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

annotate MasterLoop.c @ 137:99343ffe1918

find changesets by author, revision, files, or words in the commit message
The shutdown now uses inter master requests
author Merten Sach <msach@mailbox.tu-berlin.de>
date Mon, 19 Sep 2011 14:15:37 +0200
parents 0b49fd35afc1
children 2c8f3cf6c058
rev   line source
Me@0 1 /*
Me@38 2 * Copyright 2010 OpenSourceStewardshipFoundation
Me@43 3 *
Me@0 4 * Licensed under BSD
Me@0 5 */
Me@0 6
Me@0 7
Me@0 8
Me@0 9 #include <stdio.h>
Me@9 10 #include <stddef.h>
Me@0 11
Me@0 12 #include "VMS.h"
msach@77 13 #include "ProcrContext.h"
msach@134 14 #include "scheduling.h"
msach@135 15 #include "inter_VMS_requests.h"
msach@135 16 #include "inter_VMS_requests_handler.h"
Me@0 17
Me@55 18 //===========================================================================
Me@55 19 void inline
Me@55 20 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
msach@127 21 VirtProcr *masterPr);
msach@127 22
msach@127 23 void inline
msach@127 24 handleInterMasterReq( InterMasterReqst *currReq, void *_semEnv,
msach@127 25 VirtProcr *masterPr);
msach@127 26
msach@127 27 void inline
msach@127 28 handleInterVMSCoreReq( InterVMSCoreReqst *currReq, VirtProcr *masterPr);
Me@55 29
Me@55 30 //===========================================================================
Me@55 31
Me@55 32
msach@69 33
Me@0 34 /*This code is animated by the virtual Master processor.
Me@0 35 *
Me@11 36 *Polls each sched slot exactly once, hands any requests made by a newly
Me@11 37 * done slave to the "request handler" plug-in function
Me@0 38 *
Me@11 39 *Any slots that need a virt procr assigned are given to the "schedule"
Me@11 40 * plug-in function, which tries to assign a virt procr (slave) to it.
Me@0 41 *
Me@11 42 *When all slots needing a processor have been given to the schedule plug-in,
Me@11 43 * a fraction of the procrs successfully scheduled are put into the
Me@11 44 * work queue, then a continuation of this function is put in, then the rest
Me@11 45 * of the virt procrs that were successfully scheduled.
Me@0 46 *
Me@11 47 *The first thing the continuation does is busy-wait until the previous
Me@11 48 * animation completes. This is because an (unlikely) continuation may
Me@11 49 * sneak through queue before previous continuation is done putting second
Me@11 50 * part of scheduled slaves in, which is the only race condition.
Me@0 51 *
Me@0 52 */
Me@0 53
Me@4 54 /*May 29, 2010 -- birth a Master during init so that first core loop to
Me@11 55 * start running gets it and does all the stuff for a newly born --
Me@11 56 * from then on, will be doing continuation, but do suspension self
Me@4 57 * directly at end of master loop
Me@4 58 *So VMS__init just births the master virtual processor same way it births
Me@4 59 * all the others -- then does any extra setup needed and puts it into the
Me@4 60 * work queue.
Me@120 61 *However means have to make masterEnv a global static volatile.
Me@31 62 *
Me@31 63 *
Me@31 64 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
Me@31 65 * avoids the suspected bug in the system stack that causes bizarre faults
Me@31 66 * at random places in the system code.
Me@31 67 *
Me@31 68 *So, this function is coupled to each of the MasterVPs, -- meaning this
Me@31 69 * function can't rely on a particular stack and frame -- each MasterVP that
Me@120 70 * animates this function has a different stack.
Me@31 71 *
Me@31 72 *At this point, the masterLoop does not write itself into the queue anymore,
Me@31 73 * instead, the coreLoop acquires the masterLock when it has nothing to
Me@31 74 * animate, and then animates its own masterLoop. However, still try to put
Me@31 75 * several AppVPs into the queue to amortize the startup cost of switching
Me@31 76 * to the MasterVP. Note, don't have to worry about latency of requests much
Me@31 77 * because most requests generate work for same core -- only latency issue
Me@31 78 * is case when other cores starved and one core's requests generate work
Me@31 79 * for them -- so keep max in queue to 3 or 4..
Me@4 80 */
Me@31 81 void masterLoop( void *initData, VirtProcr *animatingPr )
Me@21 82 {
Me@55 83 int32 slotIdx, numSlotsFilled;
Me@21 84 VirtProcr *schedVirtPr;
Me@31 85 SchedSlot *currSlot, **schedSlots;
Me@0 86 MasterEnv *masterEnv;
Me@31 87 VMSQueueStruc *readyToAnimateQ;
Me@4 88
Me@0 89 SlaveScheduler slaveScheduler;
Me@0 90 RequestHandler requestHandler;
Me@31 91 void *semanticEnv;
Me@0 92
Me@55 93 int32 thisCoresIdx;
Me@31 94 VirtProcr *masterPr;
msach@69 95 volatile VirtProcr *volatileMasterPr;
msach@69 96
msach@69 97 volatileMasterPr = animatingPr;
msach@69 98 masterPr = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
msach@132 99 masterEnv = (MasterEnv*)_VMSMasterEnv;
Me@31 100
Me@31 101 //First animation of each MasterVP will in turn animate this part
Me@31 102 // of setup code.. (VP creator sets up the stack as if this function
Me@31 103 // was called normally, but actually get here by jmp)
Me@0 104
Me@120 105 //Sept 2011
Me@120 106 //Old code jumped directly to this point, but doesn't work on x64
Me@120 107 // So, just make this an endless loop, and do assembly function at end
Me@120 108 // that saves its own return addr, then jumps to core_loop.
Me@120 109 while(1)
msach@132 110 {
Me@38 111 //============================= MEASUREMENT STUFF ========================
Me@38 112 #ifdef MEAS__TIME_MASTER
Me@38 113 //Total Master time includes one coreloop time -- just assume the core
Me@120 114 // loop time is same for Master as is for AppVPs, even though it may be
Me@68 115 // smaller due to higher predictability of the fixed jmp.
Me@38 116 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
Me@38 117 #endif
Me@38 118 //========================================================================
Me@0 119
msach@132 120 //GCC may optimize so doesn't always re-define from frame-storage
msach@69 121 thisCoresIdx = masterPr->coreAnimatedBy;
msach@132 122 masterEnv->currentMasterProcrID = thisCoresIdx;
msach@69 123 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
msach@69 124 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
msach@69 125
msach@69 126 requestHandler = masterEnv->requestHandler;
msach@69 127 slaveScheduler = masterEnv->slaveScheduler;
msach@69 128 semanticEnv = masterEnv->semanticEnv;
msach@69 129
Me@119 130 //First, check for requests from other MasterVPs, and handle them
msach@135 131 InterMasterReqst* currReqst = masterEnv->interMasterRequestsFor[thisCoresIdx];
msach@135 132 while(currReqst)
msach@127 133 {
msach@135 134 handleInterMasterReq( currReqst, semanticEnv, masterPr );
msach@135 135 currReqst = currReqst->nextReqst;
msach@127 136 }
msach@135 137 masterEnv->interMasterRequestsFor[thisCoresIdx] = NULL;
msach@135 138
msach@135 139 //Second, check for own request that were handled for other MasterVPs
msach@135 140 currReqst = masterEnv->interMasterRequestsSentBy[thisCoresIdx];
msach@135 141 while(currReqst && currReqst->obsolete)
msach@135 142 {
msach@135 143 InterMasterReqst *nextReqst = currReqst->nextSentReqst;
msach@135 144 VMS__free(currReqst);
msach@135 145 currReqst = nextReqst;
msach@135 146 }
msach@135 147 masterEnv->interMasterRequestsSentBy[thisCoresIdx] = currReqst;
msach@127 148
Me@119 149 //Now, take care of the SlaveVPs
Me@120 150 //Go through the slots -- if Slave there newly suspended, handle its request
Me@120 151 // then, either way, ask assigner to fill each slot
Me@55 152 numSlotsFilled = 0;
Me@26 153 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
Me@0 154 {
Me@4 155 currSlot = schedSlots[ slotIdx ];
Me@0 156
Me@4 157 if( currSlot->workIsDone )
Me@0 158 {
Me@4 159 currSlot->workIsDone = FALSE;
Me@4 160 currSlot->needsProcrAssigned = TRUE;
Me@0 161
Me@0 162 //process requests from slave to master
Me@68 163 //====================== MEASUREMENT STUFF ===================
Me@68 164 #ifdef MEAS__TIME_PLUGIN
Me@68 165 int32 startStamp1, endStamp1;
Me@68 166 saveLowTimeStampCountInto( startStamp1 );
Me@68 167 #endif
Me@68 168 //============================================================
Me@21 169 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
Me@68 170 //====================== MEASUREMENT STUFF ===================
Me@68 171 #ifdef MEAS__TIME_PLUGIN
Me@68 172 saveLowTimeStampCountInto( endStamp1 );
Me@68 173 addIntervalToHist( startStamp1, endStamp1,
Me@68 174 _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 175 addIntervalToHist( startStamp1, endStamp1,
Me@68 176 _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 177 #endif
Me@68 178 //============================================================
Me@0 179 }
Me@4 180 if( currSlot->needsProcrAssigned )
Me@4 181 { //give slot a new virt procr
Me@21 182 schedVirtPr =
Me@31 183 (*slaveScheduler)( semanticEnv, thisCoresIdx );
Me@0 184
Me@21 185 if( schedVirtPr != NULL )
Me@21 186 { currSlot->procrAssignedToSlot = schedVirtPr;
Me@26 187 schedVirtPr->schedSlot = currSlot;
msach@132 188 schedVirtPr->coreAnimatedBy = thisCoresIdx;
Me@26 189 currSlot->needsProcrAssigned = FALSE;
Me@55 190 numSlotsFilled += 1;
Me@55 191
Me@55 192 writeVMSQ( schedVirtPr, readyToAnimateQ );
Me@0 193 }
Me@0 194 }
Me@0 195 }
Me@0 196
Me@55 197
Me@55 198 #ifdef USE_WORK_STEALING
Me@55 199 //If no slots filled, means no more work, look for work to steal.
Me@55 200 if( numSlotsFilled == 0 )
Me@55 201 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
Me@55 202 }
Me@55 203 #endif
Me@26 204
Me@21 205
Me@38 206 #ifdef MEAS__TIME_MASTER
Me@38 207 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
Me@38 208 #endif
Me@38 209
msach@71 210 masterSwitchToCoreLoop(animatingPr);
msach@71 211 flushRegisters();
Me@119 212 }//while(1) MasterLoop
Me@0 213 }
Me@0 214
Me@119 215 /*This is for inter-master communication. Either the master itself or
Me@119 216 * the plugin sends one of these requests. Some are handled here, by the
Me@119 217 * master_loop, others are handed off to the plugin.
Me@119 218 */
Me@119 219 void inline
Me@120 220 handleInterMasterReq( InterMasterReqst *currReq, void *_semEnv,
Me@120 221 VirtProcr *masterPr )
msach@127 222 {
msach@127 223
msach@127 224 switch( currReq->reqType )
msach@127 225 {
msach@127 226 case destVMSCore:
msach@127 227 handleInterVMSCoreReq( (InterVMSCoreReqst *)currReq, masterPr);
Me@119 228 break;
Me@120 229 case destPlugin:
msach@127 230 _VMSMasterEnv->interPluginReqHdlr( ((InterPluginReqst *)currReq)->pluginReq,
msach@127 231 _semEnv );
msach@127 232 break;
Me@119 233 default:
Me@119 234 break;
Me@119 235 }
Me@119 236 }
Me@55 237
Me@119 238 void inline
msach@127 239 handleInterVMSCoreReq( InterVMSCoreReqst *currReq, VirtProcr *masterPr )
Me@119 240 {
msach@137 241 switch( currReq->secondReqType )
Me@119 242 {
msach@135 243 case transfer_free_ptr:
msach@135 244 handleTransferFree( currReq, masterPr );
msach@135 245 currReq->obsolete = 1; //now the sender can free the structure
msach@135 246 break;
msach@137 247 case shutdownVP:
msach@137 248 currReq->obsolete = 1;
msach@137 249 handleShutdown(currReq, masterPr);
msach@137 250 //The Execution of the MasterLoop ends here
msach@137 251 break;
msach@127 252 default:
msach@127 253 break;
Me@119 254 }
msach@127 255 }
msach@69 256
Me@119 257 /*Work Stealing Alg -- racy one
Me@119 258 *This algorithm has a race condition -- the coreloops are accessing their
Me@119 259 * own queues at the same time that this work-stealer on a different core
Me@119 260 * is trying to.
Me@119 261 *The second stealing alg, below, protects against this.
Me@55 262 */
Me@55 263 void inline
Me@55 264 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
Me@55 265 VirtProcr *masterPr )
Me@55 266 {
Me@55 267 VirtProcr *stolenPr;
Me@55 268 int32 coreIdx, i;
Me@55 269 VMSQueueStruc *currQ;
Me@55 270
Me@55 271 stolenPr = NULL;
Me@55 272 coreIdx = masterPr->coreAnimatedBy;
Me@55 273 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 274 {
Me@55 275 if( coreIdx >= NUM_CORES -1 )
Me@55 276 { coreIdx = 0;
Me@55 277 }
Me@55 278 else
Me@55 279 { coreIdx++;
Me@55 280 }
Me@55 281 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 282 if( numInVMSQ( currQ ) > 0 )
Me@55 283 { stolenPr = readVMSQ (currQ );
Me@55 284 break;
Me@55 285 }
Me@55 286 }
Me@55 287
Me@55 288 if( stolenPr != NULL )
Me@55 289 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 290 stolenPr->schedSlot = currSlot;
Me@55 291 currSlot->needsProcrAssigned = FALSE;
Me@55 292
Me@55 293 writeVMSQ( stolenPr, readyToAnimateQ );
Me@55 294 }
Me@55 295 }
Me@55 296
Me@119 297 /*Work Stealing alg -- protected one
Me@119 298 *This algorithm makes the common case fast. Make the coreloop passive,
Me@55 299 * and show its progress. Make the stealer control a gate that coreloop
Me@55 300 * has to pass.
Me@55 301 *To avoid interference, only one stealer at a time. Use a global
Me@55 302 * stealer-lock.
Me@55 303 *
Me@55 304 *The pattern is based on a gate -- stealer shuts the gate, then monitors
Me@55 305 * to be sure any already past make it all the way out, before starting.
Me@55 306 *So, have a "progress" measure just before the gate, then have two after it,
Me@55 307 * one is in a "waiting room" outside the gate, the other is at the exit.
Me@55 308 *Then, the stealer first shuts the gate, then checks the progress measure
Me@55 309 * outside it, then looks to see if the progress measure at the exit is the
Me@55 310 * same. If yes, it knows the protected area is empty 'cause no other way
Me@55 311 * to get in and the last to get in also exited.
Me@55 312 *If the progress measure at the exit is not the same, then the stealer goes
Me@55 313 * into a loop checking both the waiting-area and the exit progress-measures
Me@55 314 * until one of them shows the same as the measure outside the gate. Might
Me@55 315 * as well re-read the measure outside the gate each go around, just to be
Me@55 316 * sure. It is guaranteed that one of the two will eventually match the one
Me@55 317 * outside the gate.
Me@55 318 *
Me@55 319 *Here's an informal proof of correctness:
Me@55 320 *The gate can be closed at any point, and have only four cases:
Me@55 321 * 1) coreloop made it past the gate-closing but not yet past the exit
Me@55 322 * 2) coreloop made it past the pre-gate progress update but not yet past
Me@55 323 * the gate,
Me@55 324 * 3) coreloop is right before the pre-gate update
Me@55 325 * 4) coreloop is past the exit and far from the pre-gate update.
Me@55 326 *
Me@55 327 * Covering the cases in reverse order,
Me@55 328 * 4) is not a problem -- stealer will read pre-gate progress, see that it
Me@55 329 * matches exit progress, and the gate is closed, so stealer can proceed.
Me@55 330 * 3) stealer will read pre-gate progress just after coreloop updates it..
Me@55 331 * so stealer goes into a loop until the coreloop causes wait-progress
Me@55 332 * to match pre-gate progress, so then stealer can proceed
Me@55 333 * 2) same as 3..
Me@55 334 * 1) stealer reads pre-gate progress, sees that it's different than exit,
Me@55 335 * so goes into loop until exit matches pre-gate, now it knows coreloop
Me@55 336 * is not in protected and cannot get back in, so can proceed.
Me@55 337 *
Me@55 338 *Implementation for the stealer:
Me@55 339 *
Me@55 340 *First, acquire the stealer lock -- only cores with no work to do will
Me@55 341 * compete to steal, so not a big performance penalty having only one --
Me@55 342 * will rarely have multiple stealers in a system with plenty of work -- and
Me@55 343 * in a system with little work, it doesn't matter.
Me@55 344 *
Me@55 345 *Note, have single-reader, single-writer pattern for all variables used to
Me@55 346 * communicate between stealer and victims
Me@55 347 *
Me@55 348 *So, scan the queues of the core loops, until find non-empty. Each core
Me@55 349 * has its own list that it scans. The list goes in order from closest to
Me@55 350 * furthest core, so it steals first from close cores. Later can add
Me@55 351 * taking info from the app about overlapping footprints, and scan all the
Me@55 352 * others then choose work with the most footprint overlap with the contents
Me@55 353 * of this core's cache.
Me@55 354 *
Me@55 355 *Now, have a victim want to take work from. So, shut the gate in that
Me@55 356 * coreloop, by setting the "gate closed" var on its stack to TRUE.
Me@55 357 *Then, read the core's pre-gate progress and compare to the core's exit
Me@55 358 * progress.
Me@55 359 *If same, can proceed to take work from the coreloop's queue. When done,
Me@55 360 * write FALSE to gate closed var.
Me@55 361 *If different, then enter a loop that reads the pre-gate progress, then
Me@55 362 * compares to exit progress then to wait progress. When one of two
Me@55 363 * matches, proceed. Take work from the coreloop's queue. When done,
Me@55 364 * write FALSE to the gate closed var.
Me@55 365 *
Me@55 366 */
Me@55 367 void inline
Me@55 368 gateProtected_stealWorkInto( SchedSlot *currSlot,
Me@55 369 VMSQueueStruc *myReadyToAnimateQ,
Me@55 370 VirtProcr *masterPr )
Me@55 371 {
Me@55 372 VirtProcr *stolenPr;
Me@55 373 int32 coreIdx, i, haveAVictim, gotLock;
Me@55 374 VMSQueueStruc *victimsQ;
Me@55 375
Me@55 376 volatile GateStruc *vicGate;
Me@55 377 int32 coreMightBeInProtected;
Me@55 378
Me@55 379
Me@55 380
Me@55 381 //see if any other cores have work available to steal
Me@55 382 haveAVictim = FALSE;
Me@55 383 coreIdx = masterPr->coreAnimatedBy;
Me@55 384 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 385 {
Me@55 386 if( coreIdx >= NUM_CORES -1 )
Me@55 387 { coreIdx = 0;
Me@55 388 }
Me@55 389 else
Me@55 390 { coreIdx++;
Me@55 391 }
Me@55 392 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 393 if( numInVMSQ( victimsQ ) > 0 )
Me@55 394 { haveAVictim = TRUE;
Me@55 395 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
Me@55 396 break;
Me@55 397 }
Me@55 398 }
Me@55 399 if( !haveAVictim ) return; //no work to steal, exit
Me@55 400
Me@55 401 //have a victim core, now get the stealer-lock
Me@55 402 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
Me@55 403 UNLOCKED, LOCKED );
Me@55 404 if( !gotLock ) return; //go back to core loop, which will re-start master
Me@55 405
Me@55 406
Me@55 407 //====== Start Gate-protection =======
Me@55 408 vicGate->gateClosed = TRUE;
Me@55 409 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
Me@55 410 while( coreMightBeInProtected )
Me@55 411 { //wait until sure
Me@55 412 if( vicGate->preGateProgress == vicGate->waitProgress )
Me@55 413 coreMightBeInProtected = FALSE;
Me@55 414 if( vicGate->preGateProgress == vicGate->exitProgress )
Me@55 415 coreMightBeInProtected = FALSE;
Me@55 416 }
Me@55 417
Me@55 418 stolenPr = readVMSQ ( victimsQ );
Me@55 419
Me@55 420 vicGate->gateClosed = FALSE;
Me@55 421 //======= End Gate-protection =======
Me@55 422
Me@55 423
Me@119 424 if( stolenPr != NULL ) //victim could have been in protected and took it
Me@55 425 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 426 stolenPr->schedSlot = currSlot;
Me@55 427 currSlot->needsProcrAssigned = FALSE;
Me@55 428
Me@55 429 writeVMSQ( stolenPr, myReadyToAnimateQ );
Me@55 430 }
Me@55 431
Me@55 432 //unlock the work stealing lock
Me@55 433 _VMSMasterEnv->workStealingLock = UNLOCKED;
Me@55 434 }