1.1 --- a/VMS.h	Thu Oct 14 17:07:23 2010 -0700
     1.2 +++ b/VMS.h	Sat Oct 30 20:54:36 2010 -0700
     1.3 @@ -13,17 +13,31 @@
     1.4  #include "VMS_primitive_data_types.h"
     1.5  #include "Queue_impl/BlockingQueue.h"
     1.6  #include "Histogram/Histogram.h"
     1.7 +#include "DynArray/DynArray.h"
     1.8 +#include "Hash_impl/PrivateHash.h"
     1.9 +#include "vmalloc.h"
    1.10 +
    1.11  #include <pthread.h>
    1.12 +#include <sys/time.h>
    1.13  
    1.14 +
    1.15 +//===============================  Debug  ===================================
    1.16     //When SEQUENTIAL is defined, VMS does sequential exe in the main thread
    1.17     // It still does co-routines and all the mechanisms are the same, it just
    1.18     // has only a single thread and animates VPs one at a time
    1.19  //#define SEQUENTIAL
    1.20  
    1.21 -#define PRINT_DEBUG(msg) //printf(msg); fflush(stdin);
    1.22 +#define PRINT_DEBUG(msg)// printf(msg); fflush(stdin);
    1.23  #define PRINT1_DEBUG(msg, param) //printf(msg, param); fflush(stdin);
    1.24  #define PRINT2_DEBUG(msg, p1, p2) //printf(msg, p1, p2); fflush(stdin);
    1.25  
    1.26 +#define PRINT_ERROR(msg) printf(msg); fflush(stdin);
    1.27 +#define PRINT1_ERROR(msg, param) printf(msg, param); fflush(stdin);
    1.28 +#define PRINT2_ERROR(msg, p1, p2) printf(msg, p1, p2); fflush(stdin);
    1.29 +
    1.30 +
    1.31 +//===========================  STATS =======================
    1.32 +
    1.33     //when MEAS__TIME_STAMP_SUSP is defined, causes code to be inserted and
    1.34     // compiled-in that saves the low part of the time stamp count just before
    1.35     // suspending a processor and just after resuming that processor.  It is
    1.36 @@ -35,6 +49,8 @@
    1.37  
    1.38  #define NUM_TSC_ROUND_TRIPS 10
    1.39  
    1.40 +
    1.41 +//=========================  Hardware related Constants =====================
    1.42     //This value is the number of hardware threads in the shared memory
    1.43     // machine
    1.44  #define NUM_CORES        4
    1.45 @@ -47,39 +63,75 @@
    1.46  #define READYTOANIMATE_RETRIES 10000
    1.47  
    1.48     // stack
    1.49 -#define VIRT_PROCR_STACK_SIZE 0x10000
    1.50 +#define VIRT_PROCR_STACK_SIZE 0x4000
    1.51  
    1.52 -   //256M of total memory for VMS__malloc
    1.53 -#define MASSIVE_MALLOC_SIZE 0x10000000
    1.54 +   // memory for VMS__malloc -- 256M
    1.55 +#define MALLOC_ADDITIONAL_MEM_FROM_OS_SIZE 0x10000000
    1.56  
    1.57 -#define NUM_PREPEND_BYTES sizeof(FreeListElem) + sizeof(ownerElem);
    1.58 +
    1.59 +//==============================
    1.60  
    1.61  #define SUCCESS 0
    1.62  
    1.63 -#define writeVMSQ     writeCASQ
    1.64 -#define readVMSQ      readCASQ
    1.65 -#define makeVMSQ      makeCASQ
    1.66 -#define VMSQueueStruc CASQueueStruc
    1.67 +#define writeVMSQ     writeSRSWQ
    1.68 +#define readVMSQ      readSRSWQ
    1.69 +#define makeVMSQ      makeSRSWQ
    1.70 +#define VMSQueueStruc SRSWQueueStruc
    1.71  
    1.72 -//#define thdAttrs NULL  //For PThreads
    1.73  
    1.74 -typedef struct _SchedSlot  SchedSlot;
    1.75 -typedef struct _VMSReqst   VMSReqst;
    1.76 -typedef struct _VirtProcr  VirtProcr;
    1.77 +
    1.78 +//===========================================================================
    1.79 +typedef unsigned long long TSCount;
    1.80 +
    1.81 +typedef struct _SchedSlot     SchedSlot;
    1.82 +typedef struct _VMSReqst      VMSReqst;
    1.83 +typedef struct _VirtProcr     VirtProcr;
    1.84 +typedef struct _IntervalProbe IntervalProbe;
    1.85  
    1.86  typedef VirtProcr * (*SlaveScheduler)  ( void *, int );   //semEnv, coreIdx
    1.87  typedef void  (*RequestHandler)  ( VirtProcr *, void * ); //prWReqst, semEnv
    1.88  typedef void  (*VirtProcrFnPtr)  ( void *, VirtProcr * ); //initData, animPr
    1.89  typedef void    VirtProcrFn      ( void *, VirtProcr * ); //initData, animPr
    1.90 +typedef void  (*ResumePrFnPtr)   ( VirtProcr *, void * );
    1.91 +
    1.92 +
    1.93 +//============= Requests ===========
    1.94 +//
    1.95 +
    1.96 +enum VMSReqstType   //avoid starting enums at 0, for debug reasons
    1.97 + {
    1.98 +   semantic = 1,
    1.99 +   createReq,
   1.100 +   dissipate,
   1.101 +   VMSSemantic      //goes with VMSSemReqst below
   1.102 + };
   1.103 +
   1.104 +struct _VMSReqst
   1.105 + {
   1.106 +   enum VMSReqstType  reqType;//used for dissipate and in future for IO requests
   1.107 +   void              *semReqData;
   1.108 +
   1.109 +   VMSReqst *nextReqst;
   1.110 + };
   1.111 +//VMSReqst
   1.112 +
   1.113 +enum VMSSemReqstType   //These are equivalent to semantic requests, but for
   1.114 + {                     // VMS's services available directly to app, like OS
   1.115 +   createProbe = 1,    // and probe services -- like a VMS-wide built-in lang
   1.116 +   openFile,
   1.117 +   otherIO
   1.118 + };
   1.119  
   1.120  typedef struct
   1.121 - {
   1.122 -   void           *endThdPt;
   1.123 -   unsigned int    coreNum;
   1.124 + { enum VMSSemReqstType reqType;
   1.125 +   VirtProcr           *requestingPr;
   1.126 +   char                *nameStr;  //for create probe
   1.127   }
   1.128 -ThdParams;
   1.129 + VMSSemReq;
   1.130  
   1.131  
   1.132 +//====================  Core data structures  ===================
   1.133 +
   1.134  struct _SchedSlot
   1.135   {
   1.136     int         workIsDone;
   1.137 @@ -87,24 +139,6 @@
   1.138     VirtProcr  *procrAssignedToSlot;
   1.139   };
   1.140  //SchedSlot
   1.141 - 
   1.142 -enum ReqstType
   1.143 - {
   1.144 -   semantic = 1,
   1.145 -   dissipate,
   1.146 -   regCreated,
   1.147 -   IO
   1.148 - };
   1.149 -
   1.150 -struct _VMSReqst
   1.151 - {
   1.152 -//   VirtProcr   *virtProcrFrom;
   1.153 -   enum ReqstType  reqType;//used for dissipate and in future for IO requests
   1.154 -   void           *semReqData;
   1.155 -
   1.156 -   VMSReqst *nextReqst;
   1.157 - };
   1.158 -//VMSReqst
   1.159  
   1.160  struct _VirtProcr
   1.161   { int         procrID;  //for debugging -- count up each time create
   1.162 @@ -123,9 +157,10 @@
   1.163     SchedSlot  *schedSlot;
   1.164     VMSReqst   *requests;
   1.165  
   1.166 -   void       *semanticData;
   1.167 +   void       *semanticData; //this lives here for the life of VP
   1.168 +   void       *dataReturnedFromReq;//values returned from plugin to VP go here
   1.169  
   1.170 -   //============================= MEASUREMENT STUFF ========================
   1.171 +      //=========== MEASUREMENT STUFF ==========
   1.172     #ifdef MEAS__TIME_STAMP_SUSP
   1.173     unsigned int preSuspTSCLow;
   1.174     unsigned int postSuspTSCLow;
   1.175 @@ -134,12 +169,12 @@
   1.176     unsigned int startMasterTSCLow;
   1.177     unsigned int endMasterTSCLow;
   1.178     #endif
   1.179 -   //========================================================================
   1.180 +   
   1.181 +   float64      createPtInSecs;  //have space but don't use on some configs
   1.182   };
   1.183  //VirtProcr
   1.184  
   1.185  
   1.186 -
   1.187  typedef struct
   1.188   {
   1.189     SlaveScheduler   slaveScheduler;
   1.190 @@ -151,35 +186,61 @@
   1.191  
   1.192     void            *semanticEnv;
   1.193     void            *OSEventStruc;   //for future, when add I/O to BLIS
   1.194 +   MallocProlog    *freeListHead;
   1.195 +   int32            amtOfOutstandingMem; //total currently allocated
   1.196  
   1.197     void            *coreLoopStartPt;//addr to jump to to re-enter coreLoop
   1.198     void            *coreLoopEndPt;  //addr to jump to to shut down a coreLoop
   1.199  
   1.200 -   int              setupComplete;
   1.201 -   int              masterLock;
   1.202 +   int32            setupComplete;
   1.203 +   int32            masterLock;
   1.204  
   1.205 +   int32            numMasterInARow[NUM_CORES];//detect back-to-back masterVP
   1.206 +   int32            numProcrsCreated; //gives ordering to processor creation
   1.207 +
   1.208 +      //=========== MEASUREMENT STUFF =============
   1.209 +   IntervalProbe  **intervalProbes;
   1.210 +   DynArrayInfo    *dynIntervalProbesInfo;
   1.211 +   HashTable       *probeNameHashTbl;
   1.212 +   int32            masterCreateProbeID;
   1.213 +   float64          createPtInSecs;
   1.214   }
   1.215  MasterEnv;
   1.216  
   1.217  
   1.218 -//==========================================================
   1.219 +
   1.220 +
   1.221 +//=======================  OS Thread related  ===============================
   1.222  
   1.223  void * coreLoop( void *paramsIn );  //standard PThreads fn prototype
   1.224  void * coreLoop_Seq( void *paramsIn );  //standard PThreads fn prototype
   1.225  void masterLoop( void *initData, VirtProcr *masterPr );
   1.226  
   1.227  
   1.228 -//=====================  Global Vars ===================
   1.229 -
   1.230 +typedef struct
   1.231 + {
   1.232 +   void           *endThdPt;
   1.233 +   unsigned int    coreNum;
   1.234 + }
   1.235 +ThdParams;
   1.236  
   1.237  pthread_t       coreLoopThdHandles[ NUM_CORES ];  //pthread's virt-procr state
   1.238  ThdParams      *coreLoopThdParams [ NUM_CORES ];
   1.239  pthread_mutex_t suspendLock;
   1.240  pthread_cond_t  suspend_cond;
   1.241  
   1.242 +
   1.243 +
   1.244 +//=====================  Global Vars ===================
   1.245 +
   1.246  volatile MasterEnv      *_VMSMasterEnv;
   1.247  
   1.248 -//==========================
   1.249 +
   1.250 +
   1.251 +
   1.252 +//===========================  Function Prototypes  =========================
   1.253 +
   1.254 +//============== Setup and shutdown =============
   1.255  void
   1.256  VMS__init();
   1.257  
   1.258 @@ -195,16 +256,28 @@
   1.259  VirtProcr *
   1.260  VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
   1.261  
   1.262 +   //Use this to create processor inside entry point & other places outside
   1.263 +   // the VMS system boundary (IE, not run in slave nor Master)
   1.264 +VirtProcr *
   1.265 +VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
   1.266 +
   1.267  VirtProcr *
   1.268  VMS__create_the_shutdown_procr();
   1.269  
   1.270 -//==========================
   1.271 +void
   1.272 +VMS__cleanup_after_shutdown();
   1.273 +
   1.274 +
   1.275 +//==============  Request Related  ===============
   1.276 +
   1.277 +void
   1.278 +VMS__suspend_procr( VirtProcr *callingPr );
   1.279 +
   1.280  inline void
   1.281  VMS__add_sem_request( void *semReqData, VirtProcr *callingPr );
   1.282  
   1.283  void
   1.284 -VMS__send_req_to_register_new_procr( VirtProcr *newPrToRegister,
   1.285 -                                      VirtProcr *reqstingPr );
   1.286 +VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr );
   1.287  
   1.288  void
   1.289  VMS__free_request( VMSReqst *req );
   1.290 @@ -216,7 +289,7 @@
   1.291  VMS__take_top_request_from( VirtProcr *reqstingPr );
   1.292  
   1.293  VMSReqst *
   1.294 -VMS__free_top_and_give_next_request_from( VirtProcr *procrWithReq );
   1.295 +VMS__take_next_request_out_of( VirtProcr *procrWithReq );
   1.296  
   1.297  inline void *
   1.298  VMS__take_sem_reqst_from( VMSReqst *req );
   1.299 @@ -232,25 +305,15 @@
   1.300  
   1.301  //==========================
   1.302  
   1.303 -void
   1.304 -VMS__suspend_procr( VirtProcr *callingPr );
   1.305 -
   1.306 -void
   1.307 +void inline
   1.308  VMS__dissipate_procr( VirtProcr *prToDissipate );
   1.309  
   1.310  void
   1.311  VMS__handle_dissipate_reqst( VirtProcr *procrToDissipate );
   1.312  
   1.313 -void
   1.314 -VMS__cleanup_after_shutdown();
   1.315  
   1.316 -//============================= Statistics ==================================
   1.317  
   1.318 -typedef unsigned long long TSCount;
   1.319 -
   1.320 -   //Frequency of TS counts
   1.321 -   //TODO: change freq for each machine
   1.322 -#define TSCOUNT_FREQ 3180000000
   1.323 +//===================== RDTSC wrapper ==================
   1.324  
   1.325  #define saveTimeStampCountInto(low, high) \
   1.326     asm volatile("RDTSC;                   \
   1.327 @@ -269,11 +332,9 @@
   1.328     /* clobber */ : "%eax", "%edx"         \
   1.329                  );
   1.330  
   1.331 -inline TSCount getTSCount();
   1.332 +//======================== STATS ======================
   1.333  
   1.334 -//===================== Debug ==========================
   1.335 -int numProcrsCreated;
   1.336 -
   1.337 +#include "probes.h"
   1.338  
   1.339  #endif	/* _VMS_H */
   1.340