annotate VMS.c @ 55:3bac84e4e56e

Works with correct matrix mult Nov 4 -- switch animators macros, many updates Changed all queues back to VMSQ variants #defines correct, protected, work-stealing, with compiler switch in and out
author Me
date Thu, 04 Nov 2010 18:13:18 -0700
parents f8508572f3de
children 26d53313a8f2
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>
Me@50 11 #include <sys/time.h>
Me@0 12
Me@0 13 #include "VMS.h"
Me@0 14 #include "Queue_impl/BlockingQueue.h"
Me@38 15 #include "Histogram/Histogram.h"
Me@0 16
Me@0 17
Me@26 18 #define thdAttrs NULL
Me@26 19
Me@22 20 //===========================================================================
Me@22 21 void
Me@22 22 shutdownFn( void *dummy, VirtProcr *dummy2 );
Me@22 23
Me@31 24 SchedSlot **
Me@31 25 create_sched_slots();
Me@22 26
Me@28 27 void
Me@28 28 create_masterEnv();
Me@28 29
Me@28 30 void
Me@28 31 create_the_coreLoop_OS_threads();
Me@28 32
Me@50 33 MallocProlog *
Me@50 34 create_free_list();
Me@50 35
Me@53 36 void
Me@53 37 endOSThreadFn( void *initData, VirtProcr *animatingPr );
Me@50 38
Me@26 39 pthread_mutex_t suspendLock = PTHREAD_MUTEX_INITIALIZER;
Me@26 40 pthread_cond_t suspend_cond = PTHREAD_COND_INITIALIZER;
Me@26 41
Me@22 42 //===========================================================================
Me@22 43
Me@0 44 /*Setup has two phases:
Me@0 45 * 1) Semantic layer first calls init_VMS, which creates masterEnv, and puts
Me@8 46 * the master virt procr into the work-queue, ready for first "call"
Me@8 47 * 2) Semantic layer then does its own init, which creates the seed virt
Me@8 48 * procr inside the semantic layer, ready to schedule it when
Me@0 49 * asked by the first run of the masterLoop.
Me@0 50 *
Me@0 51 *This part is bit weird because VMS really wants to be "always there", and
Me@0 52 * have applications attach and detach.. for now, this VMS is part of
Me@0 53 * the app, so the VMS system starts up as part of running the app.
Me@0 54 *
Me@8 55 *The semantic layer is isolated from the VMS internals by making the
Me@8 56 * semantic layer do setup to a state that it's ready with its
Me@8 57 * initial virt procrs, ready to schedule them to slots when the masterLoop
Me@0 58 * asks. Without this pattern, the semantic layer's setup would
Me@8 59 * have to modify slots directly to assign the initial virt-procrs, and put
Me@31 60 * them into the readyToAnimateQ itself, breaking the isolation completely.
Me@0 61 *
Me@0 62 *
Me@8 63 *The semantic layer creates the initial virt procr(s), and adds its
Me@8 64 * own environment to masterEnv, and fills in the pointers to
Me@0 65 * the requestHandler and slaveScheduler plug-in functions
Me@8 66 */
Me@8 67
Me@8 68 /*This allocates VMS data structures, populates the master VMSProc,
Me@0 69 * and master environment, and returns the master environment to the semantic
Me@0 70 * layer.
Me@0 71 */
Me@8 72 void
Me@8 73 VMS__init()
Me@28 74 {
Me@28 75 create_masterEnv();
Me@28 76 create_the_coreLoop_OS_threads();
Me@28 77 }
Me@28 78
Me@28 79 /*To initialize the sequential version, just don't create the threads
Me@28 80 */
Me@28 81 void
Me@28 82 VMS__init_Seq()
Me@28 83 {
Me@28 84 create_masterEnv();
Me@28 85 }
Me@28 86
Me@28 87 void
Me@28 88 create_masterEnv()
Me@31 89 { MasterEnv *masterEnv;
Me@55 90 VMSQueueStruc **readyToAnimateQs;
Me@31 91 int coreIdx;
Me@31 92 VirtProcr **masterVPs;
Me@31 93 SchedSlot ***allSchedSlots; //ptr to array of ptrs
Me@53 94
Me@53 95
Me@31 96 //Make the master env, which holds everything else
Me@1 97 _VMSMasterEnv = malloc( sizeof(MasterEnv) );
Me@53 98
Me@53 99 //Very first thing put into the master env is the free-list, seeded
Me@53 100 // with a massive initial chunk of memory.
Me@53 101 //After this, all other mallocs are VMS__malloc.
Me@53 102 _VMSMasterEnv->freeListHead = VMS_ext__create_free_list();
Me@53 103
Me@53 104 //===================== Only VMS__malloc after this ====================
Me@1 105 masterEnv = _VMSMasterEnv;
Me@31 106
Me@31 107 //Make a readyToAnimateQ for each core loop
Me@55 108 readyToAnimateQs = VMS__malloc( NUM_CORES * sizeof(VMSQueueStruc *) );
Me@53 109 masterVPs = VMS__malloc( NUM_CORES * sizeof(VirtProcr *) );
Me@0 110
Me@31 111 //One array for each core, 3 in array, core's masterVP scheds all
Me@53 112 allSchedSlots = VMS__malloc( NUM_CORES * sizeof(SchedSlot *) );
Me@0 113
Me@53 114 _VMSMasterEnv->numProcrsCreated = 0; //used by create procr
Me@31 115 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@53 116 {
Me@55 117 readyToAnimateQs[ coreIdx ] = makeVMSQ();
Me@31 118
Me@50 119 //Q: should give masterVP core-specific info as its init data?
Me@53 120 masterVPs[ coreIdx ] = VMS__create_procr( &masterLoop, masterEnv );
Me@31 121 masterVPs[ coreIdx ]->coreAnimatedBy = coreIdx;
Me@31 122 allSchedSlots[ coreIdx ] = create_sched_slots(); //makes for one core
Me@53 123 _VMSMasterEnv->numMasterInARow[ coreIdx ] = 0;
Me@55 124 _VMSMasterEnv->workStealingGates[ coreIdx ] = NULL;
Me@31 125 }
Me@31 126 _VMSMasterEnv->readyToAnimateQs = readyToAnimateQs;
Me@31 127 _VMSMasterEnv->masterVPs = masterVPs;
Me@50 128 _VMSMasterEnv->masterLock = UNLOCKED;
Me@31 129 _VMSMasterEnv->allSchedSlots = allSchedSlots;
Me@55 130 _VMSMasterEnv->workStealingLock = UNLOCKED;
Me@28 131
Me@12 132
Me@31 133 //Aug 19, 2010: no longer need to place initial masterVP into queue
Me@31 134 // because coreLoop now controls -- animates its masterVP when no work
Me@31 135
Me@30 136
Me@50 137 //============================= MEASUREMENT STUFF ========================
Me@50 138 #ifdef STATS__TURN_ON_PROBES
Me@50 139 _VMSMasterEnv->dynIntervalProbesInfo =
Me@53 140 makePrivDynArrayOfSize( &(_VMSMasterEnv->intervalProbes), 200);
Me@30 141
Me@53 142 _VMSMasterEnv->probeNameHashTbl = makeHashTable( 1000, &VMS__free );
Me@53 143
Me@53 144 //put creation time directly into master env, for fast retrieval
Me@50 145 struct timeval timeStamp;
Me@50 146 gettimeofday( &(timeStamp), NULL);
Me@50 147 _VMSMasterEnv->createPtInSecs =
Me@50 148 timeStamp.tv_sec +(timeStamp.tv_usec/1000000.0);
Me@50 149 #endif
Me@50 150 //========================================================================
Me@38 151
Me@0 152 }
Me@0 153
Me@31 154 SchedSlot **
Me@31 155 create_sched_slots()
Me@31 156 { SchedSlot **schedSlots;
Me@0 157 int i;
Me@0 158
Me@53 159 schedSlots = VMS__malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
Me@8 160
Me@1 161 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@0 162 {
Me@53 163 schedSlots[i] = VMS__malloc( sizeof(SchedSlot) );
Me@8 164
Me@1 165 //Set state to mean "handling requests done, slot needs filling"
Me@8 166 schedSlots[i]->workIsDone = FALSE;
Me@8 167 schedSlots[i]->needsProcrAssigned = TRUE;
Me@0 168 }
Me@31 169 return schedSlots;
Me@31 170 }
Me@31 171
Me@31 172
Me@31 173 void
Me@31 174 freeSchedSlots( SchedSlot **schedSlots )
Me@31 175 { int i;
Me@31 176 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
Me@31 177 {
Me@53 178 VMS__free( schedSlots[i] );
Me@31 179 }
Me@53 180 VMS__free( schedSlots );
Me@0 181 }
Me@0 182
Me@8 183
Me@28 184 void
Me@28 185 create_the_coreLoop_OS_threads()
Me@28 186 {
Me@28 187 //========================================================================
Me@28 188 // Create the Threads
Me@28 189 int coreIdx, retCode;
Me@28 190
Me@28 191 //Need the threads to be created suspended, and wait for a signal
Me@28 192 // before proceeding -- gives time after creating to initialize other
Me@28 193 // stuff before the coreLoops set off.
Me@28 194 _VMSMasterEnv->setupComplete = 0;
Me@28 195
Me@28 196 //Make the threads that animate the core loops
Me@28 197 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@53 198 { coreLoopThdParams[coreIdx] = VMS__malloc( sizeof(ThdParams) );
Me@28 199 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
Me@28 200
Me@28 201 retCode =
Me@28 202 pthread_create( &(coreLoopThdHandles[coreIdx]),
Me@28 203 thdAttrs,
Me@28 204 &coreLoop,
Me@28 205 (void *)(coreLoopThdParams[coreIdx]) );
Me@50 206 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
Me@28 207 }
Me@28 208 }
Me@28 209
Me@0 210 /*Semantic layer calls this when it want the system to start running..
Me@0 211 *
Me@24 212 *This starts the core loops running then waits for them to exit.
Me@0 213 */
Me@12 214 void
Me@24 215 VMS__start_the_work_then_wait_until_done()
Me@12 216 { int coreIdx;
Me@24 217 //Start the core loops running
Me@25 218
Me@25 219 //tell the core loop threads that setup is complete
Me@25 220 //get lock, to lock out any threads still starting up -- they'll see
Me@25 221 // that setupComplete is true before entering while loop, and so never
Me@25 222 // wait on the condition
Me@26 223 pthread_mutex_lock( &suspendLock );
Me@25 224 _VMSMasterEnv->setupComplete = 1;
Me@26 225 pthread_mutex_unlock( &suspendLock );
Me@26 226 pthread_cond_broadcast( &suspend_cond );
Me@25 227
Me@25 228
Me@24 229 //wait for all to complete
Me@8 230 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@8 231 {
Me@25 232 pthread_join( coreLoopThdHandles[coreIdx], NULL );
Me@24 233 }
Me@25 234
Me@24 235 //NOTE: do not clean up VMS env here -- semantic layer has to have
Me@24 236 // a chance to clean up its environment first, then do a call to free
Me@24 237 // the Master env and rest of VMS locations
Me@8 238 }
Me@0 239
Me@28 240 /*Only difference between version with an OS thread pinned to each core and
Me@28 241 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
Me@28 242 */
Me@28 243 void
Me@28 244 VMS__start_the_work_then_wait_until_done_Seq()
Me@28 245 {
Me@28 246 //Instead of un-suspending threads, just call the one and only
Me@28 247 // core loop (sequential version), in the main thread.
Me@28 248 coreLoop_Seq( NULL );
Me@28 249
Me@28 250 }
Me@28 251
Me@0 252
Me@0 253
Me@8 254 /*Create stack, then create __cdecl structure on it and put initialData and
Me@8 255 * pointer to the new structure instance into the parameter positions on
Me@8 256 * the stack
Me@8 257 *Then put function pointer into nextInstrPt -- the stack is setup in std
Me@8 258 * call structure, so jumping to function ptr is same as a GCC generated
Me@8 259 * function call
Me@8 260 *No need to save registers on old stack frame, because there's no old
Me@8 261 * animator state to return to --
Me@8 262 *
Me@8 263 */
Me@50 264 inline VirtProcr *
Me@50 265 create_procr_helper( VirtProcr *newPr, VirtProcrFnPtr fnPtr,
Me@50 266 void *initialData, char *stackLocs )
Me@50 267 {
Me@50 268 char *stackPtr;
Me@8 269
Me@53 270 newPr->startOfStack = stackLocs;
Me@53 271 newPr->procrID = _VMSMasterEnv->numProcrsCreated++;
Me@53 272 newPr->nextInstrPt = fnPtr;
Me@53 273 newPr->initialData = initialData;
Me@53 274 newPr->requests = NULL;
Me@53 275 newPr->schedSlot = NULL;
Me@8 276
Me@14 277 //fnPtr takes two params -- void *initData & void *animProcr
Me@8 278 //alloc stack locations, make stackPtr be the highest addr minus room
Me@14 279 // for 2 params + return addr. Return addr (NULL) is in loc pointed to
Me@14 280 // by stackPtr, initData at stackPtr + 4 bytes, animatingPr just above
Me@22 281 stackPtr = ( (char *)stackLocs + VIRT_PROCR_STACK_SIZE - 0x10 );
Me@50 282
Me@8 283 //setup __cdecl on stack -- coreloop will switch to stackPtr before jmp
Me@22 284 *( (int *)stackPtr + 2 ) = (int) newPr; //rightmost param -- 32bit pointer
Me@14 285 *( (int *)stackPtr + 1 ) = (int) initialData; //next param to left
Me@8 286 newPr->stackPtr = stackPtr; //core loop will switch to this, then
Me@8 287 newPr->framePtr = stackPtr; //suspend loop will save new stack & frame ptr
Me@8 288
Me@50 289 //============================= MEASUREMENT STUFF ========================
Me@50 290 #ifdef STATS__TURN_ON_PROBES
Me@50 291 struct timeval timeStamp;
Me@50 292 gettimeofday( &(timeStamp), NULL);
Me@54 293 newPr->createPtInSecs = timeStamp.tv_sec +(timeStamp.tv_usec/1000000.0) -
Me@54 294 _VMSMasterEnv->createPtInSecs;
Me@50 295 #endif
Me@50 296 //========================================================================
Me@50 297
Me@8 298 return newPr;
Me@8 299 }
Me@8 300
Me@50 301 inline VirtProcr *
Me@50 302 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 303 { VirtProcr *newPr;
Me@50 304 char *stackLocs;
Me@50 305
Me@50 306 newPr = VMS__malloc( sizeof(VirtProcr) );
Me@50 307 stackLocs = VMS__malloc( VIRT_PROCR_STACK_SIZE );
Me@50 308 if( stackLocs == 0 )
Me@50 309 { perror("VMS__malloc stack"); exit(1); }
Me@50 310
Me@50 311 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 312 }
Me@50 313
Me@50 314 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 315 * be called from main thread or other thread -- never from code animated by
Me@50 316 * a VMS virtual processor.
Me@50 317 */
Me@50 318 inline VirtProcr *
Me@50 319 VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
Me@50 320 { VirtProcr *newPr;
Me@50 321 char *stackLocs;
Me@50 322
Me@50 323 newPr = malloc( sizeof(VirtProcr) );
Me@50 324 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
Me@50 325 if( stackLocs == 0 )
Me@50 326 { perror("malloc stack"); exit(1); }
Me@50 327
Me@50 328 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
Me@50 329 }
Me@50 330
Me@8 331
Me@26 332 /*there is a label inside this function -- save the addr of this label in
Me@0 333 * the callingPr struc, as the pick-up point from which to start the next
Me@0 334 * work-unit for that procr. If turns out have to save registers, then
Me@0 335 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
Me@0 336 * "done with work-unit" label. The procr struc is in the request in the
Me@0 337 * slave that animated the just-ended work-unit, so all the state is saved
Me@0 338 * there, and will get passed along, inside the request handler, to the
Me@0 339 * next work-unit for that procr.
Me@0 340 */
Me@8 341 void
Me@38 342 VMS__suspend_procr( VirtProcr *animatingPr )
Me@55 343 {
Me@0 344
Me@14 345 //The request to master will cause this suspended virt procr to get
Me@14 346 // scheduled again at some future point -- to resume, core loop jumps
Me@14 347 // to the resume point (below), which causes restore of saved regs and
Me@14 348 // "return" from this call.
Me@38 349 animatingPr->nextInstrPt = &&ResumePt;
Me@1 350
Me@1 351 //return ownership of the virt procr and sched slot to Master virt pr
Me@38 352 animatingPr->schedSlot->workIsDone = TRUE;
Me@1 353
Me@41 354 //=========================== Measurement stuff ========================
Me@38 355 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 356 //record time stamp: compare to time-stamp recorded below
Me@38 357 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
Me@38 358 #endif
Me@41 359 //=======================================================================
Me@38 360
Me@1 361
Me@55 362 SwitchToCoreLoop( animatingPr )
Me@55 363
Me@55 364 //=======================================================================
Me@1 365 ResumePt:
Me@38 366 #ifdef MEAS__TIME_STAMP_SUSP
Me@41 367 //NOTE: only take low part of count -- do sanity check when take diff
Me@38 368 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
Me@38 369 #endif
Me@38 370
Me@0 371 return;
Me@0 372 }
Me@0 373
Me@22 374
Me@22 375
Me@50 376 /*For this implementation of VMS, it may not make much sense to have the
Me@50 377 * system of requests for creating a new processor done this way.. but over
Me@50 378 * the scope of single-master, multi-master, mult-tasking, OS-implementing,
Me@50 379 * distributed-memory, and so on, this gives VMS implementation a chance to
Me@50 380 * do stuff before suspend, in the AppVP, and in the Master before the plugin
Me@50 381 * is called, as well as in the lang-lib before this is called, and in the
Me@50 382 * plugin. So, this gives both VMS and language implementations a chance to
Me@50 383 * intercept at various points and do order-dependent stuff.
Me@50 384 *Having a standard VMSNewPrReqData struc allows the language to create and
Me@50 385 * free the struc, while VMS knows how to get the newPr if it wants it, and
Me@50 386 * it lets the lang have lang-specific data related to creation transported
Me@50 387 * to the plugin.
Me@50 388 */
Me@50 389 void
Me@50 390 VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr )
Me@50 391 { VMSReqst req;
Me@50 392
Me@50 393 req.reqType = createReq;
Me@50 394 req.semReqData = semReqData;
Me@50 395 req.nextReqst = reqstingPr->requests;
Me@50 396 reqstingPr->requests = &req;
Me@50 397
Me@50 398 VMS__suspend_procr( reqstingPr );
Me@50 399 }
Me@50 400
Me@22 401
Me@38 402 /*
Me@22 403 *This adds a request to dissipate, then suspends the processor so that the
Me@22 404 * request handler will receive the request. The request handler is what
Me@22 405 * does the work of freeing memory and removing the processor from the
Me@22 406 * semantic environment's data structures.
Me@22 407 *The request handler also is what figures out when to shutdown the VMS
Me@22 408 * system -- which causes all the core loop threads to die, and returns from
Me@22 409 * the call that started up VMS to perform the work.
Me@22 410 *
Me@22 411 *This form is a bit misleading to understand if one is trying to figure out
Me@22 412 * how VMS works -- it looks like a normal function call, but inside it
Me@22 413 * sends a request to the request handler and suspends the processor, which
Me@22 414 * jumps out of the VMS__dissipate_procr function, and out of all nestings
Me@22 415 * above it, transferring the work of dissipating to the request handler,
Me@22 416 * which then does the actual work -- causing the processor that animated
Me@22 417 * the call of this function to disappear and the "hanging" state of this
Me@22 418 * function to just poof into thin air -- the virtual processor's trace
Me@22 419 * never returns from this call, but instead the virtual processor's trace
Me@22 420 * gets suspended in this call and all the virt processor's state disap-
Me@22 421 * pears -- making that suspend the last thing in the virt procr's trace.
Me@8 422 */
Me@8 423 void
Me@53 424 VMS__send_dissipate_req( VirtProcr *procrToDissipate )
Me@50 425 { VMSReqst req;
Me@22 426
Me@50 427 req.reqType = dissipate;
Me@50 428 req.nextReqst = procrToDissipate->requests;
Me@50 429 procrToDissipate->requests = &req;
Me@50 430
Me@22 431 VMS__suspend_procr( procrToDissipate );
Me@50 432 }
Me@50 433
Me@50 434
Me@50 435 /* "ext" designates that it's for use outside the VMS system -- should only
Me@50 436 * be called from main thread or other thread -- never from code animated by
Me@50 437 * a VMS virtual processor.
Me@50 438 *
Me@50 439 *Use this version to dissipate VPs created outside the VMS system.
Me@50 440 */
Me@50 441 void
Me@50 442 VMS_ext__dissipate_procr( VirtProcr *procrToDissipate )
Me@50 443 {
Me@50 444 //NOTE: initialData was given to the processor, so should either have
Me@50 445 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@50 446 //So, all that's left to free here is the stack and the VirtProcr struc
Me@50 447 // itself
Me@50 448 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 449 // general that creating processor will outlive ones it creates.
Me@50 450 free( procrToDissipate->startOfStack );
Me@50 451 free( procrToDissipate );
Me@50 452 }
Me@50 453
Me@22 454
Me@22 455
Me@53 456 /*This call's name indicates that request is malloc'd -- so req handler
Me@53 457 * has to free any extra requests tacked on before a send, using this.
Me@53 458 *
Me@53 459 * This inserts the semantic-layer's request data into standard VMS carrier
Me@53 460 * request data-struct that is mallocd. The sem request doesn't need to
Me@53 461 * be malloc'd if this is called inside the same call chain before the
Me@53 462 * send of the last request is called.
Me@53 463 *
Me@53 464 *The request handler has to call VMS__free_VMSReq for any of these
Me@22 465 */
Me@22 466 inline void
Me@53 467 VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData,
Me@53 468 VirtProcr *callingPr )
Me@53 469 { VMSReqst *req;
Me@22 470
Me@53 471 req = VMS__malloc( sizeof(VMSReqst) );
Me@53 472 req->reqType = semantic;
Me@53 473 req->semReqData = semReqData;
Me@53 474 req->nextReqst = callingPr->requests;
Me@53 475 callingPr->requests = req;
Me@22 476 }
Me@22 477
Me@50 478 /*This inserts the semantic-layer's request data into standard VMS carrier
Me@50 479 * request data-struct is allocated on stack of this call & ptr to it sent
Me@50 480 * to plugin
Me@50 481 *Then it does suspend, to cause request to be sent.
Me@50 482 */
Me@50 483 inline void
Me@50 484 VMS__send_sem_request( void *semReqData, VirtProcr *callingPr )
Me@50 485 { VMSReqst req;
Me@22 486
Me@50 487 req.reqType = semantic;
Me@50 488 req.semReqData = semReqData;
Me@50 489 req.nextReqst = callingPr->requests;
Me@50 490 callingPr->requests = &req;
Me@50 491
Me@50 492 VMS__suspend_procr( callingPr );
Me@50 493 }
Me@50 494
Me@50 495
Me@50 496 inline void
Me@50 497 VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr )
Me@50 498 { VMSReqst req;
Me@50 499
Me@50 500 req.reqType = VMSSemantic;
Me@50 501 req.semReqData = semReqData;
Me@50 502 req.nextReqst = callingPr->requests; //gab any other preceeding
Me@50 503 callingPr->requests = &req;
Me@50 504
Me@50 505 VMS__suspend_procr( callingPr );
Me@50 506 }
Me@50 507
Me@50 508
Me@50 509 /*
Me@38 510 */
Me@24 511 VMSReqst *
Me@50 512 VMS__take_next_request_out_of( VirtProcr *procrWithReq )
Me@31 513 { VMSReqst *req;
Me@31 514
Me@31 515 req = procrWithReq->requests;
Me@38 516 if( req == NULL ) return NULL;
Me@31 517
Me@31 518 procrWithReq->requests = procrWithReq->requests->nextReqst;
Me@50 519 return req;
Me@24 520 }
Me@22 521
Me@24 522
Me@24 523 inline void *
Me@24 524 VMS__take_sem_reqst_from( VMSReqst *req )
Me@24 525 {
Me@24 526 return req->semReqData;
Me@24 527 }
Me@24 528
Me@24 529
Me@24 530
Me@50 531 /* This is for OS requests and VMS infrastructure requests, such as to create
Me@50 532 * a probe -- a probe is inside the heart of VMS-core, it's not part of any
Me@50 533 * language -- but it's also a semantic thing that's triggered from and used
Me@50 534 * in the application.. so it crosses abstractions.. so, need some special
Me@50 535 * pattern here for handling such requests.
Me@52 536 * Doing this just like it were a second language sharing VMS-core.
Me@52 537 *
Me@50 538 * This is called from the language's request handler when it sees a request
Me@50 539 * of type VMSSemReq
Me@52 540 *
Me@52 541 * TODO: Later change this, to give probes their own separate plugin & have
Me@52 542 * VMS-core steer the request to appropriate plugin
Me@52 543 * Do the same for OS calls -- look later at it..
Me@50 544 */
Me@50 545 void inline
Me@50 546 VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv,
Me@50 547 ResumePrFnPtr resumePrFnPtr )
Me@50 548 { VMSSemReq *semReq;
Me@50 549 IntervalProbe *newProbe;
Me@50 550 int32 nameLen;
Me@24 551
Me@50 552 semReq = req->semReqData;
Me@24 553
Me@50 554 newProbe = VMS__malloc( sizeof(IntervalProbe) );
Me@50 555 nameLen = strlen( semReq->nameStr );
Me@50 556 newProbe->nameStr = VMS__malloc( nameLen );
Me@50 557 memcpy( newProbe->nameStr, semReq->nameStr, nameLen );
Me@50 558 newProbe->hist = NULL;
Me@50 559 newProbe->schedChoiceWasRecorded = FALSE;
Me@53 560
Me@53 561 //This runs in masterVP, so no race-condition worries
Me@50 562 newProbe->probeID =
Me@50 563 addToDynArray( newProbe, _VMSMasterEnv->dynIntervalProbesInfo );
Me@50 564
Me@53 565 requestingPr->dataRetFromReq = newProbe;
Me@50 566
Me@50 567 (*resumePrFnPtr)( requestingPr, semEnv );
Me@22 568 }
Me@22 569
Me@22 570
Me@22 571
Me@24 572 /*This must be called by the request handler plugin -- it cannot be called
Me@24 573 * from the semantic library "dissipate processor" function -- instead, the
Me@50 574 * semantic layer has to generate a request, and the plug-in calls this
Me@24 575 * function.
Me@24 576 *The reason is that this frees the virtual processor's stack -- which is
Me@24 577 * still in use inside semantic library calls!
Me@24 578 *
Me@24 579 *This frees or recycles all the state owned by and comprising the VMS
Me@24 580 * portion of the animating virtual procr. The request handler must first
Me@24 581 * free any semantic data created for the processor that didn't use the
Me@24 582 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
Me@24 583 * system to disown any state that did use VMS_malloc, and then frees the
Me@24 584 * statck and the processor-struct itself.
Me@24 585 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
Me@24 586 * state, then that state gets freed (or sent to recycling) as a side-effect
Me@24 587 * of dis-owning it.
Me@24 588 */
Me@24 589 void
Me@53 590 VMS__dissipate_procr( VirtProcr *animatingPr )
Me@24 591 {
Me@24 592 //dis-own all locations owned by this processor, causing to be freed
Me@24 593 // any locations that it is (was) sole owner of
Me@29 594 //TODO: implement VMS__malloc system, including "give up ownership"
Me@24 595
Me@24 596
Me@24 597 //NOTE: initialData was given to the processor, so should either have
Me@24 598 // been alloc'd with VMS__malloc, or freed by the level above animPr.
Me@24 599 //So, all that's left to free here is the stack and the VirtProcr struc
Me@24 600 // itself
Me@50 601 //Note, should not stack-allocate initial data -- no guarantee, in
Me@50 602 // general that creating processor will outlive ones it creates.
Me@50 603 VMS__free( animatingPr->startOfStack );
Me@50 604 VMS__free( animatingPr );
Me@24 605 }
Me@24 606
Me@24 607
Me@53 608 //TODO: look at architecting cleanest separation between request handler
Me@29 609 // and master loop, for dissipate, create, shutdown, and other non-semantic
Me@29 610 // requests. Issue is chain: one removes requests from AppVP, one dispatches
Me@29 611 // on type of request, and one handles each type.. but some types require
Me@29 612 // action from both request handler and master loop -- maybe just give the
Me@29 613 // request handler calls like: VMS__handle_X_request_type
Me@24 614
Me@29 615
Me@29 616 /*This is called by the semantic layer's request handler when it decides its
Me@29 617 * time to shut down the VMS system. Calling this causes the core loop OS
Me@29 618 * threads to exit, which unblocks the entry-point function that started up
Me@29 619 * VMS, and allows it to grab the result and return to the original single-
Me@29 620 * threaded application.
Me@22 621 *
Me@29 622 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
Me@29 623 * and-wait function has to free a bunch of stuff after it detects the
Me@29 624 * threads have all died: the masterEnv, the thread-related locations,
Me@29 625 * masterVP any AppVPs that might still be allocated and sitting in the
Me@29 626 * semantic environment, or have been orphaned in the _VMSWorkQ.
Me@29 627 *
Me@53 628 *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the
Me@29 629 * locations it needs, and give ownership to masterVP. Then, they will be
Me@53 630 * automatically freed.
Me@22 631 *
Me@29 632 *In here,create one core-loop shut-down processor for each core loop and put
Me@31 633 * them all directly into the readyToAnimateQ.
Me@29 634 *Note, this function can ONLY be called after the semantic environment no
Me@29 635 * longer cares if AppVPs get animated after the point this is called. In
Me@29 636 * other words, this can be used as an abort, or else it should only be
Me@29 637 * called when all AppVPs have finished dissipate requests -- only at that
Me@29 638 * point is it sure that all results have completed.
Me@22 639 */
Me@22 640 void
Me@53 641 VMS__shutdown()
Me@8 642 { int coreIdx;
Me@14 643 VirtProcr *shutDownPr;
Me@22 644
Me@29 645 //create the shutdown processors, one for each core loop -- put them
Me@31 646 // directly into the Q -- each core will die when gets one
Me@8 647 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
Me@50 648 { //Note, this is running in the master
Me@29 649 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
Me@55 650 writeVMSQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
Me@8 651 }
Me@22 652
Me@12 653 }
Me@12 654
Me@12 655
Me@29 656 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
Me@29 657 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
Me@29 658 *This function has the sole purpose of setting the stack and framePtr
Me@29 659 * to the coreLoop's stack and framePtr.. it does that then jumps to the
Me@29 660 * core loop's shutdown point -- might be able to just call Pthread_exit
Me@30 661 * from here, but am going back to the pthread's stack and setting everything
Me@29 662 * up just as if it never jumped out, before calling pthread_exit.
Me@29 663 *The end-point of core loop will free the stack and so forth of the
Me@29 664 * processor that animates this function, (this fn is transfering the
Me@29 665 * animator of the AppVP that is in turn animating this function over
Me@29 666 * to core loop function -- note that this slices out a level of virtual
Me@29 667 * processors).
Me@29 668 */
Me@29 669 void
Me@29 670 endOSThreadFn( void *initData, VirtProcr *animatingPr )
Me@29 671 { void *jmpPt, *coreLoopStackPtr, *coreLoopFramePtr;
Me@29 672
Me@29 673 jmpPt = _VMSMasterEnv->coreLoopEndPt;
Me@29 674 coreLoopStackPtr = animatingPr->coreLoopStackPtr;
Me@29 675 coreLoopFramePtr = animatingPr->coreLoopFramePtr;
Me@29 676
Me@29 677
Me@29 678 asm volatile("movl %0, %%eax; \
Me@29 679 movl %1, %%esp; \
Me@29 680 movl %2, %%ebp; \
Me@29 681 jmp %%eax " \
Me@29 682 /* outputs */ : \
Me@29 683 /* inputs */ : "m" (jmpPt), "m"(coreLoopStackPtr), "m"(coreLoopFramePtr)\
Me@29 684 /* clobber */ : "memory", "%eax", "%ebx", "%ecx", "%edx", "%edi","%esi" \
Me@29 685 );
Me@29 686 }
Me@29 687
Me@29 688
Me@53 689 /*This is called from the startup & shutdown
Me@24 690 */
Me@24 691 void
Me@53 692 VMS__cleanup_at_end_of_shutdown()
Me@31 693 {
Me@55 694 VMSQueueStruc **readyToAnimateQs;
Me@31 695 int coreIdx;
Me@31 696 VirtProcr **masterVPs;
Me@31 697 SchedSlot ***allSchedSlots; //ptr to array of ptrs
Me@31 698
Me@53 699 //All the environment data has been allocated with VMS__malloc, so just
Me@53 700 // free its internal big-chunk and all inside it disappear.
Me@53 701 /*
Me@31 702 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
Me@31 703 masterVPs = _VMSMasterEnv->masterVPs;
Me@31 704 allSchedSlots = _VMSMasterEnv->allSchedSlots;
Me@31 705
Me@31 706 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
Me@24 707 {
Me@55 708 freeVMSQ( readyToAnimateQs[ coreIdx ] );
Me@50 709 //master VPs were created external to VMS, so use external free
Me@53 710 VMS__dissipate_procr( masterVPs[ coreIdx ] );
Me@31 711
Me@31 712 freeSchedSlots( allSchedSlots[ coreIdx ] );
Me@24 713 }
Me@31 714
Me@53 715 VMS__free( _VMSMasterEnv->readyToAnimateQs );
Me@53 716 VMS__free( _VMSMasterEnv->masterVPs );
Me@53 717 VMS__free( _VMSMasterEnv->allSchedSlots );
Me@50 718
Me@50 719 //============================= MEASUREMENT STUFF ========================
Me@50 720 #ifdef STATS__TURN_ON_PROBES
Me@53 721 freeDynArrayDeep( _VMSMasterEnv->dynIntervalProbesInfo, &VMS__free_probe);
Me@50 722 #endif
Me@50 723 //========================================================================
Me@53 724 */
Me@53 725 //These are the only two that use system free
Me@53 726 VMS_ext__free_free_list( _VMSMasterEnv->freeListHead );
Me@53 727 free( (void *)_VMSMasterEnv );
Me@24 728 }
Me@24 729
Me@54 730
Me@54 731 //================================
Me@54 732
Me@54 733
Me@54 734 /*Later, improve this -- for now, just exits the application after printing
Me@54 735 * the error message.
Me@54 736 */
Me@54 737 void
Me@54 738 VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData )
Me@54 739 {
Me@54 740 printf(msgStr);
Me@54 741 fflush(stdin);
Me@54 742 exit(1);
Me@54 743 }
Me@54 744
Me@54 745
Me@54 746