annotate VMS.c @ 110:724c7a0b687f

save counters to file in csv
author Nina Engelhardt
date Wed, 03 Aug 2011 17:05:18 +0200
parents 659299627e70
children 43ea731da94e
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@79 121 _VMSMasterEnv->mallocTimeHist = makeFixedBinHistExt( 50, 0, 100,
msach@79 122 "malloc_time_hist");
Me@68 123 _VMSMasterEnv->freeTimeHist = makeFixedBinHistExt( 50, 0, 100,
msach@79 124 "free_time_hist");
Me@65 125 #endif
Me@68 126 #ifdef MEAS__TIME_PLUGIN
Me@68 127 _VMSMasterEnv->reqHdlrLowTimeHist = makeFixedBinHistExt( 50, 0, 10,
msach@79 128 "plugin_low_time_hist");
Me@68 129 _VMSMasterEnv->reqHdlrHighTimeHist = makeFixedBinHistExt( 50, 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@109 189 #ifdef MEAS__PERF_COUNTERS
Nina@110 190 //printf("Creating HW counters...");
Nina@110 191 FILE* output;
Nina@110 192 int n;
Nina@110 193 char filename[255];
Nina@110 194 for(n=0;n<255;n++)
Nina@110 195 {
Nina@110 196 sprintf(filename, "./counters/Counters.%d.dat",n);
Nina@110 197 output = fopen(filename,"r");
Nina@110 198 if(output)
Nina@110 199 {
Nina@110 200 fclose(output);
Nina@110 201 }else{
Nina@110 202 break;
Nina@110 203 }
Nina@110 204 }
Nina@110 205 printf("Saving Counter measurements to File: %s ...\n", filename);
Nina@110 206 output = fopen(filename,"w+");
Nina@110 207 _VMSMasterEnv->counteroutput = output;
Nina@110 208
engelhardt@108 209 struct perf_event_attr hw_event;
Nina@109 210 memset(&hw_event,0,sizeof(hw_event));
engelhardt@108 211 hw_event.type = PERF_TYPE_HARDWARE;
Nina@109 212 hw_event.size = sizeof(hw_event);
engelhardt@108 213 hw_event.disabled = 1;
Nina@109 214 hw_event.freq = 0;
engelhardt@108 215 hw_event.inherit = 1; /* children inherit it */
engelhardt@108 216 hw_event.pinned = 1; /* must always be on PMU */
engelhardt@108 217 hw_event.exclusive = 0; /* only group on PMU */
engelhardt@108 218 hw_event.exclude_user = 0; /* don't count user */
engelhardt@108 219 hw_event.exclude_kernel = 1; /* ditto kernel */
engelhardt@108 220 hw_event.exclude_hv = 1; /* ditto hypervisor */
engelhardt@108 221 hw_event.exclude_idle = 0; /* don't count when idle */
engelhardt@108 222 hw_event.mmap = 0; /* include mmap data */
engelhardt@108 223 hw_event.comm = 0; /* include comm data */
engelhardt@108 224
Nina@109 225
engelhardt@108 226 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
engelhardt@108 227 {
Nina@109 228 hw_event.config = 0x0000000000000000; //cycles
engelhardt@108 229 _VMSMasterEnv->cycles_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 230 0,//pid_t pid,
Nina@109 231 -1,//int cpu,
engelhardt@108 232 -1,//int group_fd,
engelhardt@108 233 0//unsigned long flags
engelhardt@108 234 );
Nina@109 235 if (_VMSMasterEnv->cycles_counter_fd[coreIdx]<0){
Nina@109 236 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 237 perror("Failed to open cycles counter");
Nina@109 238 }
Nina@109 239 hw_event.config = 0x0000000000000001; //instrs
engelhardt@108 240 _VMSMasterEnv->instrs_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
engelhardt@108 241 0,//pid_t pid,
Nina@109 242 -1,//int cpu,
engelhardt@108 243 -1,//int group_fd,
engelhardt@108 244 0//unsigned long flags
engelhardt@108 245 );
Nina@109 246 if (_VMSMasterEnv->instrs_counter_fd[coreIdx]<0){
Nina@109 247 fprintf(stderr,"On core %d: ",coreIdx);
Nina@109 248 perror("Failed to open instrs counter");
Nina@109 249 }
engelhardt@108 250 }
engelhardt@108 251 prctl(PR_TASK_PERF_EVENTS_ENABLE);
Nina@109 252 uint64 tmpc,tmpi;
Nina@109 253 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@109 254 printf("Start: cycles = %lu, instrs = %lu\n",tmpc,tmpi);
engelhardt@108 255 #endif
engelhardt@108 256
Me@50 257 //========================================================================
Me@38 258
Me@0 259 }
Me@0 260
Me@31 261 SchedSlot **
Me@31 262 create_sched_slots()
Me@31 263 { SchedSlot **schedSlots;
Me@0 264 int i;
Me@0 265
Me@53 266 schedSlots = VMS__malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
Me@8 267
Me@1 268 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@0 269 {
Me@53 270 schedSlots[i] = VMS__malloc( sizeof(SchedSlot) );
Me@8 271
Me@1 272 //Set state to mean "handling requests done, slot needs filling"
Me@8 273 schedSlots[i]->workIsDone = FALSE;
Me@8 274 schedSlots[i]->needsProcrAssigned = TRUE;
Me@0 275 }
Me@31 276 return schedSlots;
Me@31 277 }
Me@31 278
Me@31 279
Me@31 280 void
Me@31 281 freeSchedSlots( SchedSlot **schedSlots )
Me@31 282 { int i;
Me@31 283 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@31 284 {
Me@53 285 VMS__free( schedSlots[i] );
Me@31 286 }
Me@53 287 VMS__free( schedSlots );
Me@0 288 }
Me@0 289
Me@8 290
Me@28 291 void
Me@28 292 create_the_coreLoop_OS_threads()
Me@28 293 {
Me@28 294 //========================================================================
Me@28 295 // Create the Threads
Me@28 296 int coreIdx, retCode;
Me@28 297
Me@28 298 //Need the threads to be created suspended, and wait for a signal
Me@28 299 // before proceeding -- gives time after creating to initialize other
Me@28 300 // stuff before the coreLoops set off.
Me@28 301 _VMSMasterEnv->setupComplete = 0;
Me@28 302
Me@28 303 //Make the threads that animate the core loops
Me@28 304 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@53 305 { coreLoopThdParams[coreIdx] = VMS__malloc( sizeof(ThdParams) );
Me@28 306 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
Me@28 307
Me@28 308 retCode =
Me@28 309 pthread_create( &(coreLoopThdHandles[coreIdx]),
Me@28 310 thdAttrs,
Me@28 311 &coreLoop,
Me@28 312 (void *)(coreLoopThdParams[coreIdx]) );
Me@50 313 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
Me@28 314 }
Me@28 315 }
Me@28 316
Me@0 317 /*Semantic layer calls this when it want the system to start running..
Me@0 318 *
Me@24 319 *This starts the core loops running then waits for them to exit.
Me@0 320 */
Me@12 321 void
Me@24 322 VMS__start_the_work_then_wait_until_done()
Me@12 323 { int coreIdx;
Me@24 324 //Start the core loops running
Me@25 325
Me@25 326 //tell the core loop threads that setup is complete
Me@25 327 //get lock, to lock out any threads still starting up -- they'll see
Me@25 328 // that setupComplete is true before entering while loop, and so never
Me@25 329 // wait on the condition
Me@26 330 pthread_mutex_lock( &suspendLock );
Me@25 331 _VMSMasterEnv->setupComplete = 1;
Me@26 332 pthread_mutex_unlock( &suspendLock );
Me@26 333 pthread_cond_broadcast( &suspend_cond );
Me@25 334
Me@25 335
Me@24 336 //wait for all to complete
Me@8 337 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@8 338 {
Me@25 339 pthread_join( coreLoopThdHandles[coreIdx], NULL );
Me@24 340 }
Me@25 341
Me@24 342 //NOTE: do not clean up VMS env here -- semantic layer has to have
Me@24 343 // a chance to clean up its environment first, then do a call to free
Me@24 344 // the Master env and rest of VMS locations
Me@8 345 }
Me@0 346
msach@71 347 #ifdef SEQUENTIAL
Me@28 348 /*Only difference between version with an OS thread pinned to each core and
Me@28 349 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
Me@28 350 */
Me@28 351 void
Me@28 352 VMS__start_the_work_then_wait_until_done_Seq()
Me@28 353 {
Me@28 354 //Instead of un-suspending threads, just call the one and only
Me@28 355 // core loop (sequential version), in the main thread.
Me@28 356 coreLoop_Seq( NULL );
msach@75 357 flushRegisters();
Me@28 358
Me@28 359 }
msach@71 360 #endif
Me@0 361
Me@50 362 inline VirtProcr *
Me@50 363 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 364 { VirtProcr *newPr;
msach@76 365 void *stackLocs;
Me@50 366
Me@50 367 newPr = VMS__malloc( sizeof(VirtProcr) );
Me@50 368 stackLocs = VMS__malloc( VIRT_PROCR_STACK_SIZE );
Me@50 369 if( stackLocs == 0 )
Me@50 370 { perror("VMS__malloc stack"); exit(1); }
Me@50 371
Me@50 372 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 373 }
Me@50 374
Me@50 375 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 376 * be called from main thread or other thread -- never from code animated by
Me@50 377 * a VMS virtual processor.
Me@50 378 */
Me@50 379 inline VirtProcr *
Me@50 380 VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 381 { VirtProcr *newPr;
Me@50 382 char *stackLocs;
Me@50 383
Me@50 384 newPr = malloc( sizeof(VirtProcr) );
Me@50 385 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
Me@50 386 if( stackLocs == 0 )
Me@50 387 { perror("malloc stack"); exit(1); }
Me@50 388
Me@50 389 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 390 }
Me@50 391
Me@8 392
Me@64 393 /*Anticipating multi-tasking
Me@64 394 */
Me@64 395 void *
Me@64 396 VMS__give_sem_env_for( VirtProcr *animPr )
Me@64 397 {
Me@64 398 return _VMSMasterEnv->semanticEnv;
Me@64 399 }
Me@64 400 //===========================================================================
Me@26 401 /*there is a label inside this function -- save the addr of this label in
Me@0 402 * the callingPr struc, as the pick-up point from which to start the next
Me@0 403 * work-unit for that procr. If turns out have to save registers, then
Me@0 404 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
Me@0 405 * "done with work-unit" label. The procr struc is in the request in the
Me@0 406 * slave that animated the just-ended work-unit, so all the state is saved
Me@0 407 * there, and will get passed along, inside the request handler, to the
Me@0 408 * next work-unit for that procr.
Me@0 409 */
Me@8 410 void
Me@38 411 VMS__suspend_procr( VirtProcr *animatingPr )
Me@55 412 {
Me@0 413
Me@14 414 //The request to master will cause this suspended virt procr to get
Me@14 415 // scheduled again at some future point -- to resume, core loop jumps
Me@14 416 // to the resume point (below), which causes restore of saved regs and
Me@14 417 // "return" from this call.
msach@71 418 //animatingPr->nextInstrPt = &&ResumePt;
Me@1 419
Me@1 420 //return ownership of the virt procr and sched slot to Master virt pr
Me@38 421 animatingPr->schedSlot->workIsDone = TRUE;
Me@1 422
Me@41 423 //=========================== Measurement stuff ========================
Me@38 424 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 425 //record time stamp: compare to time-stamp recorded below
Me@38 426 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
Me@38 427 #endif
Me@41 428 //=======================================================================
Me@38 429
msach@71 430 switchToCoreLoop(animatingPr);
msach@71 431 flushRegisters();
Me@55 432
Me@55 433 //=======================================================================
msach@71 434
Me@38 435 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 436 //NOTE: only take low part of count -- do sanity check when take diff
Me@38 437 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
Me@38 438 #endif
Me@38 439
Me@0 440 return;
Me@0 441 }
Me@0 442
Me@22 443
Me@22 444
Me@50 445 /*For this implementation of VMS, it may not make much sense to have the
Me@50 446 * system of requests for creating a new processor done this way.. but over
Me@50 447 * the scope of single-master, multi-master, mult-tasking, OS-implementing,
Me@50 448 * distributed-memory, and so on, this gives VMS implementation a chance to
Me@50 449 * do stuff before suspend, in the AppVP, and in the Master before the plugin
Me@50 450 * is called, as well as in the lang-lib before this is called, and in the
Me@50 451 * plugin. So, this gives both VMS and language implementations a chance to
Me@50 452 * intercept at various points and do order-dependent stuff.
Me@50 453 *Having a standard VMSNewPrReqData struc allows the language to create and
Me@50 454 * free the struc, while VMS knows how to get the newPr if it wants it, and
Me@50 455 * it lets the lang have lang-specific data related to creation transported
Me@50 456 * to the plugin.
Me@50 457 */
Nina@110 458 __attribute__ ((noinline)) void
Nina@110 459 VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr )
Nina@109 460
Me@50 461 { VMSReqst req;
Me@50 462
Me@50 463 req.reqType = createReq;
Me@50 464 req.semReqData = semReqData;
Me@50 465 req.nextReqst = reqstingPr->requests;
Me@50 466 reqstingPr->requests = &req;
Me@50 467
Me@50 468 VMS__suspend_procr( reqstingPr );
Me@50 469 }
Me@50 470
Me@22 471
Me@38 472 /*
Me@22 473 *This adds a request to dissipate, then suspends the processor so that the
Me@22 474 * request handler will receive the request. The request handler is what
Me@22 475 * does the work of freeing memory and removing the processor from the
Me@22 476 * semantic environment's data structures.
Me@22 477 *The request handler also is what figures out when to shutdown the VMS
Me@22 478 * system -- which causes all the core loop threads to die, and returns from
Me@22 479 * the call that started up VMS to perform the work.
Me@22 480 *
Me@22 481 *This form is a bit misleading to understand if one is trying to figure out
Me@22 482 * how VMS works -- it looks like a normal function call, but inside it
Me@22 483 * sends a request to the request handler and suspends the processor, which
Me@22 484 * jumps out of the VMS__dissipate_procr function, and out of all nestings
Me@22 485 * above it, transferring the work of dissipating to the request handler,
Me@22 486 * which then does the actual work -- causing the processor that animated
Me@22 487 * the call of this function to disappear and the "hanging" state of this
Me@22 488 * function to just poof into thin air -- the virtual processor's trace
Me@22 489 * never returns from this call, but instead the virtual processor's trace
Me@22 490 * gets suspended in this call and all the virt processor's state disap-
Me@22 491 * pears -- making that suspend the last thing in the virt procr's trace.
Me@8 492 */
Nina@110 493 __attribute__ ((noinline)) void
Nina@110 494 VMS__send_dissipate_req( VirtProcr *procrToDissipate )
Me@50 495 { VMSReqst req;
Me@22 496
Me@50 497 req.reqType = dissipate;
Me@50 498 req.nextReqst = procrToDissipate->requests;
Me@50 499 procrToDissipate->requests = &req;
Me@50 500
Me@22 501 VMS__suspend_procr( procrToDissipate );
Me@50 502 }
Me@50 503
Me@50 504
Me@50 505 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 506 * be called from main thread or other thread -- never from code animated by
Me@50 507 * a VMS virtual processor.
Me@50 508 *
Me@50 509 *Use this version to dissipate VPs created outside the VMS system.
Me@50 510 */
Me@50 511 void
Me@50 512 VMS_ext__dissipate_procr( VirtProcr *procrToDissipate )
Me@50 513 {
Me@50 514 //NOTE: initialData was given to the processor, so should either have
Me@50 515 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@50 516 //So, all that's left to free here is the stack and the VirtProcr struc
Me@50 517 // itself
Me@50 518 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 519 // general that creating processor will outlive ones it creates.
Me@50 520 free( procrToDissipate->startOfStack );
Me@50 521 free( procrToDissipate );
Me@50 522 }
Me@50 523
Me@22 524
Me@22 525
Me@53 526 /*This call's name indicates that request is malloc'd -- so req handler
Me@53 527 * has to free any extra requests tacked on before a send, using this.
Me@53 528 *
Me@53 529 * This inserts the semantic-layer's request data into standard VMS carrier
Me@53 530 * request data-struct that is mallocd. The sem request doesn't need to
Me@53 531 * be malloc'd if this is called inside the same call chain before the
Me@53 532 * send of the last request is called.
Me@53 533 *
Me@53 534 *The request handler has to call VMS__free_VMSReq for any of these
Me@22 535 */
Me@22 536 inline void
Me@53 537 VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData,
Me@53 538 VirtProcr *callingPr )
Me@53 539 { VMSReqst *req;
Me@22 540
Me@53 541 req = VMS__malloc( sizeof(VMSReqst) );
Me@53 542 req->reqType = semantic;
Me@53 543 req->semReqData = semReqData;
Me@53 544 req->nextReqst = callingPr->requests;
Me@53 545 callingPr->requests = req;
Me@22 546 }
Me@22 547
Me@50 548 /*This inserts the semantic-layer's request data into standard VMS carrier
Me@50 549 * request data-struct is allocated on stack of this call & ptr to it sent
Me@50 550 * to plugin
Me@50 551 *Then it does suspend, to cause request to be sent.
Me@50 552 */
Nina@110 553 /*inline*/__attribute__ ((noinline)) void
Nina@110 554 VMS__send_sem_request( void *semReqData, VirtProcr *callingPr )
Me@50 555 { VMSReqst req;
Me@22 556
Me@50 557 req.reqType = semantic;
Me@50 558 req.semReqData = semReqData;
Me@50 559 req.nextReqst = callingPr->requests;
Me@50 560 callingPr->requests = &req;
Me@50 561
Me@50 562 VMS__suspend_procr( callingPr );
Me@50 563 }
Me@50 564
Me@50 565
Nina@110 566 /*inline*/ __attribute__ ((noinline)) void
Nina@110 567 VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr )
Nina@109 568
Me@50 569 { VMSReqst req;
Me@50 570
Me@50 571 req.reqType = VMSSemantic;
Me@50 572 req.semReqData = semReqData;
Me@50 573 req.nextReqst = callingPr->requests; //gab any other preceeding
Me@50 574 callingPr->requests = &req;
Me@50 575
Me@50 576 VMS__suspend_procr( callingPr );
Me@50 577 }
Me@50 578
Me@50 579
Me@50 580 /*
Me@38 581 */
Me@24 582 VMSReqst *
Me@50 583 VMS__take_next_request_out_of( VirtProcr *procrWithReq )
Me@31 584 { VMSReqst *req;
Me@31 585
Me@31 586 req = procrWithReq->requests;
Me@38 587 if( req == NULL ) return NULL;
Me@31 588
Me@31 589 procrWithReq->requests = procrWithReq->requests->nextReqst;
Me@50 590 return req;
Me@24 591 }
Me@22 592
Me@24 593
Me@24 594 inline void *
Me@24 595 VMS__take_sem_reqst_from( VMSReqst *req )
Me@24 596 {
Me@24 597 return req->semReqData;
Me@24 598 }
Me@24 599
Me@24 600
Me@24 601
Me@50 602 /* This is for OS requests and VMS infrastructure requests, such as to create
Me@50 603 * a probe -- a probe is inside the heart of VMS-core, it's not part of any
Me@50 604 * language -- but it's also a semantic thing that's triggered from and used
Me@50 605 * in the application.. so it crosses abstractions.. so, need some special
Me@50 606 * pattern here for handling such requests.
Me@52 607 * Doing this just like it were a second language sharing VMS-core.
Me@52 608 *
Me@50 609 * This is called from the language's request handler when it sees a request
Me@50 610 * of type VMSSemReq
Me@52 611 *
Me@52 612 * TODO: Later change this, to give probes their own separate plugin & have
Me@52 613 * VMS-core steer the request to appropriate plugin
Me@52 614 * Do the same for OS calls -- look later at it..
Me@50 615 */
Me@50 616 void inline
Me@50 617 VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv,
Me@50 618 ResumePrFnPtr resumePrFnPtr )
Me@50 619 { VMSSemReq *semReq;
Me@50 620 IntervalProbe *newProbe;
Me@24 621
Me@50 622 semReq = req->semReqData;
Me@24 623
Me@50 624 newProbe = VMS__malloc( sizeof(IntervalProbe) );
Me@65 625 newProbe->nameStr = VMS__strDup( semReq->nameStr );
Me@50 626 newProbe->hist = NULL;
Me@50 627 newProbe->schedChoiceWasRecorded = FALSE;
Me@53 628
Me@53 629 //This runs in masterVP, so no race-condition worries
Me@50 630 newProbe->probeID =
Me@50 631 addToDynArray( newProbe, _VMSMasterEnv->dynIntervalProbesInfo );
Me@50 632
Me@53 633 requestingPr->dataRetFromReq = newProbe;
Me@50 634
Me@50 635 (*resumePrFnPtr)( requestingPr, semEnv );
Me@22 636 }
Me@22 637
Me@22 638
Me@22 639
Me@24 640 /*This must be called by the request handler plugin -- it cannot be called
Me@24 641 * from the semantic library "dissipate processor" function -- instead, the
Me@50 642 * semantic layer has to generate a request, and the plug-in calls this
Me@24 643 * function.
Me@24 644 *The reason is that this frees the virtual processor's stack -- which is
Me@24 645 * still in use inside semantic library calls!
Me@24 646 *
Me@24 647 *This frees or recycles all the state owned by and comprising the VMS
Me@24 648 * portion of the animating virtual procr. The request handler must first
Me@24 649 * free any semantic data created for the processor that didn't use the
Me@24 650 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
Me@24 651 * system to disown any state that did use VMS_malloc, and then frees the
Me@24 652 * statck and the processor-struct itself.
Me@24 653 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
Me@24 654 * state, then that state gets freed (or sent to recycling) as a side-effect
Me@24 655 * of dis-owning it.
Me@24 656 */
Me@24 657 void
Me@53 658 VMS__dissipate_procr( VirtProcr *animatingPr )
Me@24 659 {
Me@24 660 //dis-own all locations owned by this processor, causing to be freed
Me@24 661 // any locations that it is (was) sole owner of
Me@29 662 //TODO: implement VMS__malloc system, including "give up ownership"
Me@24 663
Me@24 664
Me@24 665 //NOTE: initialData was given to the processor, so should either have
Me@24 666 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@24 667 //So, all that's left to free here is the stack and the VirtProcr struc
Me@24 668 // itself
Me@50 669 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 670 // general that creating processor will outlive ones it creates.
Me@50 671 VMS__free( animatingPr->startOfStack );
Me@50 672 VMS__free( animatingPr );
Me@24 673 }
Me@24 674
Me@24 675
Me@53 676 //TODO: look at architecting cleanest separation between request handler
Me@29 677 // and master loop, for dissipate, create, shutdown, and other non-semantic
Me@29 678 // requests. Issue is chain: one removes requests from AppVP, one dispatches
Me@29 679 // on type of request, and one handles each type.. but some types require
Me@29 680 // action from both request handler and master loop -- maybe just give the
Me@29 681 // request handler calls like: VMS__handle_X_request_type
Me@24 682
Me@29 683
Me@29 684 /*This is called by the semantic layer's request handler when it decides its
Me@29 685 * time to shut down the VMS system. Calling this causes the core loop OS
Me@29 686 * threads to exit, which unblocks the entry-point function that started up
Me@29 687 * VMS, and allows it to grab the result and return to the original single-
Me@29 688 * threaded application.
Me@22 689 *
Me@29 690 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
Me@29 691 * and-wait function has to free a bunch of stuff after it detects the
Me@29 692 * threads have all died: the masterEnv, the thread-related locations,
Me@29 693 * masterVP any AppVPs that might still be allocated and sitting in the
Me@29 694 * semantic environment, or have been orphaned in the _VMSWorkQ.
Me@29 695 *
Me@53 696 *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the
Me@29 697 * locations it needs, and give ownership to masterVP. Then, they will be
Me@53 698 * automatically freed.
Me@22 699 *
Me@29 700 *In here,create one core-loop shut-down processor for each core loop and put
Me@31 701 * them all directly into the readyToAnimateQ.
Me@29 702 *Note, this function can ONLY be called after the semantic environment no
Me@29 703 * longer cares if AppVPs get animated after the point this is called. In
Me@29 704 * other words, this can be used as an abort, or else it should only be
Me@29 705 * called when all AppVPs have finished dissipate requests -- only at that
Me@29 706 * point is it sure that all results have completed.
Me@22 707 */
Me@22 708 void
Me@53 709 VMS__shutdown()
Me@8 710 { int coreIdx;
Me@14 711 VirtProcr *shutDownPr;
Me@22 712
Me@29 713 //create the shutdown processors, one for each core loop -- put them
Me@31 714 // directly into the Q -- each core will die when gets one
Me@8 715 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@50 716 { //Note, this is running in the master
Me@29 717 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
Me@55 718 writeVMSQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
Me@8 719 }
Nina@109 720 #ifdef MEAS__PERF_COUNTERS
Nina@109 721 uint64 tmpc,tmpi;
Nina@109 722 saveCyclesAndInstrs(0,tmpc,tmpi);
Nina@109 723 printf("End: cycles = %lu, instrs = %lu\n",tmpc,tmpi);
Nina@109 724 prctl(PR_TASK_PERF_EVENTS_DISABLE);
Nina@109 725 /*
Nina@109 726 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ){
Nina@109 727 close(_VMSMasterEnv->cycles_counter_fd[coreIdx]);
Nina@109 728 close(_VMSMasterEnv->instrs_counter_fd[coreIdx]);
Nina@109 729 }
Nina@109 730 */
Nina@109 731 #endif
Me@12 732 }
Me@12 733
Me@12 734
Me@29 735 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
Me@29 736 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
Me@29 737 *This function has the sole purpose of setting the stack and framePtr
Me@29 738 * to the coreLoop's stack and framePtr.. it does that then jumps to the
Me@29 739 * core loop's shutdown point -- might be able to just call Pthread_exit
Me@30 740 * from here, but am going back to the pthread's stack and setting everything
Me@29 741 * up just as if it never jumped out, before calling pthread_exit.
Me@29 742 *The end-point of core loop will free the stack and so forth of the
Me@29 743 * processor that animates this function, (this fn is transfering the
Me@29 744 * animator of the AppVP that is in turn animating this function over
Me@29 745 * to core loop function -- note that this slices out a level of virtual
Me@29 746 * processors).
Me@29 747 */
Me@29 748 void
Me@29 749 endOSThreadFn( void *initData, VirtProcr *animatingPr )
msach@71 750 {
msach@75 751 #ifdef SEQUENTIAL
msach@75 752 asmTerminateCoreLoopSeq(animatingPr);
msach@75 753 #else
msach@71 754 asmTerminateCoreLoop(animatingPr);
msach@75 755 #endif
Me@29 756 }
Me@29 757
Me@29 758
Me@53 759 /*This is called from the startup & shutdown
Me@24 760 */
Me@24 761 void
Me@53 762 VMS__cleanup_at_end_of_shutdown()
Me@31 763 {
msach@78 764 //unused
msach@78 765 //VMSQueueStruc **readyToAnimateQs;
msach@78 766 //int coreIdx;
msach@78 767 //VirtProcr **masterVPs;
msach@78 768 //SchedSlot ***allSchedSlots; //ptr to array of ptrs
Me@31 769
Me@65 770 //Before getting rid of everything, print out any measurements made
Nina@109 771 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&printHist );
Nina@109 772 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&saveHistToFile);
Me@68 773 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, &freeHistExt );
Me@65 774 #ifdef MEAS__TIME_PLUGIN
Me@68 775 printHist( _VMSMasterEnv->reqHdlrLowTimeHist );
msach@79 776 saveHistToFile( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 777 printHist( _VMSMasterEnv->reqHdlrHighTimeHist );
msach@79 778 saveHistToFile( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@68 779 freeHistExt( _VMSMasterEnv->reqHdlrLowTimeHist );
Me@68 780 freeHistExt( _VMSMasterEnv->reqHdlrHighTimeHist );
Me@65 781 #endif
Me@65 782 #ifdef MEAS__TIME_MALLOC
Me@65 783 printHist( _VMSMasterEnv->mallocTimeHist );
msach@79 784 saveHistToFile( _VMSMasterEnv->mallocTimeHist );
Me@65 785 printHist( _VMSMasterEnv->freeTimeHist );
msach@79 786 saveHistToFile( _VMSMasterEnv->freeTimeHist );
Me@65 787 freeHistExt( _VMSMasterEnv->mallocTimeHist );
Me@65 788 freeHistExt( _VMSMasterEnv->freeTimeHist );
Me@65 789 #endif
Me@65 790 #ifdef MEAS__TIME_MASTER_LOCK
Me@65 791 printHist( _VMSMasterEnv->masterLockLowTimeHist );
Me@65 792 printHist( _VMSMasterEnv->masterLockHighTimeHist );
Me@65 793 #endif
Me@65 794 #ifdef MEAS__TIME_MASTER
Me@65 795 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 796 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 797 {
Me@65 798 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 799 //master VPs were created external to VMS, so use external free
Me@65 800 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 801
Me@65 802 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 803 }
Me@65 804 #endif
Me@65 805 #ifdef MEAS__TIME_STAMP_SUSP
Me@65 806 printHist( _VMSMasterEnv->pluginTimeHist );
Me@65 807 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@65 808 {
Me@65 809 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@65 810 //master VPs were created external to VMS, so use external free
Me@65 811 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@65 812
Me@65 813 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@65 814 }
Me@65 815 #endif
Me@65 816
Me@53 817 //All the environment data has been allocated with VMS__malloc, so just
Me@53 818 // free its internal big-chunk and all inside it disappear.
Me@53 819 /*
Me@31 820 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
Me@31 821 masterVPs = _VMSMasterEnv->masterVPs;
Me@31 822 allSchedSlots = _VMSMasterEnv->allSchedSlots;
Me@31 823
Me@31 824 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@24 825 {
Me@55 826 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@50 827 //master VPs were created external to VMS, so use external free
Me@53 828 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@31 829
Me@31 830 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@24 831 }
Me@31 832
Me@53 833 VMS__free( _VMSMasterEnv->readyToAnimateQs );
Me@53 834 VMS__free( _VMSMasterEnv->masterVPs );
Me@53 835 VMS__free( _VMSMasterEnv->allSchedSlots );
Me@50 836
Me@50 837 //============================= MEASUREMENT STUFF ========================
Me@50 838 #ifdef STATS__TURN_ON_PROBES
Me@53 839 freeDynArrayDeep( _VMSMasterEnv->dynIntervalProbesInfo, &VMS__free_probe);
Me@50 840 #endif
Me@50 841 //========================================================================
Me@53 842 */
Me@53 843 //These are the only two that use system free
Me@53 844 VMS_ext__free_free_list( _VMSMasterEnv->freeListHead );
Me@53 845 free( (void *)_VMSMasterEnv );
Me@24 846 }
Me@24 847
Me@54 848
Me@54 849 //================================
Me@54 850
Me@54 851
Me@54 852 /*Later, improve this -- for now, just exits the application after printing
Me@54 853 * the error message.
Me@54 854 */
Me@54 855 void
Me@54 856 VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData )
Me@54 857 {
msach@69 858 printf("%s",msgStr);
Me@54 859 fflush(stdin);
Me@54 860 exit(1);
Me@54 861 }
Me@54 862