1.1 --- a/MasterLoop.c	Thu Oct 06 16:24:17 2011 +0200
     1.2 +++ b/MasterLoop.c	Wed Jan 04 16:10:11 2012 -0800
     1.3 @@ -1,373 +1,373 @@
     1.4 -/*
     1.5 - * Copyright 2010  OpenSourceStewardshipFoundation
     1.6 - * 
     1.7 - * Licensed under BSD
     1.8 - */
     1.9 -
    1.10 -
    1.11 -
    1.12 -#include <stdio.h>
    1.13 -#include <stddef.h>
    1.14 -
    1.15 -#include "VMS.h"
    1.16 -#include "ProcrContext.h"
    1.17 -
    1.18 -
    1.19 -//===========================================================================
    1.20 -void inline
    1.21 -stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
    1.22 -               VirtProcr *masterPr );
    1.23 -
    1.24 -//===========================================================================
    1.25 -
    1.26 -
    1.27 -
    1.28 -/*This code is animated by the virtual Master processor.
    1.29 - *
    1.30 - *Polls each sched slot exactly once, hands any requests made by a newly
    1.31 - * done slave to the "request handler" plug-in function
    1.32 - *
    1.33 - *Any slots that need a virt procr assigned are given to the "schedule"
    1.34 - * plug-in function, which tries to assign a virt procr (slave) to it.
    1.35 - *
    1.36 - *When all slots needing a processor have been given to the schedule plug-in,
    1.37 - * a fraction of the procrs successfully scheduled are put into the
    1.38 - * work queue, then a continuation of this function is put in, then the rest
    1.39 - * of the virt procrs that were successfully scheduled.
    1.40 - *
    1.41 - *The first thing the continuation does is busy-wait until the previous
    1.42 - * animation completes.  This is because an (unlikely) continuation may
    1.43 - * sneak through queue before previous continuation is done putting second
    1.44 - * part of scheduled slaves in, which is the only race condition.
    1.45 - *
    1.46 - */
    1.47 -
    1.48 -/*May 29, 2010 -- birth a Master during init so that first core loop to
    1.49 - * start running gets it and does all the stuff for a newly born --
    1.50 - * from then on, will be doing continuation, but do suspension self
    1.51 - * directly at end of master loop
    1.52 - *So VMS__init just births the master virtual processor same way it births
    1.53 - * all the others -- then does any extra setup needed and puts it into the
    1.54 - * work queue.
    1.55 - *However means have to make masterEnv a global static volatile the same way
    1.56 - * did with readyToAnimateQ in core loop.  -- for performance, put the
    1.57 - * jump to the core loop directly in here, and have it directly jump back.
    1.58 - *
    1.59 - *
    1.60 - *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
    1.61 - * avoids the suspected bug in the system stack that causes bizarre faults
    1.62 - * at random places in the system code.
    1.63 - *
    1.64 - *So, this function is coupled to each of the MasterVPs, -- meaning this
    1.65 - * function can't rely on a particular stack and frame -- each MasterVP that
    1.66 - * animates this function has a different one.
    1.67 - *
    1.68 - *At this point, the masterLoop does not write itself into the queue anymore,
    1.69 - * instead, the coreLoop acquires the masterLock when it has nothing to
    1.70 - * animate, and then animates its own masterLoop.  However, still try to put
    1.71 - * several AppVPs into the queue to amortize the startup cost of switching
    1.72 - * to the MasterVP.  Note, don't have to worry about latency of requests much
    1.73 - * because most requests generate work for same core -- only latency issue
    1.74 - * is case when other cores starved and one core's requests generate work
    1.75 - * for them -- so keep max in queue to 3 or 4..
    1.76 - */
    1.77 -void masterLoop( void *initData, VirtProcr *animatingPr )
    1.78 - { 
    1.79 -   int32           slotIdx, numSlotsFilled;
    1.80 -   VirtProcr      *schedVirtPr;
    1.81 -   SchedSlot      *currSlot, **schedSlots;
    1.82 -   MasterEnv      *masterEnv;
    1.83 -   VMSQueueStruc  *readyToAnimateQ;
    1.84 -   
    1.85 -   SlaveScheduler  slaveScheduler;
    1.86 -   RequestHandler  requestHandler;
    1.87 -   void           *semanticEnv;
    1.88 -
    1.89 -   int32           thisCoresIdx;
    1.90 -   VirtProcr      *masterPr;
    1.91 -   volatile        VirtProcr *volatileMasterPr;
    1.92 -   
    1.93 -   volatileMasterPr = animatingPr;
    1.94 -   masterPr         = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
    1.95 -
    1.96 -      //First animation of each MasterVP will in turn animate this part
    1.97 -      // of setup code.. (VP creator sets up the stack as if this function
    1.98 -      // was called normally, but actually get here by jmp)
    1.99 -      //So, setup values about stack ptr, jmp pt and all that
   1.100 -   //masterPr->nextInstrPt = &&masterLoopStartPt;
   1.101 -
   1.102 -
   1.103 -      //Note, got rid of writing the stack and frame ptr up here, because
   1.104 -      // only one
   1.105 -      // core can ever animate a given MasterVP, so don't need to communicate
   1.106 -      // new frame and stack ptr to the MasterVP storage before a second
   1.107 -      // version of that MasterVP can get animated on a different core.
   1.108 -      //Also got rid of the busy-wait.
   1.109 -
   1.110 -   
   1.111 -   //masterLoopStartPt:
   1.112 -   while(1){
   1.113 -       
   1.114 -   //============================= MEASUREMENT STUFF ========================
   1.115 -   #ifdef MEAS__TIME_MASTER
   1.116 -      //Total Master time includes one coreloop time -- just assume the core
   1.117 -      // loop time is same for Master as for AppVPs, even though it may be
   1.118 -      // smaller due to higher predictability of the fixed jmp.
   1.119 -   saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
   1.120 -   #endif
   1.121 -   //========================================================================
   1.122 -
   1.123 -   masterEnv        = (MasterEnv*)_VMSMasterEnv;
   1.124 -   
   1.125 -      //GCC may optimize so doesn't always re-define from frame-storage
   1.126 -   masterPr         = (VirtProcr*)volatileMasterPr;  //just to make sure after jmp
   1.127 -   thisCoresIdx     = masterPr->coreAnimatedBy;
   1.128 -   readyToAnimateQ  = masterEnv->readyToAnimateQs[thisCoresIdx];
   1.129 -   schedSlots       = masterEnv->allSchedSlots[thisCoresIdx];
   1.130 -
   1.131 -   requestHandler   = masterEnv->requestHandler;
   1.132 -   slaveScheduler   = masterEnv->slaveScheduler;
   1.133 -   semanticEnv      = masterEnv->semanticEnv;
   1.134 -
   1.135 -
   1.136 -      //Poll each slot's Done flag
   1.137 -   numSlotsFilled = 0;
   1.138 -   for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
   1.139 -    {
   1.140 -      currSlot = schedSlots[ slotIdx ];
   1.141 -
   1.142 -      if( currSlot->workIsDone )
   1.143 -       {
   1.144 -         currSlot->workIsDone         = FALSE;
   1.145 -         currSlot->needsProcrAssigned = TRUE;
   1.146 -
   1.147 -            //process requests from slave to master
   1.148 -               //====================== MEASUREMENT STUFF ===================
   1.149 -               #ifdef MEAS__TIME_PLUGIN
   1.150 -               int32 startStamp1, endStamp1;
   1.151 -               saveLowTimeStampCountInto( startStamp1 );
   1.152 -               #endif
   1.153 -               //============================================================
   1.154 -         (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
   1.155 -               //====================== MEASUREMENT STUFF ===================
   1.156 -               #ifdef MEAS__TIME_PLUGIN
   1.157 -               saveLowTimeStampCountInto( endStamp1 );
   1.158 -               addIntervalToHist( startStamp1, endStamp1,
   1.159 -                                        _VMSMasterEnv->reqHdlrLowTimeHist );
   1.160 -               addIntervalToHist( startStamp1, endStamp1,
   1.161 -                                        _VMSMasterEnv->reqHdlrHighTimeHist );
   1.162 -               #endif
   1.163 -               //============================================================
   1.164 -       }
   1.165 -      if( currSlot->needsProcrAssigned )
   1.166 -       {    //give slot a new virt procr
   1.167 -         schedVirtPr =
   1.168 -          (*slaveScheduler)( semanticEnv, thisCoresIdx );
   1.169 -         
   1.170 -         if( schedVirtPr != NULL )
   1.171 -          { currSlot->procrAssignedToSlot = schedVirtPr;
   1.172 -            schedVirtPr->schedSlot        = currSlot;
   1.173 -            currSlot->needsProcrAssigned  = FALSE;
   1.174 -            numSlotsFilled               += 1;
   1.175 -            
   1.176 -            writeVMSQ( schedVirtPr, readyToAnimateQ );
   1.177 -          }
   1.178 -       }
   1.179 -    }
   1.180 -
   1.181 -   
   1.182 -   #ifdef USE_WORK_STEALING
   1.183 -      //If no slots filled, means no more work, look for work to steal.
   1.184 -   if( numSlotsFilled == 0 )
   1.185 -    { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
   1.186 -    }
   1.187 -   #endif
   1.188 -
   1.189 -   
   1.190 -   #ifdef MEAS__TIME_MASTER
   1.191 -   saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
   1.192 -   #endif
   1.193 -
   1.194 -   masterSwitchToCoreLoop(animatingPr);
   1.195 -   flushRegisters();
   1.196 -   }//MasterLoop
   1.197 -
   1.198 -
   1.199 - }
   1.200 -
   1.201 -
   1.202 -
   1.203 -/*This has a race condition -- the coreloops are accessing their own queues
   1.204 - * at the same time that this work-stealer on a different core is trying to
   1.205 - */
   1.206 -void inline
   1.207 -stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
   1.208 -               VirtProcr *masterPr )
   1.209 - { 
   1.210 -   VirtProcr   *stolenPr;
   1.211 -   int32        coreIdx, i;
   1.212 -   VMSQueueStruc *currQ;
   1.213 -
   1.214 -   stolenPr = NULL;
   1.215 -   coreIdx = masterPr->coreAnimatedBy;
   1.216 -   for( i = 0; i < NUM_CORES -1; i++ )
   1.217 -    {
   1.218 -      if( coreIdx >= NUM_CORES -1 )
   1.219 -       { coreIdx = 0;
   1.220 -       }
   1.221 -      else
   1.222 -       { coreIdx++;
   1.223 -       }
   1.224 -      currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
   1.225 -      if( numInVMSQ( currQ ) > 0 )
   1.226 -       { stolenPr = readVMSQ (currQ );
   1.227 -         break;
   1.228 -       }
   1.229 -    }
   1.230 -
   1.231 -   if( stolenPr != NULL )
   1.232 -    { currSlot->procrAssignedToSlot = stolenPr;
   1.233 -      stolenPr->schedSlot           = currSlot;
   1.234 -      currSlot->needsProcrAssigned  = FALSE;
   1.235 -
   1.236 -      writeVMSQ( stolenPr, readyToAnimateQ );
   1.237 -    }
   1.238 - }
   1.239 -
   1.240 -/*This algorithm makes the common case fast.  Make the coreloop passive,
   1.241 - * and show its progress.  Make the stealer control a gate that coreloop
   1.242 - * has to pass.
   1.243 - *To avoid interference, only one stealer at a time.  Use a global
   1.244 - * stealer-lock.
   1.245 - *
   1.246 - *The pattern is based on a gate -- stealer shuts the gate, then monitors
   1.247 - * to be sure any already past make it all the way out, before starting.
   1.248 - *So, have a "progress" measure just before the gate, then have two after it,
   1.249 - * one is in a "waiting room" outside the gate, the other is at the exit.
   1.250 - *Then, the stealer first shuts the gate, then checks the progress measure
   1.251 - * outside it, then looks to see if the progress measure at the exit is the
   1.252 - * same.  If yes, it knows the protected area is empty 'cause no other way
   1.253 - * to get in and the last to get in also exited.
   1.254 - *If the progress measure at the exit is not the same, then the stealer goes
   1.255 - * into a loop checking both the waiting-area and the exit progress-measures
   1.256 - * until one of them shows the same as the measure outside the gate.  Might
   1.257 - * as well re-read the measure outside the gate each go around, just to be
   1.258 - * sure.  It is guaranteed that one of the two will eventually match the one
   1.259 - * outside the gate.
   1.260 - *
   1.261 - *Here's an informal proof of correctness:
   1.262 - *The gate can be closed at any point, and have only four cases:
   1.263 - *  1) coreloop made it past the gate-closing but not yet past the exit
   1.264 - *  2) coreloop made it past the pre-gate progress update but not yet past
   1.265 - *     the gate,
   1.266 - *  3) coreloop is right before the pre-gate update
   1.267 - *  4) coreloop is past the exit and far from the pre-gate update.
   1.268 - *
   1.269 - * Covering the cases in reverse order,
   1.270 - *  4) is not a problem -- stealer will read pre-gate progress, see that it
   1.271 - *     matches exit progress, and the gate is closed, so stealer can proceed.
   1.272 - *  3) stealer will read pre-gate progress just after coreloop updates it..
   1.273 - *     so stealer goes into a loop until the coreloop causes wait-progress
   1.274 - *     to match pre-gate progress, so then stealer can proceed
   1.275 - *  2) same as 3..
   1.276 - *  1) stealer reads pre-gate progress, sees that it's different than exit,
   1.277 - *     so goes into loop until exit matches pre-gate, now it knows coreloop
   1.278 - *     is not in protected and cannot get back in, so can proceed.
   1.279 - *
   1.280 - *Implementation for the stealer:
   1.281 - *
   1.282 - *First, acquire the stealer lock -- only cores with no work to do will
   1.283 - * compete to steal, so not a big performance penalty having only one --
   1.284 - * will rarely have multiple stealers in a system with plenty of work -- and
   1.285 - * in a system with little work, it doesn't matter.
   1.286 - *
   1.287 - *Note, have single-reader, single-writer pattern for all variables used to
   1.288 - * communicate between stealer and victims
   1.289 - *
   1.290 - *So, scan the queues of the core loops, until find non-empty.  Each core
   1.291 - * has its own list that it scans.  The list goes in order from closest to
   1.292 - * furthest core, so it steals first from close cores.  Later can add
   1.293 - * taking info from the app about overlapping footprints, and scan all the
   1.294 - * others then choose work with the most footprint overlap with the contents
   1.295 - * of this core's cache.
   1.296 - *
   1.297 - *Now, have a victim want to take work from.  So, shut the gate in that
   1.298 - * coreloop, by setting the "gate closed" var on its stack to TRUE.
   1.299 - *Then, read the core's pre-gate progress and compare to the core's exit
   1.300 - * progress.
   1.301 - *If same, can proceed to take work from the coreloop's queue.  When done,
   1.302 - * write FALSE to gate closed var.
   1.303 - *If different, then enter a loop that reads the pre-gate progress, then
   1.304 - * compares to exit progress then to wait progress.  When one of two
   1.305 - * matches, proceed.  Take work from the coreloop's queue.  When done,
   1.306 - * write FALSE to the gate closed var.
   1.307 - * 
   1.308 - */
   1.309 -void inline
   1.310 -gateProtected_stealWorkInto( SchedSlot *currSlot,
   1.311 -                             VMSQueueStruc *myReadyToAnimateQ,
   1.312 -                             VirtProcr *masterPr )
   1.313 - {
   1.314 -   VirtProcr     *stolenPr;
   1.315 -   int32          coreIdx, i, haveAVictim, gotLock;
   1.316 -   VMSQueueStruc *victimsQ;
   1.317 -
   1.318 -   volatile GateStruc *vicGate;
   1.319 -   int32               coreMightBeInProtected;
   1.320 -
   1.321 -
   1.322 -
   1.323 -      //see if any other cores have work available to steal
   1.324 -   haveAVictim = FALSE;
   1.325 -   coreIdx = masterPr->coreAnimatedBy;
   1.326 -   for( i = 0; i < NUM_CORES -1; i++ )
   1.327 -    {
   1.328 -      if( coreIdx >= NUM_CORES -1 )
   1.329 -       { coreIdx = 0;
   1.330 -       }
   1.331 -      else
   1.332 -       { coreIdx++;
   1.333 -       }
   1.334 -      victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
   1.335 -      if( numInVMSQ( victimsQ ) > 0 )
   1.336 -       { haveAVictim = TRUE;
   1.337 -         vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
   1.338 -         break;
   1.339 -       }
   1.340 -    }
   1.341 -   if( !haveAVictim ) return;  //no work to steal, exit
   1.342 -
   1.343 -      //have a victim core, now get the stealer-lock
   1.344 -   gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
   1.345 -                                                          UNLOCKED, LOCKED );
   1.346 -   if( !gotLock ) return; //go back to core loop, which will re-start master
   1.347 -
   1.348 -
   1.349 -   //====== Start Gate-protection =======
   1.350 -   vicGate->gateClosed = TRUE;
   1.351 -   coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
   1.352 -   while( coreMightBeInProtected )
   1.353 -    {    //wait until sure
   1.354 -      if( vicGate->preGateProgress == vicGate->waitProgress )
   1.355 -         coreMightBeInProtected = FALSE;
   1.356 -      if( vicGate->preGateProgress == vicGate->exitProgress )
   1.357 -         coreMightBeInProtected = FALSE;
   1.358 -    }
   1.359 -
   1.360 -   stolenPr = readVMSQ ( victimsQ );
   1.361 -
   1.362 -   vicGate->gateClosed = FALSE;
   1.363 -   //======= End Gate-protection  =======
   1.364 -
   1.365 -
   1.366 -   if( stolenPr != NULL )  //victim could have been in protected and taken
   1.367 -    { currSlot->procrAssignedToSlot = stolenPr;
   1.368 -      stolenPr->schedSlot           = currSlot;
   1.369 -      currSlot->needsProcrAssigned  = FALSE;
   1.370 -
   1.371 -      writeVMSQ( stolenPr, myReadyToAnimateQ );
   1.372 -    }
   1.373 -
   1.374 -      //unlock the work stealing lock
   1.375 -   _VMSMasterEnv->workStealingLock = UNLOCKED;
   1.376 - }
   1.377 +/*
   1.378 + * Copyright 2010  OpenSourceStewardshipFoundation
   1.379 + * 
   1.380 + * Licensed under BSD
   1.381 + */
   1.382 +
   1.383 +
   1.384 +
   1.385 +#include <stdio.h>
   1.386 +#include <stddef.h>
   1.387 +
   1.388 +#include "VMS.h"
   1.389 +#include "ProcrContext.h"
   1.390 +
   1.391 +
   1.392 +//===========================================================================
   1.393 +void inline
   1.394 +stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
   1.395 +               VirtProcr *masterPr );
   1.396 +
   1.397 +//===========================================================================
   1.398 +
   1.399 +
   1.400 +
   1.401 +/*This code is animated by the virtual Master processor.
   1.402 + *
   1.403 + *Polls each sched slot exactly once, hands any requests made by a newly
   1.404 + * done slave to the "request handler" plug-in function
   1.405 + *
   1.406 + *Any slots that need a virt procr assigned are given to the "schedule"
   1.407 + * plug-in function, which tries to assign a virt procr (slave) to it.
   1.408 + *
   1.409 + *When all slots needing a processor have been given to the schedule plug-in,
   1.410 + * a fraction of the procrs successfully scheduled are put into the
   1.411 + * work queue, then a continuation of this function is put in, then the rest
   1.412 + * of the virt procrs that were successfully scheduled.
   1.413 + *
   1.414 + *The first thing the continuation does is busy-wait until the previous
   1.415 + * animation completes.  This is because an (unlikely) continuation may
   1.416 + * sneak through queue before previous continuation is done putting second
   1.417 + * part of scheduled slaves in, which is the only race condition.
   1.418 + *
   1.419 + */
   1.420 +
   1.421 +/*May 29, 2010 -- birth a Master during init so that first core loop to
   1.422 + * start running gets it and does all the stuff for a newly born --
   1.423 + * from then on, will be doing continuation, but do suspension self
   1.424 + * directly at end of master loop
   1.425 + *So VMS__init just births the master virtual processor same way it births
   1.426 + * all the others -- then does any extra setup needed and puts it into the
   1.427 + * work queue.
   1.428 + *However means have to make masterEnv a global static volatile the same way
   1.429 + * did with readyToAnimateQ in core loop.  -- for performance, put the
   1.430 + * jump to the core loop directly in here, and have it directly jump back.
   1.431 + *
   1.432 + *
   1.433 + *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
   1.434 + * avoids the suspected bug in the system stack that causes bizarre faults
   1.435 + * at random places in the system code.
   1.436 + *
   1.437 + *So, this function is coupled to each of the MasterVPs, -- meaning this
   1.438 + * function can't rely on a particular stack and frame -- each MasterVP that
   1.439 + * animates this function has a different one.
   1.440 + *
   1.441 + *At this point, the masterLoop does not write itself into the queue anymore,
   1.442 + * instead, the coreLoop acquires the masterLock when it has nothing to
   1.443 + * animate, and then animates its own masterLoop.  However, still try to put
   1.444 + * several AppVPs into the queue to amortize the startup cost of switching
   1.445 + * to the MasterVP.  Note, don't have to worry about latency of requests much
   1.446 + * because most requests generate work for same core -- only latency issue
   1.447 + * is case when other cores starved and one core's requests generate work
   1.448 + * for them -- so keep max in queue to 3 or 4..
   1.449 + */
   1.450 +void masterLoop( void *initData, VirtProcr *animatingPr )
   1.451 + { 
   1.452 +   int32           slotIdx, numSlotsFilled;
   1.453 +   VirtProcr      *schedVirtPr;
   1.454 +   SchedSlot      *currSlot, **schedSlots;
   1.455 +   MasterEnv      *masterEnv;
   1.456 +   VMSQueueStruc  *readyToAnimateQ;
   1.457 +   
   1.458 +   SlaveScheduler  slaveScheduler;
   1.459 +   RequestHandler  requestHandler;
   1.460 +   void           *semanticEnv;
   1.461 +
   1.462 +   int32           thisCoresIdx;
   1.463 +   VirtProcr      *masterPr;
   1.464 +   volatile        VirtProcr *volatileMasterPr;
   1.465 +   
   1.466 +   volatileMasterPr = animatingPr;
   1.467 +   masterPr         = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
   1.468 +
   1.469 +      //First animation of each MasterVP will in turn animate this part
   1.470 +      // of setup code.. (VP creator sets up the stack as if this function
   1.471 +      // was called normally, but actually get here by jmp)
   1.472 +      //So, setup values about stack ptr, jmp pt and all that
   1.473 +   //masterPr->nextInstrPt = &&masterLoopStartPt;
   1.474 +
   1.475 +
   1.476 +      //Note, got rid of writing the stack and frame ptr up here, because
   1.477 +      // only one
   1.478 +      // core can ever animate a given MasterVP, so don't need to communicate
   1.479 +      // new frame and stack ptr to the MasterVP storage before a second
   1.480 +      // version of that MasterVP can get animated on a different core.
   1.481 +      //Also got rid of the busy-wait.
   1.482 +
   1.483 +   
   1.484 +   //masterLoopStartPt:
   1.485 +   while(1){
   1.486 +       
   1.487 +   //============================= MEASUREMENT STUFF ========================
   1.488 +   #ifdef MEAS__TIME_MASTER
   1.489 +      //Total Master time includes one coreloop time -- just assume the core
   1.490 +      // loop time is same for Master as for AppVPs, even though it may be
   1.491 +      // smaller due to higher predictability of the fixed jmp.
   1.492 +   saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
   1.493 +   #endif
   1.494 +   //========================================================================
   1.495 +
   1.496 +   masterEnv        = (MasterEnv*)_VMSMasterEnv;
   1.497 +   
   1.498 +      //GCC may optimize so doesn't always re-define from frame-storage
   1.499 +   masterPr         = (VirtProcr*)volatileMasterPr;  //just to make sure after jmp
   1.500 +   thisCoresIdx     = masterPr->coreAnimatedBy;
   1.501 +   readyToAnimateQ  = masterEnv->readyToAnimateQs[thisCoresIdx];
   1.502 +   schedSlots       = masterEnv->allSchedSlots[thisCoresIdx];
   1.503 +
   1.504 +   requestHandler   = masterEnv->requestHandler;
   1.505 +   slaveScheduler   = masterEnv->slaveScheduler;
   1.506 +   semanticEnv      = masterEnv->semanticEnv;
   1.507 +
   1.508 +
   1.509 +      //Poll each slot's Done flag
   1.510 +   numSlotsFilled = 0;
   1.511 +   for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
   1.512 +    {
   1.513 +      currSlot = schedSlots[ slotIdx ];
   1.514 +
   1.515 +      if( currSlot->workIsDone )
   1.516 +       {
   1.517 +         currSlot->workIsDone         = FALSE;
   1.518 +         currSlot->needsProcrAssigned = TRUE;
   1.519 +
   1.520 +            //process requests from slave to master
   1.521 +               //====================== MEASUREMENT STUFF ===================
   1.522 +               #ifdef MEAS__TIME_PLUGIN
   1.523 +               int32 startStamp1, endStamp1;
   1.524 +               saveLowTimeStampCountInto( startStamp1 );
   1.525 +               #endif
   1.526 +               //============================================================
   1.527 +         (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
   1.528 +               //====================== MEASUREMENT STUFF ===================
   1.529 +               #ifdef MEAS__TIME_PLUGIN
   1.530 +               saveLowTimeStampCountInto( endStamp1 );
   1.531 +               addIntervalToHist( startStamp1, endStamp1,
   1.532 +                                        _VMSMasterEnv->reqHdlrLowTimeHist );
   1.533 +               addIntervalToHist( startStamp1, endStamp1,
   1.534 +                                        _VMSMasterEnv->reqHdlrHighTimeHist );
   1.535 +               #endif
   1.536 +               //============================================================
   1.537 +       }
   1.538 +      if( currSlot->needsProcrAssigned )
   1.539 +       {    //give slot a new virt procr
   1.540 +         schedVirtPr =
   1.541 +          (*slaveScheduler)( semanticEnv, thisCoresIdx );
   1.542 +         
   1.543 +         if( schedVirtPr != NULL )
   1.544 +          { currSlot->procrAssignedToSlot = schedVirtPr;
   1.545 +            schedVirtPr->schedSlot        = currSlot;
   1.546 +            currSlot->needsProcrAssigned  = FALSE;
   1.547 +            numSlotsFilled               += 1;
   1.548 +            
   1.549 +            writeVMSQ( schedVirtPr, readyToAnimateQ );
   1.550 +          }
   1.551 +       }
   1.552 +    }
   1.553 +
   1.554 +   
   1.555 +   #ifdef USE_WORK_STEALING
   1.556 +      //If no slots filled, means no more work, look for work to steal.
   1.557 +   if( numSlotsFilled == 0 )
   1.558 +    { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
   1.559 +    }
   1.560 +   #endif
   1.561 +
   1.562 +   
   1.563 +   #ifdef MEAS__TIME_MASTER
   1.564 +   saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
   1.565 +   #endif
   1.566 +
   1.567 +   masterSwitchToCoreLoop(animatingPr);
   1.568 +   flushRegisters();
   1.569 +   }//MasterLoop
   1.570 +
   1.571 +
   1.572 + }
   1.573 +
   1.574 +
   1.575 +
   1.576 +/*This has a race condition -- the coreloops are accessing their own queues
   1.577 + * at the same time that this work-stealer on a different core is trying to
   1.578 + */
   1.579 +void inline
   1.580 +stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
   1.581 +               VirtProcr *masterPr )
   1.582 + { 
   1.583 +   VirtProcr   *stolenPr;
   1.584 +   int32        coreIdx, i;
   1.585 +   VMSQueueStruc *currQ;
   1.586 +
   1.587 +   stolenPr = NULL;
   1.588 +   coreIdx = masterPr->coreAnimatedBy;
   1.589 +   for( i = 0; i < NUM_CORES -1; i++ )
   1.590 +    {
   1.591 +      if( coreIdx >= NUM_CORES -1 )
   1.592 +       { coreIdx = 0;
   1.593 +       }
   1.594 +      else
   1.595 +       { coreIdx++;
   1.596 +       }
   1.597 +      currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
   1.598 +      if( numInVMSQ( currQ ) > 0 )
   1.599 +       { stolenPr = readVMSQ (currQ );
   1.600 +         break;
   1.601 +       }
   1.602 +    }
   1.603 +
   1.604 +   if( stolenPr != NULL )
   1.605 +    { currSlot->procrAssignedToSlot = stolenPr;
   1.606 +      stolenPr->schedSlot           = currSlot;
   1.607 +      currSlot->needsProcrAssigned  = FALSE;
   1.608 +
   1.609 +      writeVMSQ( stolenPr, readyToAnimateQ );
   1.610 +    }
   1.611 + }
   1.612 +
   1.613 +/*This algorithm makes the common case fast.  Make the coreloop passive,
   1.614 + * and show its progress.  Make the stealer control a gate that coreloop
   1.615 + * has to pass.
   1.616 + *To avoid interference, only one stealer at a time.  Use a global
   1.617 + * stealer-lock.
   1.618 + *
   1.619 + *The pattern is based on a gate -- stealer shuts the gate, then monitors
   1.620 + * to be sure any already past make it all the way out, before starting.
   1.621 + *So, have a "progress" measure just before the gate, then have two after it,
   1.622 + * one is in a "waiting room" outside the gate, the other is at the exit.
   1.623 + *Then, the stealer first shuts the gate, then checks the progress measure
   1.624 + * outside it, then looks to see if the progress measure at the exit is the
   1.625 + * same.  If yes, it knows the protected area is empty 'cause no other way
   1.626 + * to get in and the last to get in also exited.
   1.627 + *If the progress measure at the exit is not the same, then the stealer goes
   1.628 + * into a loop checking both the waiting-area and the exit progress-measures
   1.629 + * until one of them shows the same as the measure outside the gate.  Might
   1.630 + * as well re-read the measure outside the gate each go around, just to be
   1.631 + * sure.  It is guaranteed that one of the two will eventually match the one
   1.632 + * outside the gate.
   1.633 + *
   1.634 + *Here's an informal proof of correctness:
   1.635 + *The gate can be closed at any point, and have only four cases:
   1.636 + *  1) coreloop made it past the gate-closing but not yet past the exit
   1.637 + *  2) coreloop made it past the pre-gate progress update but not yet past
   1.638 + *     the gate,
   1.639 + *  3) coreloop is right before the pre-gate update
   1.640 + *  4) coreloop is past the exit and far from the pre-gate update.
   1.641 + *
   1.642 + * Covering the cases in reverse order,
   1.643 + *  4) is not a problem -- stealer will read pre-gate progress, see that it
   1.644 + *     matches exit progress, and the gate is closed, so stealer can proceed.
   1.645 + *  3) stealer will read pre-gate progress just after coreloop updates it..
   1.646 + *     so stealer goes into a loop until the coreloop causes wait-progress
   1.647 + *     to match pre-gate progress, so then stealer can proceed
   1.648 + *  2) same as 3..
   1.649 + *  1) stealer reads pre-gate progress, sees that it's different than exit,
   1.650 + *     so goes into loop until exit matches pre-gate, now it knows coreloop
   1.651 + *     is not in protected and cannot get back in, so can proceed.
   1.652 + *
   1.653 + *Implementation for the stealer:
   1.654 + *
   1.655 + *First, acquire the stealer lock -- only cores with no work to do will
   1.656 + * compete to steal, so not a big performance penalty having only one --
   1.657 + * will rarely have multiple stealers in a system with plenty of work -- and
   1.658 + * in a system with little work, it doesn't matter.
   1.659 + *
   1.660 + *Note, have single-reader, single-writer pattern for all variables used to
   1.661 + * communicate between stealer and victims
   1.662 + *
   1.663 + *So, scan the queues of the core loops, until find non-empty.  Each core
   1.664 + * has its own list that it scans.  The list goes in order from closest to
   1.665 + * furthest core, so it steals first from close cores.  Later can add
   1.666 + * taking info from the app about overlapping footprints, and scan all the
   1.667 + * others then choose work with the most footprint overlap with the contents
   1.668 + * of this core's cache.
   1.669 + *
   1.670 + *Now, have a victim want to take work from.  So, shut the gate in that
   1.671 + * coreloop, by setting the "gate closed" var on its stack to TRUE.
   1.672 + *Then, read the core's pre-gate progress and compare to the core's exit
   1.673 + * progress.
   1.674 + *If same, can proceed to take work from the coreloop's queue.  When done,
   1.675 + * write FALSE to gate closed var.
   1.676 + *If different, then enter a loop that reads the pre-gate progress, then
   1.677 + * compares to exit progress then to wait progress.  When one of two
   1.678 + * matches, proceed.  Take work from the coreloop's queue.  When done,
   1.679 + * write FALSE to the gate closed var.
   1.680 + * 
   1.681 + */
   1.682 +void inline
   1.683 +gateProtected_stealWorkInto( SchedSlot *currSlot,
   1.684 +                             VMSQueueStruc *myReadyToAnimateQ,
   1.685 +                             VirtProcr *masterPr )
   1.686 + {
   1.687 +   VirtProcr     *stolenPr;
   1.688 +   int32          coreIdx, i, haveAVictim, gotLock;
   1.689 +   VMSQueueStruc *victimsQ;
   1.690 +
   1.691 +   volatile GateStruc *vicGate;
   1.692 +   int32               coreMightBeInProtected;
   1.693 +
   1.694 +
   1.695 +
   1.696 +      //see if any other cores have work available to steal
   1.697 +   haveAVictim = FALSE;
   1.698 +   coreIdx = masterPr->coreAnimatedBy;
   1.699 +   for( i = 0; i < NUM_CORES -1; i++ )
   1.700 +    {
   1.701 +      if( coreIdx >= NUM_CORES -1 )
   1.702 +       { coreIdx = 0;
   1.703 +       }
   1.704 +      else
   1.705 +       { coreIdx++;
   1.706 +       }
   1.707 +      victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
   1.708 +      if( numInVMSQ( victimsQ ) > 0 )
   1.709 +       { haveAVictim = TRUE;
   1.710 +         vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
   1.711 +         break;
   1.712 +       }
   1.713 +    }
   1.714 +   if( !haveAVictim ) return;  //no work to steal, exit
   1.715 +
   1.716 +      //have a victim core, now get the stealer-lock
   1.717 +   gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
   1.718 +                                                          UNLOCKED, LOCKED );
   1.719 +   if( !gotLock ) return; //go back to core loop, which will re-start master
   1.720 +
   1.721 +
   1.722 +   //====== Start Gate-protection =======
   1.723 +   vicGate->gateClosed = TRUE;
   1.724 +   coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
   1.725 +   while( coreMightBeInProtected )
   1.726 +    {    //wait until sure
   1.727 +      if( vicGate->preGateProgress == vicGate->waitProgress )
   1.728 +         coreMightBeInProtected = FALSE;
   1.729 +      if( vicGate->preGateProgress == vicGate->exitProgress )
   1.730 +         coreMightBeInProtected = FALSE;
   1.731 +    }
   1.732 +
   1.733 +   stolenPr = readVMSQ ( victimsQ );
   1.734 +
   1.735 +   vicGate->gateClosed = FALSE;
   1.736 +   //======= End Gate-protection  =======
   1.737 +
   1.738 +
   1.739 +   if( stolenPr != NULL )  //victim could have been in protected and taken
   1.740 +    { currSlot->procrAssignedToSlot = stolenPr;
   1.741 +      stolenPr->schedSlot           = currSlot;
   1.742 +      currSlot->needsProcrAssigned  = FALSE;
   1.743 +
   1.744 +      writeVMSQ( stolenPr, myReadyToAnimateQ );
   1.745 +    }
   1.746 +
   1.747 +      //unlock the work stealing lock
   1.748 +   _VMSMasterEnv->workStealingLock = UNLOCKED;
   1.749 + }