1.1 --- a/VMS.h	Sat Sep 11 04:40:12 2010 -0700
     1.2 +++ b/VMS.h	Sat Sep 11 04:47:17 2010 -0700
     1.3 @@ -1,272 +1,272 @@
     1.4 -/*
     1.5 - *  Copyright 2009 OpenSourceStewardshipFoundation.org
     1.6 - *  Licensed under GNU General Public License version 2
     1.7 - *
     1.8 - * Author: seanhalle@yahoo.com
     1.9 - *
    1.10 - */
    1.11 -
    1.12 -#ifndef _VMS_H
    1.13 -#define	_VMS_H
    1.14 -#define __USE_GNU
    1.15 -
    1.16 -#include "VMS_primitive_data_types.h"
    1.17 -#include "Queue_impl/BlockingQueue.h"
    1.18 -#include "Histogram/Histogram.h"
    1.19 -#include <pthread.h>
    1.20 -
    1.21 -   //When DEBUG is defined, VMS does sequential exe in the main thread
    1.22 -   // It still does co-routines and all the mechanisms are the same, it just
    1.23 -   // has only a single thread and animates VPs one at a time
    1.24 -#define SEQUENTIAL
    1.25 -
    1.26 -   //when MEAS__TAKE_SUSP_TSC is defined, causes code to be inserted and
    1.27 -   // compiled-in that saves the low part of the time stamp count just before
    1.28 -   // suspending a processor and just after resuming that processor.  It is
    1.29 -   // saved into a field added to VirtProcr.  Have to sanity-check for
    1.30 -   // rollover of low portion into high portion.
    1.31 -#define MEAS__TIME_STAMP_SUSP
    1.32 -#define MEAS__TIME_MASTER
    1.33 -#define MEAS__NUM_TIMES_TO_RUN 100000
    1.34 -
    1.35 -   //This value is the number of hardware threads in the shared memory
    1.36 -   // machine
    1.37 -#define NUM_CORES        4
    1.38 -
    1.39 -   // make double-num-cores scheduling slots, plus extra for master
    1.40 -//#define NUM_SCHED_SLOTS  (2 * NUM_CORES + 1)
    1.41 -#define NUM_SCHED_SLOTS  3
    1.42 -
    1.43 -#define READYTOANIMATE_RETRIES 10000
    1.44 -
    1.45 -   // stack
    1.46 -#define VIRT_PROCR_STACK_SIZE 0x10000
    1.47 -
    1.48 -   //256M of total memory for VMS__malloc
    1.49 -#define MASSIVE_MALLOC_SIZE 0x10000000
    1.50 -
    1.51 -#define NUM_PREPEND_BYTES sizeof(FreeListElem) + sizeof(ownerElem);
    1.52 -
    1.53 -#define SUCCESS 0
    1.54 -
    1.55 -#define writeVMSQ     writeCASQ
    1.56 -#define readVMSQ      readCASQ
    1.57 -#define makeVMSQ      makeCASQ
    1.58 -#define VMSQueueStruc CASQueueStruc
    1.59 -
    1.60 -//#define thdAttrs NULL  //For PThreads
    1.61 -
    1.62 -typedef struct _SchedSlot  SchedSlot;
    1.63 -typedef struct _VMSReqst   VMSReqst;
    1.64 -typedef struct _VirtProcr  VirtProcr;
    1.65 -
    1.66 -typedef VirtProcr * (*SlaveScheduler)  ( void *, int );   //semEnv, coreIdx
    1.67 -typedef void  (*RequestHandler)  ( VirtProcr *, void * ); //prWReqst, semEnv
    1.68 -typedef void  (*VirtProcrFnPtr)  ( void *, VirtProcr * ); //initData, animPr
    1.69 -typedef void    VirtProcrFn      ( void *, VirtProcr * ); //initData, animPr
    1.70 -
    1.71 -typedef struct
    1.72 - {
    1.73 -   void           *endThdPt;
    1.74 -   unsigned int    coreNum;
    1.75 - }
    1.76 -ThdParams;
    1.77 -
    1.78 -
    1.79 -struct _SchedSlot
    1.80 - {
    1.81 -   int         workIsDone;
    1.82 -   int         needsProcrAssigned;
    1.83 -   VirtProcr  *procrAssignedToSlot;
    1.84 - };
    1.85 -//SchedSlot
    1.86 - 
    1.87 -enum ReqstType
    1.88 - {
    1.89 -   semantic = 1,
    1.90 -   dissipate,
    1.91 -   regCreated,
    1.92 -   IO
    1.93 - };
    1.94 -
    1.95 -struct _VMSReqst
    1.96 - {
    1.97 -//   VirtProcr   *virtProcrFrom;
    1.98 -   enum ReqstType  reqType;//used for dissipate and in future for IO requests
    1.99 -   void           *semReqData;
   1.100 -
   1.101 -   VMSReqst *nextReqst;
   1.102 - };
   1.103 -//VMSReqst
   1.104 -
   1.105 -struct _VirtProcr
   1.106 - { int         procrID;  //for debugging -- count up each time create
   1.107 -   int         coreAnimatedBy;
   1.108 -   void       *startOfStack;
   1.109 -   void       *stackPtr;
   1.110 -   void       *framePtr;
   1.111 -   void       *nextInstrPt;
   1.112 -   
   1.113 -   void       *coreLoopStartPt;  //allows proto-runtime to be linked later
   1.114 -   void       *coreLoopFramePtr; //restore before jmp back to core loop
   1.115 -   void       *coreLoopStackPtr; //restore before jmp back to core loop
   1.116 -
   1.117 -   void       *initialData;
   1.118 -
   1.119 -   SchedSlot  *schedSlot;
   1.120 -   VMSReqst   *requests;
   1.121 -
   1.122 -   void       *semanticData;
   1.123 -
   1.124 -   //============================= MEASUREMENT STUFF ========================
   1.125 -   #ifdef MEAS__TIME_STAMP_SUSP
   1.126 -   unsigned int preSuspTSCLow;
   1.127 -   unsigned int postSuspTSCLow;
   1.128 -   #endif
   1.129 -   #ifdef MEAS__TIME_MASTER
   1.130 -   unsigned int startMasterTSCLow;
   1.131 -   unsigned int endMasterTSCLow;
   1.132 -   #endif
   1.133 -   //========================================================================
   1.134 - };
   1.135 -//VirtProcr
   1.136 -
   1.137 -
   1.138 -
   1.139 -typedef struct
   1.140 - {
   1.141 -   SlaveScheduler   slaveScheduler;
   1.142 -   RequestHandler   requestHandler;
   1.143 -   
   1.144 -   SchedSlot     ***allSchedSlots;
   1.145 -   SRSWQueueStruc **readyToAnimateQs;
   1.146 -   VirtProcr      **masterVPs;
   1.147 -
   1.148 -   void            *semanticEnv;
   1.149 -   void            *OSEventStruc;   //for future, when add I/O to BLIS
   1.150 -
   1.151 -   void            *coreLoopStartPt;//addr to jump to to re-enter coreLoop
   1.152 -   void            *coreLoopEndPt;  //addr to jump to to shut down a coreLoop
   1.153 -
   1.154 -   int              setupComplete;
   1.155 -   int              masterLock;
   1.156 -
   1.157 - }
   1.158 -MasterEnv;
   1.159 -
   1.160 -
   1.161 -//==========================================================
   1.162 -
   1.163 -void * coreLoop( void *paramsIn );  //standard PThreads fn prototype
   1.164 -void * coreLoop_Seq( void *paramsIn );  //standard PThreads fn prototype
   1.165 -void masterLoop( void *initData, VirtProcr *masterPr );
   1.166 -
   1.167 -
   1.168 -//=====================  Global Vars ===================
   1.169 -
   1.170 -
   1.171 -pthread_t       coreLoopThdHandles[ NUM_CORES ];  //pthread's virt-procr state
   1.172 -ThdParams      *coreLoopThdParams [ NUM_CORES ];
   1.173 -pthread_mutex_t suspendLock;
   1.174 -pthread_cond_t  suspend_cond;
   1.175 -
   1.176 -volatile MasterEnv      *_VMSMasterEnv;
   1.177 -
   1.178 -//==========================
   1.179 -void
   1.180 -VMS__init();
   1.181 -
   1.182 -void
   1.183 -VMS__init_Seq();
   1.184 -
   1.185 -void
   1.186 -VMS__start_the_work_then_wait_until_done();
   1.187 -
   1.188 -void
   1.189 -VMS__start_the_work_then_wait_until_done_Seq();
   1.190 -
   1.191 -VirtProcr *
   1.192 -VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
   1.193 -
   1.194 -VirtProcr *
   1.195 -VMS__create_the_shutdown_procr();
   1.196 -
   1.197 -//==========================
   1.198 -inline void
   1.199 -VMS__add_sem_request( void *semReqData, VirtProcr *callingPr );
   1.200 -
   1.201 -void
   1.202 -VMS__send_req_to_register_new_procr( VirtProcr *newPrToRegister,
   1.203 -                                      VirtProcr *reqstingPr );
   1.204 -
   1.205 -void
   1.206 -VMS__free_request( VMSReqst *req );
   1.207 -
   1.208 -void
   1.209 -VMS__remove_and_free_top_request( VirtProcr *reqstingPr );
   1.210 -
   1.211 -VMSReqst *
   1.212 -VMS__take_top_request_from( VirtProcr *reqstingPr );
   1.213 -
   1.214 -VMSReqst *
   1.215 -VMS__free_top_and_give_next_request_from( VirtProcr *procrWithReq );
   1.216 -
   1.217 -inline void *
   1.218 -VMS__take_sem_reqst_from( VMSReqst *req );
   1.219 -
   1.220 -inline int
   1.221 -VMS__isSemanticReqst( VMSReqst *req );
   1.222 -
   1.223 -inline int
   1.224 -VMS__isDissipateReqst( VMSReqst *req );
   1.225 -
   1.226 -inline int
   1.227 -VMS__isCreateReqst( VMSReqst *req );
   1.228 -
   1.229 -//==========================
   1.230 -
   1.231 -void
   1.232 -VMS__suspend_procr( VirtProcr *callingPr );
   1.233 -
   1.234 -void
   1.235 -VMS__dissipate_procr( VirtProcr *prToDissipate );
   1.236 -
   1.237 -void
   1.238 -VMS__handle_dissipate_reqst( VirtProcr *procrToDissipate );
   1.239 -
   1.240 -void
   1.241 -VMS__cleanup_after_shutdown();
   1.242 -
   1.243 -//============================= Statistics ==================================
   1.244 -
   1.245 -typedef unsigned long long TSCount;
   1.246 -
   1.247 -   //Frequency of TS counts
   1.248 -   //TODO: change freq for each machine
   1.249 -#define TSCOUNT_FREQ 3180000000
   1.250 -
   1.251 -#define saveTimeStampCountInto(low, high) \
   1.252 -   asm volatile("RDTSC;                   \
   1.253 -                 movl %%eax, %0;          \
   1.254 -                 movl %%edx, %1;"         \
   1.255 -   /* outputs */ : "=m" (low), "=m" (high)\
   1.256 -   /* inputs  */ :                        \
   1.257 -   /* clobber */ : "%eax", "%edx"         \
   1.258 -                );
   1.259 -
   1.260 -#define saveLowTimeStampCountInto(low)    \
   1.261 -   asm volatile("RDTSC;                   \
   1.262 -                 movl %%eax, %0;"         \
   1.263 -   /* outputs */ : "=m" (low)             \
   1.264 -   /* inputs  */ :                        \
   1.265 -   /* clobber */ : "%eax", "%edx"         \
   1.266 -                );
   1.267 -
   1.268 -inline TSCount getTSCount();
   1.269 -
   1.270 -//===================== Debug ==========================
   1.271 -int numProcrsCreated;
   1.272 -
   1.273 -
   1.274 -#endif	/* _VMS_H */
   1.275 -
   1.276 +/*
   1.277 + *  Copyright 2009 OpenSourceStewardshipFoundation.org
   1.278 + *  Licensed under GNU General Public License version 2
   1.279 + *
   1.280 + * Author: seanhalle@yahoo.com
   1.281 + * 
   1.282 + */
   1.283 +
   1.284 +#ifndef _VMS_H
   1.285 +#define	_VMS_H
   1.286 +#define __USE_GNU
   1.287 +
   1.288 +#include "VMS_primitive_data_types.h"
   1.289 +#include "Queue_impl/BlockingQueue.h"
   1.290 +#include "Histogram/Histogram.h"
   1.291 +#include <pthread.h>
   1.292 +
   1.293 +   //When DEBUG is defined, VMS does sequential exe in the main thread
   1.294 +   // It still does co-routines and all the mechanisms are the same, it just
   1.295 +   // has only a single thread and animates VPs one at a time
   1.296 +#define SEQUENTIAL
   1.297 +
   1.298 +   //when MEAS__TAKE_SUSP_TSC is defined, causes code to be inserted and
   1.299 +   // compiled-in that saves the low part of the time stamp count just before
   1.300 +   // suspending a processor and just after resuming that processor.  It is
   1.301 +   // saved into a field added to VirtProcr.  Have to sanity-check for
   1.302 +   // rollover of low portion into high portion.
   1.303 +#define MEAS__TIME_STAMP_SUSP
   1.304 +#define MEAS__TIME_MASTER
   1.305 +#define MEAS__NUM_TIMES_TO_RUN 100000
   1.306 +
   1.307 +   //This value is the number of hardware threads in the shared memory
   1.308 +   // machine
   1.309 +#define NUM_CORES        4
   1.310 +
   1.311 +   // make double-num-cores scheduling slots, plus extra for master
   1.312 +//#define NUM_SCHED_SLOTS  (2 * NUM_CORES + 1)
   1.313 +#define NUM_SCHED_SLOTS  3
   1.314 +
   1.315 +#define READYTOANIMATE_RETRIES 10000
   1.316 +
   1.317 +   // stack
   1.318 +#define VIRT_PROCR_STACK_SIZE 0x10000
   1.319 +
   1.320 +   //256M of total memory for VMS__malloc
   1.321 +#define MASSIVE_MALLOC_SIZE 0x10000000
   1.322 +
   1.323 +#define NUM_PREPEND_BYTES sizeof(FreeListElem) + sizeof(ownerElem);
   1.324 +
   1.325 +#define SUCCESS 0
   1.326 +
   1.327 +#define writeVMSQ     writeCASQ
   1.328 +#define readVMSQ      readCASQ
   1.329 +#define makeVMSQ      makeCASQ
   1.330 +#define VMSQueueStruc CASQueueStruc
   1.331 +
   1.332 +//#define thdAttrs NULL  //For PThreads
   1.333 +
   1.334 +typedef struct _SchedSlot  SchedSlot;
   1.335 +typedef struct _VMSReqst   VMSReqst;
   1.336 +typedef struct _VirtProcr  VirtProcr;
   1.337 +
   1.338 +typedef VirtProcr * (*SlaveScheduler)  ( void *, int );   //semEnv, coreIdx
   1.339 +typedef void  (*RequestHandler)  ( VirtProcr *, void * ); //prWReqst, semEnv
   1.340 +typedef void  (*VirtProcrFnPtr)  ( void *, VirtProcr * ); //initData, animPr
   1.341 +typedef void    VirtProcrFn      ( void *, VirtProcr * ); //initData, animPr
   1.342 +
   1.343 +typedef struct
   1.344 + {
   1.345 +   void           *endThdPt;
   1.346 +   unsigned int    coreNum;
   1.347 + }
   1.348 +ThdParams;
   1.349 +
   1.350 +
   1.351 +struct _SchedSlot
   1.352 + {
   1.353 +   int         workIsDone;
   1.354 +   int         needsProcrAssigned;
   1.355 +   VirtProcr  *procrAssignedToSlot;
   1.356 + };
   1.357 +//SchedSlot
   1.358 + 
   1.359 +enum ReqstType
   1.360 + {
   1.361 +   semantic = 1,
   1.362 +   dissipate,
   1.363 +   regCreated,
   1.364 +   IO
   1.365 + };
   1.366 +
   1.367 +struct _VMSReqst
   1.368 + {
   1.369 +//   VirtProcr   *virtProcrFrom;
   1.370 +   enum ReqstType  reqType;//used for dissipate and in future for IO requests
   1.371 +   void           *semReqData;
   1.372 +
   1.373 +   VMSReqst *nextReqst;
   1.374 + };
   1.375 +//VMSReqst
   1.376 +
   1.377 +struct _VirtProcr
   1.378 + { int         procrID;  //for debugging -- count up each time create
   1.379 +   int         coreAnimatedBy;
   1.380 +   void       *startOfStack;
   1.381 +   void       *stackPtr;
   1.382 +   void       *framePtr;
   1.383 +   void       *nextInstrPt;
   1.384 +   
   1.385 +   void       *coreLoopStartPt;  //allows proto-runtime to be linked later
   1.386 +   void       *coreLoopFramePtr; //restore before jmp back to core loop
   1.387 +   void       *coreLoopStackPtr; //restore before jmp back to core loop
   1.388 +
   1.389 +   void       *initialData;
   1.390 +
   1.391 +   SchedSlot  *schedSlot;
   1.392 +   VMSReqst   *requests;
   1.393 +
   1.394 +   void       *semanticData;
   1.395 +
   1.396 +   //============================= MEASUREMENT STUFF ========================
   1.397 +   #ifdef MEAS__TIME_STAMP_SUSP
   1.398 +   unsigned int preSuspTSCLow;
   1.399 +   unsigned int postSuspTSCLow;
   1.400 +   #endif
   1.401 +   #ifdef MEAS__TIME_MASTER
   1.402 +   unsigned int startMasterTSCLow;
   1.403 +   unsigned int endMasterTSCLow;
   1.404 +   #endif
   1.405 +   //========================================================================
   1.406 + };
   1.407 +//VirtProcr
   1.408 +
   1.409 +
   1.410 +
   1.411 +typedef struct
   1.412 + {
   1.413 +   SlaveScheduler   slaveScheduler;
   1.414 +   RequestHandler   requestHandler;
   1.415 +   
   1.416 +   SchedSlot     ***allSchedSlots;
   1.417 +   SRSWQueueStruc **readyToAnimateQs;
   1.418 +   VirtProcr      **masterVPs;
   1.419 +
   1.420 +   void            *semanticEnv;
   1.421 +   void            *OSEventStruc;   //for future, when add I/O to BLIS
   1.422 +
   1.423 +   void            *coreLoopStartPt;//addr to jump to to re-enter coreLoop
   1.424 +   void            *coreLoopEndPt;  //addr to jump to to shut down a coreLoop
   1.425 +
   1.426 +   int              setupComplete;
   1.427 +   int              masterLock;
   1.428 +
   1.429 + }
   1.430 +MasterEnv;
   1.431 +
   1.432 +
   1.433 +//==========================================================
   1.434 +
   1.435 +void * coreLoop( void *paramsIn );  //standard PThreads fn prototype
   1.436 +void * coreLoop_Seq( void *paramsIn );  //standard PThreads fn prototype
   1.437 +void masterLoop( void *initData, VirtProcr *masterPr );
   1.438 +
   1.439 +
   1.440 +//=====================  Global Vars ===================
   1.441 +
   1.442 +
   1.443 +pthread_t       coreLoopThdHandles[ NUM_CORES ];  //pthread's virt-procr state
   1.444 +ThdParams      *coreLoopThdParams [ NUM_CORES ];
   1.445 +pthread_mutex_t suspendLock;
   1.446 +pthread_cond_t  suspend_cond;
   1.447 +
   1.448 +volatile MasterEnv      *_VMSMasterEnv;
   1.449 +
   1.450 +//==========================
   1.451 +void
   1.452 +VMS__init();
   1.453 +
   1.454 +void
   1.455 +VMS__init_Seq();
   1.456 +
   1.457 +void
   1.458 +VMS__start_the_work_then_wait_until_done();
   1.459 +
   1.460 +void
   1.461 +VMS__start_the_work_then_wait_until_done_Seq();
   1.462 +
   1.463 +VirtProcr *
   1.464 +VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData );
   1.465 +
   1.466 +VirtProcr *
   1.467 +VMS__create_the_shutdown_procr();
   1.468 +
   1.469 +//==========================
   1.470 +inline void
   1.471 +VMS__add_sem_request( void *semReqData, VirtProcr *callingPr );
   1.472 +
   1.473 +void
   1.474 +VMS__send_req_to_register_new_procr( VirtProcr *newPrToRegister,
   1.475 +                                      VirtProcr *reqstingPr );
   1.476 +
   1.477 +void
   1.478 +VMS__free_request( VMSReqst *req );
   1.479 +
   1.480 +void
   1.481 +VMS__remove_and_free_top_request( VirtProcr *reqstingPr );
   1.482 +
   1.483 +VMSReqst *
   1.484 +VMS__take_top_request_from( VirtProcr *reqstingPr );
   1.485 +
   1.486 +VMSReqst *
   1.487 +VMS__free_top_and_give_next_request_from( VirtProcr *procrWithReq );
   1.488 +
   1.489 +inline void *
   1.490 +VMS__take_sem_reqst_from( VMSReqst *req );
   1.491 +
   1.492 +inline int
   1.493 +VMS__isSemanticReqst( VMSReqst *req );
   1.494 +
   1.495 +inline int
   1.496 +VMS__isDissipateReqst( VMSReqst *req );
   1.497 +
   1.498 +inline int
   1.499 +VMS__isCreateReqst( VMSReqst *req );
   1.500 +
   1.501 +//==========================
   1.502 +
   1.503 +void
   1.504 +VMS__suspend_procr( VirtProcr *callingPr );
   1.505 +
   1.506 +void
   1.507 +VMS__dissipate_procr( VirtProcr *prToDissipate );
   1.508 +
   1.509 +void
   1.510 +VMS__handle_dissipate_reqst( VirtProcr *procrToDissipate );
   1.511 +
   1.512 +void
   1.513 +VMS__cleanup_after_shutdown();
   1.514 +
   1.515 +//============================= Statistics ==================================
   1.516 +
   1.517 +typedef unsigned long long TSCount;
   1.518 +
   1.519 +   //Frequency of TS counts
   1.520 +   //TODO: change freq for each machine
   1.521 +#define TSCOUNT_FREQ 3180000000
   1.522 +
   1.523 +#define saveTimeStampCountInto(low, high) \
   1.524 +   asm volatile("RDTSC;                   \
   1.525 +                 movl %%eax, %0;          \
   1.526 +                 movl %%edx, %1;"         \
   1.527 +   /* outputs */ : "=m" (low), "=m" (high)\
   1.528 +   /* inputs  */ :                        \
   1.529 +   /* clobber */ : "%eax", "%edx"         \
   1.530 +                );
   1.531 +
   1.532 +#define saveLowTimeStampCountInto(low)    \
   1.533 +   asm volatile("RDTSC;                   \
   1.534 +                 movl %%eax, %0;"         \
   1.535 +   /* outputs */ : "=m" (low)             \
   1.536 +   /* inputs  */ :                        \
   1.537 +   /* clobber */ : "%eax", "%edx"         \
   1.538 +                );
   1.539 +
   1.540 +inline TSCount getTSCount();
   1.541 +
   1.542 +//===================== Debug ==========================
   1.543 +int numProcrsCreated;
   1.544 +
   1.545 +
   1.546 +#endif	/* _VMS_H */
   1.547 +