VMS/VMS_Implementations/SSR_impls/SSR__MC_shared_impl

annotate SSR.c @ 77:833b31a8abc1

working version (w/ old hash impl)
author Nina Engelhardt <nengel@mailbox.tu-berlin.de>
date Tue, 28 Aug 2012 13:40:07 +0200
parents d20b105981b7
children ee8510009267
rev   line source
seanhalle@64 1 /*
seanhalle@64 2 * Copyright 2010 OpenSourceCodeStewardshipFoundation
seanhalle@64 3 *
seanhalle@64 4 * Licensed under BSD
seanhalle@64 5 */
seanhalle@64 6
seanhalle@64 7 #include <stdio.h>
seanhalle@64 8 #include <stdlib.h>
seanhalle@64 9 #include <malloc.h>
seanhalle@64 10
seanhalle@64 11 #include "Queue_impl/PrivateQueue.h"
seanhalle@64 12 #include "Hash_impl/PrivateHash.h"
seanhalle@64 13
seanhalle@64 14 #include "SSR.h"
seanhalle@64 15 #include "SSR_Counter_Recording.h"
seanhalle@64 16
seanhalle@64 17 //==========================================================================
seanhalle@64 18
seanhalle@64 19 void
seanhalle@64 20 SSR__init();
seanhalle@64 21
seanhalle@64 22 void
seanhalle@64 23 SSR__init_Helper();
seanhalle@64 24 //==========================================================================
seanhalle@64 25
seanhalle@64 26
seanhalle@64 27 /*TODO: Q: dealing with library f()s and DKU vs WT vs FoR
seanhalle@64 28 * (still want to do FoR, with time-lines as syntax, could be super cool)
seanhalle@64 29 * A: thinking pin the coreCtlrs for all of BLIS -- let Master arbitrate
seanhalle@64 30 * among library, DKU, WT, FoR -- all the patterns in terms of virtual
seanhalle@64 31 * processors (or equivalently work-units), so Master picks which virt procr
seanhalle@67 32 * from which portions of app (DKU, WT, FoR) onto which anim slots
seanhalle@67 33 *Might even do hierarchy of masters -- group of anim slots for each core
seanhalle@64 34 * has its own master, that keeps generated work local
seanhalle@64 35 * single-reader-single-writer sync everywhere -- no atomic primitives
seanhalle@64 36 * Might have the different assigners talk to each other, to negotiate
seanhalle@64 37 * larger-grain sharing of resources, according to predicted critical
seanhalle@64 38 * path, and expansion of work
seanhalle@64 39 */
seanhalle@64 40
seanhalle@64 41
seanhalle@64 42
seanhalle@64 43 //===========================================================================
seanhalle@64 44
seanhalle@64 45
seanhalle@64 46 /*These are the library functions *called in the application*
seanhalle@64 47 *
seanhalle@64 48 *There's a pattern for the outside sequential code to interact with the
seanhalle@64 49 * VMS_HW code.
seanhalle@64 50 *The VMS_HW system is inside a boundary.. every SSR system is in its
seanhalle@64 51 * own directory that contains the functions for each of the processor types.
seanhalle@64 52 * One of the processor types is the "seed" processor that starts the
seanhalle@64 53 * cascade of creating all the processors that do the work.
seanhalle@64 54 *So, in the directory is a file called "EntryPoint.c" that contains the
seanhalle@64 55 * function, named appropriately to the work performed, that the outside
seanhalle@64 56 * sequential code calls. This function follows a pattern:
seanhalle@64 57 *1) it calls SSR__init()
seanhalle@64 58 *2) it creates the initial data for the seed processor, which is passed
seanhalle@64 59 * in to the function
seanhalle@64 60 *3) it creates the seed SSR processor, with the data to start it with.
seanhalle@64 61 *4) it calls startSSRThenWaitUntilWorkDone
seanhalle@64 62 *5) it gets the returnValue from the transfer struc and returns that
seanhalle@64 63 * from the function
seanhalle@64 64 *
seanhalle@64 65 *For now, a new SSR system has to be created via SSR__init every
seanhalle@64 66 * time an entry point function is called -- later, might add letting the
seanhalle@64 67 * SSR system be created once, and let all the entry points just reuse
seanhalle@64 68 * it -- want to be as simple as possible now, and see by using what makes
seanhalle@64 69 * sense for later..
seanhalle@64 70 */
seanhalle@64 71
seanhalle@64 72
seanhalle@64 73
seanhalle@64 74 //===========================================================================
seanhalle@64 75
seanhalle@64 76 /*This is the "border crossing" function -- the thing that crosses from the
seanhalle@64 77 * outside world, into the VMS_HW world. It initializes and starts up the
seanhalle@64 78 * VMS system, then creates one processor from the specified function and
seanhalle@64 79 * puts it into the readyQ. From that point, that one function is resp.
seanhalle@64 80 * for creating all the other processors, that then create others, and so
seanhalle@64 81 * forth.
seanhalle@64 82 *When all the processors, including the seed, have dissipated, then this
seanhalle@64 83 * function returns. The results will have been written by side-effect via
seanhalle@64 84 * pointers read from, or written into initData.
seanhalle@64 85 *
seanhalle@64 86 *NOTE: no Threads should exist in the outside program that might touch
seanhalle@64 87 * any of the data reachable from initData passed in to here
seanhalle@64 88 */
seanhalle@64 89 void
seanhalle@64 90 SSR__create_seed_procr_and_do_work( TopLevelFnPtr fnPtr, void *initData )
seanhalle@64 91 { SSRSemEnv *semEnv;
seanhalle@64 92 SlaveVP *seedPr;
seanhalle@64 93
seanhalle@64 94 SSR__init(); //normal multi-thd
seanhalle@64 95
seanhalle@64 96 semEnv = _VMSMasterEnv->semanticEnv;
seanhalle@64 97
seanhalle@64 98 //SSR starts with one processor, which is put into initial environ,
seanhalle@64 99 // and which then calls create() to create more, thereby expanding work
seanhalle@64 100 seedPr = SSR__create_procr_helper( fnPtr, initData,
seanhalle@64 101 semEnv, semEnv->nextCoreToGetNewPr++ );
seanhalle@64 102
seanhalle@64 103 resume_slaveVP( seedPr, semEnv );
seanhalle@64 104
seanhalle@64 105 VMS_SS__start_the_work_then_wait_until_done(); //normal multi-thd
seanhalle@64 106
seanhalle@64 107 SSR__cleanup_after_shutdown();
seanhalle@64 108 }
seanhalle@64 109
seanhalle@64 110
seanhalle@64 111 int32
seanhalle@64 112 SSR__giveMinWorkUnitCycles( float32 percentOverhead )
seanhalle@64 113 {
seanhalle@64 114 return MIN_WORK_UNIT_CYCLES;
seanhalle@64 115 }
seanhalle@64 116
seanhalle@64 117 int32
seanhalle@64 118 SSR__giveIdealNumWorkUnits()
seanhalle@64 119 {
seanhalle@69 120 return NUM_ANIM_SLOTS * NUM_CORES;
seanhalle@64 121 }
seanhalle@64 122
seanhalle@64 123 int32
seanhalle@64 124 SSR__give_number_of_cores_to_schedule_onto()
seanhalle@64 125 {
seanhalle@64 126 return NUM_CORES;
seanhalle@64 127 }
seanhalle@64 128
seanhalle@64 129 /*For now, use TSC -- later, make these two macros with assembly that first
seanhalle@64 130 * saves jump point, and second jumps back several times to get reliable time
seanhalle@64 131 */
seanhalle@64 132 void
seanhalle@64 133 SSR__start_primitive()
seanhalle@64 134 { saveLowTimeStampCountInto( ((SSRSemEnv *)(_VMSMasterEnv->semanticEnv))->
seanhalle@64 135 primitiveStartTime );
seanhalle@64 136 }
seanhalle@64 137
seanhalle@64 138 /*Just quick and dirty for now -- make reliable later
seanhalle@64 139 * will want this to jump back several times -- to be sure cache is warm
seanhalle@64 140 * because don't want comm time included in calc-time measurement -- and
seanhalle@64 141 * also to throw out any "weird" values due to OS interrupt or TSC rollover
seanhalle@64 142 */
seanhalle@64 143 int32
seanhalle@64 144 SSR__end_primitive_and_give_cycles()
seanhalle@64 145 { int32 endTime, startTime;
seanhalle@64 146 //TODO: fix by repeating time-measurement
seanhalle@64 147 saveLowTimeStampCountInto( endTime );
seanhalle@64 148 startTime =((SSRSemEnv*)(_VMSMasterEnv->semanticEnv))->primitiveStartTime;
seanhalle@64 149 return (endTime - startTime);
seanhalle@64 150 }
seanhalle@64 151
seanhalle@64 152 //===========================================================================
seanhalle@64 153
seanhalle@64 154 /*Initializes all the data-structures for a SSR system -- but doesn't
seanhalle@64 155 * start it running yet!
seanhalle@64 156 *
seanhalle@64 157 *This runs in the main thread -- before VMS starts up
seanhalle@64 158 *
seanhalle@64 159 *This sets up the semantic layer over the VMS system
seanhalle@64 160 *
seanhalle@64 161 *First, calls VMS_Setup, then creates own environment, making it ready
seanhalle@64 162 * for creating the seed processor and then starting the work.
seanhalle@64 163 */
seanhalle@64 164 void
seanhalle@64 165 SSR__init()
seanhalle@64 166 {
seanhalle@64 167 VMS_SS__init();
seanhalle@64 168 //masterEnv, a global var, now is partially set up by init_VMS
seanhalle@64 169 // after this, have VMS_int__malloc and VMS_int__free available
seanhalle@64 170
seanhalle@64 171 SSR__init_Helper();
seanhalle@64 172 }
seanhalle@64 173
seanhalle@64 174
seanhalle@64 175 void idle_fn(void* data, SlaveVP *animatingSlv){
seanhalle@64 176 while(1){
seanhalle@64 177 VMS_int__suspend_slaveVP_and_send_req(animatingSlv);
seanhalle@64 178 }
seanhalle@64 179 }
seanhalle@64 180
seanhalle@64 181 void
seanhalle@64 182 SSR__init_Helper()
seanhalle@64 183 { SSRSemEnv *semanticEnv;
seanhalle@64 184 PrivQueueStruc **readyVPQs;
seanhalle@64 185 int coreIdx, i, j;
seanhalle@64 186
seanhalle@64 187 //Hook up the semantic layer's plug-ins to the Master virt procr
seanhalle@64 188 _VMSMasterEnv->requestHandler = &SSR__Request_Handler;
seanhalle@67 189 _VMSMasterEnv->slaveAssigner = &SSR__assign_slaveVP_to_slot;
seanhalle@64 190 #ifdef HOLISTIC__TURN_ON_PERF_COUNTERS
seanhalle@64 191 _VMSMasterEnv->counterHandler = &SSR__counter_handler;
seanhalle@64 192 #endif
seanhalle@64 193
seanhalle@64 194 //create the semantic layer's environment (all its data) and add to
seanhalle@64 195 // the master environment
seanhalle@64 196 semanticEnv = VMS_int__malloc( sizeof( SSRSemEnv ) );
seanhalle@64 197 _VMSMasterEnv->semanticEnv = semanticEnv;
seanhalle@64 198
seanhalle@64 199 #ifdef HOLISTIC__TURN_ON_PERF_COUNTERS
seanhalle@64 200 SSR__init_counter_data_structs();
seanhalle@64 201 #endif
nengel@72 202 #ifdef IDLE_SLAVES
nengel@66 203 semanticEnv->shutdownInitiated = FALSE;
seanhalle@64 204 for(i=0;i<NUM_CORES;++i){
seanhalle@69 205 for(j=0;j<NUM_ANIM_SLOTS;++j){
seanhalle@64 206 semanticEnv->idlePr[i][j] = VMS_int__create_slaveVP(&idle_fn,NULL);
seanhalle@64 207 semanticEnv->idlePr[i][j]->coreAnimatedBy = i;
nengel@72 208 semanticEnv->idlePr[i][j]->typeOfVP = Idle;
seanhalle@64 209 }
seanhalle@64 210 }
nengel@72 211 #endif
seanhalle@64 212 #ifdef HOLISTIC__TURN_ON_OBSERVE_UCC
seanhalle@64 213 semanticEnv->unitList = makeListOfArrays(sizeof(Unit),128);
seanhalle@64 214 semanticEnv->ctlDependenciesList = makeListOfArrays(sizeof(Dependency),128);
seanhalle@64 215 semanticEnv->commDependenciesList = makeListOfArrays(sizeof(Dependency),128);
seanhalle@64 216 semanticEnv->dynDependenciesList = makeListOfArrays(sizeof(Dependency),128);
nengel@74 217 semanticEnv->singletonDependenciesList = makeListOfArrays(sizeof(Dependency),128);
seanhalle@64 218 semanticEnv->ntonGroupsInfo = makePrivDynArrayOfSize((void***)&(semanticEnv->ntonGroups),8);
seanhalle@64 219
seanhalle@64 220 semanticEnv->hwArcs = makeListOfArrays(sizeof(Dependency),128);
seanhalle@69 221 memset(semanticEnv->last_in_slot,0,sizeof(NUM_CORES * NUM_ANIM_SLOTS * sizeof(Unit)));
seanhalle@64 222 #endif
seanhalle@64 223
seanhalle@64 224 //create the ready queue, hash tables used for pairing send to receive
seanhalle@64 225 // and so forth
seanhalle@64 226 //TODO: add hash tables for pairing sends with receives, and
seanhalle@64 227 // initialize the data ownership system
seanhalle@64 228 readyVPQs = VMS_int__malloc( NUM_CORES * sizeof(PrivQueueStruc *) );
seanhalle@64 229
seanhalle@64 230 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
seanhalle@64 231 {
seanhalle@64 232 readyVPQs[ coreIdx ] = makeVMSQ();
seanhalle@64 233 }
seanhalle@64 234
seanhalle@64 235 semanticEnv->readyVPQs = readyVPQs;
seanhalle@64 236
seanhalle@64 237 semanticEnv->nextCoreToGetNewPr = 0;
seanhalle@64 238 semanticEnv->numSlaveVP = 0;
seanhalle@64 239
seanhalle@64 240 semanticEnv->commHashTbl = makeHashTable( 1<<16, &VMS_int__free );//start big
seanhalle@64 241
seanhalle@64 242 //TODO: bug -- turn these arrays into dyn arrays to eliminate limit
seanhalle@64 243 //semanticEnv->singletonHasBeenExecutedFlags = makeDynArrayInfo( );
seanhalle@64 244 //semanticEnv->transactionStrucs = makeDynArrayInfo( );
seanhalle@64 245 for( i = 0; i < NUM_STRUCS_IN_SEM_ENV; i++ )
seanhalle@64 246 {
seanhalle@64 247 semanticEnv->fnSingletons[i].endInstrAddr = NULL;
seanhalle@64 248 semanticEnv->fnSingletons[i].hasBeenStarted = FALSE;
seanhalle@64 249 semanticEnv->fnSingletons[i].hasFinished = FALSE;
seanhalle@64 250 semanticEnv->fnSingletons[i].waitQ = makeVMSQ();
seanhalle@64 251 semanticEnv->transactionStrucs[i].waitingVPQ = makeVMSQ();
seanhalle@64 252 }
seanhalle@64 253 }
seanhalle@64 254
seanhalle@64 255
seanhalle@64 256 /*Frees any memory allocated by SSR__init() then calls VMS_int__shutdown
seanhalle@64 257 */
seanhalle@64 258 void
seanhalle@64 259 SSR__cleanup_after_shutdown()
seanhalle@64 260 { SSRSemEnv *semanticEnv;
seanhalle@64 261
seanhalle@64 262 semanticEnv = _VMSMasterEnv->semanticEnv;
seanhalle@64 263
seanhalle@64 264 #ifdef HOLISTIC__TURN_ON_OBSERVE_UCC
seanhalle@64 265 //UCC
seanhalle@64 266 FILE* output;
seanhalle@64 267 int n;
seanhalle@64 268 char filename[255];
seanhalle@64 269 for(n=0;n<255;n++)
seanhalle@64 270 {
seanhalle@64 271 sprintf(filename, "./counters/UCC.%d",n);
seanhalle@64 272 output = fopen(filename,"r");
seanhalle@64 273 if(output)
seanhalle@64 274 {
seanhalle@64 275 fclose(output);
seanhalle@64 276 }else{
seanhalle@64 277 break;
seanhalle@64 278 }
seanhalle@64 279 }
seanhalle@64 280 if(n<255){
seanhalle@64 281 printf("Saving UCC to File: %s ...\n", filename);
seanhalle@64 282 output = fopen(filename,"w+");
seanhalle@64 283 if(output!=NULL){
seanhalle@64 284 set_dependency_file(output);
seanhalle@64 285 //fprintf(output,"digraph Dependencies {\n");
seanhalle@64 286 //set_dot_file(output);
seanhalle@64 287 //FIXME: first line still depends on counters being enabled, replace w/ unit struct!
seanhalle@64 288 //forAllInDynArrayDo(_VMSMasterEnv->counter_history_array_info, &print_dot_node_info );
seanhalle@64 289 forAllInListOfArraysDo(semanticEnv->unitList, &print_unit_to_file);
seanhalle@64 290 forAllInListOfArraysDo( semanticEnv->commDependenciesList, &print_comm_dependency_to_file );
seanhalle@64 291 forAllInListOfArraysDo( semanticEnv->ctlDependenciesList, &print_ctl_dependency_to_file );
seanhalle@64 292 forAllInDynArrayDo(semanticEnv->ntonGroupsInfo,&print_nton_to_file);
seanhalle@64 293 //fprintf(output,"}\n");
seanhalle@64 294 fflush(output);
seanhalle@64 295
seanhalle@64 296 } else
seanhalle@64 297 printf("Opening UCC file failed. Please check that folder \"counters\" exists in run directory and has write permission.\n");
seanhalle@64 298 } else {
seanhalle@64 299 printf("Could not open UCC file, please clean \"counters\" folder. (Must contain less than 255 files.)\n");
seanhalle@64 300 }
seanhalle@64 301 //Loop Graph
seanhalle@64 302 for(n=0;n<255;n++)
seanhalle@64 303 {
seanhalle@64 304 sprintf(filename, "./counters/LoopGraph.%d",n);
seanhalle@64 305 output = fopen(filename,"r");
seanhalle@64 306 if(output)
seanhalle@64 307 {
seanhalle@64 308 fclose(output);
seanhalle@64 309 }else{
seanhalle@64 310 break;
seanhalle@64 311 }
seanhalle@64 312 }
seanhalle@64 313 if(n<255){
seanhalle@64 314 printf("Saving LoopGraph to File: %s ...\n", filename);
seanhalle@64 315 output = fopen(filename,"w+");
seanhalle@64 316 if(output!=NULL){
seanhalle@64 317 set_dependency_file(output);
seanhalle@64 318 //fprintf(output,"digraph Dependencies {\n");
seanhalle@64 319 //set_dot_file(output);
seanhalle@64 320 //FIXME: first line still depends on counters being enabled, replace w/ unit struct!
seanhalle@64 321 //forAllInDynArrayDo(_VMSMasterEnv->counter_history_array_info, &print_dot_node_info );
seanhalle@64 322 forAllInListOfArraysDo( semanticEnv->unitList, &print_unit_to_file );
seanhalle@64 323 forAllInListOfArraysDo( semanticEnv->commDependenciesList, &print_comm_dependency_to_file );
seanhalle@64 324 forAllInListOfArraysDo( semanticEnv->ctlDependenciesList, &print_ctl_dependency_to_file );
seanhalle@64 325 forAllInListOfArraysDo( semanticEnv->dynDependenciesList, &print_dyn_dependency_to_file );
nengel@74 326 forAllInListOfArraysDo( semanticEnv->singletonDependenciesList, &print_singleton_dependency_to_file );
seanhalle@64 327 forAllInListOfArraysDo( semanticEnv->hwArcs, &print_hw_dependency_to_file );
seanhalle@64 328 //fprintf(output,"}\n");
seanhalle@64 329 fflush(output);
seanhalle@64 330
seanhalle@64 331 } else
seanhalle@64 332 printf("Opening LoopGraph file failed. Please check that folder \"counters\" exists in run directory and has write permission.\n");
seanhalle@64 333 } else {
seanhalle@64 334 printf("Could not open LoopGraph file, please clean \"counters\" folder. (Must contain less than 255 files.)\n");
seanhalle@64 335 }
seanhalle@64 336
seanhalle@64 337
seanhalle@64 338 freeListOfArrays(semanticEnv->unitList);
seanhalle@64 339 freeListOfArrays(semanticEnv->commDependenciesList);
seanhalle@64 340 freeListOfArrays(semanticEnv->ctlDependenciesList);
seanhalle@64 341 freeListOfArrays(semanticEnv->dynDependenciesList);
nengel@74 342 freeListOfArrays(semanticEnv->singletonDependenciesList);
seanhalle@64 343 #endif
seanhalle@64 344 #ifdef HOLISTIC__TURN_ON_PERF_COUNTERS
seanhalle@64 345 for(n=0;n<255;n++)
seanhalle@64 346 {
seanhalle@64 347 sprintf(filename, "./counters/Counters.%d.csv",n);
seanhalle@64 348 output = fopen(filename,"r");
seanhalle@64 349 if(output)
seanhalle@64 350 {
seanhalle@64 351 fclose(output);
seanhalle@64 352 }else{
seanhalle@64 353 break;
seanhalle@64 354 }
seanhalle@64 355 }
seanhalle@64 356 if(n<255){
seanhalle@64 357 printf("Saving Counter measurements to File: %s ...\n", filename);
seanhalle@64 358 output = fopen(filename,"w+");
seanhalle@64 359 if(output!=NULL){
seanhalle@64 360 set_counter_file(output);
seanhalle@64 361 int i;
seanhalle@64 362 for(i=0;i<NUM_CORES;i++){
seanhalle@64 363 forAllInListOfArraysDo( semanticEnv->counterList[i], &print_counter_events_to_file );
seanhalle@64 364 fflush(output);
seanhalle@64 365 }
seanhalle@64 366
seanhalle@64 367 } else
seanhalle@64 368 printf("Opening UCC file failed. Please check that folder \"counters\" exists in run directory and has write permission.\n");
seanhalle@64 369 } else {
seanhalle@64 370 printf("Could not open UCC file, please clean \"counters\" folder. (Must contain less than 255 files.)\n");
seanhalle@64 371 }
seanhalle@64 372
seanhalle@64 373 #endif
seanhalle@64 374 /* It's all allocated inside VMS's big chunk -- that's about to be freed, so
seanhalle@64 375 * nothing to do here
seanhalle@64 376
seanhalle@64 377
seanhalle@64 378 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
seanhalle@64 379 {
seanhalle@64 380 VMS_int__free( semanticEnv->readyVPQs[coreIdx]->startOfData );
seanhalle@64 381 VMS_int__free( semanticEnv->readyVPQs[coreIdx] );
seanhalle@64 382 }
seanhalle@64 383 VMS_int__free( semanticEnv->readyVPQs );
seanhalle@64 384
seanhalle@64 385 freeHashTable( semanticEnv->commHashTbl );
seanhalle@64 386 VMS_int__free( _VMSMasterEnv->semanticEnv );
seanhalle@64 387 */
seanhalle@64 388 VMS_SS__cleanup_at_end_of_shutdown();
seanhalle@64 389 }
seanhalle@64 390
seanhalle@64 391
seanhalle@64 392 //===========================================================================
seanhalle@64 393
seanhalle@64 394 /*
seanhalle@64 395 */
seanhalle@64 396 SlaveVP *
seanhalle@64 397 SSR__create_procr_with( TopLevelFnPtr fnPtr, void *initData,
seanhalle@64 398 SlaveVP *creatingPr )
seanhalle@64 399 { SSRSemReq reqData;
seanhalle@64 400
seanhalle@64 401 //the semantic request data is on the stack and disappears when this
seanhalle@64 402 // call returns -- it's guaranteed to remain in the VP's stack for as
seanhalle@64 403 // long as the VP is suspended.
seanhalle@64 404 reqData.reqType = 0; //know type because in a VMS create req
seanhalle@64 405 reqData.coreToAssignOnto = -1; //means round-robin assign
seanhalle@64 406 reqData.fnPtr = fnPtr;
seanhalle@64 407 reqData.initData = initData;
seanhalle@64 408 reqData.sendPr = creatingPr;
seanhalle@64 409
seanhalle@64 410 VMS_WL__send_create_slaveVP_req( &reqData, creatingPr );
seanhalle@64 411
seanhalle@64 412 return creatingPr->dataRetFromReq;
seanhalle@64 413 }
seanhalle@64 414
seanhalle@64 415 SlaveVP *
seanhalle@64 416 SSR__create_procr_with_affinity( TopLevelFnPtr fnPtr, void *initData,
seanhalle@64 417 SlaveVP *creatingPr, int32 coreToAssignOnto )
seanhalle@64 418 { SSRSemReq reqData;
seanhalle@64 419
seanhalle@64 420 //the semantic request data is on the stack and disappears when this
seanhalle@64 421 // call returns -- it's guaranteed to remain in the VP's stack for as
seanhalle@64 422 // long as the VP is suspended.
seanhalle@64 423 reqData.reqType = 0; //know type because in a VMS create req
seanhalle@64 424 reqData.coreToAssignOnto = coreToAssignOnto;
seanhalle@64 425 reqData.fnPtr = fnPtr;
seanhalle@64 426 reqData.initData = initData;
seanhalle@64 427 reqData.sendPr = creatingPr;
seanhalle@64 428
seanhalle@64 429 VMS_WL__send_create_slaveVP_req( &reqData, creatingPr );
seanhalle@64 430
seanhalle@64 431 return creatingPr->dataRetFromReq;
seanhalle@64 432 }
seanhalle@64 433
seanhalle@64 434
seanhalle@64 435 void
seanhalle@64 436 SSR__dissipate_procr( SlaveVP *procrToDissipate )
seanhalle@64 437 {
seanhalle@64 438 VMS_WL__send_dissipate_req( procrToDissipate );
seanhalle@64 439 }
seanhalle@64 440
seanhalle@64 441
seanhalle@64 442 //===========================================================================
seanhalle@64 443
seanhalle@64 444 void *
seanhalle@64 445 SSR__malloc_to( int32 sizeToMalloc, SlaveVP *owningPr )
seanhalle@64 446 { SSRSemReq reqData;
seanhalle@64 447
seanhalle@64 448 reqData.reqType = malloc_req;
seanhalle@64 449 reqData.sendPr = owningPr;
seanhalle@64 450 reqData.sizeToMalloc = sizeToMalloc;
seanhalle@64 451
seanhalle@64 452 VMS_WL__send_sem_request( &reqData, owningPr );
seanhalle@64 453
seanhalle@64 454 return owningPr->dataRetFromReq;
seanhalle@64 455 }
seanhalle@64 456
seanhalle@64 457
seanhalle@64 458 /*Sends request to Master, which does the work of freeing
seanhalle@64 459 */
seanhalle@64 460 void
seanhalle@64 461 SSR__free( void *ptrToFree, SlaveVP *owningPr )
seanhalle@64 462 { SSRSemReq reqData;
seanhalle@64 463
seanhalle@64 464 reqData.reqType = free_req;
seanhalle@64 465 reqData.sendPr = owningPr;
seanhalle@64 466 reqData.ptrToFree = ptrToFree;
seanhalle@64 467
seanhalle@64 468 VMS_WL__send_sem_request( &reqData, owningPr );
seanhalle@64 469 }
seanhalle@64 470
seanhalle@64 471
seanhalle@64 472 void
seanhalle@64 473 SSR__transfer_ownership_of_from_to( void *data, SlaveVP *oldOwnerSlv,
seanhalle@64 474 SlaveVP *newOwnerPr )
seanhalle@64 475 {
seanhalle@64 476 //TODO: put in the ownership system that automatically frees when no
seanhalle@64 477 // owners of data left -- will need keeper for keeping data around when
seanhalle@64 478 // future created processors might need it but don't exist yet
seanhalle@64 479 }
seanhalle@64 480
seanhalle@64 481
seanhalle@64 482 void
seanhalle@64 483 SSR__add_ownership_by_to( SlaveVP *newOwnerSlv, void *data )
seanhalle@64 484 {
seanhalle@64 485
seanhalle@64 486 }
seanhalle@64 487
seanhalle@64 488
seanhalle@64 489 void
seanhalle@64 490 SSR__remove_ownership_by_from( SlaveVP *loserSlv, void *dataLosing )
seanhalle@64 491 {
seanhalle@64 492
seanhalle@64 493 }
seanhalle@64 494
seanhalle@64 495
seanhalle@64 496 /*Causes the SSR system to remove internal ownership, so data won't be
seanhalle@64 497 * freed when SSR shuts down, and will persist in the external program.
seanhalle@64 498 *
seanhalle@64 499 *Must be called from the processor that currently owns the data.
seanhalle@64 500 *
seanhalle@64 501 *IMPL: Transferring ownership touches two different virtual processor's
seanhalle@64 502 * state -- which means it has to be done carefully -- the VMS rules for
seanhalle@64 503 * semantic layers say that a work-unit is only allowed to touch the
seanhalle@64 504 * virtual processor it is part of, and that only a single work-unit per
seanhalle@64 505 * virtual processor be assigned to a slave at a time. So, this has to
seanhalle@64 506 * modify the virtual processor that owns the work-unit that called this
seanhalle@64 507 * function, then create a request to have the other processor modified.
seanhalle@64 508 *However, in this case, the TO processor is the outside, and transfers
seanhalle@64 509 * are only allowed to be called by the giver-upper, so can mark caller of
seanhalle@64 510 * this function as no longer owner, and return -- done.
seanhalle@64 511 */
seanhalle@64 512 void
seanhalle@64 513 SSR__transfer_ownership_to_outside( void *data )
seanhalle@64 514 {
seanhalle@64 515 //TODO: removeAllOwnersFrom( data );
seanhalle@64 516 }
seanhalle@64 517
seanhalle@64 518
seanhalle@64 519 //===========================================================================
seanhalle@64 520
seanhalle@64 521 void
seanhalle@64 522 SSR__send_of_type_to( SlaveVP *sendPr, void *msg, const int type,
seanhalle@64 523 SlaveVP *receivePr)
seanhalle@64 524 { SSRSemReq reqData;
seanhalle@64 525
seanhalle@64 526 reqData.receivePr = receivePr;
seanhalle@64 527 reqData.sendPr = sendPr;
seanhalle@64 528 reqData.reqType = send_type;
seanhalle@64 529 reqData.msgType = type;
seanhalle@64 530 reqData.msg = msg;
seanhalle@64 531 reqData.nextReqInHashEntry = NULL;
seanhalle@64 532
seanhalle@64 533 //On ownership -- remove inside the send and let ownership sit in limbo
seanhalle@64 534 // as a potential in an entry in the hash table, when this receive msg
seanhalle@64 535 // gets paired to a send, the ownership gets added to the receivePr --
seanhalle@64 536 // the next work-unit in the receivePr's trace will have ownership.
seanhalle@64 537 VMS_WL__send_sem_request( &reqData, sendPr );
seanhalle@64 538
seanhalle@64 539 //When come back from suspend, no longer own data reachable from msg
seanhalle@64 540 //TODO: release ownership here
seanhalle@64 541 }
seanhalle@64 542
seanhalle@64 543 void
seanhalle@64 544 SSR__send_from_to( void *msg, SlaveVP *sendPr, SlaveVP *receivePr )
seanhalle@64 545 { SSRSemReq reqData;
seanhalle@64 546
seanhalle@64 547 //hash on the receiver, 'cause always know it, but sometimes want to
seanhalle@64 548 // receive from anonymous sender
seanhalle@64 549
seanhalle@64 550 reqData.receivePr = receivePr;
seanhalle@64 551 reqData.sendPr = sendPr;
seanhalle@64 552 reqData.reqType = send_from_to;
seanhalle@64 553 reqData.msg = msg;
seanhalle@64 554 reqData.nextReqInHashEntry = NULL;
seanhalle@64 555
seanhalle@64 556 VMS_WL__send_sem_request( &reqData, sendPr );
seanhalle@64 557 }
seanhalle@64 558
seanhalle@64 559
seanhalle@64 560 //===========================================================================
seanhalle@64 561
seanhalle@64 562 void *
seanhalle@64 563 SSR__receive_any_to( SlaveVP *receivePr )
seanhalle@64 564 {
seanhalle@64 565
seanhalle@64 566 }
seanhalle@64 567
seanhalle@64 568 void *
seanhalle@64 569 SSR__receive_type_to( const int type, SlaveVP *receivePr )
seanhalle@67 570 { DEBUG__printf1(dbgRqstHdlr,"WL: receive type to: %d", receivePr->slaveID);
seanhalle@64 571 SSRSemReq reqData;
seanhalle@64 572
seanhalle@64 573 reqData.receivePr = receivePr;
seanhalle@64 574 reqData.reqType = receive_type;
seanhalle@64 575 reqData.msgType = type;
seanhalle@64 576 reqData.nextReqInHashEntry = NULL;
seanhalle@64 577
seanhalle@64 578 VMS_WL__send_sem_request( &reqData, receivePr );
seanhalle@64 579
seanhalle@64 580 return receivePr->dataRetFromReq;
seanhalle@64 581 }
seanhalle@64 582
seanhalle@64 583
seanhalle@64 584
seanhalle@64 585 /*Call this at point receiving virt pr wants in-coming data.
seanhalle@64 586 *
seanhalle@64 587 *The reason receivePr must call this is that it modifies the receivPr
seanhalle@64 588 * loc structure directly -- and the VMS rules state a virtual processor
seanhalle@64 589 * loc structure can only be modified by itself.
seanhalle@64 590 */
seanhalle@64 591 void *
seanhalle@64 592 SSR__receive_from_to( SlaveVP *sendPr, SlaveVP *receivePr )
seanhalle@67 593 { DEBUG__printf2(dbgRqstHdlr,"WL: receive from %d to: %d", sendPr->slaveID, receivePr->slaveID);
seanhalle@67 594 SSRSemReq reqData;
seanhalle@64 595
seanhalle@64 596 //hash on the receiver, 'cause always know it, but sometimes want to
seanhalle@64 597 // receive from anonymous sender
seanhalle@64 598
seanhalle@64 599 reqData.receivePr = receivePr;
seanhalle@64 600 reqData.sendPr = sendPr;
seanhalle@64 601 reqData.reqType = receive_from_to;
seanhalle@64 602 reqData.nextReqInHashEntry = NULL;
seanhalle@64 603
seanhalle@64 604 VMS_WL__send_sem_request( &reqData, receivePr );
seanhalle@64 605
seanhalle@64 606 return receivePr->dataRetFromReq;
seanhalle@64 607 }
seanhalle@64 608
seanhalle@64 609
seanhalle@64 610 //===========================================================================
seanhalle@64 611 //
seanhalle@64 612 /*A function singleton is a function whose body executes exactly once, on a
seanhalle@64 613 * single core, no matter how many times the fuction is called and no
seanhalle@64 614 * matter how many cores or the timing of cores calling it.
seanhalle@64 615 *
seanhalle@64 616 *A data singleton is a ticket attached to data. That ticket can be used
seanhalle@64 617 * to get the data through the function exactly once, no matter how many
seanhalle@64 618 * times the data is given to the function, and no matter the timing of
seanhalle@64 619 * trying to get the data through from different cores.
seanhalle@64 620 */
seanhalle@64 621
seanhalle@64 622 /*asm function declarations*/
seanhalle@64 623 void asm_save_ret_to_singleton(SSRSingleton *singletonPtrAddr);
seanhalle@64 624 void asm_write_ret_from_singleton(SSRSingleton *singletonPtrAddr);
seanhalle@64 625
seanhalle@64 626 /*Fn singleton uses ID as index into array of singleton structs held in the
seanhalle@64 627 * semantic environment.
seanhalle@64 628 */
seanhalle@64 629 void
seanhalle@64 630 SSR__start_fn_singleton( int32 singletonID, SlaveVP *animPr )
seanhalle@64 631 {
seanhalle@64 632 SSRSemReq reqData;
seanhalle@64 633
seanhalle@64 634 //
seanhalle@64 635 reqData.reqType = singleton_fn_start;
seanhalle@64 636 reqData.singletonID = singletonID;
seanhalle@64 637
seanhalle@64 638 VMS_WL__send_sem_request( &reqData, animPr );
seanhalle@64 639 if( animPr->dataRetFromReq ) //will be 0 or addr of label in end singleton
seanhalle@64 640 {
seanhalle@64 641 SSRSemEnv *semEnv = VMS_int__give_sem_env_for( animPr );
seanhalle@64 642 asm_write_ret_from_singleton(&(semEnv->fnSingletons[ singletonID]));
seanhalle@64 643 }
seanhalle@64 644 }
seanhalle@64 645
seanhalle@64 646 /*Data singleton hands addr of loc holding a pointer to a singleton struct.
seanhalle@64 647 * The start_data_singleton makes the structure and puts its addr into the
seanhalle@64 648 * location.
seanhalle@64 649 */
seanhalle@64 650 void
seanhalle@64 651 SSR__start_data_singleton( SSRSingleton **singletonAddr, SlaveVP *animPr )
seanhalle@64 652 {
seanhalle@64 653 SSRSemReq reqData;
seanhalle@64 654
seanhalle@64 655 if( *singletonAddr && (*singletonAddr)->hasFinished )
seanhalle@64 656 goto JmpToEndSingleton;
seanhalle@64 657
seanhalle@64 658 reqData.reqType = singleton_data_start;
seanhalle@64 659 reqData.singletonPtrAddr = singletonAddr;
seanhalle@64 660
seanhalle@64 661 VMS_WL__send_sem_request( &reqData, animPr );
seanhalle@64 662 if( animPr->dataRetFromReq ) //either 0 or end singleton's return addr
seanhalle@64 663 { //Assembly code changes the return addr on the stack to the one
seanhalle@64 664 // saved into the singleton by the end-singleton-fn
seanhalle@64 665 //The return addr is at 0x4(%%ebp)
seanhalle@64 666 JmpToEndSingleton:
seanhalle@64 667 asm_write_ret_from_singleton(*singletonAddr);
seanhalle@64 668 }
seanhalle@64 669 //now, simply return
seanhalle@64 670 //will exit either from the start singleton call or the end-singleton call
seanhalle@64 671 }
seanhalle@64 672
seanhalle@64 673 /*Uses ID as index into array of flags. If flag already set, resumes from
seanhalle@64 674 * end-label. Else, sets flag and resumes normally.
seanhalle@64 675 *
seanhalle@64 676 *Note, this call cannot be inlined because the instr addr at the label
seanhalle@64 677 * inside is shared by all invocations of a given singleton ID.
seanhalle@64 678 */
seanhalle@64 679 void
seanhalle@64 680 SSR__end_fn_singleton( int32 singletonID, SlaveVP *animPr )
seanhalle@64 681 {
seanhalle@64 682 SSRSemReq reqData;
seanhalle@64 683
seanhalle@64 684 //don't need this addr until after at least one singleton has reached
seanhalle@64 685 // this function
seanhalle@64 686 SSRSemEnv *semEnv = VMS_int__give_sem_env_for( animPr );
seanhalle@64 687 asm_write_ret_from_singleton(&(semEnv->fnSingletons[ singletonID]));
seanhalle@64 688
seanhalle@64 689 reqData.reqType = singleton_fn_end;
seanhalle@64 690 reqData.singletonID = singletonID;
seanhalle@64 691
seanhalle@64 692 VMS_WL__send_sem_request( &reqData, animPr );
seanhalle@64 693
seanhalle@64 694 EndSingletonInstrAddr:
seanhalle@64 695 return;
seanhalle@64 696 }
seanhalle@64 697
seanhalle@64 698 void
seanhalle@64 699 SSR__end_data_singleton( SSRSingleton **singletonPtrAddr, SlaveVP *animPr )
seanhalle@64 700 {
seanhalle@64 701 SSRSemReq reqData;
seanhalle@64 702
seanhalle@64 703 //don't need this addr until after singleton struct has reached
seanhalle@64 704 // this function for first time
seanhalle@64 705 //do assembly that saves the return addr of this fn call into the
seanhalle@64 706 // data singleton -- that data-singleton can only be given to exactly
seanhalle@64 707 // one instance in the code of this function. However, can use this
seanhalle@64 708 // function in different places for different data-singletons.
seanhalle@64 709 // (*(singletonAddr))->endInstrAddr = &&EndDataSingletonInstrAddr;
seanhalle@64 710
seanhalle@64 711
seanhalle@64 712 asm_save_ret_to_singleton(*singletonPtrAddr);
seanhalle@64 713
seanhalle@64 714 reqData.reqType = singleton_data_end;
seanhalle@64 715 reqData.singletonPtrAddr = singletonPtrAddr;
seanhalle@64 716
seanhalle@64 717 VMS_WL__send_sem_request( &reqData, animPr );
seanhalle@64 718 }
seanhalle@64 719
seanhalle@64 720 /*This executes the function in the masterVP, so it executes in isolation
seanhalle@64 721 * from any other copies -- only one copy of the function can ever execute
seanhalle@64 722 * at a time.
seanhalle@64 723 *
seanhalle@64 724 *It suspends to the master, and the request handler takes the function
seanhalle@64 725 * pointer out of the request and calls it, then resumes the VP.
seanhalle@64 726 *Only very short functions should be called this way -- for longer-running
seanhalle@64 727 * isolation, use transaction-start and transaction-end, which run the code
seanhalle@64 728 * between as work-code.
seanhalle@64 729 */
seanhalle@64 730 void
seanhalle@64 731 SSR__animate_short_fn_in_isolation( PtrToAtomicFn ptrToFnToExecInMaster,
seanhalle@64 732 void *data, SlaveVP *animPr )
seanhalle@64 733 {
seanhalle@64 734 SSRSemReq reqData;
seanhalle@64 735
seanhalle@64 736 //
seanhalle@64 737 reqData.reqType = atomic;
seanhalle@64 738 reqData.fnToExecInMaster = ptrToFnToExecInMaster;
seanhalle@64 739 reqData.dataForFn = data;
seanhalle@64 740
seanhalle@64 741 VMS_WL__send_sem_request( &reqData, animPr );
seanhalle@64 742 }
seanhalle@64 743
seanhalle@64 744
seanhalle@64 745 /*This suspends to the master.
seanhalle@64 746 *First, it looks at the VP's data, to see the highest transactionID that VP
seanhalle@64 747 * already has entered. If the current ID is not larger, it throws an
seanhalle@64 748 * exception stating a bug in the code. Otherwise it puts the current ID
seanhalle@64 749 * there, and adds the ID to a linked list of IDs entered -- the list is
seanhalle@64 750 * used to check that exits are properly ordered.
seanhalle@64 751 *Next it is uses transactionID as index into an array of transaction
seanhalle@64 752 * structures.
seanhalle@64 753 *If the "VP_currently_executing" field is non-null, then put requesting VP
seanhalle@64 754 * into queue in the struct. (At some point a holder will request
seanhalle@64 755 * end-transaction, which will take this VP from the queue and resume it.)
seanhalle@64 756 *If NULL, then write requesting into the field and resume.
seanhalle@64 757 */
seanhalle@64 758 void
seanhalle@64 759 SSR__start_transaction( int32 transactionID, SlaveVP *animPr )
seanhalle@64 760 {
seanhalle@64 761 SSRSemReq reqData;
seanhalle@64 762
seanhalle@64 763 //
seanhalle@64 764 reqData.sendPr = animPr;
seanhalle@64 765 reqData.reqType = trans_start;
seanhalle@64 766 reqData.transID = transactionID;
seanhalle@64 767
seanhalle@64 768 VMS_WL__send_sem_request( &reqData, animPr );
seanhalle@64 769 }
seanhalle@64 770
seanhalle@64 771 /*This suspends to the master, then uses transactionID as index into an
seanhalle@64 772 * array of transaction structures.
seanhalle@64 773 *It looks at VP_currently_executing to be sure it's same as requesting VP.
seanhalle@64 774 * If different, throws an exception, stating there's a bug in the code.
seanhalle@64 775 *Next it looks at the queue in the structure.
seanhalle@64 776 *If it's empty, it sets VP_currently_executing field to NULL and resumes.
seanhalle@64 777 *If something in, gets it, sets VP_currently_executing to that VP, then
seanhalle@64 778 * resumes both.
seanhalle@64 779 */
seanhalle@64 780 void
seanhalle@64 781 SSR__end_transaction( int32 transactionID, SlaveVP *animPr )
seanhalle@64 782 {
seanhalle@64 783 SSRSemReq reqData;
seanhalle@64 784
seanhalle@64 785 //
seanhalle@64 786 reqData.sendPr = animPr;
seanhalle@64 787 reqData.reqType = trans_end;
seanhalle@64 788 reqData.transID = transactionID;
seanhalle@64 789
seanhalle@64 790 VMS_WL__send_sem_request( &reqData, animPr );
seanhalle@64 791 }