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

annotate MasterLoop.c @ 134:a9b72021f053

find changesets by author, revision, files, or words in the commit message
Distributed memory management w/o free requests working
author Merten Sach <msach@mailbox.tu-berlin.de>
date Fri, 16 Sep 2011 16:19:24 +0200
parents dbfc8382d546
children 0b49fd35afc1
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@127 15 #include "inter_VMS_request_handlers.h"
Me@0 16
Me@55 17 //===========================================================================
Me@55 18 void inline
Me@55 19 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
msach@127 20 VirtProcr *masterPr);
msach@127 21
msach@127 22 void inline
msach@127 23 handleInterMasterReq( InterMasterReqst *currReq, void *_semEnv,
msach@127 24 VirtProcr *masterPr);
msach@127 25
msach@127 26 void inline
msach@127 27 handleInterVMSCoreReq( InterVMSCoreReqst *currReq, VirtProcr *masterPr);
Me@55 28
Me@55 29 //===========================================================================
Me@55 30
Me@55 31
msach@69 32
Me@0 33 /*This code is animated by the virtual Master processor.
Me@0 34 *
Me@11 35 *Polls each sched slot exactly once, hands any requests made by a newly
Me@11 36 * done slave to the "request handler" plug-in function
Me@0 37 *
Me@11 38 *Any slots that need a virt procr assigned are given to the "schedule"
Me@11 39 * plug-in function, which tries to assign a virt procr (slave) to it.
Me@0 40 *
Me@11 41 *When all slots needing a processor have been given to the schedule plug-in,
Me@11 42 * a fraction of the procrs successfully scheduled are put into the
Me@11 43 * work queue, then a continuation of this function is put in, then the rest
Me@11 44 * of the virt procrs that were successfully scheduled.
Me@0 45 *
Me@11 46 *The first thing the continuation does is busy-wait until the previous
Me@11 47 * animation completes. This is because an (unlikely) continuation may
Me@11 48 * sneak through queue before previous continuation is done putting second
Me@11 49 * part of scheduled slaves in, which is the only race condition.
Me@0 50 *
Me@0 51 */
Me@0 52
Me@4 53 /*May 29, 2010 -- birth a Master during init so that first core loop to
Me@11 54 * start running gets it and does all the stuff for a newly born --
Me@11 55 * from then on, will be doing continuation, but do suspension self
Me@4 56 * directly at end of master loop
Me@4 57 *So VMS__init just births the master virtual processor same way it births
Me@4 58 * all the others -- then does any extra setup needed and puts it into the
Me@4 59 * work queue.
Me@120 60 *However means have to make masterEnv a global static volatile.
Me@31 61 *
Me@31 62 *
Me@31 63 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
Me@31 64 * avoids the suspected bug in the system stack that causes bizarre faults
Me@31 65 * at random places in the system code.
Me@31 66 *
Me@31 67 *So, this function is coupled to each of the MasterVPs, -- meaning this
Me@31 68 * function can't rely on a particular stack and frame -- each MasterVP that
Me@120 69 * animates this function has a different stack.
Me@31 70 *
Me@31 71 *At this point, the masterLoop does not write itself into the queue anymore,
Me@31 72 * instead, the coreLoop acquires the masterLock when it has nothing to
Me@31 73 * animate, and then animates its own masterLoop. However, still try to put
Me@31 74 * several AppVPs into the queue to amortize the startup cost of switching
Me@31 75 * to the MasterVP. Note, don't have to worry about latency of requests much
Me@31 76 * because most requests generate work for same core -- only latency issue
Me@31 77 * is case when other cores starved and one core's requests generate work
Me@31 78 * for them -- so keep max in queue to 3 or 4..
Me@4 79 */
Me@31 80 void masterLoop( void *initData, VirtProcr *animatingPr )
Me@21 81 {
Me@55 82 int32 slotIdx, numSlotsFilled;
Me@21 83 VirtProcr *schedVirtPr;
Me@31 84 SchedSlot *currSlot, **schedSlots;
Me@0 85 MasterEnv *masterEnv;
Me@31 86 VMSQueueStruc *readyToAnimateQ;
Me@4 87
Me@0 88 SlaveScheduler slaveScheduler;
Me@0 89 RequestHandler requestHandler;
Me@31 90 void *semanticEnv;
Me@0 91
Me@55 92 int32 thisCoresIdx;
Me@31 93 VirtProcr *masterPr;
msach@69 94 volatile VirtProcr *volatileMasterPr;
msach@69 95
msach@69 96 volatileMasterPr = animatingPr;
msach@69 97 masterPr = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
msach@132 98 masterEnv = (MasterEnv*)_VMSMasterEnv;
Me@31 99
Me@31 100 //First animation of each MasterVP will in turn animate this part
Me@31 101 // of setup code.. (VP creator sets up the stack as if this function
Me@31 102 // was called normally, but actually get here by jmp)
Me@0 103
Me@120 104 //Sept 2011
Me@120 105 //Old code jumped directly to this point, but doesn't work on x64
Me@120 106 // So, just make this an endless loop, and do assembly function at end
Me@120 107 // that saves its own return addr, then jumps to core_loop.
Me@120 108 while(1)
msach@132 109 {
Me@38 110 //============================= MEASUREMENT STUFF ========================
Me@38 111 #ifdef MEAS__TIME_MASTER
Me@38 112 //Total Master time includes one coreloop time -- just assume the core
Me@120 113 // loop time is same for Master as is for AppVPs, even though it may be
Me@68 114 // smaller due to higher predictability of the fixed jmp.
Me@38 115 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
Me@38 116 #endif
Me@38 117 //========================================================================
Me@0 118
msach@132 119 //GCC may optimize so doesn't always re-define from frame-storage
msach@69 120 thisCoresIdx = masterPr->coreAnimatedBy;
msach@132 121 masterEnv->currentMasterProcrID = thisCoresIdx;
msach@69 122 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
msach@69 123 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
msach@69 124
msach@69 125 requestHandler = masterEnv->requestHandler;
msach@69 126 slaveScheduler = masterEnv->slaveScheduler;
msach@69 127 semanticEnv = masterEnv->semanticEnv;
msach@69 128
Me@119 129 //First, check for requests from other MasterVPs, and handle them
msach@127 130 InterMasterReqst* currReq = masterEnv->interMasterRequestsFor[thisCoresIdx];
msach@127 131 while(currReq)
msach@127 132 {
msach@127 133 handleInterMasterReq( currReq, semanticEnv, masterPr );
msach@127 134 currReq = currReq->nextReqst;
msach@127 135 }
msach@127 136
Me@119 137 //Now, take care of the SlaveVPs
Me@120 138 //Go through the slots -- if Slave there newly suspended, handle its request
Me@120 139 // then, either way, ask assigner to fill each slot
Me@55 140 numSlotsFilled = 0;
Me@26 141 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
Me@0 142 {
Me@4 143 currSlot = schedSlots[ slotIdx ];
Me@0 144
Me@4 145 if( currSlot->workIsDone )
Me@0 146 {
Me@4 147 currSlot->workIsDone = FALSE;
Me@4 148 currSlot->needsProcrAssigned = TRUE;
Me@0 149
Me@0 150 //process requests from slave to master
Me@68 151 //====================== MEASUREMENT STUFF ===================
Me@68 152 #ifdef MEAS__TIME_PLUGIN
Me@68 153 int32 startStamp1, endStamp1;
Me@68 154 saveLowTimeStampCountInto( startStamp1 );
Me@68 155 #endif
Me@68 156 //============================================================
Me@21 157 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
Me@68 158 //====================== MEASUREMENT STUFF ===================
Me@68 159 #ifdef MEAS__TIME_PLUGIN
Me@68 160 saveLowTimeStampCountInto( endStamp1 );
Me@68 161 addIntervalToHist( startStamp1, endStamp1,
Me@68 162 _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 163 addIntervalToHist( startStamp1, endStamp1,
Me@68 164 _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 165 #endif
Me@68 166 //============================================================
Me@0 167 }
Me@4 168 if( currSlot->needsProcrAssigned )
Me@4 169 { //give slot a new virt procr
Me@21 170 schedVirtPr =
Me@31 171 (*slaveScheduler)( semanticEnv, thisCoresIdx );
Me@0 172
Me@21 173 if( schedVirtPr != NULL )
Me@21 174 { currSlot->procrAssignedToSlot = schedVirtPr;
Me@26 175 schedVirtPr->schedSlot = currSlot;
msach@132 176 schedVirtPr->coreAnimatedBy = thisCoresIdx;
Me@26 177 currSlot->needsProcrAssigned = FALSE;
Me@55 178 numSlotsFilled += 1;
Me@55 179
Me@55 180 writeVMSQ( schedVirtPr, readyToAnimateQ );
Me@0 181 }
Me@0 182 }
Me@0 183 }
Me@0 184
Me@55 185
Me@55 186 #ifdef USE_WORK_STEALING
Me@55 187 //If no slots filled, means no more work, look for work to steal.
Me@55 188 if( numSlotsFilled == 0 )
Me@55 189 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
Me@55 190 }
Me@55 191 #endif
Me@26 192
Me@21 193
Me@38 194 #ifdef MEAS__TIME_MASTER
Me@38 195 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
Me@38 196 #endif
Me@38 197
msach@71 198 masterSwitchToCoreLoop(animatingPr);
msach@71 199 flushRegisters();
Me@119 200 }//while(1) MasterLoop
Me@0 201 }
Me@0 202
Me@119 203 /*This is for inter-master communication. Either the master itself or
Me@119 204 * the plugin sends one of these requests. Some are handled here, by the
Me@119 205 * master_loop, others are handed off to the plugin.
Me@119 206 */
Me@119 207 void inline
Me@120 208 handleInterMasterReq( InterMasterReqst *currReq, void *_semEnv,
Me@120 209 VirtProcr *masterPr )
msach@127 210 {
msach@127 211
msach@127 212 switch( currReq->reqType )
msach@127 213 {
msach@127 214 case destVMSCore:
msach@127 215 handleInterVMSCoreReq( (InterVMSCoreReqst *)currReq, masterPr);
Me@119 216 break;
Me@120 217 case destPlugin:
msach@127 218 _VMSMasterEnv->interPluginReqHdlr( ((InterPluginReqst *)currReq)->pluginReq,
msach@127 219 _semEnv );
msach@127 220 break;
Me@119 221 default:
Me@119 222 break;
Me@119 223 }
Me@119 224 }
Me@55 225
Me@119 226 void inline
msach@127 227 handleInterVMSCoreReq( InterVMSCoreReqst *currReq, VirtProcr *masterPr )
Me@119 228 {
Me@119 229 switch( currReq->reqType )
Me@119 230 {
Me@120 231 case transfer_free_ptr: handleTransferFree( currReq, masterPr );
Me@119 232 break;
msach@127 233 default:
msach@127 234 break;
Me@119 235 }
msach@127 236 }
msach@69 237
Me@119 238 /*Work Stealing Alg -- racy one
Me@119 239 *This algorithm has a race condition -- the coreloops are accessing their
Me@119 240 * own queues at the same time that this work-stealer on a different core
Me@119 241 * is trying to.
Me@119 242 *The second stealing alg, below, protects against this.
Me@55 243 */
Me@55 244 void inline
Me@55 245 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
Me@55 246 VirtProcr *masterPr )
Me@55 247 {
Me@55 248 VirtProcr *stolenPr;
Me@55 249 int32 coreIdx, i;
Me@55 250 VMSQueueStruc *currQ;
Me@55 251
Me@55 252 stolenPr = NULL;
Me@55 253 coreIdx = masterPr->coreAnimatedBy;
Me@55 254 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 255 {
Me@55 256 if( coreIdx >= NUM_CORES -1 )
Me@55 257 { coreIdx = 0;
Me@55 258 }
Me@55 259 else
Me@55 260 { coreIdx++;
Me@55 261 }
Me@55 262 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 263 if( numInVMSQ( currQ ) > 0 )
Me@55 264 { stolenPr = readVMSQ (currQ );
Me@55 265 break;
Me@55 266 }
Me@55 267 }
Me@55 268
Me@55 269 if( stolenPr != NULL )
Me@55 270 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 271 stolenPr->schedSlot = currSlot;
Me@55 272 currSlot->needsProcrAssigned = FALSE;
Me@55 273
Me@55 274 writeVMSQ( stolenPr, readyToAnimateQ );
Me@55 275 }
Me@55 276 }
Me@55 277
Me@119 278 /*Work Stealing alg -- protected one
Me@119 279 *This algorithm makes the common case fast. Make the coreloop passive,
Me@55 280 * and show its progress. Make the stealer control a gate that coreloop
Me@55 281 * has to pass.
Me@55 282 *To avoid interference, only one stealer at a time. Use a global
Me@55 283 * stealer-lock.
Me@55 284 *
Me@55 285 *The pattern is based on a gate -- stealer shuts the gate, then monitors
Me@55 286 * to be sure any already past make it all the way out, before starting.
Me@55 287 *So, have a "progress" measure just before the gate, then have two after it,
Me@55 288 * one is in a "waiting room" outside the gate, the other is at the exit.
Me@55 289 *Then, the stealer first shuts the gate, then checks the progress measure
Me@55 290 * outside it, then looks to see if the progress measure at the exit is the
Me@55 291 * same. If yes, it knows the protected area is empty 'cause no other way
Me@55 292 * to get in and the last to get in also exited.
Me@55 293 *If the progress measure at the exit is not the same, then the stealer goes
Me@55 294 * into a loop checking both the waiting-area and the exit progress-measures
Me@55 295 * until one of them shows the same as the measure outside the gate. Might
Me@55 296 * as well re-read the measure outside the gate each go around, just to be
Me@55 297 * sure. It is guaranteed that one of the two will eventually match the one
Me@55 298 * outside the gate.
Me@55 299 *
Me@55 300 *Here's an informal proof of correctness:
Me@55 301 *The gate can be closed at any point, and have only four cases:
Me@55 302 * 1) coreloop made it past the gate-closing but not yet past the exit
Me@55 303 * 2) coreloop made it past the pre-gate progress update but not yet past
Me@55 304 * the gate,
Me@55 305 * 3) coreloop is right before the pre-gate update
Me@55 306 * 4) coreloop is past the exit and far from the pre-gate update.
Me@55 307 *
Me@55 308 * Covering the cases in reverse order,
Me@55 309 * 4) is not a problem -- stealer will read pre-gate progress, see that it
Me@55 310 * matches exit progress, and the gate is closed, so stealer can proceed.
Me@55 311 * 3) stealer will read pre-gate progress just after coreloop updates it..
Me@55 312 * so stealer goes into a loop until the coreloop causes wait-progress
Me@55 313 * to match pre-gate progress, so then stealer can proceed
Me@55 314 * 2) same as 3..
Me@55 315 * 1) stealer reads pre-gate progress, sees that it's different than exit,
Me@55 316 * so goes into loop until exit matches pre-gate, now it knows coreloop
Me@55 317 * is not in protected and cannot get back in, so can proceed.
Me@55 318 *
Me@55 319 *Implementation for the stealer:
Me@55 320 *
Me@55 321 *First, acquire the stealer lock -- only cores with no work to do will
Me@55 322 * compete to steal, so not a big performance penalty having only one --
Me@55 323 * will rarely have multiple stealers in a system with plenty of work -- and
Me@55 324 * in a system with little work, it doesn't matter.
Me@55 325 *
Me@55 326 *Note, have single-reader, single-writer pattern for all variables used to
Me@55 327 * communicate between stealer and victims
Me@55 328 *
Me@55 329 *So, scan the queues of the core loops, until find non-empty. Each core
Me@55 330 * has its own list that it scans. The list goes in order from closest to
Me@55 331 * furthest core, so it steals first from close cores. Later can add
Me@55 332 * taking info from the app about overlapping footprints, and scan all the
Me@55 333 * others then choose work with the most footprint overlap with the contents
Me@55 334 * of this core's cache.
Me@55 335 *
Me@55 336 *Now, have a victim want to take work from. So, shut the gate in that
Me@55 337 * coreloop, by setting the "gate closed" var on its stack to TRUE.
Me@55 338 *Then, read the core's pre-gate progress and compare to the core's exit
Me@55 339 * progress.
Me@55 340 *If same, can proceed to take work from the coreloop's queue. When done,
Me@55 341 * write FALSE to gate closed var.
Me@55 342 *If different, then enter a loop that reads the pre-gate progress, then
Me@55 343 * compares to exit progress then to wait progress. When one of two
Me@55 344 * matches, proceed. Take work from the coreloop's queue. When done,
Me@55 345 * write FALSE to the gate closed var.
Me@55 346 *
Me@55 347 */
Me@55 348 void inline
Me@55 349 gateProtected_stealWorkInto( SchedSlot *currSlot,
Me@55 350 VMSQueueStruc *myReadyToAnimateQ,
Me@55 351 VirtProcr *masterPr )
Me@55 352 {
Me@55 353 VirtProcr *stolenPr;
Me@55 354 int32 coreIdx, i, haveAVictim, gotLock;
Me@55 355 VMSQueueStruc *victimsQ;
Me@55 356
Me@55 357 volatile GateStruc *vicGate;
Me@55 358 int32 coreMightBeInProtected;
Me@55 359
Me@55 360
Me@55 361
Me@55 362 //see if any other cores have work available to steal
Me@55 363 haveAVictim = FALSE;
Me@55 364 coreIdx = masterPr->coreAnimatedBy;
Me@55 365 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 366 {
Me@55 367 if( coreIdx >= NUM_CORES -1 )
Me@55 368 { coreIdx = 0;
Me@55 369 }
Me@55 370 else
Me@55 371 { coreIdx++;
Me@55 372 }
Me@55 373 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 374 if( numInVMSQ( victimsQ ) > 0 )
Me@55 375 { haveAVictim = TRUE;
Me@55 376 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
Me@55 377 break;
Me@55 378 }
Me@55 379 }
Me@55 380 if( !haveAVictim ) return; //no work to steal, exit
Me@55 381
Me@55 382 //have a victim core, now get the stealer-lock
Me@55 383 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
Me@55 384 UNLOCKED, LOCKED );
Me@55 385 if( !gotLock ) return; //go back to core loop, which will re-start master
Me@55 386
Me@55 387
Me@55 388 //====== Start Gate-protection =======
Me@55 389 vicGate->gateClosed = TRUE;
Me@55 390 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
Me@55 391 while( coreMightBeInProtected )
Me@55 392 { //wait until sure
Me@55 393 if( vicGate->preGateProgress == vicGate->waitProgress )
Me@55 394 coreMightBeInProtected = FALSE;
Me@55 395 if( vicGate->preGateProgress == vicGate->exitProgress )
Me@55 396 coreMightBeInProtected = FALSE;
Me@55 397 }
Me@55 398
Me@55 399 stolenPr = readVMSQ ( victimsQ );
Me@55 400
Me@55 401 vicGate->gateClosed = FALSE;
Me@55 402 //======= End Gate-protection =======
Me@55 403
Me@55 404
Me@119 405 if( stolenPr != NULL ) //victim could have been in protected and took it
Me@55 406 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 407 stolenPr->schedSlot = currSlot;
Me@55 408 currSlot->needsProcrAssigned = FALSE;
Me@55 409
Me@55 410 writeVMSQ( stolenPr, myReadyToAnimateQ );
Me@55 411 }
Me@55 412
Me@55 413 //unlock the work stealing lock
Me@55 414 _VMSMasterEnv->workStealingLock = UNLOCKED;
Me@55 415 }