annotate VMS.c @ 187:fe5ad5726e36

counters working ...sort of
author Nina Engelhardt <nengel@mailbox.tu-berlin.de>
date Fri, 03 Feb 2012 17:32:48 +0100
parents 50b29548d4f0
children 20358f56e498
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;
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@166 189
Nina@166 190 #ifdef DETECT_LOOP_GRAPH
nengel@177 191
Nina@166 192 #endif
Nina@166 193
Nina@109 194 #ifdef MEAS__PERF_COUNTERS
nengel@184 195 /*
Nina@131 196 _VMSMasterEnv->counter_history = VMS__malloc(10*sizeof(void*));
Nina@131 197 _VMSMasterEnv->counter_history_array_info = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->counter_history),10);
nengel@184 198 */
Nina@110 199 //printf("Creating HW counters...");
nengel@184 200 /*
Nina@110 201 FILE* output;
Nina@110 202 int n;
Nina@110 203 char filename[255];
Nina@110 204 for(n=0;n<255;n++)
Nina@110 205 {
Nina@113 206 sprintf(filename, "./counters/Counters.%d.csv",n);
Nina@110 207 output = fopen(filename,"r");
Nina@110 208 if(output)
Nina@110 209 {
Nina@110 210 fclose(output);
Nina@110 211 }else{
Nina@110 212 break;
Nina@110 213 }
Nina@110 214 }
Nina@110 215 printf("Saving Counter measurements to File: %s ...\n", filename);
Nina@110 216 output = fopen(filename,"w+");
Nina@110 217 _VMSMasterEnv->counteroutput = output;
nengel@184 218 */
Nina@110 219
engelhardt@108 220 struct perf_event_attr hw_event;
Nina@109 221 memset(&hw_event,0,sizeof(hw_event));
engelhardt@108 222 hw_event.type = PERF_TYPE_HARDWARE;
Nina@109 223 hw_event.size = sizeof(hw_event);
engelhardt@108 224 hw_event.disabled = 1;
Nina@109 225 hw_event.freq = 0;
engelhardt@108 226 hw_event.inherit = 1; /* children inherit it */
engelhardt@108 227 hw_event.pinned = 1; /* must always be on PMU */
engelhardt@108 228 hw_event.exclusive = 0; /* only group on PMU */
engelhardt@108 229 hw_event.exclude_user = 0; /* don't count user */
engelhardt@108 230 hw_event.exclude_kernel = 1; /* ditto kernel */
engelhardt@108 231 hw_event.exclude_hv = 1; /* ditto hypervisor */
engelhardt@108 232 hw_event.exclude_idle = 0; /* don't count when idle */
engelhardt@108 233 hw_event.mmap = 0; /* include mmap data */
engelhardt@108 234 hw_event.comm = 0; /* include comm data */
engelhardt@108 235
Nina@109 236
engelhardt@108 237 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
engelhardt@108 238 {
Nina@109 239 hw_event.config = 0x0000000000000000; //cycles
engelhardt@108 240 _VMSMasterEnv->cycles_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 241 0,//pid_t pid,
Nina@167 242 coreIdx,//int cpu,
engelhardt@108 243 -1,//int group_fd,
engelhardt@108 244 0//unsigned long flags
engelhardt@108 245 );
Nina@109 246 if (_VMSMasterEnv->cycles_counter_fd[coreIdx]<0){
Nina@109 247 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 248 perror("Failed to open cycles counter");
Nina@109 249 }
Nina@109 250 hw_event.config = 0x0000000000000001; //instrs
engelhardt@108 251 _VMSMasterEnv->instrs_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 252 0,//pid_t pid,
Nina@167 253 coreIdx,//int cpu,
engelhardt@108 254 -1,//int group_fd,
engelhardt@108 255 0//unsigned long flags
engelhardt@108 256 );
Nina@109 257 if (_VMSMasterEnv->instrs_counter_fd[coreIdx]<0){
Nina@109 258 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 259 perror("Failed to open instrs counter");
Nina@109 260 }
engelhardt@108 261 }
engelhardt@108 262 prctl(PR_TASK_PERF_EVENTS_ENABLE);
Nina@109 263 uint64 tmpc,tmpi;
Nina@109 264 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@130 265 printf("Start: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
engelhardt@108 266 #endif
engelhardt@108 267
Me@50 268 //========================================================================
Me@38 269
Me@0 270 }
Me@0 271
Me@31 272 SchedSlot **
Me@31 273 create_sched_slots()
Me@31 274 { SchedSlot **schedSlots;
Me@0 275 int i;
Me@0 276
Me@53 277 schedSlots = VMS__malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
Me@8 278
Me@1 279 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@0 280 {
Me@53 281 schedSlots[i] = VMS__malloc( sizeof(SchedSlot) );
Me@8 282
Me@1 283 //Set state to mean "handling requests done, slot needs filling"
Me@8 284 schedSlots[i]->workIsDone = FALSE;
Me@8 285 schedSlots[i]->needsProcrAssigned = TRUE;
Me@0 286 }
Me@31 287 return schedSlots;
Me@31 288 }
Me@31 289
Me@31 290
Me@31 291 void
Me@31 292 freeSchedSlots( SchedSlot **schedSlots )
Me@31 293 { int i;
Me@31 294 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@31 295 {
Me@53 296 VMS__free( schedSlots[i] );
Me@31 297 }
Me@53 298 VMS__free( schedSlots );
Me@0 299 }
Me@0 300
Me@8 301
Me@28 302 void
Me@28 303 create_the_coreLoop_OS_threads()
Me@28 304 {
Me@28 305 //========================================================================
Me@28 306 // Create the Threads
Me@28 307 int coreIdx, retCode;
Me@28 308
Me@28 309 //Need the threads to be created suspended, and wait for a signal
Me@28 310 // before proceeding -- gives time after creating to initialize other
Me@28 311 // stuff before the coreLoops set off.
Me@28 312 _VMSMasterEnv->setupComplete = 0;
Me@28 313
Me@28 314 //Make the threads that animate the core loops
Me@28 315 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@53 316 { coreLoopThdParams[coreIdx] = VMS__malloc( sizeof(ThdParams) );
Me@28 317 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
Me@28 318
Me@28 319 retCode =
Me@28 320 pthread_create( &(coreLoopThdHandles[coreIdx]),
Me@28 321 thdAttrs,
Me@28 322 &coreLoop,
Me@28 323 (void *)(coreLoopThdParams[coreIdx]) );
Me@50 324 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
Me@28 325 }
Me@28 326 }
Me@28 327
Me@0 328 /*Semantic layer calls this when it want the system to start running..
Me@0 329 *
Me@24 330 *This starts the core loops running then waits for them to exit.
Me@0 331 */
Me@12 332 void
Me@24 333 VMS__start_the_work_then_wait_until_done()
Me@12 334 { int coreIdx;
Me@24 335 //Start the core loops running
Me@25 336
Me@25 337 //tell the core loop threads that setup is complete
Me@25 338 //get lock, to lock out any threads still starting up -- they'll see
Me@25 339 // that setupComplete is true before entering while loop, and so never
Me@25 340 // wait on the condition
Me@26 341 pthread_mutex_lock( &suspendLock );
Me@25 342 _VMSMasterEnv->setupComplete = 1;
Me@26 343 pthread_mutex_unlock( &suspendLock );
Me@26 344 pthread_cond_broadcast( &suspend_cond );
Me@25 345
Me@25 346
Me@24 347 //wait for all to complete
Me@8 348 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@8 349 {
Me@25 350 pthread_join( coreLoopThdHandles[coreIdx], NULL );
Me@24 351 }
Me@25 352
Me@24 353 //NOTE: do not clean up VMS env here -- semantic layer has to have
Me@24 354 // a chance to clean up its environment first, then do a call to free
Me@24 355 // the Master env and rest of VMS locations
Me@8 356 }
Me@0 357
msach@70 358 #ifdef SEQUENTIAL
Me@28 359 /*Only difference between version with an OS thread pinned to each core and
Me@28 360 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
Me@28 361 */
Me@28 362 void
Me@28 363 VMS__start_the_work_then_wait_until_done_Seq()
Me@28 364 {
Me@28 365 //Instead of un-suspending threads, just call the one and only
Me@28 366 // core loop (sequential version), in the main thread.
Me@28 367 coreLoop_Seq( NULL );
msach@75 368 flushRegisters();
Me@28 369
Me@28 370 }
msach@70 371 #endif
Me@28 372
Me@50 373 inline VirtProcr *
Me@50 374 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 375 { VirtProcr *newPr;
msach@76 376 void *stackLocs;
Me@50 377
Me@50 378 newPr = VMS__malloc( sizeof(VirtProcr) );
Me@50 379 stackLocs = VMS__malloc( VIRT_PROCR_STACK_SIZE );
Me@50 380 if( stackLocs == 0 )
Me@50 381 { perror("VMS__malloc stack"); exit(1); }
Me@50 382
msach@69 383 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 384 }
Me@50 385
Me@50 386 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 387 * be called from main thread or other thread -- never from code animated by
Me@50 388 * a VMS virtual processor.
Me@50 389 */
Me@50 390 inline VirtProcr *
Me@50 391 VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 392 { VirtProcr *newPr;
Me@50 393 char *stackLocs;
Me@50 394
Me@50 395 newPr = malloc( sizeof(VirtProcr) );
Me@50 396 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
Me@50 397 if( stackLocs == 0 )
Me@50 398 { perror("malloc stack"); exit(1); }
Me@50 399
msach@69 400 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 401 }
Me@50 402
Me@8 403
Me@64 404 /*Anticipating multi-tasking
Me@64 405 */
Me@64 406 void *
Me@64 407 VMS__give_sem_env_for( VirtProcr *animPr )
Me@64 408 {
Me@64 409 return _VMSMasterEnv->semanticEnv;
Me@64 410 }
Me@64 411 //===========================================================================
Me@26 412 /*there is a label inside this function -- save the addr of this label in
Me@0 413 * the callingPr struc, as the pick-up point from which to start the next
Me@0 414 * work-unit for that procr. If turns out have to save registers, then
Me@0 415 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
Me@0 416 * "done with work-unit" label. The procr struc is in the request in the
Me@0 417 * slave that animated the just-ended work-unit, so all the state is saved
Me@0 418 * there, and will get passed along, inside the request handler, to the
Me@0 419 * next work-unit for that procr.
Me@0 420 */
Me@8 421 void
Me@38 422 VMS__suspend_procr( VirtProcr *animatingPr )
Me@55 423 {
Me@30 424
Me@30 425 //The request to master will cause this suspended virt procr to get
Me@30 426 // scheduled again at some future point -- to resume, core loop jumps
Me@30 427 // to the resume point (below), which causes restore of saved regs and
Me@30 428 // "return" from this call.
msach@71 429 //animatingPr->nextInstrPt = &&ResumePt;
Me@30 430
Me@30 431 //return ownership of the virt procr and sched slot to Master virt pr
Me@38 432 animatingPr->schedSlot->workIsDone = TRUE;
Me@1 433
Me@41 434 //=========================== Measurement stuff ========================
Me@38 435 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 436 //record time stamp: compare to time-stamp recorded below
Me@38 437 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
Me@38 438 #endif
nengel@186 439 #ifdef MEAS__PERF_COUNTERS
nengel@186 440 //start work
nengel@186 441 uint64 cycles,instrs;
nengel@186 442 saveCyclesAndInstrs(animatingPr->coreAnimatedBy,cycles, instrs);
nengel@186 443 (*(_VMSMasterEnv->counterHandler))(Procr_suspend,animatingPr,cycles,instrs);
nengel@186 444 #endif
Me@41 445 //=======================================================================
Me@30 446
msach@71 447 switchToCoreLoop(animatingPr);
msach@71 448 flushRegisters();
Me@55 449
Me@55 450 //=======================================================================
Me@30 451
Me@38 452 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 453 //NOTE: only take low part of count -- do sanity check when take diff
Me@38 454 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
Me@38 455 #endif
Me@30 456
Me@0 457 return;
Me@0 458 }
Me@0 459
Me@22 460
Me@22 461
Me@50 462 /*For this implementation of VMS, it may not make much sense to have the
Me@50 463 * system of requests for creating a new processor done this way.. but over
Me@50 464 * the scope of single-master, multi-master, mult-tasking, OS-implementing,
Me@50 465 * distributed-memory, and so on, this gives VMS implementation a chance to
Me@50 466 * do stuff before suspend, in the AppVP, and in the Master before the plugin
Me@50 467 * is called, as well as in the lang-lib before this is called, and in the
Me@50 468 * plugin. So, this gives both VMS and language implementations a chance to
Me@50 469 * intercept at various points and do order-dependent stuff.
Me@50 470 *Having a standard VMSNewPrReqData struc allows the language to create and
Me@50 471 * free the struc, while VMS knows how to get the newPr if it wants it, and
Me@50 472 * it lets the lang have lang-specific data related to creation transported
Me@50 473 * to the plugin.
Me@50 474 */
Me@50 475 void
Me@50 476 VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr )
Me@50 477 { VMSReqst req;
Me@50 478
Me@50 479 req.reqType = createReq;
Me@50 480 req.semReqData = semReqData;
Me@50 481 req.nextReqst = reqstingPr->requests;
Me@50 482 reqstingPr->requests = &req;
Me@50 483
Me@50 484 VMS__suspend_procr( reqstingPr );
Me@50 485 }
Me@50 486
Me@22 487
Me@38 488 /*
Me@22 489 *This adds a request to dissipate, then suspends the processor so that the
Me@22 490 * request handler will receive the request. The request handler is what
Me@22 491 * does the work of freeing memory and removing the processor from the
Me@22 492 * semantic environment's data structures.
Me@22 493 *The request handler also is what figures out when to shutdown the VMS
Me@22 494 * system -- which causes all the core loop threads to die, and returns from
Me@22 495 * the call that started up VMS to perform the work.
Me@22 496 *
Me@22 497 *This form is a bit misleading to understand if one is trying to figure out
Me@22 498 * how VMS works -- it looks like a normal function call, but inside it
Me@22 499 * sends a request to the request handler and suspends the processor, which
Me@22 500 * jumps out of the VMS__dissipate_procr function, and out of all nestings
Me@22 501 * above it, transferring the work of dissipating to the request handler,
Me@22 502 * which then does the actual work -- causing the processor that animated
Me@22 503 * the call of this function to disappear and the "hanging" state of this
Me@22 504 * function to just poof into thin air -- the virtual processor's trace
Me@22 505 * never returns from this call, but instead the virtual processor's trace
Me@22 506 * gets suspended in this call and all the virt processor's state disap-
Me@22 507 * pears -- making that suspend the last thing in the virt procr's trace.
Me@8 508 */
Me@8 509 void
Me@53 510 VMS__send_dissipate_req( VirtProcr *procrToDissipate )
Me@50 511 { VMSReqst req;
Me@22 512
Me@50 513 req.reqType = dissipate;
Me@50 514 req.nextReqst = procrToDissipate->requests;
Me@50 515 procrToDissipate->requests = &req;
Me@50 516
Me@22 517 VMS__suspend_procr( procrToDissipate );
Me@50 518 }
Me@50 519
Me@50 520
Me@50 521 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 522 * be called from main thread or other thread -- never from code animated by
Me@50 523 * a VMS virtual processor.
Me@50 524 *
Me@50 525 *Use this version to dissipate VPs created outside the VMS system.
Me@50 526 */
Me@50 527 void
Me@50 528 VMS_ext__dissipate_procr( VirtProcr *procrToDissipate )
Me@50 529 {
Me@50 530 //NOTE: initialData was given to the processor, so should either have
Me@50 531 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@50 532 //So, all that's left to free here is the stack and the VirtProcr struc
Me@50 533 // itself
Me@50 534 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 535 // general that creating processor will outlive ones it creates.
Me@50 536 free( procrToDissipate->startOfStack );
Me@50 537 free( procrToDissipate );
Me@50 538 }
Me@50 539
Me@22 540
Me@22 541
Me@53 542 /*This call's name indicates that request is malloc'd -- so req handler
Me@53 543 * has to free any extra requests tacked on before a send, using this.
Me@53 544 *
Me@53 545 * This inserts the semantic-layer's request data into standard VMS carrier
Me@53 546 * request data-struct that is mallocd. The sem request doesn't need to
Me@53 547 * be malloc'd if this is called inside the same call chain before the
Me@53 548 * send of the last request is called.
Me@53 549 *
Me@53 550 *The request handler has to call VMS__free_VMSReq for any of these
Me@22 551 */
Me@22 552 inline void
Me@53 553 VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData,
Me@53 554 VirtProcr *callingPr )
Me@53 555 { VMSReqst *req;
Me@22 556
Me@53 557 req = VMS__malloc( sizeof(VMSReqst) );
Me@53 558 req->reqType = semantic;
Me@53 559 req->semReqData = semReqData;
Me@53 560 req->nextReqst = callingPr->requests;
Me@53 561 callingPr->requests = req;
Me@22 562 }
Me@22 563
Me@50 564 /*This inserts the semantic-layer's request data into standard VMS carrier
Me@50 565 * request data-struct is allocated on stack of this call & ptr to it sent
Me@50 566 * to plugin
Me@50 567 *Then it does suspend, to cause request to be sent.
Me@50 568 */
Me@50 569 inline void
Me@50 570 VMS__send_sem_request( void *semReqData, VirtProcr *callingPr )
Me@50 571 { VMSReqst req;
Me@22 572
Me@50 573 req.reqType = semantic;
Me@50 574 req.semReqData = semReqData;
Me@50 575 req.nextReqst = callingPr->requests;
Me@50 576 callingPr->requests = &req;
Me@50 577
Me@50 578 VMS__suspend_procr( callingPr );
Me@50 579 }
Me@50 580
Me@50 581
Me@50 582 inline void
Me@50 583 VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr )
Me@50 584 { VMSReqst req;
Me@50 585
Me@50 586 req.reqType = VMSSemantic;
Me@50 587 req.semReqData = semReqData;
Me@50 588 req.nextReqst = callingPr->requests; //gab any other preceeding
Me@50 589 callingPr->requests = &req;
Me@50 590
Me@50 591 VMS__suspend_procr( callingPr );
Me@50 592 }
Me@50 593
Me@50 594
Me@50 595 /*
Me@38 596 */
Me@24 597 VMSReqst *
Me@50 598 VMS__take_next_request_out_of( VirtProcr *procrWithReq )
Me@31 599 { VMSReqst *req;
Me@31 600
Me@31 601 req = procrWithReq->requests;
Me@38 602 if( req == NULL ) return NULL;
Me@31 603
Me@31 604 procrWithReq->requests = procrWithReq->requests->nextReqst;
Me@50 605 return req;
Me@24 606 }
Me@22 607
Me@24 608
Me@24 609 inline void *
Me@24 610 VMS__take_sem_reqst_from( VMSReqst *req )
Me@24 611 {
Me@24 612 return req->semReqData;
Me@24 613 }
Me@24 614
Me@24 615
Me@24 616
Me@50 617 /* This is for OS requests and VMS infrastructure requests, such as to create
Me@50 618 * a probe -- a probe is inside the heart of VMS-core, it's not part of any
Me@50 619 * language -- but it's also a semantic thing that's triggered from and used
Me@50 620 * in the application.. so it crosses abstractions.. so, need some special
Me@50 621 * pattern here for handling such requests.
Me@52 622 * Doing this just like it were a second language sharing VMS-core.
Me@52 623 *
Me@50 624 * This is called from the language's request handler when it sees a request
Me@50 625 * of type VMSSemReq
Me@52 626 *
Me@52 627 * TODO: Later change this, to give probes their own separate plugin & have
Me@52 628 * VMS-core steer the request to appropriate plugin
Me@52 629 * Do the same for OS calls -- look later at it..
Me@50 630 */
Me@50 631 void inline
Me@50 632 VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv,
Me@50 633 ResumePrFnPtr resumePrFnPtr )
Me@50 634 { VMSSemReq *semReq;
Me@50 635 IntervalProbe *newProbe;
Me@24 636
Me@50 637 semReq = req->semReqData;
Me@24 638
Me@50 639 newProbe = VMS__malloc( sizeof(IntervalProbe) );
Me@65 640 newProbe->nameStr = VMS__strDup( semReq->nameStr );
Me@50 641 newProbe->hist = NULL;
Me@50 642 newProbe->schedChoiceWasRecorded = FALSE;
Me@53 643
Me@53 644 //This runs in masterVP, so no race-condition worries
Me@50 645 newProbe->probeID =
Me@50 646 addToDynArray( newProbe, _VMSMasterEnv->dynIntervalProbesInfo );
Me@50 647
Me@53 648 requestingPr->dataRetFromReq = newProbe;
Me@50 649
Me@50 650 (*resumePrFnPtr)( requestingPr, semEnv );
Me@22 651 }
Me@22 652
Me@22 653
Me@22 654
Me@24 655 /*This must be called by the request handler plugin -- it cannot be called
Me@24 656 * from the semantic library "dissipate processor" function -- instead, the
Me@50 657 * semantic layer has to generate a request, and the plug-in calls this
Me@24 658 * function.
Me@24 659 *The reason is that this frees the virtual processor's stack -- which is
Me@24 660 * still in use inside semantic library calls!
Me@24 661 *
Me@24 662 *This frees or recycles all the state owned by and comprising the VMS
Me@24 663 * portion of the animating virtual procr. The request handler must first
Me@24 664 * free any semantic data created for the processor that didn't use the
Me@24 665 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
Me@24 666 * system to disown any state that did use VMS_malloc, and then frees the
Me@24 667 * statck and the processor-struct itself.
Me@24 668 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
Me@24 669 * state, then that state gets freed (or sent to recycling) as a side-effect
Me@24 670 * of dis-owning it.
Me@24 671 */
Me@24 672 void
Me@53 673 VMS__dissipate_procr( VirtProcr *animatingPr )
Me@24 674 {
Me@24 675 //dis-own all locations owned by this processor, causing to be freed
Me@24 676 // any locations that it is (was) sole owner of
Me@29 677 //TODO: implement VMS__malloc system, including "give up ownership"
Me@24 678
Me@24 679
Me@24 680 //NOTE: initialData was given to the processor, so should either have
Me@24 681 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@24 682 //So, all that's left to free here is the stack and the VirtProcr struc
Me@24 683 // itself
Me@50 684 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 685 // general that creating processor will outlive ones it creates.
Me@50 686 VMS__free( animatingPr->startOfStack );
Me@50 687 VMS__free( animatingPr );
Me@24 688 }
Me@24 689
Me@24 690
Me@53 691 //TODO: look at architecting cleanest separation between request handler
Me@29 692 // and master loop, for dissipate, create, shutdown, and other non-semantic
Me@29 693 // requests. Issue is chain: one removes requests from AppVP, one dispatches
Me@29 694 // on type of request, and one handles each type.. but some types require
Me@29 695 // action from both request handler and master loop -- maybe just give the
Me@29 696 // request handler calls like: VMS__handle_X_request_type
Me@24 697
Me@29 698
Me@29 699 /*This is called by the semantic layer's request handler when it decides its
Me@29 700 * time to shut down the VMS system. Calling this causes the core loop OS
Me@29 701 * threads to exit, which unblocks the entry-point function that started up
Me@29 702 * VMS, and allows it to grab the result and return to the original single-
Me@29 703 * threaded application.
Me@22 704 *
Me@29 705 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
Me@29 706 * and-wait function has to free a bunch of stuff after it detects the
Me@29 707 * threads have all died: the masterEnv, the thread-related locations,
Me@29 708 * masterVP any AppVPs that might still be allocated and sitting in the
Me@29 709 * semantic environment, or have been orphaned in the _VMSWorkQ.
Me@29 710 *
Me@53 711 *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the
Me@29 712 * locations it needs, and give ownership to masterVP. Then, they will be
Me@53 713 * automatically freed.
Me@22 714 *
Me@29 715 *In here,create one core-loop shut-down processor for each core loop and put
Me@31 716 * them all directly into the readyToAnimateQ.
Me@29 717 *Note, this function can ONLY be called after the semantic environment no
Me@29 718 * longer cares if AppVPs get animated after the point this is called. In
Me@29 719 * other words, this can be used as an abort, or else it should only be
Me@29 720 * called when all AppVPs have finished dissipate requests -- only at that
Me@29 721 * point is it sure that all results have completed.
Me@22 722 */
Me@22 723 void
Me@53 724 VMS__shutdown()
Me@8 725 { int coreIdx;
Me@14 726 VirtProcr *shutDownPr;
Me@22 727
Me@29 728 //create the shutdown processors, one for each core loop -- put them
Me@31 729 // directly into the Q -- each core will die when gets one
Me@8 730 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@50 731 { //Note, this is running in the master
Me@29 732 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
Me@55 733 writeVMSQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
Me@8 734 }
Nina@109 735 #ifdef MEAS__PERF_COUNTERS
Nina@109 736 uint64 tmpc,tmpi;
Nina@109 737 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@130 738 printf("End: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
Nina@109 739 prctl(PR_TASK_PERF_EVENTS_DISABLE);
Nina@109 740 /*
Nina@109 741 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ){
Nina@109 742 close(_VMSMasterEnv->cycles_counter_fd[coreIdx]);
Nina@109 743 close(_VMSMasterEnv->instrs_counter_fd[coreIdx]);
Nina@109 744 }
Nina@109 745 */
Nina@109 746 #endif
Me@12 747 }
Me@12 748
Me@12 749
Me@29 750 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
Me@29 751 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
Me@29 752 *This function has the sole purpose of setting the stack and framePtr
Me@29 753 * to the coreLoop's stack and framePtr.. it does that then jumps to the
Me@29 754 * core loop's shutdown point -- might be able to just call Pthread_exit
Me@30 755 * from here, but am going back to the pthread's stack and setting everything
Me@29 756 * up just as if it never jumped out, before calling pthread_exit.
Me@29 757 *The end-point of core loop will free the stack and so forth of the
Me@29 758 * processor that animates this function, (this fn is transfering the
Me@29 759 * animator of the AppVP that is in turn animating this function over
Me@29 760 * to core loop function -- note that this slices out a level of virtual
Me@29 761 * processors).
Me@29 762 */
Me@29 763 void
Me@29 764 endOSThreadFn( void *initData, VirtProcr *animatingPr )
msach@71 765 {
msach@75 766 #ifdef SEQUENTIAL
msach@75 767 asmTerminateCoreLoopSeq(animatingPr);
msach@75 768 #else
msach@71 769 asmTerminateCoreLoop(animatingPr);
msach@75 770 #endif
Me@30 771 }
Me@29 772
Me@29 773
Me@53 774 /*This is called from the startup & shutdown
Me@24 775 */
Me@24 776 void
Me@53 777 VMS__cleanup_at_end_of_shutdown()
Me@31 778 {
msach@78 779 //unused
msach@78 780 //VMSQueueStruc **readyToAnimateQs;
msach@78 781 //int coreIdx;
msach@78 782 //VirtProcr **masterVPs;
msach@78 783 //SchedSlot ***allSchedSlots; //ptr to array of ptrs
Me@31 784
Me@65 785 //Before getting rid of everything, print out any measurements made
Nina@109 786 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&printHist );
Nina@109 787 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&saveHistToFile);
nengel@182 788 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, &freeHist );
Nina@167 789
nengel@177 790
Me@65 791 #ifdef MEAS__TIME_PLUGIN
Me@68 792 printHist( _VMSMasterEnv->reqHdlrLowTimeHist );
msach@84 793 saveHistToFile( _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 794 printHist( _VMSMasterEnv->reqHdlrHighTimeHist );
msach@79 795 saveHistToFile( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 796 freeHistExt( _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 797 freeHistExt( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@65 798 #endif
Me@65 799 #ifdef MEAS__TIME_MALLOC
Me@65 800 printHist( _VMSMasterEnv->mallocTimeHist );
msach@79 801 saveHistToFile( _VMSMasterEnv->mallocTimeHist );
Me@65 802 printHist( _VMSMasterEnv->freeTimeHist );
msach@79 803 saveHistToFile( _VMSMasterEnv->freeTimeHist );
Me@65 804 freeHistExt( _VMSMasterEnv->mallocTimeHist );
Me@65 805 freeHistExt( _VMSMasterEnv->freeTimeHist );
Me@65 806 #endif
Me@65 807 #ifdef MEAS__TIME_MASTER_LOCK
Me@65 808 printHist( _VMSMasterEnv->masterLockLowTimeHist );
Me@65 809 printHist( _VMSMasterEnv->masterLockHighTimeHist );
Me@65 810 #endif
Me@65 811 #ifdef MEAS__TIME_MASTER
Me@65 812 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 813 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 814 {
Me@65 815 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 816 //master VPs were created external to VMS, so use external free
Me@65 817 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 818
Me@65 819 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 820 }
Me@65 821 #endif
Me@65 822 #ifdef MEAS__TIME_STAMP_SUSP
Me@65 823 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 824 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 825 {
Me@65 826 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 827 //master VPs were created external to VMS, so use external free
Me@65 828 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 829
Me@65 830 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 831 }
Me@65 832 #endif
Me@65 833
Me@53 834 //All the environment data has been allocated with VMS__malloc, so just
Me@53 835 // free its internal big-chunk and all inside it disappear.
Me@53 836 /*
Me@31 837 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
Me@31 838 masterVPs = _VMSMasterEnv->masterVPs;
Me@31 839 allSchedSlots = _VMSMasterEnv->allSchedSlots;
Me@31 840
Me@31 841 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@24 842 {
Me@55 843 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@50 844 //master VPs were created external to VMS, so use external free
Me@53 845 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@31 846
Me@31 847 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@24 848 }
Me@31 849
Me@53 850 VMS__free( _VMSMasterEnv->readyToAnimateQs );
Me@53 851 VMS__free( _VMSMasterEnv->masterVPs );
Me@53 852 VMS__free( _VMSMasterEnv->allSchedSlots );
Me@50 853
Me@50 854 //============================= MEASUREMENT STUFF ========================
Me@50 855 #ifdef STATS__TURN_ON_PROBES
Me@53 856 freeDynArrayDeep( _VMSMasterEnv->dynIntervalProbesInfo, &VMS__free_probe);
Me@50 857 #endif
Me@50 858 //========================================================================
Me@53 859 */
Me@53 860 //These are the only two that use system free
Me@53 861 VMS_ext__free_free_list( _VMSMasterEnv->freeListHead );
Me@53 862 free( (void *)_VMSMasterEnv );
Me@24 863 }
Me@24 864
Me@54 865
Me@54 866 //================================
Me@54 867
Me@54 868
Me@54 869 /*Later, improve this -- for now, just exits the application after printing
Me@54 870 * the error message.
Me@54 871 */
Me@54 872 void
Me@54 873 VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData )
Me@54 874 {
msach@69 875 printf("%s",msgStr);
Me@54 876 fflush(stdin);
Me@54 877 exit(1);
Me@54 878 }
Me@54 879