1.1 --- a/VMS.c	Fri Feb 10 18:35:00 2012 +0100
     1.2 +++ b/VMS.c	Thu Mar 08 19:02:16 2012 +0100
     1.3 @@ -193,31 +193,6 @@
     1.4     #endif
     1.5  
     1.6     #ifdef MEAS__PERF_COUNTERS
     1.7 -/*
     1.8 -   _VMSMasterEnv->counter_history = VMS__malloc(10*sizeof(void*));
     1.9 -   _VMSMasterEnv->counter_history_array_info = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->counter_history),10);
    1.10 -*/
    1.11 -   //printf("Creating HW counters...");
    1.12 -/*
    1.13 -   FILE* output;
    1.14 -   int n;
    1.15 -   char filename[255];    
    1.16 -    for(n=0;n<255;n++)
    1.17 -    {
    1.18 -        sprintf(filename, "./counters/Counters.%d.csv",n);
    1.19 -        output = fopen(filename,"r");
    1.20 -        if(output)
    1.21 -        {
    1.22 -            fclose(output);
    1.23 -        }else{
    1.24 -            break;
    1.25 -        }
    1.26 -    }
    1.27 -    printf("Saving Counter measurements to File: %s ...\n", filename);
    1.28 -    output = fopen(filename,"w+");
    1.29 -   _VMSMasterEnv->counteroutput = output;
    1.30 -*/
    1.31 -    
    1.32     struct perf_event_attr hw_event;
    1.33     memset(&hw_event,0,sizeof(hw_event));
    1.34     	hw_event.type = PERF_TYPE_HARDWARE;
    1.35 @@ -228,8 +203,8 @@
    1.36  	hw_event.pinned = 1; /* must always be on PMU */
    1.37  	hw_event.exclusive = 0; /* only group on PMU     */
    1.38  	hw_event.exclude_user = 0; /* don't count user      */
    1.39 -	hw_event.exclude_kernel = 1; /* ditto kernel          */
    1.40 -	hw_event.exclude_hv = 1; /* ditto hypervisor      */
    1.41 +	hw_event.exclude_kernel = 0; /* ditto kernel          */
    1.42 +	hw_event.exclude_hv = 0; /* ditto hypervisor      */
    1.43  	hw_event.exclude_idle = 0; /* don't count when idle */
    1.44  	hw_event.mmap = 0; /* include mmap data     */
    1.45  	hw_event.comm = 0; /* include comm data     */
    1.46 @@ -260,10 +235,9 @@
    1.47              perror("Failed to open instrs counter");
    1.48          }
    1.49     }
    1.50 -   prctl(PR_TASK_PERF_EVENTS_ENABLE);
    1.51 -   uint64 tmpc,tmpi;
    1.52 -   saveCyclesAndInstrs(0,tmpc,tmpi);
    1.53 -   printf("Start: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
    1.54 +   //uint64 tmpc,tmpi;
    1.55 +   //saveCyclesAndInstrs(0,tmpc,tmpi);
    1.56 +   //printf("Start: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
    1.57     #endif
    1.58     
    1.59     //========================================================================
    1.60 @@ -314,8 +288,10 @@
    1.61  
    1.62        //Make the threads that animate the core loops
    1.63     for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
    1.64 -    { coreLoopThdParams[coreIdx]          = VMS__malloc( sizeof(ThdParams) );
    1.65 +    { coreLoopThdParams[coreIdx]          = VMS__malloc( sizeof(ThdParams) + CACHE_LINE ); //make sure there is no false sharing
    1.66        coreLoopThdParams[coreIdx]->coreNum = coreIdx;
    1.67 +      coreLoopThdParams[coreIdx]->sent_ctr = 0;
    1.68 +      coreLoopThdParams[coreIdx]->ret_tsc = 0;
    1.69  
    1.70        retCode =
    1.71        pthread_create( &(coreLoopThdHandles[coreIdx]),
    1.72 @@ -324,6 +300,7 @@
    1.73                 (void *)(coreLoopThdParams[coreIdx]) );
    1.74        if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
    1.75      }
    1.76 +      prctl(PR_TASK_PERF_EVENTS_ENABLE);
    1.77   }
    1.78  
    1.79  /*Semantic layer calls this when it want the system to start running..
    1.80 @@ -737,8 +714,7 @@
    1.81  #ifdef MEAS__PERF_COUNTERS 
    1.82     uint64 tmpc,tmpi;
    1.83     saveCyclesAndInstrs(0,tmpc,tmpi);
    1.84 -   printf("End: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
    1.85 -   prctl(PR_TASK_PERF_EVENTS_DISABLE);
    1.86 +   //printf("End: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
    1.87  /*
    1.88     for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ){
    1.89         close(_VMSMasterEnv->cycles_counter_fd[coreIdx]);
    1.90 @@ -789,7 +765,7 @@
    1.91     //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&saveHistToFile);
    1.92     //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, &freeHist );
    1.93  
    1.94 -
    1.95 +   prctl(PR_TASK_PERF_EVENTS_DISABLE);
    1.96     #ifdef MEAS__TIME_PLUGIN
    1.97     printHist( _VMSMasterEnv->reqHdlrLowTimeHist );
    1.98     saveHistToFile( _VMSMasterEnv->reqHdlrLowTimeHist );
    1.99 @@ -879,3 +855,74 @@
   1.100     exit(1);
   1.101   }
   1.102  
   1.103 + __inline__ uint64_t rdtsc(void){
   1.104 +    uint32_t lo, hi;
   1.105 +    __asm__ __volatile__ (      // serialize
   1.106 +    "xorl %%eax,%%eax \n        cpuid"
   1.107 +    ::: "%rax", "%rbx", "%rcx", "%rdx");
   1.108 +    __asm__ __volatile__ ("rdtsc" : "=a" (lo), "=d" (hi)); 
   1.109 +   /* asm volatile("RDTSC;"                   
   1.110 +                 "movl %%eax, %0;"         
   1.111 +                 "movl %%edx, %1;"         
   1.112 +               : "=m" (lo), "=m" (hi)
   1.113 +               :                        
   1.114 +               : "%eax", "%edx"         
   1.115 +                ); */
   1.116 +    return (uint64_t)hi << 32 | lo;
   1.117 +  }
   1.118 + 
   1.119 + uint64 tsc_offset_send(ThdParams* thisCoresThdParams, uint64 initval){
   1.120 +     uint64 ret_tsc_curr; //local copy of coreLoopThdParams->ret_tsc
   1.121 +     uint64 ret_tsc_prev;
   1.122 +     uint64 local_before;
   1.123 +     uint64 local_after;
   1.124 +     
   1.125 +     ret_tsc_prev = initval;
   1.126 +     ret_tsc_curr = initval;
   1.127 +     local_before = rdtsc();
   1.128 +     thisCoresThdParams->sent_ctr++;
   1.129 +     while(ret_tsc_curr == ret_tsc_prev)
   1.130 +         ret_tsc_curr = thisCoresThdParams->ret_tsc;
   1.131 +     local_after = rdtsc();
   1.132 +     ret_tsc_prev = ret_tsc_curr;
   1.133 +     
   1.134 +     int i;
   1.135 +     for(i=0;i<3;++i){
   1.136 +         local_before = rdtsc();
   1.137 +         thisCoresThdParams->sent_ctr++;
   1.138 +         while(ret_tsc_curr == ret_tsc_prev)
   1.139 +             ret_tsc_curr = thisCoresThdParams->ret_tsc;
   1.140 +         local_after = rdtsc();
   1.141 +         int64 midpoint = local_before + (local_after-local_before)/2;
   1.142 +         int64 difference;
   1.143 +         if (midpoint > ret_tsc_curr)
   1.144 +             difference = midpoint - (int64)ret_tsc_curr;
   1.145 +         else
   1.146 +             difference = (int64)ret_tsc_curr - midpoint;
   1.147 +         //printf("TSC: %llu (Core %d) =  %llu (Core %d) // difference=%llu\n",midpoint,thisCoresThdParams->coreNum,ret_tsc_curr,thisCoresThdParams->coreNum + 1,difference);
   1.148 +         ret_tsc_prev = ret_tsc_curr;
   1.149 +     }
   1.150 +     
   1.151 +     return ret_tsc_curr;
   1.152 + }
   1.153 + 
   1.154 + int tsc_offset_resp(ThdParams* sendCoresThdParams, int initialctrval){
   1.155 +
   1.156 +       int send_ctr_curr = initialctrval;
   1.157 +       int send_ctr_prev = initialctrval;
   1.158 +
   1.159 +       
   1.160 +       while(send_ctr_curr == send_ctr_prev)
   1.161 +           send_ctr_curr = sendCoresThdParams->sent_ctr;
   1.162 +       sendCoresThdParams->ret_tsc = rdtsc();
   1.163 +       send_ctr_prev = send_ctr_curr;
   1.164 +       
   1.165 +       int i;
   1.166 +       for(i=0;i<3;++i){
   1.167 +           while(send_ctr_curr == send_ctr_prev)
   1.168 +               send_ctr_curr = sendCoresThdParams->sent_ctr;
   1.169 +           sendCoresThdParams->ret_tsc = rdtsc();
   1.170 +           send_ctr_prev = send_ctr_curr;
   1.171 +       }
   1.172 +       return send_ctr_curr;
   1.173 + }
   1.174 \ No newline at end of file