source: trunk/hal/tsar_mips32/core/hal_gpt.c

/*
 * hal_gpt.c - implementation of the Generic Page Table API for TSAR-MIPS32
 *
 * Author   Alain Greiner (2016,2017,2018,2019,2020)
 *
 * Copyright (c) UPMC Sorbonne Universites
 *
 * This file is part of ALMOS-MKH.
 *
 * ALMOS-MKH.is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License as published by
 * the Free Software Foundation; version 2.0 of the License.
 *
 * ALMOS-MKH.is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with ALMOS-MKH.; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
 */

#include <hal_kernel_types.h>
#include <hal_gpt.h>
#include <hal_vmm.h>
#include <hal_special.h>
#include <hal_irqmask.h>
#include <printk.h>
#include <bits.h>
#include <process.h>
#include <kmem.h>
#include <thread.h>
#include <cluster.h>
#include <ppm.h>
#include <page.h>

////////////////////////////////////////////////////////////////////////////////////////
// The Page Table for the TSAR-MIPS32 MMU is defined as a two levels radix tree.
//
// It defines two page sizes : 4 Kbytes pages, and 2 Mbytes pages.
// The virtual address space size is 4 Gbytes (32 bits virtual addresses). 
// The physical address space is limited to 1 Tbytes (40 bits physical addresses).
// - For a 4 Kbytes page, the VPN uses 20 bits, and the PPN requires 28 bits.
// - For a 2 Mbytes page, the PPN uses 11 bits, and the PPN requires 19 bits.
//
// The first level array (PT1) contains 2048 entries, each entry contains 4 bytes,
// and this array is aligned on a 8K bytes boundary.
//
// The second level array (PT2) contains 512 entries, each entry contains 8 bytes,
// and this array is ligned on a 4K bytes boundary. 
////////////////////////////////////////////////////////////////////////////////////////


////////////////////////////////////////////////////////////////////////////////////////
// This define the masks for the TSAR MMU PTE attributes (from TSAR MMU specification)
////////////////////////////////////////////////////////////////////////////////////////

#define TSAR_PTE_MAPPED         0x80000000 
#define TSAR_PTE_SMALL          0x40000000 
#define TSAR_PTE_LOCAL          0x20000000 
#define TSAR_PTE_REMOTE         0x10000000 
#define TSAR_PTE_CACHABLE       0x08000000 
#define TSAR_PTE_WRITABLE       0x04000000 
#define TSAR_PTE_EXECUTABLE     0x02000000 
#define TSAR_PTE_USER           0x01000000 
#define TSAR_PTE_GLOBAL         0x00800000 
#define TSAR_PTE_DIRTY          0x00400000 

#define TSAR_PTE_COW            0x00000001       // only for small pages
#define TSAR_PTE_SWAP           0x00000004       // only for small pages
#define TSAR_PTE_LOCKED         0x00000008       // only for small pages

////////////////////////////////////////////////////////////////////////////////////////
//       TSAR MMU related macros  (from the TSAR MMU specification)
// - IX1  on 11 bits
// - IX2  on  9 bits
// - PPN  on 28 bits
////////////////////////////////////////////////////////////////////////////////////////

#define TSAR_MMU_IX1_WIDTH                 11
#define TSAR_MMU_IX2_WIDTH                 9
#define TSAR_MMU_PPN_WIDTH                 28

#define TSAR_MMU_PTE1_ATTR_MASK            0xFFC00000
#define TSAR_MMU_PTE1_PPN_MASK             0x0007FFFF

#define TSAR_MMU_IX1_FROM_VPN( vpn )       ((vpn >> 9) & 0x7FF)
#define TSAR_MMU_IX2_FROM_VPN( vpn )       (vpn & 0x1FF)

#define TSAR_MMU_PPN2_FROM_PTE1( pte1 )    (pte1 & 0x0FFFFFFF)
#define TSAR_MMU_PPN1_FROM_PTE1( pte1 )    ((pte1 & 0x0007FFFF)<<9) 
#define TSAR_MMU_ATTR_FROM_PTE1( pte1 )    (pte1 & 0xFFC00000)

#define TSAR_MMU_PPN_FROM_PTE2( pte2 )     (pte2 & 0x0FFFFFFF) 
#define TSAR_MMU_ATTR_FROM_PTE2( pte2 )    (pte2 & 0xFFC000FF)

///////////////////////////////////////////////////////////////////////////////////////
// This static function translates the GPT attributes to the TSAR attributes
///////////////////////////////////////////////////////////////////////////////////////
static inline uint32_t gpt2tsar( uint32_t gpt_attr )
{
    uint32_t tsar_attr = 0;

    if( gpt_attr & GPT_MAPPED     ) tsar_attr |= TSAR_PTE_MAPPED;
    if( gpt_attr & GPT_SMALL      ) tsar_attr |= TSAR_PTE_SMALL;
    if( gpt_attr & GPT_WRITABLE   ) tsar_attr |= TSAR_PTE_WRITABLE;
    if( gpt_attr & GPT_EXECUTABLE ) tsar_attr |= TSAR_PTE_EXECUTABLE;
    if( gpt_attr & GPT_CACHABLE   ) tsar_attr |= TSAR_PTE_CACHABLE; 
    if( gpt_attr & GPT_USER       ) tsar_attr |= TSAR_PTE_USER;
    if( gpt_attr & GPT_DIRTY      ) tsar_attr |= TSAR_PTE_DIRTY;
    if( gpt_attr & GPT_ACCESSED   ) tsar_attr |= TSAR_PTE_LOCAL;
    if( gpt_attr & GPT_GLOBAL     ) tsar_attr |= TSAR_PTE_GLOBAL;
    if( gpt_attr & GPT_COW        ) tsar_attr |= TSAR_PTE_COW;
    if( gpt_attr & GPT_SWAP       ) tsar_attr |= TSAR_PTE_SWAP;
    if( gpt_attr & GPT_LOCKED     ) tsar_attr |= TSAR_PTE_LOCKED;

    return tsar_attr;
}

///////////////////////////////////////////////////////////////////////////////////////
// This static function translates the TSAR attributes to the GPT attributes
///////////////////////////////////////////////////////////////////////////////////////
static inline uint32_t tsar2gpt( uint32_t tsar_attr )
{
    uint32_t gpt_attr = 0;

    if( tsar_attr & TSAR_PTE_MAPPED     ) gpt_attr |= GPT_MAPPED;
    if( tsar_attr & TSAR_PTE_MAPPED     ) gpt_attr |= GPT_READABLE;
    if( tsar_attr & TSAR_PTE_SMALL      ) gpt_attr |= GPT_SMALL;
    if( tsar_attr & TSAR_PTE_WRITABLE   ) gpt_attr |= GPT_WRITABLE;
    if( tsar_attr & TSAR_PTE_EXECUTABLE ) gpt_attr |= GPT_EXECUTABLE;
    if( tsar_attr & TSAR_PTE_CACHABLE   ) gpt_attr |= GPT_CACHABLE; 
    if( tsar_attr & TSAR_PTE_USER       ) gpt_attr |= GPT_USER;
    if( tsar_attr & TSAR_PTE_DIRTY      ) gpt_attr |= GPT_DIRTY;
    if( tsar_attr & TSAR_PTE_LOCAL      ) gpt_attr |= GPT_ACCESSED;
    if( tsar_attr & TSAR_PTE_REMOTE     ) gpt_attr |= GPT_ACCESSED;
    if( tsar_attr & TSAR_PTE_GLOBAL     ) gpt_attr |= GPT_GLOBAL;
    if( tsar_attr & TSAR_PTE_COW        ) gpt_attr |= GPT_COW;
    if( tsar_attr & TSAR_PTE_SWAP       ) gpt_attr |= GPT_SWAP;
    if( tsar_attr & TSAR_PTE_LOCKED     ) gpt_attr |= GPT_LOCKED;

    return gpt_attr;
}

///////////////////////////////////////////////////////////////////////////////////////
// The blocking hal_gpt_lock_pte() function implements a busy-waiting policy to get
// exclusive access to a specific GPT entry.
// - when non zero, the following variable defines the max number of iterations
//   in the busy waiting loop.
// - when zero, the watchdog mechanism is deactivated.
///////////////////////////////////////////////////////////////////////////////////////

#define GPT_LOCK_WATCHDOG   100000

/////////////////////////////////////
error_t hal_gpt_create( gpt_t * gpt )
{
    void * base;

    thread_t * this = CURRENT_THREAD;

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

// check page size
assert( __FUNCTION__, (CONFIG_PPM_PAGE_SIZE == 4096) , 
"the TSAR page size must be 4 Kbytes\n" );

    // allocates 8 Kbytes for PT1
        base = kmem_alloc( 13 , AF_ZERO );

        if( base == NULL ) 
    {
        printk("\n[PANIC] in %s : no memory for PT1 / process %x / cluster %x\n",
        __FUNCTION__, this->process->pid, local_cxy );
        return ENOMEM;
    }

    // initialze the GPT descriptor
    gpt->ptr              = base;
    gpt->pte1_wait_events = 0;
    gpt->pte1_wait_iters  = 0;
    gpt->pte2_wait_events = 0;
    gpt->pte2_wait_iters  = 0;

#if DEBUG_HAL_GPT_CREATE
cycle = (uint32_t)hal_get_cycles();
if( DEBUG_HAL_GPT_CREATE < cycle )
printk("\n[%s] thread[%x,%x] exit / pt1_base %x / pt1_ppn %x / cycle %d\n", 
__FUNCTION__, this->process->pid, this->trdid, 
base, ppm_base2ppn( XPTR( local_cxy , base ) ), cycle );
#endif

        return 0;

} // end hal_gpt_create()

///////////////////////////////////
void hal_gpt_destroy( gpt_t * gpt )
{
        uint32_t     ix1;
        uint32_t     ix2;
        uint32_t   * pt1;
    uint32_t     pte1;
    ppn_t        pt2_ppn;
    uint32_t   * pt2;
    uint32_t     attr;

    thread_t * this = CURRENT_THREAD;

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

    // get pointer on PT1
    pt1 = (uint32_t *)gpt->ptr;

    // scan the PT1
        for( ix1 = 0 ; ix1 < 2048 ; ix1++ )
        {
        pte1 = pt1[ix1];

                if( (pte1 & TSAR_PTE_MAPPED) != 0 )  // PTE1 mapped
        {
            if( (pte1 & TSAR_PTE_SMALL) == 0 )   // BIG page
            {
                printk("\n[WARNING] %s : valid PTE1 / thread[%x,%x] / ix1 %x\n", 
                __FUNCTION__, this->process->pid, this->trdid, ix1 );
            }
            else                             // PT2 exist
            {
                // get local pointer on PT2
                pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
                pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );

                // scan the PT2 
                for( ix2 = 0 ; ix2 < 512 ; ix2++ )
                {
                    attr = TSAR_MMU_ATTR_FROM_PTE2( pt2[2 * ix2] );

                    if( (attr & TSAR_PTE_MAPPED) != 0 )  // PTE2 mapped
                    {
                        printk("\n[WARNING] %s : valid PTE2 / thread[%x,%x] / ix1 %x / ix2 %x\n", 
                        __FUNCTION__, this->process->pid, this->trdid, ix1, ix2 );
                    }
                }

                // release the 4K bytes allocated for the PT2
                kmem_free( pt2 , 12 );
            }
        }
        }

    // release the 8K bytes allocated for PT1
    kmem_free( pt1 , 13 );

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

} // end hal_gpt_destroy()

////////////////////////////////////////////
error_t hal_gpt_lock_pte( xptr_t     gpt_xp,
                          vpn_t      vpn,
                          uint32_t * attr,
                          ppn_t    * ppn )
{
    uint32_t          * pt1;             // local pointer on PT1 base 
    xptr_t              pte1_xp;         // extended pointer on PT1[x1] entry
        uint32_t            pte1;            // value of PT1[x1] entry
    uint32_t          * pt2;             // local pointer on PT2 base
        ppn_t               pt2_ppn;         // PPN of page containing PT2
        xptr_t              pte2_xp;         // extended pointer on PT2[ix2].attr
    uint32_t            pte2_attr;       // PT2[ix2].attr current value   
    uint32_t            pte2_ppn;        // PT2[ix2].ppn current value   
        bool_t              success;         // used for both PTE1 and PTE2 mapping 
    uint32_t            count;           // watchdog
    uint32_t            sr_save;         // for critical section

    // get cluster and local pointer on GPT
    cxy_t   gpt_cxy = GET_CXY( gpt_xp );
    gpt_t * gpt_ptr = GET_PTR( gpt_xp );

#if DEBUG_HAL_GPT_LOCK_PTE
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
// if( DEBUG_HAL_GPT_LOCK_PTE < cycle )
if( (vpn == 0xc1fff) && (gpt_cxy == 0x1) )
printk("\n[%s] thread[%x,%x] enters / vpn %x in cluster %x / cycle %d\n", 
__FUNCTION__, this->process->pid, this->trdid, vpn, gpt_cxy, cycle );
#endif

    // get indexes in PTI & PT2 from vpn
    uint32_t  ix1 = TSAR_MMU_IX1_FROM_VPN( vpn );
    uint32_t  ix2 = TSAR_MMU_IX2_FROM_VPN( vpn );

    // get local pointer on PT1
    pt1 = hal_remote_lpt( XPTR( gpt_cxy , &gpt_ptr->ptr ) );

    // build extended pointer on PTE1 == PT1[ix1]
        pte1_xp = XPTR( gpt_cxy , &pt1[ix1] );

    // get current PT1 entry value
    pte1 = hal_remote_l32( pte1_xp );

    // If PTE1 is unmapped, the calling thread try to map this PTE1.
    // To prevent multiple concurrent PT2 allocations, only the thread that 
    // successfully locked the PTE1 allocates a new PT2 and updates the PTE1.
    // All other threads simply wait until the missing PTE1 is mapped.

    if( (pte1 & TSAR_PTE_MAPPED) == 0 )  
        {
        if( (pte1 & TSAR_PTE_LOCKED)  == 0 )
        { 
            // try to atomically lock the PTE1
            success = hal_remote_atomic_cas( pte1_xp,
                                             pte1,
                                             TSAR_PTE_LOCKED );
        }
        else
        {
            success = false;
        }
    
        if( success )       // winner thread allocates one 4 Kbytes page for PT2
        {
            // enter critical section
            hal_disable_irq( &sr_save );

            // allocate a 4K bytes PT2
            pt2       = kmem_remote_alloc( gpt_cxy , 12 , AF_ZERO );

            if( pt2 == NULL )
            {
                printk("\n[ERROR] in %s : cannot allocate memory for PT2 in cluster %d\n",
                __FUNCTION__, gpt_cxy );
                return -1;
            }

            // get the PT2 PPN
            pt2_ppn = ppm_base2ppn( XPTR( gpt_cxy , pt2 ) );

            // build  PTE1
            pte1 = TSAR_PTE_MAPPED | TSAR_PTE_SMALL | pt2_ppn;

            // set the PTE1 value in PT1 / this unlocks the PTE1
            hal_remote_s32( pte1_xp , pte1 );
            hal_fence();

            // exit critical section
            hal_restore_irq( sr_save );

#if DEBUG_HAL_GPT_LOCK_PTE
// if( DEBUG_HAL_GPT_LOCK_PTE < cycle )
if( (vpn == 0xc1fff) && (gpt_cxy == 0x1) )
printk("\n[%s] PTE1 unmapped : winner thread[%x,%x] allocates a PT2 for vpn %x in cluster %x\n", 
__FUNCTION__, this->process->pid, this->trdid, vpn, gpt_cxy );
#endif

        }
        else                 // other threads wait until PTE1 mapped by the winner
        {

#if DEBUG_HAL_GPT_LOCK_PTE
// if( DEBUG_HAL_GPT_LOCK_PTE < cycle )
if( (vpn == 0xc1fff) && (gpt_cxy == 0x1) )
printk("\n[%s] PTE1 unmapped : loser thread[%x,%x] wait PTE1 for vpn %x in cluster %x\n", 
__FUNCTION__, this->process->pid, this->trdid, vpn, gpt_cxy );
#endif

            count = 0;
            do
            {
                // get current pte1 value
                pte1 = hal_remote_l32( pte1_xp );

                // check iterations number
                if( count > GPT_LOCK_WATCHDOG )
                {
                    thread_t * this  = CURRENT_THREAD;
                    uint32_t   cycle = (uint32_t)hal_get_cycles();
                    printk("\n[PANIC] in %s for PTE1 after %d iterations\n"
                    "  thread[%x,%x] / vpn %x / cluster %x / pte1 %x / cycle %d\n",
                    __FUNCTION__, count, this->process->pid, this->trdid,
                    vpn, gpt_cxy, pte1, cycle );

                    xptr_t process_xp = cluster_get_process_from_pid_in_cxy( gpt_cxy,
                                                                             this->process->pid );
                    hal_vmm_display( process_xp , true );
 
                    hal_core_sleep();
                }

                // increment watchdog
                count++;
            }
            while( (pte1 & TSAR_PTE_MAPPED) == 0 ); 

#if CONFIG_INSTRUMENTATION_GPT
hal_remote_atomic_add( XPTR( gpt_cxy , &gpt_ptr->pte1_wait_events ) , 1 );
hal_remote_atomic_add( XPTR( gpt_cxy , &gpt_ptr->pte1_wait_iters  ) , count );
#endif


#if DEBUG_HAL_GPT_LOCK_PTE
// if( DEBUG_HAL_GPT_LOCK_PTE < cycle )
if( (vpn == 0xc1fff) && (gpt_cxy == 0x1) )
printk("\n[%s] PTE1 unmapped : loser thread[%x,%x] get PTE1 for vpn %x in cluster %x\n", 
__FUNCTION__, this->process->pid, this->trdid, vpn, gpt_cxy );
#endif
        }
    }  // end if pte1 unmapped

    // This code is executed by all calling threads
 
// check PTE1 : only small and mapped pages can be locked
assert( __FUNCTION__, (pte1 & (TSAR_PTE_SMALL | TSAR_PTE_MAPPED)) , "cannot lock a big or unmapped page\n");

#if DEBUG_HAL_GPT_LOCK_PTE 
// if( DEBUG_HAL_GPT_LOCK_PTE < cycle )
if( (vpn == 0xc1fff) && (gpt_cxy == 0x1) )
printk("\n[%s] thread[%x,%x] get pte1 %x for vpn %x in cluster %x\n", 
__FUNCTION__, this->process->pid, this->trdid, pte1, vpn, gpt_cxy );
#endif

    // get pointer on PT2 base 
    pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
    pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );

    // build extended pointers on PT2[ix2].attr  
    pte2_xp = XPTR( gpt_cxy , &pt2[2 * ix2] );

    // initialize external loop watchdog
    count = 0;

    // in this busy waiting loop, each thread try to atomically
    // lock the PTE2, after checking that the PTE2 is not locked 
    
    do
    {
        // get current value of pte2_attr
        pte2_attr = hal_remote_l32( pte2_xp );

        // check loop watchdog
        if( count > GPT_LOCK_WATCHDOG )
        {
            thread_t * this  = CURRENT_THREAD;
            uint32_t   cycle = (uint32_t)hal_get_cycles();
            printk("\n[PANIC] in %s for PTE2 after %d iterations\n"
            "  thread[%x,%x] / vpn %x / cluster %x / pte2_attr %x / cycle %d\n",
            __FUNCTION__, count, this->process->pid, this->trdid,
            vpn, gpt_cxy, pte2_attr, cycle );

                    xptr_t process_xp = cluster_get_process_from_pid_in_cxy( gpt_cxy,
                                                                             this->process->pid );
                    hal_vmm_display( process_xp , true );
 
            hal_core_sleep();
        }

        // increment loop watchdog
        count++;

        if( (pte2_attr & TSAR_PTE_LOCKED) == 0 )
        {
            // try to atomically set the TSAR_PTE_LOCKED attribute    
                    success = hal_remote_atomic_cas( pte2_xp,
                                             pte2_attr,
                                             (pte2_attr | TSAR_PTE_LOCKED) );
        }
        else
        {
            success = false;
        }
    }
    while( success == false );

#if CONFIG_INSTRUMENTATION_GPT
hal_remote_atomic_add( XPTR( gpt_cxy , &gpt_ptr->pte2_wait_events ) , 1 );
hal_remote_atomic_add( XPTR( gpt_cxy , &gpt_ptr->pte2_wait_iters  ) , count );
#endif

    // get PTE2.ppn
    pte2_ppn = hal_remote_l32( pte2_xp + 4 );

#if DEBUG_HAL_GPT_LOCK_PTE
cycle = (uint32_t)hal_get_cycles();
// if( DEBUG_HAL_GPT_LOCK_PTE < cycle )
if( (vpn == 0xc1fff) && (gpt_cxy == 0x1) )
printk("\n[%s] thread[%x,%x] success / vpn %x in cluster %x / attr %x / ppn %x / cycle %d\n", 
__FUNCTION__, this->process->pid, this->trdid, vpn, gpt_cxy, pte2_attr, pte2_ppn, cycle );
#endif
    
    // return PPN and GPT attributes
    *ppn  = pte2_ppn & ((1<<TSAR_MMU_PPN_WIDTH)-1);
    *attr = tsar2gpt( pte2_attr );
    return 0;

}  // end hal_gpt_lock_pte()

////////////////////////////////////////
void hal_gpt_unlock_pte( xptr_t  gpt_xp,
                         vpn_t   vpn )
{
    uint32_t * pt1;             // local pointer on PT1 base
    xptr_t     pte1_xp;         // extended pointer on PT1[ix1]
        uint32_t   pte1;            // value of PT1[ix1] entry

    uint32_t * pt2;             // PT2 base address
        ppn_t      pt2_ppn;         // PPN of page containing PT2
        xptr_t     pte2_xp;         // extended pointer on PT2[ix2].attr 
        uint32_t   pte2_attr;       // PTE2 attribute

    // get cluster and local pointer on GPT
    cxy_t   gpt_cxy = GET_CXY( gpt_xp );
    gpt_t * gpt_ptr = GET_PTR( gpt_xp );

    // compute indexes in P1 and PT2
    uint32_t  ix1 = TSAR_MMU_IX1_FROM_VPN( vpn ); 
    uint32_t  ix2 = TSAR_MMU_IX2_FROM_VPN( vpn );

    // get local pointer on PT1
    pt1 = hal_remote_lpt( XPTR( gpt_cxy , &gpt_ptr->ptr ) );

    // build extended pointer on PTE1 == PT1[ix1]
        pte1_xp = XPTR( gpt_cxy , &pt1[ix1] );

    // get current pte1 value
    pte1 = hal_remote_l32( pte1_xp );

assert( __FUNCTION__, ((pte1 & TSAR_PTE_MAPPED) != 0),
"PTE1 for vpn %x in cluster %x is unmapped / pte1 = %x\n", vpn, gpt_cxy, pte1 );

assert( __FUNCTION__, ((pte1 & TSAR_PTE_SMALL ) != 0),
"PTE1 for vpn %x in cluster %x is not small / pte1 = %x\n", vpn, gpt_cxy, pte1 );

    // get pointer on PT2 base  
        pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
        pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );

    // build extended pointers on PT2[ix2].attr  
    pte2_xp = XPTR( gpt_cxy , &pt2[2 * ix2] );

    // get PT2[ix2].attr
    pte2_attr = hal_remote_l32( pte2_xp );

assert( __FUNCTION__, ((pte2_attr & TSAR_PTE_LOCKED) != 0),
"PTE2 for vpn %x in cluster %x is unlocked / pte2_attr = %x\n", vpn, gpt_cxy, pte2_attr );

    // reset TSAR_PTE_LOCKED attribute 
    hal_remote_s32( pte2_xp , pte2_attr & ~TSAR_PTE_LOCKED );

#if DEBUG_HAL_GPT_LOCK_PTE
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
// if( DEBUG_HAL_GPT_LOCK_PTE < cycle )
if( (vpn == 0xc5fff) && (gpt_cxy == 0x1) )
printk("\n[%s] thread[%x,%x] unlocks vpn %x in cluster %x / cycle %d\n", 
__FUNCTION__, this->process->pid, this->trdid, vpn, gpt_cxy, cycle );
#endif

}  // end hal_gpt_unlock_pte()


///////////////////////////////////////
void hal_gpt_set_pte( xptr_t    gpt_xp,
                      vpn_t     vpn,
                      uint32_t  attr,  
                      ppn_t     ppn )
{
    cxy_t               gpt_cxy;             // target GPT cluster
    gpt_t             * gpt_ptr;             // target GPT local pointer

    uint32_t          * pt1;                 // local pointer on PT1 base
        xptr_t              pte1_xp;             // extended pointer on PT1 entry
        uint32_t            pte1;                // PT1 entry value if PTE1

        uint32_t          * pt2;                 // local pointer on PT2 base
        ppn_t               pt2_ppn;             // PPN of PT2
    xptr_t              pte2_attr_xp;        // extended pointer on PT2[ix2].attr
    xptr_t              pte2_ppn_xp;         // extended pointer on PT2[ix2].ppn
    uint32_t            pte2_attr;           // current value of PT2[ix2].attr

    uint32_t            ix1;                 // index in PT1
    uint32_t            ix2;                 // index in PT2

    uint32_t            tsar_attr;           // PTE attributes for TSAR MMU
        uint32_t            small;               // requested PTE is for a small page

    // get cluster and local pointer on GPT
    gpt_cxy = GET_CXY( gpt_xp );
    gpt_ptr = GET_PTR( gpt_xp );

    // compute indexes in PT1 and PT2
    ix1 = TSAR_MMU_IX1_FROM_VPN( vpn );
    ix2 = TSAR_MMU_IX2_FROM_VPN( vpn );

#if DEBUG_HAL_GPT_SET_PTE
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( DEBUG_HAL_GPT_SET_PTE < cycle )
printk("\n[%s] thread[%x,%x] enter gpt (%x,%x) / vpn %x / attr %x / ppn %x\n", 
__FUNCTION__, this->process->pid, this->trdid, gpt_cxy, &gpt_ptr->ptr, vpn, attr, ppn );
#endif

        small = attr & GPT_SMALL;

    // get local pointer on PT1
    pt1 = hal_remote_lpt( XPTR( gpt_cxy , &gpt_ptr->ptr ) );

    // compute tsar attributes from generic attributes
    tsar_attr = gpt2tsar( attr );

    // build extended pointer on PTE1 = PT1[ix1]
        pte1_xp = XPTR( gpt_cxy , &pt1[ix1] );

    // get current pte1 value
    pte1 = hal_remote_l32( pte1_xp );

        if( small == 0 )  ///////////////// map a big page in PT1 
    {

// check PT1 entry not mapped
assert( __FUNCTION__, (pte1 == 0) , "try to set a big page in an already mapped PTE1\n" );

// check VPN aligned
assert( __FUNCTION__, (ix2 == 0) , "illegal vpn for a big page\n" );

// check PPN aligned
assert( __FUNCTION__, ((ppn & 0x1FF) == 0) , "illegal ppn for a big page\n" );

        // set the PTE1 value in PT1
        pte1 = (tsar_attr  & TSAR_MMU_PTE1_ATTR_MASK) | ((ppn >> 9) & TSAR_MMU_PTE1_PPN_MASK);
        hal_remote_s32( pte1_xp , pte1 );
                hal_fence();

#if DEBUG_HAL_GPT_SET_PTE
if( DEBUG_HAL_GPT_SET_PTE < cycle )
printk("\n[%s] thread[%x,%x] map PTE1 / cxy %x / ix1 %x / pt1 %x / pte1 %x\n", 
__FUNCTION__, this->process->pid, this->trdid, gpt_cxy, ix1, pt1, pte1 );
#endif

        }
    else      ///////////////// map a small page in PT2
    {

// PTE1 must be mapped because PTE2 must be locked
assert( __FUNCTION__, (pte1 & TSAR_PTE_MAPPED),
"PTE1 for vpn %x in cluster %x must be mapped / pte1 = %x\n", vpn, gpt_cxy, pte1 );

        // get PT2 base 
            pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
            pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );

        // build extended pointers on PT2[ix2].attr and PT2[ix2].ppn
        pte2_attr_xp = XPTR( gpt_cxy , &pt2[2 * ix2] );
        pte2_ppn_xp  = XPTR( gpt_cxy , &pt2[2 * ix2 + 1] );

        // get current value of PTE2.attr
        pte2_attr = hal_remote_l32( pte2_attr_xp );

// PTE2 must be locked
assert( __FUNCTION__, (pte2_attr & TSAR_PTE_LOCKED),
"PTE2 for vpn %x in cluster %x must be locked / pte2_attr = %x\n", vpn, gpt_cxy, pte2_attr );
 
        // set PTE2 in PT2 (in this order)
            hal_remote_s32( pte2_ppn_xp  , ppn );
            hal_fence();
            hal_remote_s32( pte2_attr_xp , tsar_attr );
            hal_fence();

#if DEBUG_HAL_GPT_SET_PTE
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( DEBUG_HAL_GPT_SET_PTE < cycle )
printk("\n[%s] thread[%x,%x] map PTE2 / cxy %x / ix2 %x / pt2 %x / attr %x / ppn %x\n", 
__FUNCTION__, this->process->pid, this->trdid, gpt_cxy, ix2, pt2, tsar_attr, ppn );
#endif

    }
} // end of hal_gpt_set_pte()

///////////////////////////////////////
void hal_gpt_reset_pte( xptr_t  gpt_xp,
                        vpn_t   vpn )
{
    cxy_t      gpt_cxy;        // target GPT cluster
    gpt_t    * gpt_ptr;        // target GPT local pointer

    uint32_t   ix1;            // index in PT1 
    uint32_t   ix2;            // index in PT2

    uint32_t * pt1;            // PT1 base address
    xptr_t     pte1_xp;        // extended pointer on PT1[ix1]
    uint32_t   pte1;           // PT1 entry value

    uint32_t * pt2;            // PT2 base address
    ppn_t      pt2_ppn;        // PPN of PT2 
    xptr_t     pte2_attr_xp;   // extended pointer on PT2[ix2].attr
    xptr_t     pte2_ppn_xp;    // extended pointer on PT2[ix2].ppn

    // get cluster and local pointer on GPT
    gpt_cxy = GET_CXY( gpt_xp );
    gpt_ptr = GET_PTR( gpt_xp );

    // get ix1 & ix2 indexes
    ix1 = TSAR_MMU_IX1_FROM_VPN( vpn );
    ix2 = TSAR_MMU_IX2_FROM_VPN( vpn );

    // get local pointer on PT1 base
    pt1 = hal_remote_lpt( XPTR( gpt_cxy , &gpt_ptr->ptr ) );

    // build extended pointer on PTE1 = PT1[ix1]
        pte1_xp = XPTR( gpt_cxy , &pt1[ix1] );

    // get current PTE1 value
    pte1 = hal_remote_l32( pte1_xp );

        if( (pte1 & TSAR_PTE_MAPPED) == 0 )     // PTE1 unmapped => do nothing
        {
                return;
        }

        if( (pte1 & TSAR_PTE_SMALL) == 0 )      // it's a PTE1 => unmap it from PT1
        {
        hal_remote_s32( pte1_xp , 0 );
            hal_fence();

#if DEBUG_HAL_GPT_RESET_PTE
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( DEBUG_HAL_GPT_RESET_PTE < cycle )
printk("\n[%s] thread[%x,%x] unmap PTE1 / cxy %x / vpn %x / ix1 %x\n", 
__FUNCTION__, this->process->pid, this->trdid, gpt_cxy, vpn, ix1 );
#endif

        return;
        }
    else                                    // it's a PTE2 => unmap it from PT2
    {
        // get PT2 base
        pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
        pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );
        
        // build extended pointer on PT2[ix2].attr and PT2[ix2].ppn
        pte2_attr_xp = XPTR( gpt_cxy , &pt2[2 * ix2] );
        pte2_ppn_xp  = XPTR( gpt_cxy , &pt2[2 * ix2 + 1] );

        // unmap the PTE2
        hal_remote_s32( pte2_attr_xp , 0 );
            hal_fence();        
        hal_remote_s32( pte2_ppn_xp  , 0 );
            hal_fence();        

#if DEBUG_HAL_GPT_RESET_PTE
thread_t * this  = CURRENT_THREAD;
uint32_t   cycle = (uint32_t)hal_get_cycles();
if( DEBUG_HAL_GPT_RESET_PTE < cycle )
printk("\n[%s] thread[%x,%x] unmap PTE2 / cxy %x / vpn %x / ix2 %x\n", 
__FUNCTION__, this->process->pid, this->trdid, gpt_cxy, vpn, ix2 );
#endif

        return;
    }
}  // end hal_gpt_reset_pte()

////////////////////////////////////////
void hal_gpt_get_pte( xptr_t     gpt_xp,
                      vpn_t      vpn,
                      uint32_t * attr,
                      ppn_t    * ppn )
{
    uint32_t * pt1;            // local pointer on PT1 base
    uint32_t   pte1;           // PTE1 value

    uint32_t * pt2;            // local pointer on PT2 base
    ppn_t      pt2_ppn;        // PPN of page containing the PT2
    xptr_t     pte2_attr_xp;   // extended pointer on PT2[ix2].attr
    xptr_t     pte2_ppn_xp;    // extended pointer on PT2[ix2].ppn
    uint32_t   pte2_attr;      // current value of PT2[ix2].attr
    ppn_t      pte2_ppn;       // current value of PT2[ix2].ppn 

    // get cluster and local pointer on GPT
    cxy_t   gpt_cxy = GET_CXY( gpt_xp );
    gpt_t * gpt_ptr = GET_PTR( gpt_xp );

    // compute indexes in PT1 and PT2
    uint32_t   ix1 = TSAR_MMU_IX1_FROM_VPN( vpn );
    uint32_t   ix2 = TSAR_MMU_IX2_FROM_VPN( vpn );

    // get PT1 base
    pt1 = hal_remote_lpt( XPTR( gpt_cxy , &gpt_ptr->ptr ) );
    
    // get pte1
    pte1 = hal_remote_l32( XPTR( gpt_cxy , &pt1[ix1] ) );

    // check PTE1 mapped
        if( (pte1 & TSAR_PTE_MAPPED) == 0 )   // PTE1 unmapped
        {
                *attr = 0;
                *ppn  = 0;
        return;
        }

    // access GPT
        if( (pte1 & TSAR_PTE_SMALL) == 0 )     // it's a PTE1
        {
        // get PPN & ATTR
                *attr = tsar2gpt( TSAR_MMU_ATTR_FROM_PTE1( pte1 ) );
        *ppn  = TSAR_MMU_PPN1_FROM_PTE1( pte1 ) | (vpn & ((1<<TSAR_MMU_IX2_WIDTH)-1));
        }
    else                                  // it's a PTE2
    {
        // compute PT2 base address
        pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
        pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );

        // build extended pointer on PT2[ix2].attr and PT2[ix2].ppn
        pte2_attr_xp = XPTR( gpt_cxy , &pt2[2 * ix2] );
        pte2_ppn_xp  = XPTR( gpt_cxy , &pt2[2 * ix2 + 1] );

        // get current value of PTE2.attr & PTE2.ppn
        pte2_attr = hal_remote_l32( pte2_attr_xp );
        pte2_ppn  = hal_remote_l32( pte2_ppn_xp );

        // return PPN & GPT attributes 
        *ppn  = pte2_ppn & ((1<<TSAR_MMU_PPN_WIDTH)-1);
        *attr = tsar2gpt( pte2_attr );
    }
} // end hal_gpt_get_pte()


///////////////////////////////////////////
error_t hal_gpt_pte_copy( gpt_t  * dst_gpt,
                          vpn_t    dst_vpn,
                          xptr_t   src_gpt_xp,
                          vpn_t    src_vpn,
                          bool_t   cow,
                          ppn_t  * ppn,
                          bool_t * mapped )
{
    uint32_t     src_ix1;   // index in SRC PT1
    uint32_t     src_ix2;   // index in SRC PT2

    uint32_t     dst_ix1;   // index in DST PT1
    uint32_t     dst_ix2;   // index in DST PT2

    cxy_t        src_cxy;   // SRC GPT cluster
    gpt_t      * src_gpt;   // SRC GPT local pointer

        uint32_t   * src_pt1;   // local pointer on SRC PT1 
        uint32_t   * dst_pt1;   // local pointer on DST PT1

    uint32_t   * src_pt2;   // local pointer on SRC PT2
    uint32_t   * dst_pt2;   // local pointer on DST PT2

    uint32_t     src_pte1;
    uint32_t     dst_pte1;

    uint32_t     src_pte2_attr;
    uint32_t     src_pte2_ppn;

    page_t     * page;
    xptr_t       page_xp;

    ppn_t        src_pt2_ppn;
    ppn_t        dst_pt2_ppn;

    // get remote src_gpt cluster and local pointer
    src_cxy = GET_CXY( src_gpt_xp );
    src_gpt = GET_PTR( src_gpt_xp );

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

    // get remote src_pt1 and local dst_pt1
    src_pt1 = (uint32_t *)hal_remote_lpt( XPTR( src_cxy , &src_gpt->ptr ) );
    dst_pt1 = (uint32_t *)dst_gpt->ptr;

    // check src_pt1 and dst_pt1 existence
    assert( __FUNCTION__, (src_pt1 != NULL) , "src_pt1 does not exist\n");
    assert( __FUNCTION__, (dst_pt1 != NULL) , "dst_pt1 does not exist\n");

    // compute SRC indexes
    src_ix1 = TSAR_MMU_IX1_FROM_VPN( src_vpn );
    src_ix2 = TSAR_MMU_IX2_FROM_VPN( src_vpn );

    // compute DST indexes
    dst_ix1 = TSAR_MMU_IX1_FROM_VPN( dst_vpn );
    dst_ix2 = TSAR_MMU_IX2_FROM_VPN( dst_vpn );

    // get src_pte1 
    src_pte1 = hal_remote_l32( XPTR( src_cxy , &src_pt1[src_ix1] ) );

    // do nothing if src_pte1 not MAPPED or not SMALL
        if( (src_pte1 & TSAR_PTE_MAPPED) && (src_pte1 & TSAR_PTE_SMALL) )    
    {
        // get dst_pt1 entry
        dst_pte1 = dst_pt1[dst_ix1];

        // map dst_pte1 when this entry is not mapped
        if( (dst_pte1 & TSAR_PTE_MAPPED) == 0 ) 
        { 
            // allocate one 4K bytes physical page for a new PT2
                dst_pt2   = kmem_alloc( 12 , AF_ZERO );

            if( dst_pt2 == NULL )
            {
                        printk("\n[ERROR] in %s : cannot allocate PT2\n", __FUNCTION__ );
                return -1;
            }

            // build extended pointer on page descriptor
            page_xp = XPTR( local_cxy , page );

            // get PPN for this new PT2 
            dst_pt2_ppn = ppm_base2ppn( XPTR( local_cxy , dst_pt2 ) );

            // build new dst_pte1
            dst_pte1 = TSAR_PTE_MAPPED | TSAR_PTE_SMALL | dst_pt2_ppn;

            // register it in DST_GPT
            dst_pt1[dst_ix1] = dst_pte1;
        }

        // get pointer on src_pt2
        src_pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( src_pte1 );
        src_pt2     = GET_PTR( ppm_ppn2base( src_pt2_ppn ) );

        // get pointer on dst_pt2
        dst_pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( dst_pte1 );
        dst_pt2     = GET_PTR( ppm_ppn2base( dst_pt2_ppn ) );

        // get attr and ppn from SRC_PT2
        src_pte2_attr = hal_remote_l32( XPTR( src_cxy , &src_pt2[2 * src_ix2]     ) );
        src_pte2_ppn  = hal_remote_l32( XPTR( src_cxy , &src_pt2[2 * src_ix2 + 1] ) );

        // do nothing if src_pte2 not MAPPED
        if( (src_pte2_attr & TSAR_PTE_MAPPED) != 0 )  
        {
            // set PPN in DST PTE2
            dst_pt2[2 * dst_ix2 + 1] = src_pte2_ppn;
                        
            // set attributes in DST PTE2         
            if( cow && (src_pte2_attr & TSAR_PTE_WRITABLE) ) 
            {
                dst_pt2[2 * dst_ix2] = (src_pte2_attr | TSAR_PTE_COW) & (~TSAR_PTE_WRITABLE);
            } 
            else
            {
                dst_pt2[2 * dst_ix2] = src_pte2_attr;
            }

            // return "successfully copied"
            *mapped = true;
            *ppn    = src_pte2_ppn;
        
#if DEBUG_HAL_GPT_COPY
cycle = (uint32_t)hal_get_cycles;
if( DEBUG_HAL_GPT_COPY < cycle )
printk("\n[%s] thread[%x,%x] exit / copy done for src_vpn %x / dst_vpn %x / cycle %d\n", 
__FUNCTION__, this->process->pid, this->trdid, src_vpn, dst_vpn, cycle );
#endif

            hal_fence();

            return 0;
        }   // end if PTE2 mapped
    }   // end if PTE1 mapped

    // return "nothing done"
    *mapped = false;
    *ppn    = 0;
   
#if DEBUG_HAL_GPT_COPY
cycle = (uint32_t)hal_get_cycles;
if( DEBUG_HAL_GPT_COPY < cycle )
printk("\n[%s] thread[%x,%x] exit / nothing done / cycle %d\n", 
__FUNCTION__, this->process->pid, this->trdid, cycle );
#endif

    hal_fence();

    return 0;

}  // end hal_gpt_pte_copy()

/////////////////////////////////////////
void hal_gpt_set_cow( xptr_t  gpt_xp,
                      vpn_t   vpn_base,
                      vpn_t   vpn_size )
{
    cxy_t      gpt_cxy;
    gpt_t    * gpt_ptr;

    uint32_t   ix1;       // current
    uint32_t   ix2;       // current

    vpn_t      vpn_min;
    vpn_t      vpn_max;   // included

    uint32_t   ix1_min;
    uint32_t   ix1_max;   // included

    uint32_t   ix2_min;
    uint32_t   ix2_max;   // included

    uint32_t * pt1;
    uint32_t   pte1;

    uint32_t * pt2;
    ppn_t      pt2_ppn;
    uint32_t   attr;

    // get GPT cluster and local pointer
    gpt_cxy = GET_CXY( gpt_xp );
    gpt_ptr = GET_PTR( gpt_xp );

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

    // get PT1 pointer
    pt1 = (uint32_t *)hal_remote_lpt( XPTR( gpt_cxy , &gpt_ptr->ptr ) );

#if (DEBUG_HAL_GPT_SET_COW & 1)
if(DEBUG_HAL_GPT_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] get pt1 = %x\n", 
__FUNCTION__, this->process->pid, this->trdid, pt1 );
#endif

    vpn_min = vpn_base;
    vpn_max = vpn_base + vpn_size - 1;

    ix1_min = TSAR_MMU_IX1_FROM_VPN( vpn_base );
    ix1_max = TSAR_MMU_IX1_FROM_VPN( vpn_max );

    for( ix1 = ix1_min ; ix1 <= ix1_max ; ix1++ )
    {

#if (DEBUG_HAL_GPT_SET_COW & 1)
if(DEBUG_HAL_GPT_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] : &pt1[%x] = %x\n", 
__FUNCTION__, this->process->pid, this->trdid, ix1,  &pt1[ix1] );
#endif
        // get PTE1 value
        pte1 = hal_remote_l32( XPTR( gpt_cxy , &pt1[ix1] ) );

#if (DEBUG_HAL_GPT_SET_COW & 1)
if(DEBUG_HAL_GPT_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] : pt1[%x] = %x\n", 
__FUNCTION__, this->process->pid, this->trdid, ix1, pte1 );
#endif

        // only MAPPED & SMALL PTEs are modified
            if( (pte1 & TSAR_PTE_MAPPED) && (pte1 & TSAR_PTE_SMALL) )
        {
            // get PT2 pointer
            pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
            pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );

#if (DEBUG_HAL_GPT_SET_COW & 1)
if(DEBUG_HAL_GPT_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] : get pt2 = %x\n", 
__FUNCTION__, this->process->pid, this->trdid, pt2 );
#endif
            ix2_min = (ix1 == ix1_min) ? TSAR_MMU_IX2_FROM_VPN(vpn_min) : 0;
            ix2_max = (ix1 == ix1_max) ? TSAR_MMU_IX2_FROM_VPN(vpn_max) : 511;
 
            for( ix2 = ix2_min ; ix2 <= ix2_max ; ix2++ )
            {

#if (DEBUG_HAL_GPT_SET_COW & 1)
if(DEBUG_HAL_GPT_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] : &pte2[%x] = %x\n", 
__FUNCTION__, this->process->pid, this->trdid, 2*ix2, &pt2[2*ix2] );
#endif
                // get current PTE2 attributes
                attr = hal_remote_l32( XPTR( gpt_cxy , &pt2[2*ix2] ) );

#if (DEBUG_HAL_GPT_SET_COW & 1)
if(DEBUG_HAL_GPT_SET_COW < cycle )
printk("\n[%s] thread[%x,%x] : pte2[%x] (attr) = %x\n", 
__FUNCTION__, this->process->pid, this->trdid, 2*ix2, attr );
#endif
                // only MAPPED PTEs are modified        
                if( attr & TSAR_PTE_MAPPED ) 
                {
                    attr = (attr | TSAR_PTE_COW) & (~TSAR_PTE_WRITABLE);
                    hal_remote_s32( XPTR( gpt_cxy , &pt2[2*ix2] ) , attr );
                }
            }  // end loop on ix2
        }
    }  // end loop on ix1

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

}  // end hal_gpt_set_cow()

//////////////////////////////////////////
void hal_gpt_update_pte( xptr_t    gpt_xp,
                         vpn_t     vpn,
                         uint32_t  attr,     // generic GPT attributes
                         ppn_t     ppn )
{
    uint32_t          * pt1;                 // PT1 base addres
        uint32_t            pte1;                // PT1 entry value

        ppn_t               pt2_ppn;             // PPN of PT2
        uint32_t          * pt2;                 // PT2 base address
    xptr_t              pte2_attr_xp;        // exended pointer on pte2.attr
    xptr_t              pte2_ppn_xp;         // exended pointer on pte2.ppn

    uint32_t            ix1;                 // index in PT1
    uint32_t            ix2;                 // index in PT2

// check MAPPED, SMALL, and not LOCKED in attr argument
assert( __FUNCTION__, ((attr & GPT_MAPPED) != 0), "attribute MAPPED must be set in new attributes\n" );
assert( __FUNCTION__, ((attr & GPT_SMALL ) != 0), "attribute SMALL  must be set in new attributes\n" );
assert( __FUNCTION__, ((attr & GPT_LOCKED) == 0), "attribute LOCKED must not be set in new attributes\n" );

    // get cluster and local pointer on remote GPT
    cxy_t   gpt_cxy = GET_CXY( gpt_xp );
    gpt_t * gpt_ptr = GET_PTR( gpt_xp );

    // compute indexes in PT1 and PT2
    ix1 = TSAR_MMU_IX1_FROM_VPN( vpn );
    ix2 = TSAR_MMU_IX2_FROM_VPN( vpn );

    // get PT1 base
    pt1 = (uint32_t *)hal_remote_lpt( XPTR( gpt_cxy , &gpt_ptr->ptr ) );

    // get PTE1 value
    pte1 = hal_remote_l32( XPTR( gpt_cxy , &pt1[ix1] ) );

// check MAPPED and SMALL in target PTE1
assert( __FUNCTION__, ((pte1 & TSAR_PTE_MAPPED) != 0), "attribute MAPPED must be set in target PTE1\n" );
assert( __FUNCTION__, ((pte1 & TSAR_PTE_SMALL ) != 0), "attribute SMALL  must be set in target PTE1\n" );

    // get PT2 base 
    pt2_ppn = TSAR_MMU_PPN2_FROM_PTE1( pte1 );
    pt2     = GET_PTR( ppm_ppn2base( pt2_ppn ) );

    // build extended pointers on PT2[ix2].attr and PT2[ix2].ppn
    pte2_attr_xp = XPTR( gpt_cxy , &pt2[2 * ix2] );
    pte2_ppn_xp  = XPTR( gpt_cxy , &pt2[2 * ix2 + 1] );

    
// check MAPPED in target PTE2
assert( __FUNCTION__, ((hal_remote_l32(pte2_attr_xp) & TSAR_PTE_MAPPED) != 0),
"attribute MAPPED must be set in target PTE2\n" );

    // set PTE2 in this order
        hal_remote_s32( pte2_ppn_xp , ppn );
        hal_fence();
        hal_remote_s32( pte2_attr_xp , gpt2tsar( attr ) );
        hal_fence();

} // end hal_gpt_update_pte()