annotate VMS.c @ 194:0072a555f59c

remove subrepos
author Nina Engelhardt <nengel@mailbox.tu-berlin.de>
date Thu, 09 Feb 2012 16:48:42 +0100
parents 69eb54ce9c4b
children d83f59e6e2db
rev   line source
Me@0 1 /*
Me@38 2 * Copyright 2010 OpenSourceStewardshipFoundation
Me@0 3 *
Me@0 4 * Licensed under BSD
Me@0 5 */
Me@0 6
Me@0 7 #include <stdio.h>
Me@0 8 #include <stdlib.h>
Me@50 9 #include <string.h>
Me@0 10 #include <malloc.h>
msach@76 11 #include <inttypes.h>
Me@50 12 #include <sys/time.h>
Me@0 13
Me@0 14 #include "VMS.h"
msach@77 15 #include "ProcrContext.h"
Me@0 16 #include "Queue_impl/BlockingQueue.h"
Me@38 17 #include "Histogram/Histogram.h"
Me@0 18
Nina@109 19 #include <unistd.h>
Nina@109 20 #include <fcntl.h>
Nina@109 21 #include <linux/types.h>
engelhardt@108 22 #include <linux/perf_event.h>
Nina@109 23 #include <errno.h>
Nina@109 24 #include <sys/syscall.h>
Nina@109 25 #include <linux/prctl.h>
Nina@109 26
Me@0 27
Me@26 28 #define thdAttrs NULL
Me@26 29
Me@22 30 //===========================================================================
Me@22 31 void
Me@22 32 shutdownFn( void *dummy, VirtProcr *dummy2 );
Me@22 33
Me@31 34 SchedSlot **
Me@31 35 create_sched_slots();
Me@22 36
Me@28 37 void
Me@28 38 create_masterEnv();
Me@28 39
Me@28 40 void
Me@28 41 create_the_coreLoop_OS_threads();
Me@28 42
Me@50 43 MallocProlog *
Me@50 44 create_free_list();
Me@50 45
Me@53 46 void
Me@53 47 endOSThreadFn( void *initData, VirtProcr *animatingPr );
Me@50 48
Me@26 49 pthread_mutex_t suspendLock = PTHREAD_MUTEX_INITIALIZER;
Me@26 50 pthread_cond_t suspend_cond = PTHREAD_COND_INITIALIZER;
Me@26 51
Me@22 52 //===========================================================================
Me@22 53
Me@0 54 /*Setup has two phases:
Me@0 55 * 1) Semantic layer first calls init_VMS, which creates masterEnv, and puts
Me@8 56 * the master virt procr into the work-queue, ready for first "call"
Me@8 57 * 2) Semantic layer then does its own init, which creates the seed virt
Me@8 58 * procr inside the semantic layer, ready to schedule it when
Me@0 59 * asked by the first run of the masterLoop.
Me@0 60 *
Me@0 61 *This part is bit weird because VMS really wants to be "always there", and
Me@0 62 * have applications attach and detach.. for now, this VMS is part of
Me@0 63 * the app, so the VMS system starts up as part of running the app.
Me@0 64 *
Me@8 65 *The semantic layer is isolated from the VMS internals by making the
Me@8 66 * semantic layer do setup to a state that it's ready with its
Me@8 67 * initial virt procrs, ready to schedule them to slots when the masterLoop
Me@0 68 * asks. Without this pattern, the semantic layer's setup would
Me@8 69 * have to modify slots directly to assign the initial virt-procrs, and put
Me@31 70 * them into the readyToAnimateQ itself, breaking the isolation completely.
Me@0 71 *
Me@0 72 *
Me@8 73 *The semantic layer creates the initial virt procr(s), and adds its
Me@8 74 * own environment to masterEnv, and fills in the pointers to
Me@0 75 * the requestHandler and slaveScheduler plug-in functions
Me@8 76 */
Me@8 77
Me@8 78 /*This allocates VMS data structures, populates the master VMSProc,
Me@0 79 * and master environment, and returns the master environment to the semantic
Me@0 80 * layer.
Me@0 81 */
Me@8 82 void
Me@8 83 VMS__init()
Me@28 84 {
Me@28 85 create_masterEnv();
Me@28 86 create_the_coreLoop_OS_threads();
Me@28 87 }
Me@28 88
msach@71 89 #ifdef SEQUENTIAL
msach@71 90
Me@28 91 /*To initialize the sequential version, just don't create the threads
Me@28 92 */
Me@28 93 void
Me@28 94 VMS__init_Seq()
Me@28 95 {
Me@28 96 create_masterEnv();
Me@28 97 }
Me@28 98
msach@71 99 #endif
msach@71 100
Me@28 101 void
Me@28 102 create_masterEnv()
Me@31 103 { MasterEnv *masterEnv;
Me@55 104 VMSQueueStruc **readyToAnimateQs;
Me@31 105 int coreIdx;
Me@31 106 VirtProcr **masterVPs;
Me@31 107 SchedSlot ***allSchedSlots; //ptr to array of ptrs
Me@53 108
Me@53 109
Me@31 110 //Make the master env, which holds everything else
Me@1 111 _VMSMasterEnv = malloc( sizeof(MasterEnv) );
Me@53 112
Me@53 113 //Very first thing put into the master env is the free-list, seeded
Me@53 114 // with a massive initial chunk of memory.
Me@53 115 //After this, all other mallocs are VMS__malloc.
Me@53 116 _VMSMasterEnv->freeListHead = VMS_ext__create_free_list();
Me@53 117
Me@65 118
Me@65 119 //============================= MEASUREMENT STUFF ========================
Me@65 120 #ifdef MEAS__TIME_MALLOC
msach@160 121 _VMSMasterEnv->mallocTimeHist = makeFixedBinHistExt( 100, 0, 30,
msach@79 122 "malloc_time_hist");
msach@160 123 _VMSMasterEnv->freeTimeHist = makeFixedBinHistExt( 100, 0, 30,
msach@79 124 "free_time_hist");
Me@65 125 #endif
Me@68 126 #ifdef MEAS__TIME_PLUGIN
msach@160 127 _VMSMasterEnv->reqHdlrLowTimeHist = makeFixedBinHistExt( 100, 0, 200,
msach@79 128 "plugin_low_time_hist");
msach@160 129 _VMSMasterEnv->reqHdlrHighTimeHist = makeFixedBinHistExt( 100, 0, 200,
msach@79 130 "plugin_high_time_hist");
Me@68 131 #endif
Me@65 132 //========================================================================
Me@65 133
Me@53 134 //===================== Only VMS__malloc after this ====================
msach@69 135 masterEnv = (MasterEnv*)_VMSMasterEnv;
Me@31 136
Me@31 137 //Make a readyToAnimateQ for each core loop
Me@55 138 readyToAnimateQs = VMS__malloc( NUM_CORES * sizeof(VMSQueueStruc *) );
Me@53 139 masterVPs = VMS__malloc( NUM_CORES * sizeof(VirtProcr *) );
Me@0 140
Me@31 141 //One array for each core, 3 in array, core's masterVP scheds all
Me@53 142 allSchedSlots = VMS__malloc( NUM_CORES * sizeof(SchedSlot *) );
Me@0 143
Me@53 144 _VMSMasterEnv->numProcrsCreated = 0; //used by create procr
Me@31 145 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@53 146 {
Me@55 147 readyToAnimateQs[ coreIdx ] = makeVMSQ();
Me@31 148
Me@50 149 //Q: should give masterVP core-specific info as its init data?
msach@76 150 masterVPs[ coreIdx ] = VMS__create_procr( (VirtProcrFnPtr)&masterLoop, (void*)masterEnv );
Me@31 151 masterVPs[ coreIdx ]->coreAnimatedBy = coreIdx;
nengel@193 152 masterVPs[ coreIdx ]->isMasterVP = TRUE;
Me@31 153 allSchedSlots[ coreIdx ] = create_sched_slots(); //makes for one core
Me@53 154 _VMSMasterEnv->numMasterInARow[ coreIdx ] = 0;
Me@55 155 _VMSMasterEnv->workStealingGates[ coreIdx ] = NULL;
Me@31 156 }
Me@31 157 _VMSMasterEnv->readyToAnimateQs = readyToAnimateQs;
Me@31 158 _VMSMasterEnv->masterVPs = masterVPs;
Me@50 159 _VMSMasterEnv->masterLock = UNLOCKED;
Me@31 160 _VMSMasterEnv->allSchedSlots = allSchedSlots;
Me@55 161 _VMSMasterEnv->workStealingLock = UNLOCKED;
Me@28 162
Me@12 163
Me@31 164 //Aug 19, 2010: no longer need to place initial masterVP into queue
Me@31 165 // because coreLoop now controls -- animates its masterVP when no work
Me@31 166
Me@30 167
Me@50 168 //============================= MEASUREMENT STUFF ========================
Me@50 169 #ifdef STATS__TURN_ON_PROBES
Me@50 170 _VMSMasterEnv->dynIntervalProbesInfo =
msach@69 171 makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->intervalProbes), 200);
Me@30 172
Me@53 173 _VMSMasterEnv->probeNameHashTbl = makeHashTable( 1000, &VMS__free );
Me@53 174
Me@53 175 //put creation time directly into master env, for fast retrieval
Me@50 176 struct timeval timeStamp;
Me@50 177 gettimeofday( &(timeStamp), NULL);
Me@50 178 _VMSMasterEnv->createPtInSecs =
Me@50 179 timeStamp.tv_sec +(timeStamp.tv_usec/1000000.0);
Me@50 180 #endif
Me@65 181 #ifdef MEAS__TIME_MASTER_LOCK
Me@65 182 _VMSMasterEnv->masterLockLowTimeHist = makeFixedBinHist( 50, 0, 2,
Me@65 183 "master lock low time hist");
Me@68 184 _VMSMasterEnv->masterLockHighTimeHist = makeFixedBinHist( 50, 0, 100,
Me@65 185 "master lock high time hist");
Me@65 186 #endif
Me@68 187
msach@76 188 MakeTheMeasHists();
engelhardt@108 189
Nina@166 190
Nina@166 191 #ifdef DETECT_LOOP_GRAPH
nengel@177 192
Nina@166 193 #endif
Nina@166 194
Nina@109 195 #ifdef MEAS__PERF_COUNTERS
nengel@184 196 /*
Nina@131 197 _VMSMasterEnv->counter_history = VMS__malloc(10*sizeof(void*));
Nina@131 198 _VMSMasterEnv->counter_history_array_info = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->counter_history),10);
nengel@184 199 */
Nina@110 200 //printf("Creating HW counters...");
nengel@184 201 /*
Nina@110 202 FILE* output;
Nina@110 203 int n;
Nina@110 204 char filename[255];
Nina@110 205 for(n=0;n<255;n++)
Nina@110 206 {
Nina@113 207 sprintf(filename, "./counters/Counters.%d.csv",n);
Nina@110 208 output = fopen(filename,"r");
Nina@110 209 if(output)
Nina@110 210 {
Nina@110 211 fclose(output);
Nina@110 212 }else{
Nina@110 213 break;
Nina@110 214 }
Nina@110 215 }
Nina@110 216 printf("Saving Counter measurements to File: %s ...\n", filename);
Nina@110 217 output = fopen(filename,"w+");
Nina@110 218 _VMSMasterEnv->counteroutput = output;
nengel@184 219 */
Nina@110 220
engelhardt@108 221 struct perf_event_attr hw_event;
Nina@109 222 memset(&hw_event,0,sizeof(hw_event));
engelhardt@108 223 hw_event.type = PERF_TYPE_HARDWARE;
Nina@109 224 hw_event.size = sizeof(hw_event);
engelhardt@108 225 hw_event.disabled = 1;
Nina@109 226 hw_event.freq = 0;
engelhardt@108 227 hw_event.inherit = 1; /* children inherit it */
engelhardt@108 228 hw_event.pinned = 1; /* must always be on PMU */
engelhardt@108 229 hw_event.exclusive = 0; /* only group on PMU */
engelhardt@108 230 hw_event.exclude_user = 0; /* don't count user */
engelhardt@108 231 hw_event.exclude_kernel = 1; /* ditto kernel */
engelhardt@108 232 hw_event.exclude_hv = 1; /* ditto hypervisor */
engelhardt@108 233 hw_event.exclude_idle = 0; /* don't count when idle */
engelhardt@108 234 hw_event.mmap = 0; /* include mmap data */
engelhardt@108 235 hw_event.comm = 0; /* include comm data */
engelhardt@108 236
Nina@109 237
engelhardt@108 238 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
engelhardt@108 239 {
Nina@109 240 hw_event.config = 0x0000000000000000; //cycles
engelhardt@108 241 _VMSMasterEnv->cycles_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 242 0,//pid_t pid,
Nina@167 243 coreIdx,//int cpu,
engelhardt@108 244 -1,//int group_fd,
engelhardt@108 245 0//unsigned long flags
engelhardt@108 246 );
Nina@109 247 if (_VMSMasterEnv->cycles_counter_fd[coreIdx]<0){
Nina@109 248 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 249 perror("Failed to open cycles counter");
Nina@109 250 }
Nina@109 251 hw_event.config = 0x0000000000000001; //instrs
engelhardt@108 252 _VMSMasterEnv->instrs_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 253 0,//pid_t pid,
Nina@167 254 coreIdx,//int cpu,
engelhardt@108 255 -1,//int group_fd,
engelhardt@108 256 0//unsigned long flags
engelhardt@108 257 );
Nina@109 258 if (_VMSMasterEnv->instrs_counter_fd[coreIdx]<0){
Nina@109 259 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 260 perror("Failed to open instrs counter");
Nina@109 261 }
engelhardt@108 262 }
engelhardt@108 263 prctl(PR_TASK_PERF_EVENTS_ENABLE);
Nina@109 264 uint64 tmpc,tmpi;
Nina@109 265 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@130 266 printf("Start: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
engelhardt@108 267 #endif
engelhardt@108 268
Me@50 269 //========================================================================
Me@38 270
Me@0 271 }
Me@0 272
Me@31 273 SchedSlot **
Me@31 274 create_sched_slots()
Me@31 275 { SchedSlot **schedSlots;
Me@0 276 int i;
Me@0 277
Me@53 278 schedSlots = VMS__malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
Me@8 279
Me@1 280 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@0 281 {
Me@53 282 schedSlots[i] = VMS__malloc( sizeof(SchedSlot) );
Me@8 283
Me@1 284 //Set state to mean "handling requests done, slot needs filling"
Me@8 285 schedSlots[i]->workIsDone = FALSE;
Me@8 286 schedSlots[i]->needsProcrAssigned = TRUE;
Me@0 287 }
Me@31 288 return schedSlots;
Me@31 289 }
Me@31 290
Me@31 291
Me@31 292 void
Me@31 293 freeSchedSlots( SchedSlot **schedSlots )
Me@31 294 { int i;
Me@31 295 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@31 296 {
Me@53 297 VMS__free( schedSlots[i] );
Me@31 298 }
Me@53 299 VMS__free( schedSlots );
Me@0 300 }
Me@0 301
Me@8 302
Me@28 303 void
Me@28 304 create_the_coreLoop_OS_threads()
Me@28 305 {
Me@28 306 //========================================================================
Me@28 307 // Create the Threads
Me@28 308 int coreIdx, retCode;
Me@28 309
Me@28 310 //Need the threads to be created suspended, and wait for a signal
Me@28 311 // before proceeding -- gives time after creating to initialize other
Me@28 312 // stuff before the coreLoops set off.
Me@28 313 _VMSMasterEnv->setupComplete = 0;
Me@28 314
Me@28 315 //Make the threads that animate the core loops
Me@28 316 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@53 317 { coreLoopThdParams[coreIdx] = VMS__malloc( sizeof(ThdParams) );
Me@28 318 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
Me@28 319
Me@28 320 retCode =
Me@28 321 pthread_create( &(coreLoopThdHandles[coreIdx]),
Me@28 322 thdAttrs,
Me@28 323 &coreLoop,
Me@28 324 (void *)(coreLoopThdParams[coreIdx]) );
Me@50 325 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
Me@28 326 }
Me@28 327 }
Me@28 328
Me@0 329 /*Semantic layer calls this when it want the system to start running..
Me@0 330 *
Me@24 331 *This starts the core loops running then waits for them to exit.
Me@0 332 */
Me@12 333 void
Me@24 334 VMS__start_the_work_then_wait_until_done()
Me@12 335 { int coreIdx;
Me@24 336 //Start the core loops running
Me@25 337
Me@25 338 //tell the core loop threads that setup is complete
Me@25 339 //get lock, to lock out any threads still starting up -- they'll see
Me@25 340 // that setupComplete is true before entering while loop, and so never
Me@25 341 // wait on the condition
Me@26 342 pthread_mutex_lock( &suspendLock );
Me@25 343 _VMSMasterEnv->setupComplete = 1;
Me@26 344 pthread_mutex_unlock( &suspendLock );
Me@26 345 pthread_cond_broadcast( &suspend_cond );
Me@25 346
Me@25 347
Me@24 348 //wait for all to complete
Me@8 349 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@8 350 {
Me@25 351 pthread_join( coreLoopThdHandles[coreIdx], NULL );
Me@24 352 }
Me@25 353
Me@24 354 //NOTE: do not clean up VMS env here -- semantic layer has to have
Me@24 355 // a chance to clean up its environment first, then do a call to free
Me@24 356 // the Master env and rest of VMS locations
Me@8 357 }
Me@0 358
msach@70 359 #ifdef SEQUENTIAL
Me@28 360 /*Only difference between version with an OS thread pinned to each core and
Me@28 361 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
Me@28 362 */
Me@28 363 void
Me@28 364 VMS__start_the_work_then_wait_until_done_Seq()
Me@28 365 {
Me@28 366 //Instead of un-suspending threads, just call the one and only
Me@28 367 // core loop (sequential version), in the main thread.
Me@28 368 coreLoop_Seq( NULL );
msach@75 369 flushRegisters();
Me@28 370
Me@28 371 }
msach@70 372 #endif
Me@28 373
Me@50 374 inline VirtProcr *
Me@50 375 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 376 { VirtProcr *newPr;
msach@76 377 void *stackLocs;
Me@50 378
Me@50 379 newPr = VMS__malloc( sizeof(VirtProcr) );
Me@50 380 stackLocs = VMS__malloc( VIRT_PROCR_STACK_SIZE );
Me@50 381 if( stackLocs == 0 )
Me@50 382 { perror("VMS__malloc stack"); exit(1); }
Me@50 383
msach@69 384 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 385 }
Me@50 386
Me@50 387 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 388 * be called from main thread or other thread -- never from code animated by
Me@50 389 * a VMS virtual processor.
Me@50 390 */
Me@50 391 inline VirtProcr *
Me@50 392 VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 393 { VirtProcr *newPr;
Me@50 394 char *stackLocs;
Me@50 395
Me@50 396 newPr = malloc( sizeof(VirtProcr) );
Me@50 397 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
Me@50 398 if( stackLocs == 0 )
Me@50 399 { perror("malloc stack"); exit(1); }
Me@50 400
msach@69 401 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 402 }
Me@50 403
Me@8 404
Me@64 405 /*Anticipating multi-tasking
Me@64 406 */
Me@64 407 void *
Me@64 408 VMS__give_sem_env_for( VirtProcr *animPr )
Me@64 409 {
Me@64 410 return _VMSMasterEnv->semanticEnv;
Me@64 411 }
Me@64 412 //===========================================================================
Me@26 413 /*there is a label inside this function -- save the addr of this label in
Me@0 414 * the callingPr struc, as the pick-up point from which to start the next
Me@0 415 * work-unit for that procr. If turns out have to save registers, then
Me@0 416 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
Me@0 417 * "done with work-unit" label. The procr struc is in the request in the
Me@0 418 * slave that animated the just-ended work-unit, so all the state is saved
Me@0 419 * there, and will get passed along, inside the request handler, to the
Me@0 420 * next work-unit for that procr.
Me@0 421 */
Me@8 422 void
Me@38 423 VMS__suspend_procr( VirtProcr *animatingPr )
Me@55 424 {
Me@30 425
Me@30 426 //The request to master will cause this suspended virt procr to get
Me@30 427 // scheduled again at some future point -- to resume, core loop jumps
Me@30 428 // to the resume point (below), which causes restore of saved regs and
Me@30 429 // "return" from this call.
msach@71 430 //animatingPr->nextInstrPt = &&ResumePt;
Me@30 431
Me@30 432 //return ownership of the virt procr and sched slot to Master virt pr
Me@38 433 animatingPr->schedSlot->workIsDone = TRUE;
Me@1 434
Me@41 435 //=========================== Measurement stuff ========================
Me@38 436 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 437 //record time stamp: compare to time-stamp recorded below
Me@38 438 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
Me@38 439 #endif
nengel@186 440 #ifdef MEAS__PERF_COUNTERS
nengel@186 441 //start work
nengel@186 442 uint64 cycles,instrs;
nengel@186 443 saveCyclesAndInstrs(animatingPr->coreAnimatedBy,cycles, instrs);
nengel@193 444 (*(_VMSMasterEnv->counterHandler))(HwResponderInvocation_start,animatingPr->procrID,animatingPr->numTimesScheduled,animatingPr,cycles,instrs);
nengel@186 445 #endif
Me@41 446 //=======================================================================
Me@30 447
msach@71 448 switchToCoreLoop(animatingPr);
msach@71 449 flushRegisters();
Me@55 450
Me@55 451 //=======================================================================
Me@30 452
Me@38 453 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 454 //NOTE: only take low part of count -- do sanity check when take diff
Me@38 455 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
Me@38 456 #endif
Me@30 457
Me@0 458 return;
Me@0 459 }
Me@0 460
Me@22 461
Me@22 462
Me@50 463 /*For this implementation of VMS, it may not make much sense to have the
Me@50 464 * system of requests for creating a new processor done this way.. but over
Me@50 465 * the scope of single-master, multi-master, mult-tasking, OS-implementing,
Me@50 466 * distributed-memory, and so on, this gives VMS implementation a chance to
Me@50 467 * do stuff before suspend, in the AppVP, and in the Master before the plugin
Me@50 468 * is called, as well as in the lang-lib before this is called, and in the
Me@50 469 * plugin. So, this gives both VMS and language implementations a chance to
Me@50 470 * intercept at various points and do order-dependent stuff.
Me@50 471 *Having a standard VMSNewPrReqData struc allows the language to create and
Me@50 472 * free the struc, while VMS knows how to get the newPr if it wants it, and
Me@50 473 * it lets the lang have lang-specific data related to creation transported
Me@50 474 * to the plugin.
Me@50 475 */
Me@50 476 void
Me@50 477 VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr )
Me@50 478 { VMSReqst req;
Me@50 479
Me@50 480 req.reqType = createReq;
Me@50 481 req.semReqData = semReqData;
Me@50 482 req.nextReqst = reqstingPr->requests;
Me@50 483 reqstingPr->requests = &req;
Me@50 484
Me@50 485 VMS__suspend_procr( reqstingPr );
Me@50 486 }
Me@50 487
Me@22 488
Me@38 489 /*
Me@22 490 *This adds a request to dissipate, then suspends the processor so that the
Me@22 491 * request handler will receive the request. The request handler is what
Me@22 492 * does the work of freeing memory and removing the processor from the
Me@22 493 * semantic environment's data structures.
Me@22 494 *The request handler also is what figures out when to shutdown the VMS
Me@22 495 * system -- which causes all the core loop threads to die, and returns from
Me@22 496 * the call that started up VMS to perform the work.
Me@22 497 *
Me@22 498 *This form is a bit misleading to understand if one is trying to figure out
Me@22 499 * how VMS works -- it looks like a normal function call, but inside it
Me@22 500 * sends a request to the request handler and suspends the processor, which
Me@22 501 * jumps out of the VMS__dissipate_procr function, and out of all nestings
Me@22 502 * above it, transferring the work of dissipating to the request handler,
Me@22 503 * which then does the actual work -- causing the processor that animated
Me@22 504 * the call of this function to disappear and the "hanging" state of this
Me@22 505 * function to just poof into thin air -- the virtual processor's trace
Me@22 506 * never returns from this call, but instead the virtual processor's trace
Me@22 507 * gets suspended in this call and all the virt processor's state disap-
Me@22 508 * pears -- making that suspend the last thing in the virt procr's trace.
Me@8 509 */
Me@8 510 void
Me@53 511 VMS__send_dissipate_req( VirtProcr *procrToDissipate )
Me@50 512 { VMSReqst req;
Me@22 513
Me@50 514 req.reqType = dissipate;
Me@50 515 req.nextReqst = procrToDissipate->requests;
Me@50 516 procrToDissipate->requests = &req;
Me@50 517
Me@22 518 VMS__suspend_procr( procrToDissipate );
Me@50 519 }
Me@50 520
Me@50 521
Me@50 522 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 523 * be called from main thread or other thread -- never from code animated by
Me@50 524 * a VMS virtual processor.
Me@50 525 *
Me@50 526 *Use this version to dissipate VPs created outside the VMS system.
Me@50 527 */
Me@50 528 void
Me@50 529 VMS_ext__dissipate_procr( VirtProcr *procrToDissipate )
Me@50 530 {
Me@50 531 //NOTE: initialData was given to the processor, so should either have
Me@50 532 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@50 533 //So, all that's left to free here is the stack and the VirtProcr struc
Me@50 534 // itself
Me@50 535 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 536 // general that creating processor will outlive ones it creates.
Me@50 537 free( procrToDissipate->startOfStack );
Me@50 538 free( procrToDissipate );
Me@50 539 }
Me@50 540
Me@22 541
Me@22 542
Me@53 543 /*This call's name indicates that request is malloc'd -- so req handler
Me@53 544 * has to free any extra requests tacked on before a send, using this.
Me@53 545 *
Me@53 546 * This inserts the semantic-layer's request data into standard VMS carrier
Me@53 547 * request data-struct that is mallocd. The sem request doesn't need to
Me@53 548 * be malloc'd if this is called inside the same call chain before the
Me@53 549 * send of the last request is called.
Me@53 550 *
Me@53 551 *The request handler has to call VMS__free_VMSReq for any of these
Me@22 552 */
Me@22 553 inline void
Me@53 554 VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData,
Me@53 555 VirtProcr *callingPr )
Me@53 556 { VMSReqst *req;
Me@22 557
Me@53 558 req = VMS__malloc( sizeof(VMSReqst) );
Me@53 559 req->reqType = semantic;
Me@53 560 req->semReqData = semReqData;
Me@53 561 req->nextReqst = callingPr->requests;
Me@53 562 callingPr->requests = req;
Me@22 563 }
Me@22 564
Me@50 565 /*This inserts the semantic-layer's request data into standard VMS carrier
Me@50 566 * request data-struct is allocated on stack of this call & ptr to it sent
Me@50 567 * to plugin
Me@50 568 *Then it does suspend, to cause request to be sent.
Me@50 569 */
Me@50 570 inline void
Me@50 571 VMS__send_sem_request( void *semReqData, VirtProcr *callingPr )
Me@50 572 { VMSReqst req;
Me@22 573
Me@50 574 req.reqType = semantic;
Me@50 575 req.semReqData = semReqData;
Me@50 576 req.nextReqst = callingPr->requests;
Me@50 577 callingPr->requests = &req;
Me@50 578
Me@50 579 VMS__suspend_procr( callingPr );
Me@50 580 }
Me@50 581
Me@50 582
Me@50 583 inline void
Me@50 584 VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr )
Me@50 585 { VMSReqst req;
Me@50 586
Me@50 587 req.reqType = VMSSemantic;
Me@50 588 req.semReqData = semReqData;
Me@50 589 req.nextReqst = callingPr->requests; //gab any other preceeding
Me@50 590 callingPr->requests = &req;
Me@50 591
Me@50 592 VMS__suspend_procr( callingPr );
Me@50 593 }
Me@50 594
Me@50 595
Me@50 596 /*
Me@38 597 */
Me@24 598 VMSReqst *
Me@50 599 VMS__take_next_request_out_of( VirtProcr *procrWithReq )
Me@31 600 { VMSReqst *req;
Me@31 601
Me@31 602 req = procrWithReq->requests;
Me@38 603 if( req == NULL ) return NULL;
Me@31 604
Me@31 605 procrWithReq->requests = procrWithReq->requests->nextReqst;
Me@50 606 return req;
Me@24 607 }
Me@22 608
Me@24 609
Me@24 610 inline void *
Me@24 611 VMS__take_sem_reqst_from( VMSReqst *req )
Me@24 612 {
Me@24 613 return req->semReqData;
Me@24 614 }
Me@24 615
Me@24 616
Me@24 617
Me@50 618 /* This is for OS requests and VMS infrastructure requests, such as to create
Me@50 619 * a probe -- a probe is inside the heart of VMS-core, it's not part of any
Me@50 620 * language -- but it's also a semantic thing that's triggered from and used
Me@50 621 * in the application.. so it crosses abstractions.. so, need some special
Me@50 622 * pattern here for handling such requests.
Me@52 623 * Doing this just like it were a second language sharing VMS-core.
Me@52 624 *
Me@50 625 * This is called from the language's request handler when it sees a request
Me@50 626 * of type VMSSemReq
Me@52 627 *
Me@52 628 * TODO: Later change this, to give probes their own separate plugin & have
Me@52 629 * VMS-core steer the request to appropriate plugin
Me@52 630 * Do the same for OS calls -- look later at it..
Me@50 631 */
Me@50 632 void inline
Me@50 633 VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv,
Me@50 634 ResumePrFnPtr resumePrFnPtr )
Me@50 635 { VMSSemReq *semReq;
Me@50 636 IntervalProbe *newProbe;
Me@24 637
Me@50 638 semReq = req->semReqData;
Me@24 639
Me@50 640 newProbe = VMS__malloc( sizeof(IntervalProbe) );
Me@65 641 newProbe->nameStr = VMS__strDup( semReq->nameStr );
Me@50 642 newProbe->hist = NULL;
Me@50 643 newProbe->schedChoiceWasRecorded = FALSE;
Me@53 644
Me@53 645 //This runs in masterVP, so no race-condition worries
Me@50 646 newProbe->probeID =
Me@50 647 addToDynArray( newProbe, _VMSMasterEnv->dynIntervalProbesInfo );
Me@50 648
Me@53 649 requestingPr->dataRetFromReq = newProbe;
Me@50 650
Me@50 651 (*resumePrFnPtr)( requestingPr, semEnv );
Me@22 652 }
Me@22 653
Me@22 654
Me@22 655
Me@24 656 /*This must be called by the request handler plugin -- it cannot be called
Me@24 657 * from the semantic library "dissipate processor" function -- instead, the
Me@50 658 * semantic layer has to generate a request, and the plug-in calls this
Me@24 659 * function.
Me@24 660 *The reason is that this frees the virtual processor's stack -- which is
Me@24 661 * still in use inside semantic library calls!
Me@24 662 *
Me@24 663 *This frees or recycles all the state owned by and comprising the VMS
Me@24 664 * portion of the animating virtual procr. The request handler must first
Me@24 665 * free any semantic data created for the processor that didn't use the
Me@24 666 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
Me@24 667 * system to disown any state that did use VMS_malloc, and then frees the
Me@24 668 * statck and the processor-struct itself.
Me@24 669 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
Me@24 670 * state, then that state gets freed (or sent to recycling) as a side-effect
Me@24 671 * of dis-owning it.
Me@24 672 */
Me@24 673 void
Me@53 674 VMS__dissipate_procr( VirtProcr *animatingPr )
Me@24 675 {
Me@24 676 //dis-own all locations owned by this processor, causing to be freed
Me@24 677 // any locations that it is (was) sole owner of
Me@29 678 //TODO: implement VMS__malloc system, including "give up ownership"
Me@24 679
Me@24 680
Me@24 681 //NOTE: initialData was given to the processor, so should either have
Me@24 682 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@24 683 //So, all that's left to free here is the stack and the VirtProcr struc
Me@24 684 // itself
Me@50 685 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 686 // general that creating processor will outlive ones it creates.
Me@50 687 VMS__free( animatingPr->startOfStack );
Me@50 688 VMS__free( animatingPr );
Me@24 689 }
Me@24 690
Me@24 691
Me@53 692 //TODO: look at architecting cleanest separation between request handler
Me@29 693 // and master loop, for dissipate, create, shutdown, and other non-semantic
Me@29 694 // requests. Issue is chain: one removes requests from AppVP, one dispatches
Me@29 695 // on type of request, and one handles each type.. but some types require
Me@29 696 // action from both request handler and master loop -- maybe just give the
Me@29 697 // request handler calls like: VMS__handle_X_request_type
Me@24 698
Me@29 699
Me@29 700 /*This is called by the semantic layer's request handler when it decides its
Me@29 701 * time to shut down the VMS system. Calling this causes the core loop OS
Me@29 702 * threads to exit, which unblocks the entry-point function that started up
Me@29 703 * VMS, and allows it to grab the result and return to the original single-
Me@29 704 * threaded application.
Me@22 705 *
Me@29 706 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
Me@29 707 * and-wait function has to free a bunch of stuff after it detects the
Me@29 708 * threads have all died: the masterEnv, the thread-related locations,
Me@29 709 * masterVP any AppVPs that might still be allocated and sitting in the
Me@29 710 * semantic environment, or have been orphaned in the _VMSWorkQ.
Me@29 711 *
Me@53 712 *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the
Me@29 713 * locations it needs, and give ownership to masterVP. Then, they will be
Me@53 714 * automatically freed.
Me@22 715 *
Me@29 716 *In here,create one core-loop shut-down processor for each core loop and put
Me@31 717 * them all directly into the readyToAnimateQ.
Me@29 718 *Note, this function can ONLY be called after the semantic environment no
Me@29 719 * longer cares if AppVPs get animated after the point this is called. In
Me@29 720 * other words, this can be used as an abort, or else it should only be
Me@29 721 * called when all AppVPs have finished dissipate requests -- only at that
Me@29 722 * point is it sure that all results have completed.
Me@22 723 */
Me@22 724 void
Me@53 725 VMS__shutdown()
Me@8 726 { int coreIdx;
Me@14 727 VirtProcr *shutDownPr;
Me@22 728
Me@29 729 //create the shutdown processors, one for each core loop -- put them
Me@31 730 // directly into the Q -- each core will die when gets one
Me@8 731 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@50 732 { //Note, this is running in the master
Me@29 733 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
nengel@193 734 shutDownPr->isShutdownVP = TRUE;
Me@55 735 writeVMSQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
Me@8 736 }
Nina@109 737 #ifdef MEAS__PERF_COUNTERS
Nina@109 738 uint64 tmpc,tmpi;
Nina@109 739 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@130 740 printf("End: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
Nina@109 741 prctl(PR_TASK_PERF_EVENTS_DISABLE);
Nina@109 742 /*
Nina@109 743 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ){
Nina@109 744 close(_VMSMasterEnv->cycles_counter_fd[coreIdx]);
Nina@109 745 close(_VMSMasterEnv->instrs_counter_fd[coreIdx]);
Nina@109 746 }
Nina@109 747 */
Nina@109 748 #endif
Me@12 749 }
Me@12 750
Me@12 751
Me@29 752 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
Me@29 753 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
Me@29 754 *This function has the sole purpose of setting the stack and framePtr
Me@29 755 * to the coreLoop's stack and framePtr.. it does that then jumps to the
Me@29 756 * core loop's shutdown point -- might be able to just call Pthread_exit
Me@30 757 * from here, but am going back to the pthread's stack and setting everything
Me@29 758 * up just as if it never jumped out, before calling pthread_exit.
Me@29 759 *The end-point of core loop will free the stack and so forth of the
Me@29 760 * processor that animates this function, (this fn is transfering the
Me@29 761 * animator of the AppVP that is in turn animating this function over
Me@29 762 * to core loop function -- note that this slices out a level of virtual
Me@29 763 * processors).
Me@29 764 */
Me@29 765 void
Me@29 766 endOSThreadFn( void *initData, VirtProcr *animatingPr )
msach@71 767 {
msach@75 768 #ifdef SEQUENTIAL
msach@75 769 asmTerminateCoreLoopSeq(animatingPr);
msach@75 770 #else
msach@71 771 asmTerminateCoreLoop(animatingPr);
msach@75 772 #endif
Me@30 773 }
Me@29 774
Me@29 775
Me@53 776 /*This is called from the startup & shutdown
Me@24 777 */
Me@24 778 void
Me@53 779 VMS__cleanup_at_end_of_shutdown()
Me@31 780 {
msach@78 781 //unused
msach@78 782 //VMSQueueStruc **readyToAnimateQs;
msach@78 783 //int coreIdx;
msach@78 784 //VirtProcr **masterVPs;
msach@78 785 //SchedSlot ***allSchedSlots; //ptr to array of ptrs
Me@31 786
Me@65 787 //Before getting rid of everything, print out any measurements made
Nina@109 788 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&printHist );
Nina@109 789 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&saveHistToFile);
nengel@182 790 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, &freeHist );
Nina@167 791
nengel@177 792
Me@65 793 #ifdef MEAS__TIME_PLUGIN
Me@68 794 printHist( _VMSMasterEnv->reqHdlrLowTimeHist );
msach@84 795 saveHistToFile( _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 796 printHist( _VMSMasterEnv->reqHdlrHighTimeHist );
msach@79 797 saveHistToFile( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 798 freeHistExt( _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 799 freeHistExt( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@65 800 #endif
Me@65 801 #ifdef MEAS__TIME_MALLOC
Me@65 802 printHist( _VMSMasterEnv->mallocTimeHist );
msach@79 803 saveHistToFile( _VMSMasterEnv->mallocTimeHist );
Me@65 804 printHist( _VMSMasterEnv->freeTimeHist );
msach@79 805 saveHistToFile( _VMSMasterEnv->freeTimeHist );
Me@65 806 freeHistExt( _VMSMasterEnv->mallocTimeHist );
Me@65 807 freeHistExt( _VMSMasterEnv->freeTimeHist );
Me@65 808 #endif
Me@65 809 #ifdef MEAS__TIME_MASTER_LOCK
Me@65 810 printHist( _VMSMasterEnv->masterLockLowTimeHist );
Me@65 811 printHist( _VMSMasterEnv->masterLockHighTimeHist );
Me@65 812 #endif
Me@65 813 #ifdef MEAS__TIME_MASTER
Me@65 814 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 815 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 816 {
Me@65 817 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 818 //master VPs were created external to VMS, so use external free
Me@65 819 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 820
Me@65 821 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 822 }
Me@65 823 #endif
Me@65 824 #ifdef MEAS__TIME_STAMP_SUSP
Me@65 825 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 826 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 827 {
Me@65 828 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 829 //master VPs were created external to VMS, so use external free
Me@65 830 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 831
Me@65 832 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 833 }
Me@65 834 #endif
Me@65 835
Me@53 836 //All the environment data has been allocated with VMS__malloc, so just
Me@53 837 // free its internal big-chunk and all inside it disappear.
Me@53 838 /*
Me@31 839 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
Me@31 840 masterVPs = _VMSMasterEnv->masterVPs;
Me@31 841 allSchedSlots = _VMSMasterEnv->allSchedSlots;
Me@31 842
Me@31 843 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@24 844 {
Me@55 845 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@50 846 //master VPs were created external to VMS, so use external free
Me@53 847 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@31 848
Me@31 849 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@24 850 }
Me@31 851
Me@53 852 VMS__free( _VMSMasterEnv->readyToAnimateQs );
Me@53 853 VMS__free( _VMSMasterEnv->masterVPs );
Me@53 854 VMS__free( _VMSMasterEnv->allSchedSlots );
Me@50 855
Me@50 856 //============================= MEASUREMENT STUFF ========================
Me@50 857 #ifdef STATS__TURN_ON_PROBES
Me@53 858 freeDynArrayDeep( _VMSMasterEnv->dynIntervalProbesInfo, &VMS__free_probe);
Me@50 859 #endif
Me@50 860 //========================================================================
Me@53 861 */
Me@53 862 //These are the only two that use system free
Me@53 863 VMS_ext__free_free_list( _VMSMasterEnv->freeListHead );
Me@53 864 free( (void *)_VMSMasterEnv );
Me@24 865 }
Me@24 866
Me@54 867
Me@54 868 //================================
Me@54 869
Me@54 870
Me@54 871 /*Later, improve this -- for now, just exits the application after printing
Me@54 872 * the error message.
Me@54 873 */
Me@54 874 void
Me@54 875 VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData )
Me@54 876 {
msach@69 877 printf("%s",msgStr);
Me@54 878 fflush(stdin);
Me@54 879 exit(1);
Me@54 880 }
Me@54 881