Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl

annotate MasterLoop.c @ 187:fe5ad5726e36

find changesets by author, revision, files, or words in the commit message
counters working ...sort of
author Nina Engelhardt <nengel@mailbox.tu-berlin.de>
date Fri, 03 Feb 2012 17:32:48 +0100
parents 69eb54ce9c4b
children 20358f56e498
rev   line source
msach@94 1 /*
msach@94 2 * Copyright 2010 OpenSourceStewardshipFoundation
msach@94 3 *
msach@94 4 * Licensed under BSD
msach@94 5 */
msach@94 6
msach@94 7
msach@94 8
msach@94 9 #include <stdio.h>
msach@94 10 #include <stddef.h>
msach@94 11
msach@94 12 #include "VMS.h"
msach@94 13 #include "ProcrContext.h"
msach@94 14
msach@94 15
msach@94 16 //===========================================================================
msach@94 17 void inline
msach@94 18 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
msach@94 19 VirtProcr *masterPr );
msach@94 20
msach@94 21 //===========================================================================
msach@94 22
msach@94 23
msach@94 24
msach@94 25 /*This code is animated by the virtual Master processor.
msach@94 26 *
msach@94 27 *Polls each sched slot exactly once, hands any requests made by a newly
msach@94 28 * done slave to the "request handler" plug-in function
msach@94 29 *
msach@94 30 *Any slots that need a virt procr assigned are given to the "schedule"
msach@94 31 * plug-in function, which tries to assign a virt procr (slave) to it.
msach@94 32 *
msach@94 33 *When all slots needing a processor have been given to the schedule plug-in,
msach@94 34 * a fraction of the procrs successfully scheduled are put into the
msach@94 35 * work queue, then a continuation of this function is put in, then the rest
msach@94 36 * of the virt procrs that were successfully scheduled.
msach@94 37 *
msach@94 38 *The first thing the continuation does is busy-wait until the previous
msach@94 39 * animation completes. This is because an (unlikely) continuation may
msach@94 40 * sneak through queue before previous continuation is done putting second
msach@94 41 * part of scheduled slaves in, which is the only race condition.
msach@94 42 *
msach@94 43 */
msach@94 44
msach@94 45 /*May 29, 2010 -- birth a Master during init so that first core loop to
msach@94 46 * start running gets it and does all the stuff for a newly born --
msach@94 47 * from then on, will be doing continuation, but do suspension self
msach@94 48 * directly at end of master loop
msach@94 49 *So VMS__init just births the master virtual processor same way it births
msach@94 50 * all the others -- then does any extra setup needed and puts it into the
msach@94 51 * work queue.
msach@94 52 *However means have to make masterEnv a global static volatile the same way
msach@94 53 * did with readyToAnimateQ in core loop. -- for performance, put the
msach@94 54 * jump to the core loop directly in here, and have it directly jump back.
msach@94 55 *
msach@94 56 *
msach@94 57 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
msach@94 58 * avoids the suspected bug in the system stack that causes bizarre faults
msach@94 59 * at random places in the system code.
msach@94 60 *
msach@94 61 *So, this function is coupled to each of the MasterVPs, -- meaning this
msach@94 62 * function can't rely on a particular stack and frame -- each MasterVP that
msach@94 63 * animates this function has a different one.
msach@94 64 *
msach@94 65 *At this point, the masterLoop does not write itself into the queue anymore,
msach@94 66 * instead, the coreLoop acquires the masterLock when it has nothing to
msach@94 67 * animate, and then animates its own masterLoop. However, still try to put
msach@94 68 * several AppVPs into the queue to amortize the startup cost of switching
msach@94 69 * to the MasterVP. Note, don't have to worry about latency of requests much
msach@94 70 * because most requests generate work for same core -- only latency issue
msach@94 71 * is case when other cores starved and one core's requests generate work
msach@94 72 * for them -- so keep max in queue to 3 or 4..
msach@94 73 */
msach@94 74 void masterLoop( void *initData, VirtProcr *animatingPr )
msach@94 75 {
msach@94 76 int32 slotIdx, numSlotsFilled;
msach@94 77 VirtProcr *schedVirtPr;
msach@94 78 SchedSlot *currSlot, **schedSlots;
msach@94 79 MasterEnv *masterEnv;
msach@94 80 VMSQueueStruc *readyToAnimateQ;
msach@94 81
msach@94 82 SlaveScheduler slaveScheduler;
msach@94 83 RequestHandler requestHandler;
msach@94 84 void *semanticEnv;
msach@94 85
msach@94 86 int32 thisCoresIdx;
msach@94 87 VirtProcr *masterPr;
msach@94 88 volatile VirtProcr *volatileMasterPr;
msach@94 89
msach@94 90 volatileMasterPr = animatingPr;
msach@94 91 masterPr = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
msach@94 92
msach@94 93 //First animation of each MasterVP will in turn animate this part
msach@94 94 // of setup code.. (VP creator sets up the stack as if this function
msach@94 95 // was called normally, but actually get here by jmp)
msach@94 96 //So, setup values about stack ptr, jmp pt and all that
msach@94 97 //masterPr->nextInstrPt = &&masterLoopStartPt;
msach@94 98
msach@94 99
msach@94 100 //Note, got rid of writing the stack and frame ptr up here, because
msach@94 101 // only one
msach@94 102 // core can ever animate a given MasterVP, so don't need to communicate
msach@94 103 // new frame and stack ptr to the MasterVP storage before a second
msach@94 104 // version of that MasterVP can get animated on a different core.
msach@94 105 //Also got rid of the busy-wait.
msach@94 106
msach@94 107
msach@94 108 //masterLoopStartPt:
msach@94 109 while(1){
msach@94 110
msach@94 111 //============================= MEASUREMENT STUFF ========================
msach@94 112 #ifdef MEAS__TIME_MASTER
msach@94 113 //Total Master time includes one coreloop time -- just assume the core
msach@94 114 // loop time is same for Master as for AppVPs, even though it may be
msach@94 115 // smaller due to higher predictability of the fixed jmp.
msach@94 116 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
msach@94 117 #endif
nengel@186 118
msach@94 119 //========================================================================
msach@94 120
msach@94 121 masterEnv = (MasterEnv*)_VMSMasterEnv;
msach@94 122
msach@94 123 //GCC may optimize so doesn't always re-define from frame-storage
msach@94 124 masterPr = (VirtProcr*)volatileMasterPr; //just to make sure after jmp
msach@94 125 thisCoresIdx = masterPr->coreAnimatedBy;
msach@94 126 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
msach@94 127 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
msach@94 128
msach@94 129 requestHandler = masterEnv->requestHandler;
msach@94 130 slaveScheduler = masterEnv->slaveScheduler;
msach@94 131 semanticEnv = masterEnv->semanticEnv;
nengel@184 132
nengel@186 133 #ifdef MEAS__PERF_COUNTERS
nengel@186 134 CounterHandler counterHandler = masterEnv->counterHandler;
nengel@186 135 #endif
msach@94 136
msach@94 137 //Poll each slot's Done flag
msach@94 138 numSlotsFilled = 0;
msach@94 139 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
msach@94 140 {
msach@94 141 currSlot = schedSlots[ slotIdx ];
msach@94 142
msach@94 143 if( currSlot->workIsDone )
msach@94 144 {
msach@94 145 currSlot->workIsDone = FALSE;
msach@94 146 currSlot->needsProcrAssigned = TRUE;
msach@94 147
msach@94 148 //process requests from slave to master
msach@94 149 //====================== MEASUREMENT STUFF ===================
msach@94 150 #ifdef MEAS__TIME_PLUGIN
msach@94 151 int32 startStamp1, endStamp1;
msach@94 152 saveLowTimeStampCountInto( startStamp1 );
msach@94 153 #endif
Nina@109 154 #ifdef MEAS__PERF_COUNTERS
nengel@187 155 VirtProcr p_sav = *(currSlot->procrAssignedToSlot);
nengel@185 156 uint64 cycles, instrs;
nengel@184 157 saveCyclesAndInstrs(thisCoresIdx,cycles, instrs);
nengel@187 158 (*counterHandler)(MasterLoop_beforeReqHdlr,&p_sav,cycles,instrs);
engelhardt@108 159 #endif
msach@94 160 //============================================================
msach@94 161 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
msach@94 162 //====================== MEASUREMENT STUFF ===================
msach@94 163 #ifdef MEAS__TIME_PLUGIN
msach@94 164 saveLowTimeStampCountInto( endStamp1 );
msach@94 165 addIntervalToHist( startStamp1, endStamp1,
msach@94 166 _VMSMasterEnv->reqHdlrLowTimeHist );
msach@94 167 addIntervalToHist( startStamp1, endStamp1,
msach@94 168 _VMSMasterEnv->reqHdlrHighTimeHist );
msach@94 169 #endif
Nina@109 170 #ifdef MEAS__PERF_COUNTERS
Nina@109 171 //done with constraints check
nengel@185 172 uint64 cycles2,instrs2;
nengel@185 173 saveCyclesAndInstrs(thisCoresIdx,cycles2, instrs2);
nengel@187 174 (*counterHandler)(MasterLoop_afterReqHdlr,&p_sav,cycles2,instrs2);
engelhardt@108 175 #endif
msach@94 176 //============================================================
msach@94 177 }
msach@94 178 if( currSlot->needsProcrAssigned )
msach@94 179 { //give slot a new virt procr
Nina@109 180 #ifdef MEAS__PERF_COUNTERS
engelhardt@108 181 //start assigner
engelhardt@108 182 uint64 tmp_cycles;
engelhardt@108 183 uint64 tmp_instrs;
Nina@109 184 saveCyclesAndInstrs(thisCoresIdx,tmp_cycles,tmp_instrs);
nengel@186 185 //FIXME WTF AM I DOING WHY DOES THIS EVEN WORK
nengel@186 186 //(*counterHandler)(MasterLoop_beforeNextAssign,schedVirtPr,tmp_cycles,tmp_instrs);
engelhardt@108 187 #endif
msach@94 188 schedVirtPr =
nengel@177 189 (*slaveScheduler)( semanticEnv, thisCoresIdx, slotIdx );
Nina@109 190
msach@94 191 if( schedVirtPr != NULL )
msach@94 192 { currSlot->procrAssignedToSlot = schedVirtPr;
msach@94 193 schedVirtPr->schedSlot = currSlot;
msach@94 194 currSlot->needsProcrAssigned = FALSE;
msach@94 195 numSlotsFilled += 1;
nengel@177 196
nengel@185 197 writeVMSQ( schedVirtPr, readyToAnimateQ );
nengel@185 198
Nina@109 199 #ifdef MEAS__PERF_COUNTERS
nengel@184 200 uint64 cycles;
nengel@184 201 uint64 instrs;
nengel@184 202 saveCyclesAndInstrs(thisCoresIdx,cycles,instrs);
nengel@185 203 (*counterHandler)(MasterLoop_beforeAssign,schedVirtPr,tmp_cycles,tmp_instrs);
nengel@185 204 (*counterHandler)(MasterLoop_afterAssign,schedVirtPr,cycles,instrs);
engelhardt@108 205 #endif
msach@94 206 }
msach@94 207 }
nengel@177 208
msach@94 209 }
msach@94 210
msach@94 211
msach@94 212 #ifdef USE_WORK_STEALING
msach@94 213 //If no slots filled, means no more work, look for work to steal.
msach@94 214 if( numSlotsFilled == 0 )
msach@94 215 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
msach@94 216 }
msach@94 217 #endif
msach@94 218
msach@94 219
msach@94 220 #ifdef MEAS__TIME_MASTER
msach@94 221 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
msach@94 222 #endif
msach@94 223
msach@94 224 masterSwitchToCoreLoop(animatingPr);
msach@94 225 flushRegisters();
msach@94 226 }//MasterLoop
msach@94 227
msach@94 228
msach@94 229 }
msach@94 230
msach@94 231
msach@94 232
msach@94 233 /*This has a race condition -- the coreloops are accessing their own queues
msach@94 234 * at the same time that this work-stealer on a different core is trying to
msach@94 235 */
msach@94 236 void inline
msach@94 237 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
msach@94 238 VirtProcr *masterPr )
msach@94 239 {
msach@94 240 VirtProcr *stolenPr;
msach@94 241 int32 coreIdx, i;
msach@94 242 VMSQueueStruc *currQ;
msach@94 243
msach@94 244 stolenPr = NULL;
msach@94 245 coreIdx = masterPr->coreAnimatedBy;
msach@94 246 for( i = 0; i < NUM_CORES -1; i++ )
msach@94 247 {
msach@94 248 if( coreIdx >= NUM_CORES -1 )
msach@94 249 { coreIdx = 0;
msach@94 250 }
msach@94 251 else
msach@94 252 { coreIdx++;
msach@94 253 }
msach@94 254 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
msach@94 255 if( numInVMSQ( currQ ) > 0 )
msach@94 256 { stolenPr = readVMSQ (currQ );
msach@94 257 break;
msach@94 258 }
msach@94 259 }
msach@94 260
msach@94 261 if( stolenPr != NULL )
msach@94 262 { currSlot->procrAssignedToSlot = stolenPr;
msach@94 263 stolenPr->schedSlot = currSlot;
msach@94 264 currSlot->needsProcrAssigned = FALSE;
msach@94 265
msach@94 266 writeVMSQ( stolenPr, readyToAnimateQ );
msach@94 267 }
msach@94 268 }
msach@94 269
msach@94 270 /*This algorithm makes the common case fast. Make the coreloop passive,
msach@94 271 * and show its progress. Make the stealer control a gate that coreloop
msach@94 272 * has to pass.
msach@94 273 *To avoid interference, only one stealer at a time. Use a global
msach@94 274 * stealer-lock.
msach@94 275 *
msach@94 276 *The pattern is based on a gate -- stealer shuts the gate, then monitors
msach@94 277 * to be sure any already past make it all the way out, before starting.
msach@94 278 *So, have a "progress" measure just before the gate, then have two after it,
msach@94 279 * one is in a "waiting room" outside the gate, the other is at the exit.
msach@94 280 *Then, the stealer first shuts the gate, then checks the progress measure
msach@94 281 * outside it, then looks to see if the progress measure at the exit is the
msach@94 282 * same. If yes, it knows the protected area is empty 'cause no other way
msach@94 283 * to get in and the last to get in also exited.
msach@94 284 *If the progress measure at the exit is not the same, then the stealer goes
msach@94 285 * into a loop checking both the waiting-area and the exit progress-measures
msach@94 286 * until one of them shows the same as the measure outside the gate. Might
msach@94 287 * as well re-read the measure outside the gate each go around, just to be
msach@94 288 * sure. It is guaranteed that one of the two will eventually match the one
msach@94 289 * outside the gate.
msach@94 290 *
msach@94 291 *Here's an informal proof of correctness:
msach@94 292 *The gate can be closed at any point, and have only four cases:
msach@94 293 * 1) coreloop made it past the gate-closing but not yet past the exit
msach@94 294 * 2) coreloop made it past the pre-gate progress update but not yet past
msach@94 295 * the gate,
msach@94 296 * 3) coreloop is right before the pre-gate update
msach@94 297 * 4) coreloop is past the exit and far from the pre-gate update.
msach@94 298 *
msach@94 299 * Covering the cases in reverse order,
msach@94 300 * 4) is not a problem -- stealer will read pre-gate progress, see that it
msach@94 301 * matches exit progress, and the gate is closed, so stealer can proceed.
msach@94 302 * 3) stealer will read pre-gate progress just after coreloop updates it..
msach@94 303 * so stealer goes into a loop until the coreloop causes wait-progress
msach@94 304 * to match pre-gate progress, so then stealer can proceed
msach@94 305 * 2) same as 3..
msach@94 306 * 1) stealer reads pre-gate progress, sees that it's different than exit,
msach@94 307 * so goes into loop until exit matches pre-gate, now it knows coreloop
msach@94 308 * is not in protected and cannot get back in, so can proceed.
msach@94 309 *
msach@94 310 *Implementation for the stealer:
msach@94 311 *
msach@94 312 *First, acquire the stealer lock -- only cores with no work to do will
msach@94 313 * compete to steal, so not a big performance penalty having only one --
msach@94 314 * will rarely have multiple stealers in a system with plenty of work -- and
msach@94 315 * in a system with little work, it doesn't matter.
msach@94 316 *
msach@94 317 *Note, have single-reader, single-writer pattern for all variables used to
msach@94 318 * communicate between stealer and victims
msach@94 319 *
msach@94 320 *So, scan the queues of the core loops, until find non-empty. Each core
msach@94 321 * has its own list that it scans. The list goes in order from closest to
msach@94 322 * furthest core, so it steals first from close cores. Later can add
msach@94 323 * taking info from the app about overlapping footprints, and scan all the
msach@94 324 * others then choose work with the most footprint overlap with the contents
msach@94 325 * of this core's cache.
msach@94 326 *
msach@94 327 *Now, have a victim want to take work from. So, shut the gate in that
msach@94 328 * coreloop, by setting the "gate closed" var on its stack to TRUE.
msach@94 329 *Then, read the core's pre-gate progress and compare to the core's exit
msach@94 330 * progress.
msach@94 331 *If same, can proceed to take work from the coreloop's queue. When done,
msach@94 332 * write FALSE to gate closed var.
msach@94 333 *If different, then enter a loop that reads the pre-gate progress, then
msach@94 334 * compares to exit progress then to wait progress. When one of two
msach@94 335 * matches, proceed. Take work from the coreloop's queue. When done,
msach@94 336 * write FALSE to the gate closed var.
msach@94 337 *
msach@94 338 */
msach@94 339 void inline
msach@94 340 gateProtected_stealWorkInto( SchedSlot *currSlot,
msach@94 341 VMSQueueStruc *myReadyToAnimateQ,
msach@94 342 VirtProcr *masterPr )
msach@94 343 {
msach@94 344 VirtProcr *stolenPr;
msach@94 345 int32 coreIdx, i, haveAVictim, gotLock;
msach@94 346 VMSQueueStruc *victimsQ;
msach@94 347
msach@94 348 volatile GateStruc *vicGate;
msach@94 349 int32 coreMightBeInProtected;
msach@94 350
msach@94 351
msach@94 352
msach@94 353 //see if any other cores have work available to steal
msach@94 354 haveAVictim = FALSE;
msach@94 355 coreIdx = masterPr->coreAnimatedBy;
msach@94 356 for( i = 0; i < NUM_CORES -1; i++ )
msach@94 357 {
msach@94 358 if( coreIdx >= NUM_CORES -1 )
msach@94 359 { coreIdx = 0;
msach@94 360 }
msach@94 361 else
msach@94 362 { coreIdx++;
msach@94 363 }
msach@94 364 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
msach@94 365 if( numInVMSQ( victimsQ ) > 0 )
msach@94 366 { haveAVictim = TRUE;
msach@94 367 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
msach@94 368 break;
msach@94 369 }
msach@94 370 }
msach@94 371 if( !haveAVictim ) return; //no work to steal, exit
msach@94 372
msach@94 373 //have a victim core, now get the stealer-lock
msach@94 374 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
msach@94 375 UNLOCKED, LOCKED );
msach@94 376 if( !gotLock ) return; //go back to core loop, which will re-start master
msach@94 377
msach@94 378
msach@94 379 //====== Start Gate-protection =======
msach@94 380 vicGate->gateClosed = TRUE;
msach@94 381 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
msach@94 382 while( coreMightBeInProtected )
msach@94 383 { //wait until sure
msach@94 384 if( vicGate->preGateProgress == vicGate->waitProgress )
msach@94 385 coreMightBeInProtected = FALSE;
msach@94 386 if( vicGate->preGateProgress == vicGate->exitProgress )
msach@94 387 coreMightBeInProtected = FALSE;
msach@94 388 }
msach@94 389
msach@94 390 stolenPr = readVMSQ ( victimsQ );
msach@94 391
msach@94 392 vicGate->gateClosed = FALSE;
msach@94 393 //======= End Gate-protection =======
msach@94 394
msach@94 395
msach@94 396 if( stolenPr != NULL ) //victim could have been in protected and taken
msach@94 397 { currSlot->procrAssignedToSlot = stolenPr;
msach@94 398 stolenPr->schedSlot = currSlot;
msach@94 399 currSlot->needsProcrAssigned = FALSE;
msach@94 400
msach@94 401 writeVMSQ( stolenPr, myReadyToAnimateQ );
msach@94 402 }
msach@94 403
msach@94 404 //unlock the work stealing lock
msach@94 405 _VMSMasterEnv->workStealingLock = UNLOCKED;
msach@94 406 }