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

annotate MasterLoop.c @ 70:f9b60012fd74

find changesets by author, revision, files, or words in the commit message
working ucontext version
author Merten Sach <msach@mailbox.tu-berlin.de>
date Fri, 27 May 2011 12:35:40 +0200
parents 11bfe9d136ed
children efb55f1b5fb9
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"
Me@0 13
Me@0 14
Me@55 15 //===========================================================================
Me@55 16 void inline
Me@55 17 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
Me@55 18 VirtProcr *masterPr );
Me@55 19
Me@55 20 //===========================================================================
Me@55 21
Me@55 22
Me@0 23 /*This code is animated by the virtual Master processor.
Me@0 24 *
Me@11 25 *Polls each sched slot exactly once, hands any requests made by a newly
Me@11 26 * done slave to the "request handler" plug-in function
Me@0 27 *
Me@11 28 *Any slots that need a virt procr assigned are given to the "schedule"
Me@11 29 * plug-in function, which tries to assign a virt procr (slave) to it.
Me@0 30 *
Me@11 31 *When all slots needing a processor have been given to the schedule plug-in,
Me@11 32 * a fraction of the procrs successfully scheduled are put into the
Me@11 33 * work queue, then a continuation of this function is put in, then the rest
Me@11 34 * of the virt procrs that were successfully scheduled.
Me@0 35 *
Me@11 36 *The first thing the continuation does is busy-wait until the previous
Me@11 37 * animation completes. This is because an (unlikely) continuation may
Me@11 38 * sneak through queue before previous continuation is done putting second
Me@11 39 * part of scheduled slaves in, which is the only race condition.
Me@0 40 *
Me@0 41 */
Me@0 42
Me@4 43 /*May 29, 2010 -- birth a Master during init so that first core loop to
Me@11 44 * start running gets it and does all the stuff for a newly born --
Me@11 45 * from then on, will be doing continuation, but do suspension self
Me@4 46 * directly at end of master loop
Me@4 47 *So VMS__init just births the master virtual processor same way it births
Me@4 48 * all the others -- then does any extra setup needed and puts it into the
Me@4 49 * work queue.
Me@4 50 *However means have to make masterEnv a global static volatile the same way
Me@31 51 * did with readyToAnimateQ in core loop. -- for performance, put the
Me@11 52 * jump to the core loop directly in here, and have it directly jump back.
Me@31 53 *
Me@31 54 *
Me@31 55 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
Me@31 56 * avoids the suspected bug in the system stack that causes bizarre faults
Me@31 57 * at random places in the system code.
Me@31 58 *
Me@31 59 *So, this function is coupled to each of the MasterVPs, -- meaning this
Me@31 60 * function can't rely on a particular stack and frame -- each MasterVP that
Me@31 61 * animates this function has a different one.
Me@31 62 *
Me@31 63 *At this point, the masterLoop does not write itself into the queue anymore,
Me@31 64 * instead, the coreLoop acquires the masterLock when it has nothing to
Me@31 65 * animate, and then animates its own masterLoop. However, still try to put
Me@31 66 * several AppVPs into the queue to amortize the startup cost of switching
Me@31 67 * to the MasterVP. Note, don't have to worry about latency of requests much
Me@31 68 * because most requests generate work for same core -- only latency issue
Me@31 69 * is case when other cores starved and one core's requests generate work
Me@31 70 * for them -- so keep max in queue to 3 or 4..
Me@4 71 */
Me@31 72 void masterLoop( void *initData, VirtProcr *animatingPr )
Me@21 73 {
Me@55 74 int32 slotIdx, numSlotsFilled;
Me@21 75 VirtProcr *schedVirtPr;
Me@31 76 SchedSlot *currSlot, **schedSlots;
Me@0 77 MasterEnv *masterEnv;
Me@31 78 VMSQueueStruc *readyToAnimateQ;
Me@4 79
Me@0 80 SlaveScheduler slaveScheduler;
Me@0 81 RequestHandler requestHandler;
Me@31 82 void *semanticEnv;
Me@0 83
Me@55 84 int32 thisCoresIdx;
Me@31 85 VirtProcr *masterPr;
Me@31 86
Me@31 87 //First animation of each MasterVP will in turn animate this part
Me@31 88 // of setup code.. (VP creator sets up the stack as if this function
Me@31 89 // was called normally, but actually get here by jmp)
Me@31 90 //So, setup values about stack ptr, jmp pt and all that
msach@70 91 masterPr = animatingPr; //just to make sure after jmp
Me@0 92
Me@26 93
Me@31 94 //Note, got rid of writing the stack and frame ptr up here, because
Me@31 95 // only one
Me@31 96 // core can ever animate a given MasterVP, so don't need to communicate
Me@31 97 // new frame and stack ptr to the MasterVP storage before a second
Me@31 98 // version of that MasterVP can get animated on a different core.
Me@31 99 //Also got rid of the busy-wait.
msach@70 100
msach@70 101 masterEnv = (MasterEnv*)_VMSMasterEnv;
msach@70 102
msach@70 103 thisCoresIdx = masterPr->coreAnimatedBy;
msach@70 104 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
msach@70 105 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
msach@70 106
msach@70 107 requestHandler = masterEnv->requestHandler;
msach@70 108 slaveScheduler = masterEnv->slaveScheduler;
msach@70 109 semanticEnv = masterEnv->semanticEnv;
msach@70 110
Me@26 111
Me@31 112
msach@70 113 while(1){
msach@70 114
Me@38 115 //============================= MEASUREMENT STUFF ========================
Me@38 116 #ifdef MEAS__TIME_MASTER
Me@38 117 //Total Master time includes one coreloop time -- just assume the core
Me@68 118 // loop time is same for Master as for AppVPs, even though it may be
Me@68 119 // smaller due to higher predictability of the fixed jmp.
Me@38 120 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
Me@38 121 #endif
Me@38 122 //========================================================================
Me@0 123
msach@70 124 //Poll each slot's Done flag
Me@55 125 numSlotsFilled = 0;
Me@26 126 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
Me@0 127 {
Me@4 128 currSlot = schedSlots[ slotIdx ];
Me@0 129
Me@4 130 if( currSlot->workIsDone )
Me@0 131 {
Me@4 132 currSlot->workIsDone = FALSE;
Me@4 133 currSlot->needsProcrAssigned = TRUE;
Me@0 134
Me@0 135 //process requests from slave to master
Me@68 136 //====================== MEASUREMENT STUFF ===================
Me@68 137 #ifdef MEAS__TIME_PLUGIN
Me@68 138 int32 startStamp1, endStamp1;
Me@68 139 saveLowTimeStampCountInto( startStamp1 );
Me@68 140 #endif
Me@68 141 //============================================================
Me@21 142 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
Me@68 143 //====================== MEASUREMENT STUFF ===================
Me@68 144 #ifdef MEAS__TIME_PLUGIN
Me@68 145 saveLowTimeStampCountInto( endStamp1 );
Me@68 146 addIntervalToHist( startStamp1, endStamp1,
Me@68 147 _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 148 addIntervalToHist( startStamp1, endStamp1,
Me@68 149 _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 150 #endif
Me@68 151 //============================================================
Me@0 152 }
Me@4 153 if( currSlot->needsProcrAssigned )
Me@4 154 { //give slot a new virt procr
Me@21 155 schedVirtPr =
Me@31 156 (*slaveScheduler)( semanticEnv, thisCoresIdx );
Me@0 157
Me@21 158 if( schedVirtPr != NULL )
Me@21 159 { currSlot->procrAssignedToSlot = schedVirtPr;
Me@26 160 schedVirtPr->schedSlot = currSlot;
Me@26 161 currSlot->needsProcrAssigned = FALSE;
Me@55 162 numSlotsFilled += 1;
Me@55 163
Me@55 164 writeVMSQ( schedVirtPr, readyToAnimateQ );
Me@0 165 }
Me@0 166 }
Me@0 167 }
Me@0 168
Me@55 169
Me@55 170 #ifdef USE_WORK_STEALING
Me@55 171 //If no slots filled, means no more work, look for work to steal.
Me@55 172 if( numSlotsFilled == 0 )
Me@55 173 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
Me@55 174 }
Me@55 175 #endif
Me@26 176
Me@21 177
Me@38 178 #ifdef MEAS__TIME_MASTER
Me@38 179 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
Me@38 180 #endif
Me@38 181
msach@70 182 //Clear lock
msach@70 183 _VMSMasterEnv->masterLock = 0;
msach@70 184
msach@70 185 //Swap to coreLoop
msach@70 186 swapcontext(masterPr->savedVPStatus,masterPr->savedCoreLoopStatus);
msach@70 187 }
Me@62 188
Me@0 189 }
Me@0 190
Me@55 191
Me@55 192 /*This has a race condition -- the coreloops are accessing their own queues
Me@55 193 * at the same time that this work-stealer on a different core is trying to
Me@55 194 */
Me@55 195 void inline
Me@55 196 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
Me@55 197 VirtProcr *masterPr )
Me@55 198 {
Me@55 199 VirtProcr *stolenPr;
Me@55 200 int32 coreIdx, i;
Me@55 201 VMSQueueStruc *currQ;
Me@55 202
Me@55 203 stolenPr = NULL;
Me@55 204 coreIdx = masterPr->coreAnimatedBy;
Me@55 205 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 206 {
Me@55 207 if( coreIdx >= NUM_CORES -1 )
Me@55 208 { coreIdx = 0;
Me@55 209 }
Me@55 210 else
Me@55 211 { coreIdx++;
Me@55 212 }
Me@55 213 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 214 if( numInVMSQ( currQ ) > 0 )
Me@55 215 { stolenPr = readVMSQ (currQ );
Me@55 216 break;
Me@55 217 }
Me@55 218 }
Me@55 219
Me@55 220 if( stolenPr != NULL )
Me@55 221 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 222 stolenPr->schedSlot = currSlot;
Me@55 223 currSlot->needsProcrAssigned = FALSE;
Me@55 224
Me@55 225 writeVMSQ( stolenPr, readyToAnimateQ );
Me@55 226 }
Me@55 227 }
Me@55 228
Me@55 229 /*This algorithm makes the common case fast. Make the coreloop passive,
Me@55 230 * and show its progress. Make the stealer control a gate that coreloop
Me@55 231 * has to pass.
Me@55 232 *To avoid interference, only one stealer at a time. Use a global
Me@55 233 * stealer-lock.
Me@55 234 *
Me@55 235 *The pattern is based on a gate -- stealer shuts the gate, then monitors
Me@55 236 * to be sure any already past make it all the way out, before starting.
Me@55 237 *So, have a "progress" measure just before the gate, then have two after it,
Me@55 238 * one is in a "waiting room" outside the gate, the other is at the exit.
Me@55 239 *Then, the stealer first shuts the gate, then checks the progress measure
Me@55 240 * outside it, then looks to see if the progress measure at the exit is the
Me@55 241 * same. If yes, it knows the protected area is empty 'cause no other way
Me@55 242 * to get in and the last to get in also exited.
Me@55 243 *If the progress measure at the exit is not the same, then the stealer goes
Me@55 244 * into a loop checking both the waiting-area and the exit progress-measures
Me@55 245 * until one of them shows the same as the measure outside the gate. Might
Me@55 246 * as well re-read the measure outside the gate each go around, just to be
Me@55 247 * sure. It is guaranteed that one of the two will eventually match the one
Me@55 248 * outside the gate.
Me@55 249 *
Me@55 250 *Here's an informal proof of correctness:
Me@55 251 *The gate can be closed at any point, and have only four cases:
Me@55 252 * 1) coreloop made it past the gate-closing but not yet past the exit
Me@55 253 * 2) coreloop made it past the pre-gate progress update but not yet past
Me@55 254 * the gate,
Me@55 255 * 3) coreloop is right before the pre-gate update
Me@55 256 * 4) coreloop is past the exit and far from the pre-gate update.
Me@55 257 *
Me@55 258 * Covering the cases in reverse order,
Me@55 259 * 4) is not a problem -- stealer will read pre-gate progress, see that it
Me@55 260 * matches exit progress, and the gate is closed, so stealer can proceed.
Me@55 261 * 3) stealer will read pre-gate progress just after coreloop updates it..
Me@55 262 * so stealer goes into a loop until the coreloop causes wait-progress
Me@55 263 * to match pre-gate progress, so then stealer can proceed
Me@55 264 * 2) same as 3..
Me@55 265 * 1) stealer reads pre-gate progress, sees that it's different than exit,
Me@55 266 * so goes into loop until exit matches pre-gate, now it knows coreloop
Me@55 267 * is not in protected and cannot get back in, so can proceed.
Me@55 268 *
Me@55 269 *Implementation for the stealer:
Me@55 270 *
Me@55 271 *First, acquire the stealer lock -- only cores with no work to do will
Me@55 272 * compete to steal, so not a big performance penalty having only one --
Me@55 273 * will rarely have multiple stealers in a system with plenty of work -- and
Me@55 274 * in a system with little work, it doesn't matter.
Me@55 275 *
Me@55 276 *Note, have single-reader, single-writer pattern for all variables used to
Me@55 277 * communicate between stealer and victims
Me@55 278 *
Me@55 279 *So, scan the queues of the core loops, until find non-empty. Each core
Me@55 280 * has its own list that it scans. The list goes in order from closest to
Me@55 281 * furthest core, so it steals first from close cores. Later can add
Me@55 282 * taking info from the app about overlapping footprints, and scan all the
Me@55 283 * others then choose work with the most footprint overlap with the contents
Me@55 284 * of this core's cache.
Me@55 285 *
Me@55 286 *Now, have a victim want to take work from. So, shut the gate in that
Me@55 287 * coreloop, by setting the "gate closed" var on its stack to TRUE.
Me@55 288 *Then, read the core's pre-gate progress and compare to the core's exit
Me@55 289 * progress.
Me@55 290 *If same, can proceed to take work from the coreloop's queue. When done,
Me@55 291 * write FALSE to gate closed var.
Me@55 292 *If different, then enter a loop that reads the pre-gate progress, then
Me@55 293 * compares to exit progress then to wait progress. When one of two
Me@55 294 * matches, proceed. Take work from the coreloop's queue. When done,
Me@55 295 * write FALSE to the gate closed var.
Me@55 296 *
Me@55 297 */
Me@55 298 void inline
Me@55 299 gateProtected_stealWorkInto( SchedSlot *currSlot,
Me@55 300 VMSQueueStruc *myReadyToAnimateQ,
Me@55 301 VirtProcr *masterPr )
Me@55 302 {
Me@55 303 VirtProcr *stolenPr;
Me@55 304 int32 coreIdx, i, haveAVictim, gotLock;
Me@55 305 VMSQueueStruc *victimsQ;
Me@55 306
Me@55 307 volatile GateStruc *vicGate;
Me@55 308 int32 coreMightBeInProtected;
Me@55 309
Me@55 310
Me@55 311
Me@55 312 //see if any other cores have work available to steal
Me@55 313 haveAVictim = FALSE;
Me@55 314 coreIdx = masterPr->coreAnimatedBy;
Me@55 315 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 316 {
Me@55 317 if( coreIdx >= NUM_CORES -1 )
Me@55 318 { coreIdx = 0;
Me@55 319 }
Me@55 320 else
Me@55 321 { coreIdx++;
Me@55 322 }
Me@55 323 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 324 if( numInVMSQ( victimsQ ) > 0 )
Me@55 325 { haveAVictim = TRUE;
Me@55 326 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
Me@55 327 break;
Me@55 328 }
Me@55 329 }
Me@55 330 if( !haveAVictim ) return; //no work to steal, exit
Me@55 331
Me@55 332 //have a victim core, now get the stealer-lock
Me@55 333 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
Me@55 334 UNLOCKED, LOCKED );
Me@55 335 if( !gotLock ) return; //go back to core loop, which will re-start master
Me@55 336
Me@55 337
Me@55 338 //====== Start Gate-protection =======
Me@55 339 vicGate->gateClosed = TRUE;
Me@55 340 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
Me@55 341 while( coreMightBeInProtected )
Me@55 342 { //wait until sure
Me@55 343 if( vicGate->preGateProgress == vicGate->waitProgress )
Me@55 344 coreMightBeInProtected = FALSE;
Me@55 345 if( vicGate->preGateProgress == vicGate->exitProgress )
Me@55 346 coreMightBeInProtected = FALSE;
Me@55 347 }
Me@55 348
Me@55 349 stolenPr = readVMSQ ( victimsQ );
Me@55 350
Me@55 351 vicGate->gateClosed = FALSE;
Me@55 352 //======= End Gate-protection =======
Me@55 353
Me@55 354
Me@55 355 if( stolenPr != NULL ) //victim could have been in protected and taken
Me@55 356 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 357 stolenPr->schedSlot = currSlot;
Me@55 358 currSlot->needsProcrAssigned = FALSE;
Me@55 359
Me@55 360 writeVMSQ( stolenPr, myReadyToAnimateQ );
Me@55 361 }
Me@55 362
Me@55 363 //unlock the work stealing lock
Me@55 364 _VMSMasterEnv->workStealingLock = UNLOCKED;
Me@55 365 }