Me@42: /*
Me@42:  *  Copyright 2009 OpenSourceStewardshipFoundation.org
Me@42:  *  Licensed under GNU General Public License version 2
Me@42:  *
Me@42:  * Author: seanhalle@yahoo.com
Me@42:  * 
Me@42:  */
Me@42: 
Me@42: #ifndef _VMS_H
Me@42: #define	_VMS_H
Me@42: #define __USE_GNU
Me@42: 
Me@42: #include "VMS_primitive_data_types.h"
Me@55: #include "Queue_impl/PrivateQueue.h"
Me@42: #include "Histogram/Histogram.h"
Me@50: #include "DynArray/DynArray.h"
Me@50: #include "Hash_impl/PrivateHash.h"
Me@50: #include "vmalloc.h"
Me@50: 
Me@42: #include <pthread.h>
Me@50: #include <sys/time.h>
Me@42: 
Me@50: 
Me@50: //===============================  Debug  ===================================
Me@55: //
Me@45:    //When SEQUENTIAL is defined, VMS does sequential exe in the main thread
Me@42:    // It still does co-routines and all the mechanisms are the same, it just
Me@42:    // has only a single thread and animates VPs one at a time
Me@45: //#define SEQUENTIAL
Me@42: 
Me@55: //#define USE_WORK_STEALING
Me@55: 
Me@52:    //turns on the probe-instrumentation in the application -- when not
Me@52:    // defined, the calls to the probe functions turn into comments
Me@52: #define STATS__ENABLE_PROBES
Me@60: //#define TURN_ON_DEBUG_PROBES
Me@52: 
Me@55:    //These defines turn types of bug messages on and off
Me@55:    // be sure debug messages are un-commented (next block of defines)
Me@60: #define dbgProbes    FALSE /* for issues inside probes themselves*/
Me@60: #define dbgAppFlow   FALSE /* Top level flow of application code -- general*/
Me@60: #define dbgB2BMaster FALSE /* in coreloop, back to back master VPs*/
Me@60: #define dbgRqstHdlr  FALSE /* in request handler code*/
Me@52: 
Me@55:    //Comment or un- the substitute half to turn on/off types of debug message
Me@55: #define DEBUG(  bool, msg)         \
Me@60: //   if( bool){ printf(msg); fflush(stdin);}
Me@55: #define DEBUG1( bool, msg, param)  \
Me@60: //   if(bool){printf(msg, param); fflush(stdin);}
Me@55: #define DEBUG2( bool, msg, p1, p2) \
Me@60: //   if(bool) {printf(msg, p1, p2); fflush(stdin);}
Me@45: 
Me@55: #define ERROR(msg) printf(msg); fflush(stdin);
Me@55: #define ERROR1(msg, param) printf(msg, param); fflush(stdin);
Me@55: #define ERROR2(msg, p1, p2) printf(msg, p1, p2); fflush(stdin);
Me@50: 
Me@50: //===========================  STATS =======================
Me@50: 
Me@45:    //when MEAS__TIME_STAMP_SUSP is defined, causes code to be inserted and
Me@42:    // compiled-in that saves the low part of the time stamp count just before
Me@42:    // suspending a processor and just after resuming that processor.  It is
Me@42:    // saved into a field added to VirtProcr.  Have to sanity-check for
Me@42:    // rollover of low portion into high portion.
Me@42: #define MEAS__TIME_STAMP_SUSP
Me@42: #define MEAS__TIME_MASTER
Me@42: #define MEAS__NUM_TIMES_TO_RUN 100000
Me@42: 
Me@55:    //For code that calculates normalization-offset between TSC counts of
Me@55:    // different cores.
Me@45: #define NUM_TSC_ROUND_TRIPS 10
Me@45: 
Me@50: 
Me@50: //=========================  Hardware related Constants =====================
Me@42:    //This value is the number of hardware threads in the shared memory
Me@42:    // machine
Me@42: #define NUM_CORES        4
Me@42: 
Me@55:    // tradeoff amortizing master fixed overhead vs imbalance potential
Me@55:    // when work-stealing, can make bigger, at risk of losing cache affinity
Me@55: #define NUM_SCHED_SLOTS  5
Me@42: 
Me@45: #define MIN_WORK_UNIT_CYCLES 20000
Me@45: 
Me@53: #define MASTERLOCK_RETRIES 10000
Me@42: 
Me@54:    // stack size in virtual processors created
Me@54: #define VIRT_PROCR_STACK_SIZE 0x4000 /* 16K */
Me@42: 
Me@54:    // memory for VMS__malloc
Me@54: #define MALLOC_ADDITIONAL_MEM_FROM_OS_SIZE 0x10000000 /* 256M */
Me@42: 
Me@50: 
Me@50: //==============================
Me@42: 
Me@42: #define SUCCESS 0
Me@42: 
Me@55: #define writeVMSQ     writePrivQ
Me@55: #define readVMSQ      readPrivQ
Me@55: #define makeVMSQ      makePrivQ
Me@55: #define numInVMSQ     numInPrivQ
Me@55: #define VMSQueueStruc PrivQueueStruc
Me@42: 
Me@42: 
Me@50: 
Me@50: //===========================================================================
Me@50: typedef unsigned long long TSCount;
Me@50: 
Me@50: typedef struct _SchedSlot     SchedSlot;
Me@50: typedef struct _VMSReqst      VMSReqst;
Me@50: typedef struct _VirtProcr     VirtProcr;
Me@50: typedef struct _IntervalProbe IntervalProbe;
Me@55: typedef struct _GateStruc     GateStruc;
Me@55: 
Me@42: 
Me@42: typedef VirtProcr * (*SlaveScheduler)  ( void *, int );   //semEnv, coreIdx
Me@42: typedef void  (*RequestHandler)  ( VirtProcr *, void * ); //prWReqst, semEnv
Me@42: typedef void  (*VirtProcrFnPtr)  ( void *, VirtProcr * ); //initData, animPr
Me@42: typedef void    VirtProcrFn      ( void *, VirtProcr * ); //initData, animPr
Me@50: typedef void  (*ResumePrFnPtr)   ( VirtProcr *, void * );
Me@50: 
Me@50: 
Me@50: //============= Requests ===========
Me@50: //
Me@50: 
Me@50: enum VMSReqstType   //avoid starting enums at 0, for debug reasons
Me@50:  {
Me@50:    semantic = 1,
Me@50:    createReq,
Me@50:    dissipate,
Me@50:    VMSSemantic      //goes with VMSSemReqst below
Me@50:  };
Me@50: 
Me@50: struct _VMSReqst
Me@50:  {
Me@50:    enum VMSReqstType  reqType;//used for dissipate and in future for IO requests
Me@50:    void              *semReqData;
Me@50: 
Me@50:    VMSReqst *nextReqst;
Me@50:  };
Me@50: //VMSReqst
Me@50: 
Me@50: enum VMSSemReqstType   //These are equivalent to semantic requests, but for
Me@50:  {                     // VMS's services available directly to app, like OS
Me@50:    createProbe = 1,    // and probe services -- like a VMS-wide built-in lang
Me@50:    openFile,
Me@50:    otherIO
Me@50:  };
Me@42: 
Me@42: typedef struct
Me@50:  { enum VMSSemReqstType reqType;
Me@50:    VirtProcr           *requestingPr;
Me@50:    char                *nameStr;  //for create probe
Me@42:  }
Me@50:  VMSSemReq;
Me@42: 
Me@42: 
Me@50: //====================  Core data structures  ===================
Me@50: 
Me@42: struct _SchedSlot
Me@42:  {
Me@42:    int         workIsDone;
Me@42:    int         needsProcrAssigned;
Me@42:    VirtProcr  *procrAssignedToSlot;
Me@42:  };
Me@42: //SchedSlot
Me@42: 
Me@62: /*WARNING: re-arranging this data structure could cause VP switching
Me@62:  *         assembly code to fail -- hard-codes offsets of fields
Me@62:  */
Me@42: struct _VirtProcr
Me@42:  { int         procrID;  //for debugging -- count up each time create
Me@42:    int         coreAnimatedBy;
Me@42:    void       *startOfStack;
Me@42:    void       *stackPtr;
Me@42:    void       *framePtr;
Me@42:    void       *nextInstrPt;
Me@42:    
Me@42:    void       *coreLoopStartPt;  //allows proto-runtime to be linked later
Me@42:    void       *coreLoopFramePtr; //restore before jmp back to core loop
Me@42:    void       *coreLoopStackPtr; //restore before jmp back to core loop
Me@42: 
Me@42:    void       *initialData;
Me@42: 
Me@42:    SchedSlot  *schedSlot;
Me@42:    VMSReqst   *requests;
Me@42: 
Me@50:    void       *semanticData; //this lives here for the life of VP
Me@53:    void       *dataRetFromReq;//values returned from plugin to VP go here
Me@42: 
Me@50:       //=========== MEASUREMENT STUFF ==========
Me@42:    #ifdef MEAS__TIME_STAMP_SUSP
Me@42:    unsigned int preSuspTSCLow;
Me@42:    unsigned int postSuspTSCLow;
Me@42:    #endif
Me@42:    #ifdef MEAS__TIME_MASTER
Me@42:    unsigned int startMasterTSCLow;
Me@42:    unsigned int endMasterTSCLow;
Me@42:    #endif
Me@50:    
Me@50:    float64      createPtInSecs;  //have space but don't use on some configs
Me@42:  };
Me@42: //VirtProcr
Me@42: 
Me@42: 
Me@62: /*WARNING: re-arranging this data structure could cause VP-switching
Me@62:  *         assembly code to fail -- hard-codes offsets of fields
Me@62:  *         (because -O3 messes with things otherwise)
Me@62:  */
Me@42: typedef struct
Me@42:  {
Me@42:    SlaveScheduler   slaveScheduler;
Me@42:    RequestHandler   requestHandler;
Me@42:    
Me@42:    SchedSlot     ***allSchedSlots;
Me@55:    VMSQueueStruc **readyToAnimateQs;
Me@42:    VirtProcr      **masterVPs;
Me@42: 
Me@42:    void            *semanticEnv;
Me@42:    void            *OSEventStruc;   //for future, when add I/O to BLIS
Me@50:    MallocProlog    *freeListHead;
Me@50:    int32            amtOfOutstandingMem; //total currently allocated
Me@42: 
Me@42:    void            *coreLoopStartPt;//addr to jump to to re-enter coreLoop
Me@42:    void            *coreLoopEndPt;  //addr to jump to to shut down a coreLoop
Me@42: 
Me@50:    int32            setupComplete;
Me@50:    int32            masterLock;
Me@42: 
Me@50:    int32            numMasterInARow[NUM_CORES];//detect back-to-back masterVP
Me@59:    GateStruc       *workStealingGates[ NUM_CORES ]; //concurrent work-steal
Me@55:    int32            workStealingLock;
Me@55:    
Me@50:    int32            numProcrsCreated; //gives ordering to processor creation
Me@50: 
Me@50:       //=========== MEASUREMENT STUFF =============
Me@50:    IntervalProbe  **intervalProbes;
Me@53:    PrivDynArrayInfo    *dynIntervalProbesInfo;
Me@50:    HashTable       *probeNameHashTbl;
Me@50:    int32            masterCreateProbeID;
Me@50:    float64          createPtInSecs;
Me@42:  }
Me@42: MasterEnv;
Me@42: 
Me@55: //=========================  Extra Stuff Data Strucs  =======================
Me@54: typedef struct
Me@54:  {
Me@42: 
Me@54:  }
Me@54: VMSExcp;
Me@50: 
Me@55: struct _GateStruc
Me@55:  {
Me@55:    int32 gateClosed;
Me@55:    int32 preGateProgress;
Me@55:    int32 waitProgress;
Me@55:    int32 exitProgress;
Me@55:  };
Me@55: //GateStruc
Me@50: 
Me@50: //=======================  OS Thread related  ===============================
Me@42: 
Me@42: void * coreLoop( void *paramsIn );  //standard PThreads fn prototype
Me@42: void * coreLoop_Seq( void *paramsIn );  //standard PThreads fn prototype
Me@42: void masterLoop( void *initData, VirtProcr *masterPr );
Me@42: 
Me@42: 
Me@50: typedef struct
Me@50:  {
Me@50:    void           *endThdPt;
Me@50:    unsigned int    coreNum;
Me@50:  }
Me@50: ThdParams;
Me@42: 
Me@42: pthread_t       coreLoopThdHandles[ NUM_CORES ];  //pthread's virt-procr state
Me@42: ThdParams      *coreLoopThdParams [ NUM_CORES ];
Me@42: pthread_mutex_t suspendLock;
Me@42: pthread_cond_t  suspend_cond;
Me@42: 
Me@50: 
Me@50: 
Me@50: //=====================  Global Vars ===================
Me@50: 
Me@42: volatile MasterEnv      *_VMSMasterEnv;
Me@42: 
Me@50: 
Me@50: 
Me@50: 
Me@50: //===========================  Function Prototypes  =========================
Me@50: 
Me@53: 
Me@53: //========== Setup and shutdown ==========
Me@42: void
Me@42: VMS__init();
Me@42: 
Me@42: void
Me@42: VMS__init_Seq();
Me@42: 
Me@42: void
Me@42: VMS__start_the_work_then_wait_until_done();
Me@42: 
Me@42: void
Me@42: VMS__start_the_work_then_wait_until_done_Seq();
Me@42: 
Me@42: VirtProcr *
Me@42: VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
Me@42: 
Me@53: void
Me@53: VMS__dissipate_procr( VirtProcr *procrToDissipate );
Me@53: 
Me@50:    //Use this to create processor inside entry point & other places outside
Me@50:    // the VMS system boundary (IE, not run in slave nor Master)
Me@50: VirtProcr *
Me@50: VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
Me@50: 
Me@53: void
Me@53: VMS_ext__dissipate_procr( VirtProcr *procrToDissipate );
Me@42: 
Me@50: void
Me@54: VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData );
Me@54: 
Me@54: void
Me@53: VMS__shutdown();
Me@53: 
Me@53: void
Me@53: VMS__cleanup_at_end_of_shutdown();
Me@50: 
Me@50: 
Me@50: //==============  Request Related  ===============
Me@50: 
Me@50: void
Me@50: VMS__suspend_procr( VirtProcr *callingPr );
Me@50: 
Me@42: inline void
Me@53: VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData, VirtProcr *callingPr );
Me@53: 
Me@53: inline void
Me@53: VMS__send_sem_request( void *semReqData, VirtProcr *callingPr );
Me@42: 
Me@42: void
Me@50: VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr );
Me@42: 
Me@53: void inline
Me@53: VMS__send_dissipate_req( VirtProcr *prToDissipate );
Me@53: 
Me@52: inline void
Me@52: VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr );
Me@52: 
Me@42: VMSReqst *
Me@50: VMS__take_next_request_out_of( VirtProcr *procrWithReq );
Me@42: 
Me@42: inline void *
Me@42: VMS__take_sem_reqst_from( VMSReqst *req );
Me@42: 
Me@53: //======================== STATS ======================
Me@42: 
Me@53: //===== RDTSC wrapper =====
Me@42: 
Me@42: #define saveTimeStampCountInto(low, high) \
Me@42:    asm volatile("RDTSC;                   \
Me@42:                  movl %%eax, %0;          \
Me@42:                  movl %%edx, %1;"         \
Me@42:    /* outputs */ : "=m" (low), "=m" (high)\
Me@42:    /* inputs  */ :                        \
Me@42:    /* clobber */ : "%eax", "%edx"         \
Me@42:                 );
Me@42: 
Me@42: #define saveLowTimeStampCountInto(low)    \
Me@42:    asm volatile("RDTSC;                   \
Me@42:                  movl %%eax, %0;"         \
Me@42:    /* outputs */ : "=m" (low)             \
Me@42:    /* inputs  */ :                        \
Me@42:    /* clobber */ : "%eax", "%edx"         \
Me@42:                 );
Me@53: //=====
Me@42: 
Me@55: #include "SwitchAnimators.h"
Me@50: #include "probes.h"
Me@42: 
Me@42: #endif	/* _VMS_H */
Me@42: