt;
        gpt_xp  = XPTR( gpt_cxy , gpt_ptr );
    }
    else
    {
        gpt_cxy = ref_cxy;
        gpt_ptr = &ref_ptr->vmm.gpt;
        gpt_xp  = XPTR( gpt_cxy , gpt_ptr );
    }

    // lock target PTE in relevant GPT (local or reference) 
    // and get current PTE value
    error = hal_gpt_lock_pte( gpt_xp,
                              vpn,
                              &old_attr,
                              &old_ppn );
    if( error )
    {
        printk("\n[PANIC] in %s : cannot lock PTE in GPT / cxy %x / vpn %x / process %x\n",
        __FUNCTION__ , gpt_cxy, vpn , process->pid );
        
        return EXCP_KERNEL_PANIC;
    }

#if DEBUG_VMM_HANDLE_COW 
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] get pte for vpn %x : ppn %x / attr %x\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, old_ppn, old_attr );
#endif

    // return user error if COW attribute not set or PTE2 unmapped
    if( ((old_attr & GPT_COW) == 0) || ((old_attr & GPT_MAPPED) == 0) )
    {
        hal_gpt_unlock_pte( gpt_xp , vpn );

        return EXCP_USER_ERROR;
    }

    // get pointers on physical page descriptor
    xptr_t   page_xp  = ppm_ppn2page( old_ppn );
    cxy_t    page_cxy = GET_CXY( page_xp );
    page_t * page_ptr = GET_PTR( page_xp );

    // get extended pointers on forks and lock field in page descriptor
    xptr_t forks_xp       = XPTR( page_cxy , &page_ptr->forks );
    xptr_t forks_lock_xp  = XPTR( page_cxy , &page_ptr->lock );

    // take lock protecting "forks" counter
    remote_busylock_acquire( forks_lock_xp );

    // get number of pending forks from page descriptor
    uint32_t forks = hal_remote_l32( forks_xp );

#if DEBUG_VMM_HANDLE_COW 
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] get forks = %d for vpn %x\n",
__FUNCTION__, this->process->pid, this->trdid, forks, vpn );
#endif

    if( forks )        // pending fork => allocate a new page, and copy old to new
    {
        // decrement pending forks counter in page descriptor
        hal_remote_atomic_add( forks_xp , -1 );

        // release lock protecting "forks" counter
        remote_busylock_release( forks_lock_xp );

        // allocate a new physical page depending on vseg type
        page_xp = vmm_page_allocate( vseg , vpn );

        if( page_xp == XPTR_NULL ) 
        {
            printk("\n[PANIC] in %s : no memory for vpn %x in process %x\n",
            __FUNCTION__ , vpn, process->pid );

            hal_gpt_unlock_pte( gpt_xp , vpn ); 

            return EXCP_KERNEL_PANIC;
        }

        // compute allocated page PPN 
        new_ppn = ppm_page2ppn( page_xp );

#if DEBUG_VMM_HANDLE_COW 
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] get new ppn %x for vpn %x\n",
__FUNCTION__, this->process->pid, this->trdid, new_ppn, vpn );
#endif

        // copy old page content to new page
        hal_remote_memcpy( ppm_ppn2base( new_ppn ),
                           ppm_ppn2base( old_ppn ),
                           CONFIG_PPM_PAGE_SIZE );

#if DEBUG_VMM_HANDLE_COW
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] copied old page to new page\n",
__FUNCTION__, this->process->pid, this->trdid );
#endif

    }             
    else               // no pending fork => keep the existing page
    {
        // release lock protecting "forks" counter
        remote_busylock_release( forks_lock_xp );

#if(DEBUG_VMM_HANDLE_COW & 1)
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] no pending forks / keep existing PPN %x\n",
__FUNCTION__, this->process->pid, this->trdid, old_ppn );
#endif
        new_ppn = old_ppn;
    }

    // build new_attr : set WRITABLE, reset COW, reset LOCKED
    new_attr = (((old_attr | GPT_WRITABLE) & (~GPT_COW)) & (~GPT_LOCKED));

#if(DEBUG_VMM_HANDLE_COW & 1)
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] new_attr %x / new_ppn %x\n",
__FUNCTION__, this->process->pid, this->trdid, new_attr, new_ppn );
#endif

    // update the relevant GPT(s)
    // - private vseg => update only the local GPT 
    // - public vseg => update the reference GPT AND all the GPT copies
    if( (vseg->type == VSEG_TYPE_STACK) || (vseg->type == VSEG_TYPE_CODE) )
    {
        // set new PTE in local gpt
        hal_gpt_set_pte( gpt_xp,
                         vpn,
                         new_attr,
                         new_ppn );
    }
    else
    {
        // set new PTE in all GPT copies
        vmm_global_update_pte( process,
                               vpn,
                               new_attr,
                               new_ppn );
    }

#if DEBUG_VMM_HANDLE_COW
cycle = (uint32_t)hal_get_cycles();
if( (DEBUG_VMM_HANDLE_COW < cycle) && (vpn > 0) )
printk("\n[%s] thread[%x,%x] exit for vpn %x / core[%x,%d] / cycle %d\n",
__FUNCTION__, this->process->pid, this->trdid, vpn, local_cxy, this->core->lid, cycle );
#endif

#if ((DEBUG_VMM_HANDLE_COW & 3) == 3)
hal_vmm_display( XPTR( local_cxy , process ) , true );
#endif

     return EXCP_NON_FATAL;

}   // end vmm_handle_cow()