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

annotate MasterLoop.c @ 55:3bac84e4e56e

find changesets by author, revision, files, or words in the commit message
Works with correct matrix mult Nov 4 -- switch animators macros, many updates Changed all queues back to VMSQ variants #defines correct, protected, work-stealing, with compiler switch in and out
author Me
date Thu, 04 Nov 2010 18:13:18 -0700
parents 42dd44df1bb0
children 984f7d78bfdf dd3e60aeae26
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
Me@0 24 /*This code is animated by the virtual Master processor.
Me@0 25 *
Me@11 26 *Polls each sched slot exactly once, hands any requests made by a newly
Me@11 27 * done slave to the "request handler" plug-in function
Me@0 28 *
Me@11 29 *Any slots that need a virt procr assigned are given to the "schedule"
Me@11 30 * plug-in function, which tries to assign a virt procr (slave) to it.
Me@0 31 *
Me@11 32 *When all slots needing a processor have been given to the schedule plug-in,
Me@11 33 * a fraction of the procrs successfully scheduled are put into the
Me@11 34 * work queue, then a continuation of this function is put in, then the rest
Me@11 35 * of the virt procrs that were successfully scheduled.
Me@0 36 *
Me@11 37 *The first thing the continuation does is busy-wait until the previous
Me@11 38 * animation completes. This is because an (unlikely) continuation may
Me@11 39 * sneak through queue before previous continuation is done putting second
Me@11 40 * part of scheduled slaves in, which is the only race condition.
Me@0 41 *
Me@0 42 */
Me@0 43
Me@4 44 /*May 29, 2010 -- birth a Master during init so that first core loop to
Me@11 45 * start running gets it and does all the stuff for a newly born --
Me@11 46 * from then on, will be doing continuation, but do suspension self
Me@4 47 * directly at end of master loop
Me@4 48 *So VMS__init just births the master virtual processor same way it births
Me@4 49 * all the others -- then does any extra setup needed and puts it into the
Me@4 50 * work queue.
Me@4 51 *However means have to make masterEnv a global static volatile the same way
Me@31 52 * did with readyToAnimateQ in core loop. -- for performance, put the
Me@11 53 * jump to the core loop directly in here, and have it directly jump back.
Me@31 54 *
Me@31 55 *
Me@31 56 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
Me@31 57 * avoids the suspected bug in the system stack that causes bizarre faults
Me@31 58 * at random places in the system code.
Me@31 59 *
Me@31 60 *So, this function is coupled to each of the MasterVPs, -- meaning this
Me@31 61 * function can't rely on a particular stack and frame -- each MasterVP that
Me@31 62 * animates this function has a different one.
Me@31 63 *
Me@31 64 *At this point, the masterLoop does not write itself into the queue anymore,
Me@31 65 * instead, the coreLoop acquires the masterLock when it has nothing to
Me@31 66 * animate, and then animates its own masterLoop. However, still try to put
Me@31 67 * several AppVPs into the queue to amortize the startup cost of switching
Me@31 68 * to the MasterVP. Note, don't have to worry about latency of requests much
Me@31 69 * because most requests generate work for same core -- only latency issue
Me@31 70 * is case when other cores starved and one core's requests generate work
Me@31 71 * for them -- so keep max in queue to 3 or 4..
Me@4 72 */
Me@31 73 void masterLoop( void *initData, VirtProcr *animatingPr )
Me@21 74 {
Me@55 75 int32 slotIdx, numSlotsFilled;
Me@21 76 VirtProcr *schedVirtPr;
Me@31 77 SchedSlot *currSlot, **schedSlots;
Me@0 78 MasterEnv *masterEnv;
Me@31 79 VMSQueueStruc *readyToAnimateQ;
Me@4 80
Me@0 81 SlaveScheduler slaveScheduler;
Me@0 82 RequestHandler requestHandler;
Me@31 83 void *semanticEnv;
Me@0 84
Me@55 85 int32 thisCoresIdx;
Me@31 86 VirtProcr *masterPr;
Me@31 87 volatile VirtProcr *volatileMasterPr;
Me@31 88
Me@31 89 volatileMasterPr = animatingPr;
Me@31 90 masterPr = volatileMasterPr; //used to force re-define after jmp
Me@31 91
Me@31 92 //First animation of each MasterVP will in turn animate this part
Me@31 93 // of setup code.. (VP creator sets up the stack as if this function
Me@31 94 // was called normally, but actually get here by jmp)
Me@31 95 //So, setup values about stack ptr, jmp pt and all that
Me@4 96 masterPr->nextInstrPt = &&masterLoopStartPt;
Me@0 97
Me@26 98
Me@31 99 //Note, got rid of writing the stack and frame ptr up here, because
Me@31 100 // only one
Me@31 101 // core can ever animate a given MasterVP, so don't need to communicate
Me@31 102 // new frame and stack ptr to the MasterVP storage before a second
Me@31 103 // version of that MasterVP can get animated on a different core.
Me@31 104 //Also got rid of the busy-wait.
Me@26 105
Me@31 106
Me@4 107 masterLoopStartPt:
Me@38 108 //============================= MEASUREMENT STUFF ========================
Me@38 109 #ifdef MEAS__TIME_MASTER
Me@38 110 //Total Master time includes one coreloop time -- just assume the core
Me@38 111 // loop time is same for Master as for AppVPs, even though it will be
Me@41 112 // smaller due to high predictability of the fixed jmp.
Me@38 113 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
Me@38 114 #endif
Me@38 115 //========================================================================
Me@0 116
Me@31 117 masterEnv = _VMSMasterEnv;
Me@4 118
Me@55 119 //GCC may optimize so doesn't always re-define from frame-storage
Me@31 120 masterPr = volatileMasterPr; //just to make sure after jmp
Me@31 121 thisCoresIdx = masterPr->coreAnimatedBy;
Me@31 122 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
Me@31 123 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
Me@4 124
Me@0 125 requestHandler = masterEnv->requestHandler;
Me@0 126 slaveScheduler = masterEnv->slaveScheduler;
Me@21 127 semanticEnv = masterEnv->semanticEnv;
Me@0 128
Me@0 129
Me@31 130 //Poll each slot's Done flag
Me@55 131 numSlotsFilled = 0;
Me@26 132 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
Me@0 133 {
Me@4 134 currSlot = schedSlots[ slotIdx ];
Me@0 135
Me@4 136 if( currSlot->workIsDone )
Me@0 137 {
Me@4 138 currSlot->workIsDone = FALSE;
Me@4 139 currSlot->needsProcrAssigned = TRUE;
Me@0 140
Me@0 141 //process requests from slave to master
Me@21 142 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
Me@0 143 }
Me@4 144 if( currSlot->needsProcrAssigned )
Me@4 145 { //give slot a new virt procr
Me@21 146 schedVirtPr =
Me@31 147 (*slaveScheduler)( semanticEnv, thisCoresIdx );
Me@0 148
Me@21 149 if( schedVirtPr != NULL )
Me@21 150 { currSlot->procrAssignedToSlot = schedVirtPr;
Me@26 151 schedVirtPr->schedSlot = currSlot;
Me@26 152 currSlot->needsProcrAssigned = FALSE;
Me@55 153 numSlotsFilled += 1;
Me@55 154
Me@55 155 writeVMSQ( schedVirtPr, readyToAnimateQ );
Me@0 156 }
Me@0 157 }
Me@0 158 }
Me@0 159
Me@55 160
Me@55 161 #ifdef USE_WORK_STEALING
Me@55 162 //If no slots filled, means no more work, look for work to steal.
Me@55 163 if( numSlotsFilled == 0 )
Me@55 164 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
Me@55 165 }
Me@55 166 #endif
Me@26 167
Me@21 168
Me@38 169 #ifdef MEAS__TIME_MASTER
Me@38 170 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
Me@38 171 #endif
Me@38 172
Me@55 173
Me@55 174 masterSwitchToCoreLoop( masterPr )
Me@0 175 }
Me@0 176
Me@55 177
Me@55 178
Me@55 179 /*This has a race condition -- the coreloops are accessing their own queues
Me@55 180 * at the same time that this work-stealer on a different core is trying to
Me@55 181 */
Me@55 182 void inline
Me@55 183 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
Me@55 184 VirtProcr *masterPr )
Me@55 185 {
Me@55 186 VirtProcr *stolenPr;
Me@55 187 int32 coreIdx, i;
Me@55 188 VMSQueueStruc *currQ;
Me@55 189
Me@55 190 stolenPr = NULL;
Me@55 191 coreIdx = masterPr->coreAnimatedBy;
Me@55 192 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 193 {
Me@55 194 if( coreIdx >= NUM_CORES -1 )
Me@55 195 { coreIdx = 0;
Me@55 196 }
Me@55 197 else
Me@55 198 { coreIdx++;
Me@55 199 }
Me@55 200 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 201 if( numInVMSQ( currQ ) > 0 )
Me@55 202 { stolenPr = readVMSQ (currQ );
Me@55 203 break;
Me@55 204 }
Me@55 205 }
Me@55 206
Me@55 207 if( stolenPr != NULL )
Me@55 208 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 209 stolenPr->schedSlot = currSlot;
Me@55 210 currSlot->needsProcrAssigned = FALSE;
Me@55 211
Me@55 212 writeVMSQ( stolenPr, readyToAnimateQ );
Me@55 213 }
Me@55 214 }
Me@55 215
Me@55 216 /*This algorithm makes the common case fast. Make the coreloop passive,
Me@55 217 * and show its progress. Make the stealer control a gate that coreloop
Me@55 218 * has to pass.
Me@55 219 *To avoid interference, only one stealer at a time. Use a global
Me@55 220 * stealer-lock.
Me@55 221 *
Me@55 222 *The pattern is based on a gate -- stealer shuts the gate, then monitors
Me@55 223 * to be sure any already past make it all the way out, before starting.
Me@55 224 *So, have a "progress" measure just before the gate, then have two after it,
Me@55 225 * one is in a "waiting room" outside the gate, the other is at the exit.
Me@55 226 *Then, the stealer first shuts the gate, then checks the progress measure
Me@55 227 * outside it, then looks to see if the progress measure at the exit is the
Me@55 228 * same. If yes, it knows the protected area is empty 'cause no other way
Me@55 229 * to get in and the last to get in also exited.
Me@55 230 *If the progress measure at the exit is not the same, then the stealer goes
Me@55 231 * into a loop checking both the waiting-area and the exit progress-measures
Me@55 232 * until one of them shows the same as the measure outside the gate. Might
Me@55 233 * as well re-read the measure outside the gate each go around, just to be
Me@55 234 * sure. It is guaranteed that one of the two will eventually match the one
Me@55 235 * outside the gate.
Me@55 236 *
Me@55 237 *Here's an informal proof of correctness:
Me@55 238 *The gate can be closed at any point, and have only four cases:
Me@55 239 * 1) coreloop made it past the gate-closing but not yet past the exit
Me@55 240 * 2) coreloop made it past the pre-gate progress update but not yet past
Me@55 241 * the gate,
Me@55 242 * 3) coreloop is right before the pre-gate update
Me@55 243 * 4) coreloop is past the exit and far from the pre-gate update.
Me@55 244 *
Me@55 245 * Covering the cases in reverse order,
Me@55 246 * 4) is not a problem -- stealer will read pre-gate progress, see that it
Me@55 247 * matches exit progress, and the gate is closed, so stealer can proceed.
Me@55 248 * 3) stealer will read pre-gate progress just after coreloop updates it..
Me@55 249 * so stealer goes into a loop until the coreloop causes wait-progress
Me@55 250 * to match pre-gate progress, so then stealer can proceed
Me@55 251 * 2) same as 3..
Me@55 252 * 1) stealer reads pre-gate progress, sees that it's different than exit,
Me@55 253 * so goes into loop until exit matches pre-gate, now it knows coreloop
Me@55 254 * is not in protected and cannot get back in, so can proceed.
Me@55 255 *
Me@55 256 *Implementation for the stealer:
Me@55 257 *
Me@55 258 *First, acquire the stealer lock -- only cores with no work to do will
Me@55 259 * compete to steal, so not a big performance penalty having only one --
Me@55 260 * will rarely have multiple stealers in a system with plenty of work -- and
Me@55 261 * in a system with little work, it doesn't matter.
Me@55 262 *
Me@55 263 *Note, have single-reader, single-writer pattern for all variables used to
Me@55 264 * communicate between stealer and victims
Me@55 265 *
Me@55 266 *So, scan the queues of the core loops, until find non-empty. Each core
Me@55 267 * has its own list that it scans. The list goes in order from closest to
Me@55 268 * furthest core, so it steals first from close cores. Later can add
Me@55 269 * taking info from the app about overlapping footprints, and scan all the
Me@55 270 * others then choose work with the most footprint overlap with the contents
Me@55 271 * of this core's cache.
Me@55 272 *
Me@55 273 *Now, have a victim want to take work from. So, shut the gate in that
Me@55 274 * coreloop, by setting the "gate closed" var on its stack to TRUE.
Me@55 275 *Then, read the core's pre-gate progress and compare to the core's exit
Me@55 276 * progress.
Me@55 277 *If same, can proceed to take work from the coreloop's queue. When done,
Me@55 278 * write FALSE to gate closed var.
Me@55 279 *If different, then enter a loop that reads the pre-gate progress, then
Me@55 280 * compares to exit progress then to wait progress. When one of two
Me@55 281 * matches, proceed. Take work from the coreloop's queue. When done,
Me@55 282 * write FALSE to the gate closed var.
Me@55 283 *
Me@55 284 */
Me@55 285 void inline
Me@55 286 gateProtected_stealWorkInto( SchedSlot *currSlot,
Me@55 287 VMSQueueStruc *myReadyToAnimateQ,
Me@55 288 VirtProcr *masterPr )
Me@55 289 {
Me@55 290 VirtProcr *stolenPr;
Me@55 291 int32 coreIdx, i, haveAVictim, gotLock;
Me@55 292 VMSQueueStruc *victimsQ;
Me@55 293
Me@55 294 volatile GateStruc *vicGate;
Me@55 295 int32 coreMightBeInProtected;
Me@55 296
Me@55 297
Me@55 298
Me@55 299 //see if any other cores have work available to steal
Me@55 300 haveAVictim = FALSE;
Me@55 301 coreIdx = masterPr->coreAnimatedBy;
Me@55 302 for( i = 0; i < NUM_CORES -1; i++ )
Me@55 303 {
Me@55 304 if( coreIdx >= NUM_CORES -1 )
Me@55 305 { coreIdx = 0;
Me@55 306 }
Me@55 307 else
Me@55 308 { coreIdx++;
Me@55 309 }
Me@55 310 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
Me@55 311 if( numInVMSQ( victimsQ ) > 0 )
Me@55 312 { haveAVictim = TRUE;
Me@55 313 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
Me@55 314 break;
Me@55 315 }
Me@55 316 }
Me@55 317 if( !haveAVictim ) return; //no work to steal, exit
Me@55 318
Me@55 319 //have a victim core, now get the stealer-lock
Me@55 320 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
Me@55 321 UNLOCKED, LOCKED );
Me@55 322 if( !gotLock ) return; //go back to core loop, which will re-start master
Me@55 323
Me@55 324
Me@55 325 //====== Start Gate-protection =======
Me@55 326 vicGate->gateClosed = TRUE;
Me@55 327 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
Me@55 328 while( coreMightBeInProtected )
Me@55 329 { //wait until sure
Me@55 330 if( vicGate->preGateProgress == vicGate->waitProgress )
Me@55 331 coreMightBeInProtected = FALSE;
Me@55 332 if( vicGate->preGateProgress == vicGate->exitProgress )
Me@55 333 coreMightBeInProtected = FALSE;
Me@55 334 }
Me@55 335
Me@55 336 stolenPr = readVMSQ ( victimsQ );
Me@55 337
Me@55 338 vicGate->gateClosed = FALSE;
Me@55 339 //======= End Gate-protection =======
Me@55 340
Me@55 341
Me@55 342 if( stolenPr != NULL ) //victim could have been in protected and taken
Me@55 343 { currSlot->procrAssignedToSlot = stolenPr;
Me@55 344 stolenPr->schedSlot = currSlot;
Me@55 345 currSlot->needsProcrAssigned = FALSE;
Me@55 346
Me@55 347 writeVMSQ( stolenPr, myReadyToAnimateQ );
Me@55 348 }
Me@55 349
Me@55 350 //unlock the work stealing lock
Me@55 351 _VMSMasterEnv->workStealingLock = UNLOCKED;
Me@55 352 }