msach@146: /*
msach@146:  *  Copyright 2009 OpenSourceStewardshipFoundation.org
msach@146:  *  Licensed under GNU General Public License version 2
msach@146:  *
msach@146:  * Author: seanhalle@yahoo.com
msach@146:  * 
msach@146:  */
msach@146: 
msach@146: #ifndef _VMS_H
msach@146: #define	_VMS_H
msach@146: #define __USE_GNU
msach@146: 
msach@146: #include "VMS_primitive_data_types.h"
msach@146: #include "Queue_impl/PrivateQueue.h"
msach@146: #include "Histogram/Histogram.h"
msach@146: #include "DynArray/DynArray.h"
msach@146: #include "Hash_impl/PrivateHash.h"
msach@146: #include "vmalloc.h"
msach@146: 
msach@146: #include <pthread.h>
msach@146: #include <sys/time.h>
msach@146: 
msach@146: 
msach@146: //===============================  Debug  ===================================
msach@146: //
msach@146:    //When SEQUENTIAL is defined, VMS does sequential exe in the main thread
msach@146:    // It still does co-routines and all the mechanisms are the same, it just
msach@146:    // has only a single thread and animates VPs one at a time
msach@146: //#define SEQUENTIAL
msach@146: 
msach@146: //#define USE_WORK_STEALING
msach@146: 
msach@146:    //turns on the probe-instrumentation in the application -- when not
msach@146:    // defined, the calls to the probe functions turn into comments
msach@146: #define STATS__ENABLE_PROBES
msach@146: //#define TURN_ON_DEBUG_PROBES
msach@146: 
msach@146:    //These defines turn types of bug messages on and off
msach@146:    // be sure debug messages are un-commented (next block of defines)
msach@146: #define dbgAppFlow   TRUE /* Top level flow of application code -- general*/
msach@146: #define dbgProbes    FALSE /* for issues inside probes themselves*/
msach@146: #define dbgB2BMaster FALSE /* in coreloop, back to back master VPs*/
msach@146: #define dbgRqstHdlr  FALSE /* in request handler code*/
msach@146: 
msach@146:    //Comment or un- the substitute half to turn on/off types of debug message
msach@146: #define DEBUG(  bool, msg)         \
msach@146: //  if( bool){ printf(msg); fflush(stdin);}
msach@146: #define DEBUG1( bool, msg, param)  \
msach@146: //   if(bool){printf(msg, param); fflush(stdin);}
msach@146: #define DEBUG2( bool, msg, p1, p2) \
msach@146: //   if(bool) {printf(msg, p1, p2); fflush(stdin);}
msach@146: 
msach@146: #define ERROR(msg) printf(msg);
msach@146: #define ERROR1(msg, param) printf(msg, param);
msach@146: #define ERROR2(msg, p1, p2) printf(msg, p1, p2);
msach@146: 
msach@146: //===========================  STATS =======================
msach@146: 
msach@146:    //when MEAS__TIME_STAMP_SUSP is defined, causes code to be inserted and
msach@146:    // compiled-in that saves the low part of the time stamp count just before
msach@146:    // suspending a processor and just after resuming that processor.  It is
msach@146:    // saved into a field added to VirtProcr.  Have to sanity-check for
msach@146:    // rollover of low portion into high portion.
msach@146: //#define MEAS__TIME_STAMP_SUSP
msach@146: //#define MEAS__TIME_MASTER
msach@146: #define MEAS__TIME_PLUGIN
msach@146: #define MEAS__TIME_MALLOC
msach@146: //#define MEAS__TIME_MASTER_LOCK
msach@146: #define MEAS__NUM_TIMES_TO_RUN 100000
msach@146: 
msach@146:    //For code that calculates normalization-offset between TSC counts of
msach@146:    // different cores.
msach@146: #define NUM_TSC_ROUND_TRIPS 10
msach@146: 
msach@146: 
msach@146: //=========================  Hardware related Constants =====================
msach@146:    //This value is the number of hardware threads in the shared memory
msach@146:    // machine
msach@146: #define NUM_CORES        4
msach@146: 
msach@146:    // tradeoff amortizing master fixed overhead vs imbalance potential
msach@146:    // when work-stealing, can make bigger, at risk of losing cache affinity
msach@146: #define NUM_SCHED_SLOTS  5
msach@146: 
msach@146: #define MIN_WORK_UNIT_CYCLES 20000
msach@146: 
msach@146: #define MASTERLOCK_RETRIES 10000
msach@146: 
msach@146:    // stack size in virtual processors created
msach@146: #define VIRT_PROCR_STACK_SIZE 0x4000 /* 16K */
msach@146: 
msach@146:    // memory for VMS__malloc
msach@146: #define MALLOC_ADDITIONAL_MEM_FROM_OS_SIZE 0x10000000 /* 256M */
msach@146: 
msach@146: 
msach@146: //==============================
msach@146: 
msach@146: #define SUCCESS 0
msach@146: 
msach@146: #define writeVMSQ     writePrivQ
msach@146: #define readVMSQ      readPrivQ
msach@146: #define makeVMSQ      makeVMSPrivQ
msach@146: #define numInVMSQ     numInPrivQ
msach@146: #define VMSQueueStruc PrivQueueStruc
msach@146: 
msach@146: 
msach@146: 
msach@146: //===========================================================================
msach@146: typedef unsigned long long TSCount;
msach@146: 
msach@146: typedef struct _SchedSlot     SchedSlot;
msach@146: typedef struct _VMSReqst      VMSReqst;
msach@146: typedef struct _VirtProcr     VirtProcr;
msach@146: typedef struct _IntervalProbe IntervalProbe;
msach@146: typedef struct _GateStruc     GateStruc;
msach@146: 
msach@146: 
msach@146: typedef VirtProcr * (*SlaveScheduler)  ( void *, int );   //semEnv, coreIdx
msach@146: typedef void  (*RequestHandler)  ( VirtProcr *, void * ); //prWReqst, semEnv
msach@146: typedef void  (*VirtProcrFnPtr)  ( void *, VirtProcr * ); //initData, animPr
msach@146: typedef void    VirtProcrFn      ( void *, VirtProcr * ); //initData, animPr
msach@146: typedef void  (*ResumePrFnPtr)   ( VirtProcr *, void * );
msach@146: 
msach@146: 
msach@146: //============= Requests ===========
msach@146: //
msach@146: 
msach@146: enum VMSReqstType   //avoid starting enums at 0, for debug reasons
msach@146:  {
msach@146:    semantic = 1,
msach@146:    createReq,
msach@146:    dissipate,
msach@146:    VMSSemantic      //goes with VMSSemReqst below
msach@146:  };
msach@146: 
msach@146: struct _VMSReqst
msach@146:  {
msach@146:    enum VMSReqstType  reqType;//used for dissipate and in future for IO requests
msach@146:    void              *semReqData;
msach@146: 
msach@146:    VMSReqst *nextReqst;
msach@146:  };
msach@146: //VMSReqst
msach@146: 
msach@146: enum VMSSemReqstType   //These are equivalent to semantic requests, but for
msach@146:  {                     // VMS's services available directly to app, like OS
msach@146:    createProbe = 1,    // and probe services -- like a VMS-wide built-in lang
msach@146:    openFile,
msach@146:    otherIO
msach@146:  };
msach@146: 
msach@146: typedef struct
msach@146:  { enum VMSSemReqstType reqType;
msach@146:    VirtProcr           *requestingPr;
msach@146:    char                *nameStr;  //for create probe
msach@146:  }
msach@146:  VMSSemReq;
msach@146: 
msach@146: 
msach@146: //====================  Core data structures  ===================
msach@146: 
msach@146: struct _SchedSlot
msach@146:  {
msach@146:    int         workIsDone;
msach@146:    int         needsProcrAssigned;
msach@146:    VirtProcr  *procrAssignedToSlot;
msach@146:  };
msach@146: //SchedSlot
msach@146: 
msach@146: /*WARNING: re-arranging this data structure could cause VP switching
msach@146:  *         assembly code to fail -- hard-codes offsets of fields
msach@146:  */
msach@146: struct _VirtProcr
msach@146:  { int         procrID;  //for debugging -- count up each time create
msach@146:    int         coreAnimatedBy;
msach@146:    void       *startOfStack;
msach@146:    void       *stackPtr;
msach@146:    void       *framePtr;
msach@146:    void       *nextInstrPt;
msach@146:    
msach@146:    void       *coreLoopStartPt;  //allows proto-runtime to be linked later
msach@146:    void       *coreLoopFramePtr; //restore before jmp back to core loop
msach@146:    void       *coreLoopStackPtr; //restore before jmp back to core loop
msach@146: 
msach@146:    void       *initialData;
msach@146: 
msach@146:    SchedSlot  *schedSlot;
msach@146:    VMSReqst   *requests;
msach@146: 
msach@146:    void       *semanticData; //this lives here for the life of VP
msach@146:    void       *dataRetFromReq;//values returned from plugin to VP go here
msach@146: 
msach@146:       //=========== MEASUREMENT STUFF ==========
msach@146:    #ifdef MEAS__TIME_STAMP_SUSP
msach@146:    unsigned int preSuspTSCLow;
msach@146:    unsigned int postSuspTSCLow;
msach@146:    #endif
msach@146:    #ifdef MEAS__TIME_MASTER /* in VirtProcr because multiple masterVPs*/
msach@146:    unsigned int startMasterTSCLow;
msach@146:    unsigned int endMasterTSCLow;
msach@146:    #endif
msach@146:       //========================================
msach@146:    
msach@146:    float64      createPtInSecs;  //have space but don't use on some configs
msach@146:  };
msach@146: //VirtProcr
msach@146: 
msach@146: 
msach@146: /*WARNING: re-arranging this data structure could cause VP-switching
msach@146:  *         assembly code to fail -- hard-codes offsets of fields
msach@146:  *         (because -O3 messes with things otherwise)
msach@146:  */
msach@146: typedef struct
msach@146:  {
msach@146:    SlaveScheduler   slaveScheduler;
msach@146:    RequestHandler   requestHandler;
msach@146:    
msach@146:    SchedSlot     ***allSchedSlots;
msach@146:    VMSQueueStruc **readyToAnimateQs;
msach@146:    VirtProcr      **masterVPs;
msach@146: 
msach@146:    void            *semanticEnv;
msach@146:    void            *OSEventStruc;   //for future, when add I/O to BLIS
msach@146:    MallocProlog    *freeListHead;
msach@146:    int32            amtOfOutstandingMem; //total currently allocated
msach@146: 
msach@146:    void            *coreLoopStartPt;//addr to jump to to re-enter coreLoop
msach@146:    void            *coreLoopEndPt;  //addr to jump to to shut down a coreLoop
msach@146: 
msach@146:    int32            setupComplete;
msach@146:    int32            masterLock;
msach@146: 
msach@146:    int32            numMasterInARow[NUM_CORES];//detect back-to-back masterVP
msach@146:    GateStruc       *workStealingGates[ NUM_CORES ]; //concurrent work-steal
msach@146:    int32            workStealingLock;
msach@146:    
msach@146:    int32            numProcrsCreated; //gives ordering to processor creation
msach@146: 
msach@146:       //=========== MEASUREMENT STUFF =============
msach@146:    IntervalProbe  **intervalProbes;
msach@146:    PrivDynArrayInfo    *dynIntervalProbesInfo;
msach@146:    HashTable       *probeNameHashTbl;
msach@146:    int32            masterCreateProbeID;
msach@146:    float64          createPtInSecs;
msach@146:    Histogram      **measHists;
msach@146:    PrivDynArrayInfo *measHistsInfo;
msach@146:    #ifdef MEAS__TIME_PLUGIN
msach@146:    Histogram       *reqHdlrLowTimeHist;
msach@146:    Histogram       *reqHdlrHighTimeHist;
msach@146:    #endif
msach@146:    #ifdef MEAS__TIME_MALLOC
msach@146:    Histogram       *mallocTimeHist;
msach@146:    Histogram       *freeTimeHist;
msach@146:    #endif
msach@146:    #ifdef MEAS__TIME_MASTER_LOCK
msach@146:    Histogram       *masterLockLowTimeHist;
msach@146:    Histogram       *masterLockHighTimeHist;
msach@146:    #endif
msach@146:  }
msach@146: MasterEnv;
msach@146: 
msach@146: //=========================  Extra Stuff Data Strucs  =======================
msach@146: typedef struct
msach@146:  {
msach@146: 
msach@146:  }
msach@146: VMSExcp;
msach@146: 
msach@146: struct _GateStruc
msach@146:  {
msach@146:    int32 gateClosed;
msach@146:    int32 preGateProgress;
msach@146:    int32 waitProgress;
msach@146:    int32 exitProgress;
msach@146:  };
msach@146: //GateStruc
msach@146: 
msach@146: //=======================  OS Thread related  ===============================
msach@146: 
msach@146: void * coreLoop( void *paramsIn );  //standard PThreads fn prototype
msach@146: void * coreLoop_Seq( void *paramsIn );  //standard PThreads fn prototype
msach@146: void masterLoop( void *initData, VirtProcr *masterPr );
msach@146: 
msach@146: 
msach@146: typedef struct
msach@146:  {
msach@146:    void           *endThdPt;
msach@146:    unsigned int    coreNum;
msach@146:  }
msach@146: ThdParams;
msach@146: 
msach@146: pthread_t       coreLoopThdHandles[ NUM_CORES ];  //pthread's virt-procr state
msach@146: ThdParams      *coreLoopThdParams [ NUM_CORES ];
msach@146: pthread_mutex_t suspendLock;
msach@146: pthread_cond_t  suspend_cond;
msach@146: 
msach@146: 
msach@146: 
msach@146: //=====================  Global Vars ===================
msach@146: 
msach@146: volatile MasterEnv      *_VMSMasterEnv;
msach@146: 
msach@146: 
msach@146: 
msach@146: 
msach@146: //===========================  Function Prototypes  =========================
msach@146: 
msach@146: 
msach@146: //========== Setup and shutdown ==========
msach@146: void
msach@146: VMS__init();
msach@146: 
msach@146: void
msach@146: VMS__init_Seq();
msach@146: 
msach@146: void
msach@146: VMS__start_the_work_then_wait_until_done();
msach@146: 
msach@146: void
msach@146: VMS__start_the_work_then_wait_until_done_Seq();
msach@146: 
msach@146: VirtProcr *
msach@146: VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
msach@146: 
msach@146: void
msach@146: VMS__dissipate_procr( VirtProcr *procrToDissipate );
msach@146: 
msach@146:    //Use this to create processor inside entry point & other places outside
msach@146:    // the VMS system boundary (IE, not run in slave nor Master)
msach@146: VirtProcr *
msach@146: VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
msach@146: 
msach@146: void
msach@146: VMS_ext__dissipate_procr( VirtProcr *procrToDissipate );
msach@146: 
msach@146: void
msach@146: VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData );
msach@146: 
msach@146: void
msach@146: VMS__shutdown();
msach@146: 
msach@146: void
msach@146: VMS__cleanup_at_end_of_shutdown();
msach@146: 
msach@146: void *
msach@146: VMS__give_sem_env_for( VirtProcr *animPr );
msach@146: 
msach@146: 
msach@146: //==============  Request Related  ===============
msach@146: 
msach@146: void
msach@146: VMS__suspend_procr( VirtProcr *callingPr );
msach@146: 
msach@146: inline void
msach@146: VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData, VirtProcr *callingPr );
msach@146: 
msach@146: inline void
msach@146: VMS__send_sem_request( void *semReqData, VirtProcr *callingPr );
msach@146: 
msach@146: void
msach@146: VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr );
msach@146: 
msach@146: void inline
msach@146: VMS__send_dissipate_req( VirtProcr *prToDissipate );
msach@146: 
msach@146: inline void
msach@146: VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr );
msach@146: 
msach@146: VMSReqst *
msach@146: VMS__take_next_request_out_of( VirtProcr *procrWithReq );
msach@146: 
msach@146: inline void *
msach@146: VMS__take_sem_reqst_from( VMSReqst *req );
msach@146: 
msach@146: //======================== STATS ======================
msach@146: 
msach@146: //===== RDTSC wrapper =====
msach@146: 
msach@146: #define saveTimeStampCountInto(low, high) \
msach@146:    asm volatile("RDTSC;                   \
msach@146:                  movl %%eax, %0;          \
msach@146:                  movl %%edx, %1;"         \
msach@146:    /* outputs */ : "=m" (low), "=m" (high)\
msach@146:    /* inputs  */ :                        \
msach@146:    /* clobber */ : "%eax", "%edx"         \
msach@146:                 );
msach@146: 
msach@146: #define saveLowTimeStampCountInto(low)    \
msach@146:    asm volatile("RDTSC;                   \
msach@146:                  movl %%eax, %0;"         \
msach@146:    /* outputs */ : "=m" (low)             \
msach@146:    /* inputs  */ :                        \
msach@146:    /* clobber */ : "%eax", "%edx"         \
msach@146:                 );
msach@146: 
msach@146: //====================
msach@146: #define makeAMeasHist( idx, name, numBins, startVal, binWidth ) \
msach@146:    makeHighestDynArrayIndexBeAtLeast( _VMSMasterEnv->measHistsInfo, idx ); \
msach@146:    _VMSMasterEnv->measHists[idx] =  \
msach@146:                        makeFixedBinHist( numBins, startVal, binWidth, name );
msach@146: 
msach@146: 
msach@146: #define MEAS__SUB_CREATE  /*turn on/off subtraction of create from plugin*/
msach@146: #define createHistIdx      0
msach@146: #define mutexLockHistIdx   1
msach@146: #define mutexUnlockHistIdx 2
msach@146: #define condWaitHistIdx    3
msach@146: #define condSignalHistIdx  4
msach@146: 
msach@146: 
msach@146: #define MakeTheMeasHists \
msach@146:    _VMSMasterEnv->measHistsInfo = \
msach@146:               makePrivDynArrayOfSize( &(_VMSMasterEnv->measHists), 200);\
msach@146:    makeAMeasHist( createHistIdx,      "Create",        50, 0, 200 ) \
msach@146:    makeAMeasHist( mutexLockHistIdx,   "mutex lock",    50, 0, 100 ) \
msach@146:    makeAMeasHist( mutexUnlockHistIdx, "mutex unlock",  50, 0, 100 ) \
msach@146:    makeAMeasHist( condWaitHistIdx,    "cond wait",     50, 0, 100 ) \
msach@146:    makeAMeasHist( condSignalHistIdx,  "cond signal",   50, 0, 100 )
msach@146: 
msach@146: #define Meas_startCreate \
msach@146:     int32 startStamp, endStamp; \
msach@146:     saveLowTimeStampCountInto( startStamp ); \
msach@146: 
msach@146: #define Meas_endCreate \
msach@146:     saveLowTimeStampCountInto( endStamp ); \
msach@146:     addIntervalToHist( startStamp, endStamp, \
msach@146:                                  _VMSMasterEnv->measHists[ createHistIdx ] );
msach@146: 
msach@146: #define Meas_startMutexLock \
msach@146:     int32 startStamp, endStamp; \
msach@146:     saveLowTimeStampCountInto( startStamp ); \
msach@146: 
msach@146: #define Meas_endMutexLock \
msach@146:     saveLowTimeStampCountInto( endStamp ); \
msach@146:     addIntervalToHist( startStamp, endStamp, \
msach@146:                               _VMSMasterEnv->measHists[ mutexLockHistIdx ] );
msach@146: 
msach@146: #define Meas_startMutexUnlock \
msach@146:     int32 startStamp, endStamp; \
msach@146:     saveLowTimeStampCountInto( startStamp ); \
msach@146: 
msach@146: #define Meas_endMutexUnlock \
msach@146:     saveLowTimeStampCountInto( endStamp ); \
msach@146:     addIntervalToHist( startStamp, endStamp, \
msach@146:                             _VMSMasterEnv->measHists[ mutexUnlockHistIdx ] );
msach@146: 
msach@146: #define Meas_startCondWait \
msach@146:     int32 startStamp, endStamp; \
msach@146:     saveLowTimeStampCountInto( startStamp ); \
msach@146: 
msach@146: #define Meas_endCondWait \
msach@146:     saveLowTimeStampCountInto( endStamp ); \
msach@146:     addIntervalToHist( startStamp, endStamp, \
msach@146:                                _VMSMasterEnv->measHists[ condWaitHistIdx ] );
msach@146: 
msach@146: #define Meas_startCondSignal \
msach@146:     int32 startStamp, endStamp; \
msach@146:     saveLowTimeStampCountInto( startStamp ); \
msach@146: 
msach@146: #define Meas_endCondSignal \
msach@146:     saveLowTimeStampCountInto( endStamp ); \
msach@146:     addIntervalToHist( startStamp, endStamp, \
msach@146:                              _VMSMasterEnv->measHists[ condSignalHistIdx ] );
msach@146: 
msach@146: 
msach@146: //=====
msach@146: 
msach@146: #include "SwitchAnimators.h"
msach@146: #include "probes.h"
msach@146: #include "vutilities.h"
msach@146: 
msach@146: #endif	/* _VMS_H */
msach@146: