annotate VMS.c @ 130:5475f90c248a

fix outputs (dependency file creation, counter length)
author Nina Engelhardt
date Wed, 07 Sep 2011 13:26:30 +0200
parents ce02441b77cf
children 395f58384a5c
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@84 121 _VMSMasterEnv->mallocTimeHist = makeFixedBinHistExt( 100, 0, 100,
msach@79 122 "malloc_time_hist");
msach@84 123 _VMSMasterEnv->freeTimeHist = makeFixedBinHistExt( 80, 0, 100,
msach@79 124 "free_time_hist");
Me@65 125 #endif
Me@68 126 #ifdef MEAS__TIME_PLUGIN
msach@84 127 _VMSMasterEnv->reqHdlrLowTimeHist = makeFixedBinHistExt( 1000, 0, 100,
msach@79 128 "plugin_low_time_hist");
msach@84 129 _VMSMasterEnv->reqHdlrHighTimeHist = makeFixedBinHistExt( 1000, 0, 100,
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;
Me@31 152 allSchedSlots[ coreIdx ] = create_sched_slots(); //makes for one core
Me@53 153 _VMSMasterEnv->numMasterInARow[ coreIdx ] = 0;
Me@55 154 _VMSMasterEnv->workStealingGates[ coreIdx ] = NULL;
Me@31 155 }
Me@31 156 _VMSMasterEnv->readyToAnimateQs = readyToAnimateQs;
Me@31 157 _VMSMasterEnv->masterVPs = masterVPs;
Me@50 158 _VMSMasterEnv->masterLock = UNLOCKED;
Me@31 159 _VMSMasterEnv->allSchedSlots = allSchedSlots;
Me@55 160 _VMSMasterEnv->workStealingLock = UNLOCKED;
Me@28 161
Me@12 162
Me@31 163 //Aug 19, 2010: no longer need to place initial masterVP into queue
Me@31 164 // because coreLoop now controls -- animates its masterVP when no work
Me@31 165
Me@30 166
Me@50 167 //============================= MEASUREMENT STUFF ========================
Me@50 168 #ifdef STATS__TURN_ON_PROBES
Me@50 169 _VMSMasterEnv->dynIntervalProbesInfo =
msach@69 170 makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->intervalProbes), 200);
Me@30 171
Me@53 172 _VMSMasterEnv->probeNameHashTbl = makeHashTable( 1000, &VMS__free );
Me@53 173
Me@53 174 //put creation time directly into master env, for fast retrieval
Me@50 175 struct timeval timeStamp;
Me@50 176 gettimeofday( &(timeStamp), NULL);
Me@50 177 _VMSMasterEnv->createPtInSecs =
Me@50 178 timeStamp.tv_sec +(timeStamp.tv_usec/1000000.0);
Me@50 179 #endif
Me@65 180 #ifdef MEAS__TIME_MASTER_LOCK
Me@65 181 _VMSMasterEnv->masterLockLowTimeHist = makeFixedBinHist( 50, 0, 2,
Me@65 182 "master lock low time hist");
Me@68 183 _VMSMasterEnv->masterLockHighTimeHist = makeFixedBinHist( 50, 0, 100,
Me@65 184 "master lock high time hist");
Me@65 185 #endif
Me@68 186
msach@76 187 MakeTheMeasHists();
engelhardt@108 188
Nina@129 189 #ifdef DETECT_DEPENDENCIES
Nina@129 190 _VMSMasterEnv->dependencies = VMS__malloc(10*sizeof(void*));
Nina@129 191 _VMSMasterEnv->dependenciesInfo = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->dependencies),10);
Nina@129 192 #endif
Nina@129 193
Nina@109 194 #ifdef MEAS__PERF_COUNTERS
Nina@110 195 //printf("Creating HW counters...");
Nina@110 196 FILE* output;
Nina@110 197 int n;
Nina@110 198 char filename[255];
Nina@110 199 for(n=0;n<255;n++)
Nina@110 200 {
Nina@113 201 sprintf(filename, "./counters/Counters.%d.csv",n);
Nina@110 202 output = fopen(filename,"r");
Nina@110 203 if(output)
Nina@110 204 {
Nina@110 205 fclose(output);
Nina@110 206 }else{
Nina@110 207 break;
Nina@110 208 }
Nina@110 209 }
Nina@110 210 printf("Saving Counter measurements to File: %s ...\n", filename);
Nina@110 211 output = fopen(filename,"w+");
Nina@110 212 _VMSMasterEnv->counteroutput = output;
Nina@110 213
engelhardt@108 214 struct perf_event_attr hw_event;
Nina@109 215 memset(&hw_event,0,sizeof(hw_event));
engelhardt@108 216 hw_event.type = PERF_TYPE_HARDWARE;
Nina@109 217 hw_event.size = sizeof(hw_event);
engelhardt@108 218 hw_event.disabled = 1;
Nina@109 219 hw_event.freq = 0;
engelhardt@108 220 hw_event.inherit = 1; /* children inherit it */
engelhardt@108 221 hw_event.pinned = 1; /* must always be on PMU */
engelhardt@108 222 hw_event.exclusive = 0; /* only group on PMU */
engelhardt@108 223 hw_event.exclude_user = 0; /* don't count user */
engelhardt@108 224 hw_event.exclude_kernel = 1; /* ditto kernel */
engelhardt@108 225 hw_event.exclude_hv = 1; /* ditto hypervisor */
engelhardt@108 226 hw_event.exclude_idle = 0; /* don't count when idle */
engelhardt@108 227 hw_event.mmap = 0; /* include mmap data */
engelhardt@108 228 hw_event.comm = 0; /* include comm data */
engelhardt@108 229
Nina@109 230
engelhardt@108 231 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
engelhardt@108 232 {
Nina@109 233 hw_event.config = 0x0000000000000000; //cycles
engelhardt@108 234 _VMSMasterEnv->cycles_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 235 0,//pid_t pid,
Nina@109 236 -1,//int cpu,
engelhardt@108 237 -1,//int group_fd,
engelhardt@108 238 0//unsigned long flags
engelhardt@108 239 );
Nina@109 240 if (_VMSMasterEnv->cycles_counter_fd[coreIdx]<0){
Nina@109 241 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 242 perror("Failed to open cycles counter");
Nina@109 243 }
Nina@109 244 hw_event.config = 0x0000000000000001; //instrs
engelhardt@108 245 _VMSMasterEnv->instrs_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 246 0,//pid_t pid,
Nina@109 247 -1,//int cpu,
engelhardt@108 248 -1,//int group_fd,
engelhardt@108 249 0//unsigned long flags
engelhardt@108 250 );
Nina@109 251 if (_VMSMasterEnv->instrs_counter_fd[coreIdx]<0){
Nina@109 252 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 253 perror("Failed to open instrs counter");
Nina@109 254 }
engelhardt@108 255 }
engelhardt@108 256 prctl(PR_TASK_PERF_EVENTS_ENABLE);
Nina@109 257 uint64 tmpc,tmpi;
Nina@109 258 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@130 259 printf("Start: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
engelhardt@108 260 #endif
engelhardt@108 261
Me@50 262 //========================================================================
Me@38 263
Me@0 264 }
Me@0 265
Me@31 266 SchedSlot **
Me@31 267 create_sched_slots()
Me@31 268 { SchedSlot **schedSlots;
Me@0 269 int i;
Me@0 270
Me@53 271 schedSlots = VMS__malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
Me@8 272
Me@1 273 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@0 274 {
Me@53 275 schedSlots[i] = VMS__malloc( sizeof(SchedSlot) );
Me@8 276
Me@1 277 //Set state to mean "handling requests done, slot needs filling"
Me@8 278 schedSlots[i]->workIsDone = FALSE;
Me@8 279 schedSlots[i]->needsProcrAssigned = TRUE;
Me@0 280 }
Me@31 281 return schedSlots;
Me@31 282 }
Me@31 283
Me@31 284
Me@31 285 void
Me@31 286 freeSchedSlots( SchedSlot **schedSlots )
Me@31 287 { int i;
Me@31 288 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@31 289 {
Me@53 290 VMS__free( schedSlots[i] );
Me@31 291 }
Me@53 292 VMS__free( schedSlots );
Me@0 293 }
Me@0 294
Me@8 295
Me@28 296 void
Me@28 297 create_the_coreLoop_OS_threads()
Me@28 298 {
Me@28 299 //========================================================================
Me@28 300 // Create the Threads
Me@28 301 int coreIdx, retCode;
Me@28 302
Me@28 303 //Need the threads to be created suspended, and wait for a signal
Me@28 304 // before proceeding -- gives time after creating to initialize other
Me@28 305 // stuff before the coreLoops set off.
Me@28 306 _VMSMasterEnv->setupComplete = 0;
Me@28 307
Me@28 308 //Make the threads that animate the core loops
Me@28 309 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@53 310 { coreLoopThdParams[coreIdx] = VMS__malloc( sizeof(ThdParams) );
Me@28 311 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
Me@28 312
Me@28 313 retCode =
Me@28 314 pthread_create( &(coreLoopThdHandles[coreIdx]),
Me@28 315 thdAttrs,
Me@28 316 &coreLoop,
Me@28 317 (void *)(coreLoopThdParams[coreIdx]) );
Me@50 318 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
Me@28 319 }
Me@28 320 }
Me@28 321
Me@0 322 /*Semantic layer calls this when it want the system to start running..
Me@0 323 *
Me@24 324 *This starts the core loops running then waits for them to exit.
Me@0 325 */
Me@12 326 void
Me@24 327 VMS__start_the_work_then_wait_until_done()
Me@12 328 { int coreIdx;
Me@24 329 //Start the core loops running
Me@25 330
Me@25 331 //tell the core loop threads that setup is complete
Me@25 332 //get lock, to lock out any threads still starting up -- they'll see
Me@25 333 // that setupComplete is true before entering while loop, and so never
Me@25 334 // wait on the condition
Me@26 335 pthread_mutex_lock( &suspendLock );
Me@25 336 _VMSMasterEnv->setupComplete = 1;
Me@26 337 pthread_mutex_unlock( &suspendLock );
Me@26 338 pthread_cond_broadcast( &suspend_cond );
Me@25 339
Me@25 340
Me@24 341 //wait for all to complete
Me@8 342 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@8 343 {
Me@25 344 pthread_join( coreLoopThdHandles[coreIdx], NULL );
Me@24 345 }
Me@25 346
Me@24 347 //NOTE: do not clean up VMS env here -- semantic layer has to have
Me@24 348 // a chance to clean up its environment first, then do a call to free
Me@24 349 // the Master env and rest of VMS locations
Me@8 350 }
Me@0 351
msach@70 352 #ifdef SEQUENTIAL
Me@28 353 /*Only difference between version with an OS thread pinned to each core and
Me@28 354 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
Me@28 355 */
Me@28 356 void
Me@28 357 VMS__start_the_work_then_wait_until_done_Seq()
Me@28 358 {
Me@28 359 //Instead of un-suspending threads, just call the one and only
Me@28 360 // core loop (sequential version), in the main thread.
Me@28 361 coreLoop_Seq( NULL );
msach@75 362 flushRegisters();
Me@28 363
Me@28 364 }
msach@70 365 #endif
Me@28 366
Me@50 367 inline VirtProcr *
Me@50 368 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 369 { VirtProcr *newPr;
msach@76 370 void *stackLocs;
Me@50 371
Me@50 372 newPr = VMS__malloc( sizeof(VirtProcr) );
Me@50 373 stackLocs = VMS__malloc( VIRT_PROCR_STACK_SIZE );
Me@50 374 if( stackLocs == 0 )
Me@50 375 { perror("VMS__malloc stack"); exit(1); }
Me@50 376
msach@69 377 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 378 }
Me@50 379
Me@50 380 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 381 * be called from main thread or other thread -- never from code animated by
Me@50 382 * a VMS virtual processor.
Me@50 383 */
Me@50 384 inline VirtProcr *
Me@50 385 VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 386 { VirtProcr *newPr;
Me@50 387 char *stackLocs;
Me@50 388
Me@50 389 newPr = malloc( sizeof(VirtProcr) );
Me@50 390 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
Me@50 391 if( stackLocs == 0 )
Me@50 392 { perror("malloc stack"); exit(1); }
Me@50 393
msach@69 394 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 395 }
Me@50 396
Me@8 397
Me@64 398 /*Anticipating multi-tasking
Me@64 399 */
Me@64 400 void *
Me@64 401 VMS__give_sem_env_for( VirtProcr *animPr )
Me@64 402 {
Me@64 403 return _VMSMasterEnv->semanticEnv;
Me@64 404 }
Me@64 405 //===========================================================================
Me@26 406 /*there is a label inside this function -- save the addr of this label in
Me@0 407 * the callingPr struc, as the pick-up point from which to start the next
Me@0 408 * work-unit for that procr. If turns out have to save registers, then
Me@0 409 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
Me@0 410 * "done with work-unit" label. The procr struc is in the request in the
Me@0 411 * slave that animated the just-ended work-unit, so all the state is saved
Me@0 412 * there, and will get passed along, inside the request handler, to the
Me@0 413 * next work-unit for that procr.
Me@0 414 */
Me@8 415 void
Me@38 416 VMS__suspend_procr( VirtProcr *animatingPr )
Me@55 417 {
Me@30 418
Me@30 419 //The request to master will cause this suspended virt procr to get
Me@30 420 // scheduled again at some future point -- to resume, core loop jumps
Me@30 421 // to the resume point (below), which causes restore of saved regs and
Me@30 422 // "return" from this call.
msach@71 423 //animatingPr->nextInstrPt = &&ResumePt;
Me@30 424
Me@30 425 //return ownership of the virt procr and sched slot to Master virt pr
Me@38 426 animatingPr->schedSlot->workIsDone = TRUE;
Me@1 427
Me@41 428 //=========================== Measurement stuff ========================
Me@38 429 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 430 //record time stamp: compare to time-stamp recorded below
Me@38 431 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
Me@38 432 #endif
Me@41 433 //=======================================================================
Me@30 434
msach@71 435 switchToCoreLoop(animatingPr);
msach@71 436 flushRegisters();
Me@55 437
Me@55 438 //=======================================================================
Me@30 439
Me@38 440 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 441 //NOTE: only take low part of count -- do sanity check when take diff
Me@38 442 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
Me@38 443 #endif
Me@30 444
Me@0 445 return;
Me@0 446 }
Me@0 447
Me@22 448
Me@22 449
Me@50 450 /*For this implementation of VMS, it may not make much sense to have the
Me@50 451 * system of requests for creating a new processor done this way.. but over
Me@50 452 * the scope of single-master, multi-master, mult-tasking, OS-implementing,
Me@50 453 * distributed-memory, and so on, this gives VMS implementation a chance to
Me@50 454 * do stuff before suspend, in the AppVP, and in the Master before the plugin
Me@50 455 * is called, as well as in the lang-lib before this is called, and in the
Me@50 456 * plugin. So, this gives both VMS and language implementations a chance to
Me@50 457 * intercept at various points and do order-dependent stuff.
Me@50 458 *Having a standard VMSNewPrReqData struc allows the language to create and
Me@50 459 * free the struc, while VMS knows how to get the newPr if it wants it, and
Me@50 460 * it lets the lang have lang-specific data related to creation transported
Me@50 461 * to the plugin.
Me@50 462 */
Nina@110 463 __attribute__ ((noinline)) void
Nina@110 464 VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr )
Nina@109 465
Me@50 466 { VMSReqst req;
Me@50 467
Me@50 468 req.reqType = createReq;
Me@50 469 req.semReqData = semReqData;
Me@50 470 req.nextReqst = reqstingPr->requests;
Me@50 471 reqstingPr->requests = &req;
Me@50 472
Me@50 473 VMS__suspend_procr( reqstingPr );
Me@50 474 }
Me@50 475
Me@22 476
Me@38 477 /*
Me@22 478 *This adds a request to dissipate, then suspends the processor so that the
Me@22 479 * request handler will receive the request. The request handler is what
Me@22 480 * does the work of freeing memory and removing the processor from the
Me@22 481 * semantic environment's data structures.
Me@22 482 *The request handler also is what figures out when to shutdown the VMS
Me@22 483 * system -- which causes all the core loop threads to die, and returns from
Me@22 484 * the call that started up VMS to perform the work.
Me@22 485 *
Me@22 486 *This form is a bit misleading to understand if one is trying to figure out
Me@22 487 * how VMS works -- it looks like a normal function call, but inside it
Me@22 488 * sends a request to the request handler and suspends the processor, which
Me@22 489 * jumps out of the VMS__dissipate_procr function, and out of all nestings
Me@22 490 * above it, transferring the work of dissipating to the request handler,
Me@22 491 * which then does the actual work -- causing the processor that animated
Me@22 492 * the call of this function to disappear and the "hanging" state of this
Me@22 493 * function to just poof into thin air -- the virtual processor's trace
Me@22 494 * never returns from this call, but instead the virtual processor's trace
Me@22 495 * gets suspended in this call and all the virt processor's state disap-
Me@22 496 * pears -- making that suspend the last thing in the virt procr's trace.
Me@8 497 */
Nina@110 498 __attribute__ ((noinline)) void
Nina@110 499 VMS__send_dissipate_req( VirtProcr *procrToDissipate )
Me@50 500 { VMSReqst req;
Me@22 501
Me@50 502 req.reqType = dissipate;
Me@50 503 req.nextReqst = procrToDissipate->requests;
Me@50 504 procrToDissipate->requests = &req;
Me@50 505
Me@22 506 VMS__suspend_procr( procrToDissipate );
Me@50 507 }
Me@50 508
Me@50 509
Me@50 510 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 511 * be called from main thread or other thread -- never from code animated by
Me@50 512 * a VMS virtual processor.
Me@50 513 *
Me@50 514 *Use this version to dissipate VPs created outside the VMS system.
Me@50 515 */
Me@50 516 void
Me@50 517 VMS_ext__dissipate_procr( VirtProcr *procrToDissipate )
Me@50 518 {
Me@50 519 //NOTE: initialData was given to the processor, so should either have
Me@50 520 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@50 521 //So, all that's left to free here is the stack and the VirtProcr struc
Me@50 522 // itself
Me@50 523 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 524 // general that creating processor will outlive ones it creates.
Me@50 525 free( procrToDissipate->startOfStack );
Me@50 526 free( procrToDissipate );
Me@50 527 }
Me@50 528
Me@22 529
Me@22 530
Me@53 531 /*This call's name indicates that request is malloc'd -- so req handler
Me@53 532 * has to free any extra requests tacked on before a send, using this.
Me@53 533 *
Me@53 534 * This inserts the semantic-layer's request data into standard VMS carrier
Me@53 535 * request data-struct that is mallocd. The sem request doesn't need to
Me@53 536 * be malloc'd if this is called inside the same call chain before the
Me@53 537 * send of the last request is called.
Me@53 538 *
Me@53 539 *The request handler has to call VMS__free_VMSReq for any of these
Me@22 540 */
Me@22 541 inline void
Me@53 542 VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData,
Me@53 543 VirtProcr *callingPr )
Me@53 544 { VMSReqst *req;
Me@22 545
Me@53 546 req = VMS__malloc( sizeof(VMSReqst) );
Me@53 547 req->reqType = semantic;
Me@53 548 req->semReqData = semReqData;
Me@53 549 req->nextReqst = callingPr->requests;
Me@53 550 callingPr->requests = req;
Me@22 551 }
Me@22 552
Me@50 553 /*This inserts the semantic-layer's request data into standard VMS carrier
Me@50 554 * request data-struct is allocated on stack of this call & ptr to it sent
Me@50 555 * to plugin
Me@50 556 *Then it does suspend, to cause request to be sent.
Me@50 557 */
Nina@110 558 /*inline*/__attribute__ ((noinline)) void
Nina@110 559 VMS__send_sem_request( void *semReqData, VirtProcr *callingPr )
Me@50 560 { VMSReqst req;
Me@22 561
Me@50 562 req.reqType = semantic;
Me@50 563 req.semReqData = semReqData;
Me@50 564 req.nextReqst = callingPr->requests;
Me@50 565 callingPr->requests = &req;
Me@50 566
Me@50 567 VMS__suspend_procr( callingPr );
Me@50 568 }
Me@50 569
Me@50 570
Nina@110 571 /*inline*/ __attribute__ ((noinline)) void
Nina@110 572 VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr )
Nina@109 573
Me@50 574 { VMSReqst req;
Me@50 575
Me@50 576 req.reqType = VMSSemantic;
Me@50 577 req.semReqData = semReqData;
Me@50 578 req.nextReqst = callingPr->requests; //gab any other preceeding
Me@50 579 callingPr->requests = &req;
Me@50 580
Me@50 581 VMS__suspend_procr( callingPr );
Me@50 582 }
Me@50 583
Me@50 584
Me@50 585 /*
Me@38 586 */
Me@24 587 VMSReqst *
Me@50 588 VMS__take_next_request_out_of( VirtProcr *procrWithReq )
Me@31 589 { VMSReqst *req;
Me@31 590
Me@31 591 req = procrWithReq->requests;
Me@38 592 if( req == NULL ) return NULL;
Me@31 593
Me@31 594 procrWithReq->requests = procrWithReq->requests->nextReqst;
Me@50 595 return req;
Me@24 596 }
Me@22 597
Me@24 598
Me@24 599 inline void *
Me@24 600 VMS__take_sem_reqst_from( VMSReqst *req )
Me@24 601 {
Me@24 602 return req->semReqData;
Me@24 603 }
Me@24 604
Me@24 605
Me@24 606
Me@50 607 /* This is for OS requests and VMS infrastructure requests, such as to create
Me@50 608 * a probe -- a probe is inside the heart of VMS-core, it's not part of any
Me@50 609 * language -- but it's also a semantic thing that's triggered from and used
Me@50 610 * in the application.. so it crosses abstractions.. so, need some special
Me@50 611 * pattern here for handling such requests.
Me@52 612 * Doing this just like it were a second language sharing VMS-core.
Me@52 613 *
Me@50 614 * This is called from the language's request handler when it sees a request
Me@50 615 * of type VMSSemReq
Me@52 616 *
Me@52 617 * TODO: Later change this, to give probes their own separate plugin & have
Me@52 618 * VMS-core steer the request to appropriate plugin
Me@52 619 * Do the same for OS calls -- look later at it..
Me@50 620 */
Me@50 621 void inline
Me@50 622 VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv,
Me@50 623 ResumePrFnPtr resumePrFnPtr )
Me@50 624 { VMSSemReq *semReq;
Me@50 625 IntervalProbe *newProbe;
Me@24 626
Me@50 627 semReq = req->semReqData;
Me@24 628
Me@50 629 newProbe = VMS__malloc( sizeof(IntervalProbe) );
Me@65 630 newProbe->nameStr = VMS__strDup( semReq->nameStr );
Me@50 631 newProbe->hist = NULL;
Me@50 632 newProbe->schedChoiceWasRecorded = FALSE;
Me@53 633
Me@53 634 //This runs in masterVP, so no race-condition worries
Me@50 635 newProbe->probeID =
Me@50 636 addToDynArray( newProbe, _VMSMasterEnv->dynIntervalProbesInfo );
Me@50 637
Me@53 638 requestingPr->dataRetFromReq = newProbe;
Me@50 639
Me@50 640 (*resumePrFnPtr)( requestingPr, semEnv );
Me@22 641 }
Me@22 642
Me@22 643
Me@22 644
Me@24 645 /*This must be called by the request handler plugin -- it cannot be called
Me@24 646 * from the semantic library "dissipate processor" function -- instead, the
Me@50 647 * semantic layer has to generate a request, and the plug-in calls this
Me@24 648 * function.
Me@24 649 *The reason is that this frees the virtual processor's stack -- which is
Me@24 650 * still in use inside semantic library calls!
Me@24 651 *
Me@24 652 *This frees or recycles all the state owned by and comprising the VMS
Me@24 653 * portion of the animating virtual procr. The request handler must first
Me@24 654 * free any semantic data created for the processor that didn't use the
Me@24 655 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
Me@24 656 * system to disown any state that did use VMS_malloc, and then frees the
Me@24 657 * statck and the processor-struct itself.
Me@24 658 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
Me@24 659 * state, then that state gets freed (or sent to recycling) as a side-effect
Me@24 660 * of dis-owning it.
Me@24 661 */
Me@24 662 void
Me@53 663 VMS__dissipate_procr( VirtProcr *animatingPr )
Me@24 664 {
Me@24 665 //dis-own all locations owned by this processor, causing to be freed
Me@24 666 // any locations that it is (was) sole owner of
Me@29 667 //TODO: implement VMS__malloc system, including "give up ownership"
Me@24 668
Me@24 669
Me@24 670 //NOTE: initialData was given to the processor, so should either have
Me@24 671 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@24 672 //So, all that's left to free here is the stack and the VirtProcr struc
Me@24 673 // itself
Me@50 674 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 675 // general that creating processor will outlive ones it creates.
Me@50 676 VMS__free( animatingPr->startOfStack );
Me@50 677 VMS__free( animatingPr );
Me@24 678 }
Me@24 679
Me@24 680
Me@53 681 //TODO: look at architecting cleanest separation between request handler
Me@29 682 // and master loop, for dissipate, create, shutdown, and other non-semantic
Me@29 683 // requests. Issue is chain: one removes requests from AppVP, one dispatches
Me@29 684 // on type of request, and one handles each type.. but some types require
Me@29 685 // action from both request handler and master loop -- maybe just give the
Me@29 686 // request handler calls like: VMS__handle_X_request_type
Me@24 687
Me@29 688
Me@29 689 /*This is called by the semantic layer's request handler when it decides its
Me@29 690 * time to shut down the VMS system. Calling this causes the core loop OS
Me@29 691 * threads to exit, which unblocks the entry-point function that started up
Me@29 692 * VMS, and allows it to grab the result and return to the original single-
Me@29 693 * threaded application.
Me@22 694 *
Me@29 695 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
Me@29 696 * and-wait function has to free a bunch of stuff after it detects the
Me@29 697 * threads have all died: the masterEnv, the thread-related locations,
Me@29 698 * masterVP any AppVPs that might still be allocated and sitting in the
Me@29 699 * semantic environment, or have been orphaned in the _VMSWorkQ.
Me@29 700 *
Me@53 701 *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the
Me@29 702 * locations it needs, and give ownership to masterVP. Then, they will be
Me@53 703 * automatically freed.
Me@22 704 *
Me@29 705 *In here,create one core-loop shut-down processor for each core loop and put
Me@31 706 * them all directly into the readyToAnimateQ.
Me@29 707 *Note, this function can ONLY be called after the semantic environment no
Me@29 708 * longer cares if AppVPs get animated after the point this is called. In
Me@29 709 * other words, this can be used as an abort, or else it should only be
Me@29 710 * called when all AppVPs have finished dissipate requests -- only at that
Me@29 711 * point is it sure that all results have completed.
Me@22 712 */
Me@22 713 void
Me@53 714 VMS__shutdown()
Me@8 715 { int coreIdx;
Me@14 716 VirtProcr *shutDownPr;
Me@22 717
Me@29 718 //create the shutdown processors, one for each core loop -- put them
Me@31 719 // directly into the Q -- each core will die when gets one
Me@8 720 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@50 721 { //Note, this is running in the master
Me@29 722 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
Me@55 723 writeVMSQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
Me@8 724 }
Nina@109 725 #ifdef MEAS__PERF_COUNTERS
Nina@109 726 uint64 tmpc,tmpi;
Nina@109 727 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@130 728 printf("End: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
Nina@109 729 prctl(PR_TASK_PERF_EVENTS_DISABLE);
Nina@109 730 /*
Nina@109 731 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ){
Nina@109 732 close(_VMSMasterEnv->cycles_counter_fd[coreIdx]);
Nina@109 733 close(_VMSMasterEnv->instrs_counter_fd[coreIdx]);
Nina@109 734 }
Nina@109 735 */
Nina@109 736 #endif
Me@12 737 }
Me@12 738
Me@12 739
Me@29 740 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
Me@29 741 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
Me@29 742 *This function has the sole purpose of setting the stack and framePtr
Me@29 743 * to the coreLoop's stack and framePtr.. it does that then jumps to the
Me@29 744 * core loop's shutdown point -- might be able to just call Pthread_exit
Me@30 745 * from here, but am going back to the pthread's stack and setting everything
Me@29 746 * up just as if it never jumped out, before calling pthread_exit.
Me@29 747 *The end-point of core loop will free the stack and so forth of the
Me@29 748 * processor that animates this function, (this fn is transfering the
Me@29 749 * animator of the AppVP that is in turn animating this function over
Me@29 750 * to core loop function -- note that this slices out a level of virtual
Me@29 751 * processors).
Me@29 752 */
Me@29 753 void
Me@29 754 endOSThreadFn( void *initData, VirtProcr *animatingPr )
msach@71 755 {
msach@75 756 #ifdef SEQUENTIAL
msach@75 757 asmTerminateCoreLoopSeq(animatingPr);
msach@75 758 #else
msach@71 759 asmTerminateCoreLoop(animatingPr);
msach@75 760 #endif
Me@30 761 }
Me@29 762
Me@29 763
Me@53 764 /*This is called from the startup & shutdown
Me@24 765 */
Me@24 766 void
Me@53 767 VMS__cleanup_at_end_of_shutdown()
Me@31 768 {
msach@78 769 //unused
msach@78 770 //VMSQueueStruc **readyToAnimateQs;
msach@78 771 //int coreIdx;
msach@78 772 //VirtProcr **masterVPs;
msach@78 773 //SchedSlot ***allSchedSlots; //ptr to array of ptrs
Me@31 774
Me@65 775 //Before getting rid of everything, print out any measurements made
Nina@109 776 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&printHist );
Nina@109 777 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&saveHistToFile);
Me@68 778 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, &freeHistExt );
Nina@129 779 #ifdef DETECT_DEPENDENCIES
Nina@129 780 FILE* output;
Nina@129 781 int n;
Nina@129 782 char filename[255];
Nina@129 783 for(n=0;n<255;n++)
Nina@129 784 {
Nina@130 785 sprintf(filename, "./counters/Dependencies.%d.dot",n);
Nina@129 786 output = fopen(filename,"r");
Nina@129 787 if(output)
Nina@129 788 {
Nina@129 789 fclose(output);
Nina@129 790 }else{
Nina@129 791 break;
Nina@129 792 }
Nina@129 793 }
Nina@130 794 if(n<255){
Nina@129 795 printf("Saving Dependencies to File: %s ...\n", filename);
Nina@129 796 output = fopen(filename,"w+");
Nina@130 797 if(output!=NULL){
Nina@130 798 set_dependency_file(output);
Nina@130 799 fprintf(output,"digraph Dependencies {\n");
Nina@130 800
Nina@130 801 forAllInDynArrayDo( _VMSMasterEnv->dependenciesInfo, &print_dependency_to_file );
Nina@130 802 fprintf(output,"}\n");
Nina@130 803 } else
Nina@130 804 printf("Opening Dependencies file failed. Please check that folder \"counters\" exists in run directory.\n");
Nina@130 805 } else {
Nina@130 806 printf("Could not open Dependencies file, please clean \"counters\" folder. (Must contain less than 255 files.)\n");
Nina@130 807 }
Nina@129 808 #endif
Me@65 809 #ifdef MEAS__TIME_PLUGIN
Me@68 810 printHist( _VMSMasterEnv->reqHdlrLowTimeHist );
msach@84 811 saveHistToFile( _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 812 printHist( _VMSMasterEnv->reqHdlrHighTimeHist );
msach@79 813 saveHistToFile( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 814 freeHistExt( _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 815 freeHistExt( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@65 816 #endif
Me@65 817 #ifdef MEAS__TIME_MALLOC
Me@65 818 printHist( _VMSMasterEnv->mallocTimeHist );
msach@79 819 saveHistToFile( _VMSMasterEnv->mallocTimeHist );
Me@65 820 printHist( _VMSMasterEnv->freeTimeHist );
msach@79 821 saveHistToFile( _VMSMasterEnv->freeTimeHist );
Me@65 822 freeHistExt( _VMSMasterEnv->mallocTimeHist );
Me@65 823 freeHistExt( _VMSMasterEnv->freeTimeHist );
Me@65 824 #endif
Me@65 825 #ifdef MEAS__TIME_MASTER_LOCK
Me@65 826 printHist( _VMSMasterEnv->masterLockLowTimeHist );
Me@65 827 printHist( _VMSMasterEnv->masterLockHighTimeHist );
Me@65 828 #endif
Me@65 829 #ifdef MEAS__TIME_MASTER
Me@65 830 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 831 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 832 {
Me@65 833 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 834 //master VPs were created external to VMS, so use external free
Me@65 835 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 836
Me@65 837 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 838 }
Me@65 839 #endif
Me@65 840 #ifdef MEAS__TIME_STAMP_SUSP
Me@65 841 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 842 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 843 {
Me@65 844 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 845 //master VPs were created external to VMS, so use external free
Me@65 846 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 847
Me@65 848 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 849 }
Me@65 850 #endif
Me@65 851
Me@53 852 //All the environment data has been allocated with VMS__malloc, so just
Me@53 853 // free its internal big-chunk and all inside it disappear.
Me@53 854 /*
Me@31 855 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
Me@31 856 masterVPs = _VMSMasterEnv->masterVPs;
Me@31 857 allSchedSlots = _VMSMasterEnv->allSchedSlots;
Me@31 858
Me@31 859 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@24 860 {
Me@55 861 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@50 862 //master VPs were created external to VMS, so use external free
Me@53 863 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@31 864
Me@31 865 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@24 866 }
Me@31 867
Me@53 868 VMS__free( _VMSMasterEnv->readyToAnimateQs );
Me@53 869 VMS__free( _VMSMasterEnv->masterVPs );
Me@53 870 VMS__free( _VMSMasterEnv->allSchedSlots );
Me@50 871
Me@50 872 //============================= MEASUREMENT STUFF ========================
Me@50 873 #ifdef STATS__TURN_ON_PROBES
Me@53 874 freeDynArrayDeep( _VMSMasterEnv->dynIntervalProbesInfo, &VMS__free_probe);
Me@50 875 #endif
Me@50 876 //========================================================================
Me@53 877 */
Me@53 878 //These are the only two that use system free
Me@53 879 VMS_ext__free_free_list( _VMSMasterEnv->freeListHead );
Me@53 880 free( (void *)_VMSMasterEnv );
Me@24 881 }
Me@24 882
Me@54 883
Me@54 884 //================================
Me@54 885
Me@54 886
Me@54 887 /*Later, improve this -- for now, just exits the application after printing
Me@54 888 * the error message.
Me@54 889 */
Me@54 890 void
Me@54 891 VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData )
Me@54 892 {
msach@69 893 printf("%s",msgStr);
Me@54 894 fflush(stdin);
Me@54 895 exit(1);
Me@54 896 }
Me@54 897