view VMS.c @ 185:28cc465f7eb7

new counter interface works now
author Nina Engelhardt <nengel@mailbox.tu-berlin.de>
date Fri, 13 Jan 2012 18:34:31 +0100
parents 7523ee70d66c
children 69eb54ce9c4b
line source
1 /*
2 * Copyright 2010 OpenSourceStewardshipFoundation
3 *
4 * Licensed under BSD
5 */
7 #include <stdio.h>
8 #include <stdlib.h>
9 #include <string.h>
10 #include <malloc.h>
11 #include <inttypes.h>
12 #include <sys/time.h>
14 #include "VMS.h"
15 #include "ProcrContext.h"
16 #include "Queue_impl/BlockingQueue.h"
17 #include "Histogram/Histogram.h"
19 #include <unistd.h>
20 #include <fcntl.h>
21 #include <linux/types.h>
22 #include <linux/perf_event.h>
23 #include <errno.h>
24 #include <sys/syscall.h>
25 #include <linux/prctl.h>
28 #define thdAttrs NULL
30 //===========================================================================
31 void
32 shutdownFn( void *dummy, VirtProcr *dummy2 );
34 SchedSlot **
35 create_sched_slots();
37 void
38 create_masterEnv();
40 void
41 create_the_coreLoop_OS_threads();
43 MallocProlog *
44 create_free_list();
46 void
47 endOSThreadFn( void *initData, VirtProcr *animatingPr );
49 pthread_mutex_t suspendLock = PTHREAD_MUTEX_INITIALIZER;
50 pthread_cond_t suspend_cond = PTHREAD_COND_INITIALIZER;
52 //===========================================================================
54 /*Setup has two phases:
55 * 1) Semantic layer first calls init_VMS, which creates masterEnv, and puts
56 * the master virt procr into the work-queue, ready for first "call"
57 * 2) Semantic layer then does its own init, which creates the seed virt
58 * procr inside the semantic layer, ready to schedule it when
59 * asked by the first run of the masterLoop.
60 *
61 *This part is bit weird because VMS really wants to be "always there", and
62 * have applications attach and detach.. for now, this VMS is part of
63 * the app, so the VMS system starts up as part of running the app.
64 *
65 *The semantic layer is isolated from the VMS internals by making the
66 * semantic layer do setup to a state that it's ready with its
67 * initial virt procrs, ready to schedule them to slots when the masterLoop
68 * asks. Without this pattern, the semantic layer's setup would
69 * have to modify slots directly to assign the initial virt-procrs, and put
70 * them into the readyToAnimateQ itself, breaking the isolation completely.
71 *
72 *
73 *The semantic layer creates the initial virt procr(s), and adds its
74 * own environment to masterEnv, and fills in the pointers to
75 * the requestHandler and slaveScheduler plug-in functions
76 */
78 /*This allocates VMS data structures, populates the master VMSProc,
79 * and master environment, and returns the master environment to the semantic
80 * layer.
81 */
82 void
83 VMS__init()
84 {
85 create_masterEnv();
86 create_the_coreLoop_OS_threads();
87 }
89 #ifdef SEQUENTIAL
91 /*To initialize the sequential version, just don't create the threads
92 */
93 void
94 VMS__init_Seq()
95 {
96 create_masterEnv();
97 }
99 #endif
101 void
102 create_masterEnv()
103 { MasterEnv *masterEnv;
104 VMSQueueStruc **readyToAnimateQs;
105 int coreIdx;
106 VirtProcr **masterVPs;
107 SchedSlot ***allSchedSlots; //ptr to array of ptrs
110 //Make the master env, which holds everything else
111 _VMSMasterEnv = malloc( sizeof(MasterEnv) );
113 //Very first thing put into the master env is the free-list, seeded
114 // with a massive initial chunk of memory.
115 //After this, all other mallocs are VMS__malloc.
116 _VMSMasterEnv->freeListHead = VMS_ext__create_free_list();
119 //============================= MEASUREMENT STUFF ========================
120 #ifdef MEAS__TIME_MALLOC
121 _VMSMasterEnv->mallocTimeHist = makeFixedBinHistExt( 100, 0, 30,
122 "malloc_time_hist");
123 _VMSMasterEnv->freeTimeHist = makeFixedBinHistExt( 100, 0, 30,
124 "free_time_hist");
125 #endif
126 #ifdef MEAS__TIME_PLUGIN
127 _VMSMasterEnv->reqHdlrLowTimeHist = makeFixedBinHistExt( 100, 0, 200,
128 "plugin_low_time_hist");
129 _VMSMasterEnv->reqHdlrHighTimeHist = makeFixedBinHistExt( 100, 0, 200,
130 "plugin_high_time_hist");
131 #endif
132 //========================================================================
134 //===================== Only VMS__malloc after this ====================
135 masterEnv = (MasterEnv*)_VMSMasterEnv;
137 //Make a readyToAnimateQ for each core loop
138 readyToAnimateQs = VMS__malloc( NUM_CORES * sizeof(VMSQueueStruc *) );
139 masterVPs = VMS__malloc( NUM_CORES * sizeof(VirtProcr *) );
141 //One array for each core, 3 in array, core's masterVP scheds all
142 allSchedSlots = VMS__malloc( NUM_CORES * sizeof(SchedSlot *) );
144 _VMSMasterEnv->numProcrsCreated = 0; //used by create procr
145 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
146 {
147 readyToAnimateQs[ coreIdx ] = makeVMSQ();
149 //Q: should give masterVP core-specific info as its init data?
150 masterVPs[ coreIdx ] = VMS__create_procr( (VirtProcrFnPtr)&masterLoop, (void*)masterEnv );
151 masterVPs[ coreIdx ]->coreAnimatedBy = coreIdx;
152 allSchedSlots[ coreIdx ] = create_sched_slots(); //makes for one core
153 _VMSMasterEnv->numMasterInARow[ coreIdx ] = 0;
154 _VMSMasterEnv->workStealingGates[ coreIdx ] = NULL;
155 }
156 _VMSMasterEnv->readyToAnimateQs = readyToAnimateQs;
157 _VMSMasterEnv->masterVPs = masterVPs;
158 _VMSMasterEnv->masterLock = UNLOCKED;
159 _VMSMasterEnv->allSchedSlots = allSchedSlots;
160 _VMSMasterEnv->workStealingLock = UNLOCKED;
163 //Aug 19, 2010: no longer need to place initial masterVP into queue
164 // because coreLoop now controls -- animates its masterVP when no work
167 //============================= MEASUREMENT STUFF ========================
168 #ifdef STATS__TURN_ON_PROBES
169 _VMSMasterEnv->dynIntervalProbesInfo =
170 makePrivDynArrayOfSize( (void***)&(_VMSMasterEnv->intervalProbes), 200);
172 _VMSMasterEnv->probeNameHashTbl = makeHashTable( 1000, &VMS__free );
174 //put creation time directly into master env, for fast retrieval
175 struct timeval timeStamp;
176 gettimeofday( &(timeStamp), NULL);
177 _VMSMasterEnv->createPtInSecs =
178 timeStamp.tv_sec +(timeStamp.tv_usec/1000000.0);
179 #endif
180 #ifdef MEAS__TIME_MASTER_LOCK
181 _VMSMasterEnv->masterLockLowTimeHist = makeFixedBinHist( 50, 0, 2,
182 "master lock low time hist");
183 _VMSMasterEnv->masterLockHighTimeHist = makeFixedBinHist( 50, 0, 100,
184 "master lock high time hist");
185 #endif
187 MakeTheMeasHists();
190 #ifdef DETECT_LOOP_GRAPH
192 #endif
194 #ifdef MEAS__PERF_COUNTERS
195 /*
196 _VMSMasterEnv->counter_history = VMS__malloc(10*sizeof(void*));
197 _VMSMasterEnv->counter_history_array_info = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->counter_history),10);
198 */
199 //printf("Creating HW counters...");
200 /*
201 FILE* output;
202 int n;
203 char filename[255];
204 for(n=0;n<255;n++)
205 {
206 sprintf(filename, "./counters/Counters.%d.csv",n);
207 output = fopen(filename,"r");
208 if(output)
209 {
210 fclose(output);
211 }else{
212 break;
213 }
214 }
215 printf("Saving Counter measurements to File: %s ...\n", filename);
216 output = fopen(filename,"w+");
217 _VMSMasterEnv->counteroutput = output;
218 */
220 struct perf_event_attr hw_event;
221 memset(&hw_event,0,sizeof(hw_event));
222 hw_event.type = PERF_TYPE_HARDWARE;
223 hw_event.size = sizeof(hw_event);
224 hw_event.disabled = 1;
225 hw_event.freq = 0;
226 hw_event.inherit = 1; /* children inherit it */
227 hw_event.pinned = 1; /* must always be on PMU */
228 hw_event.exclusive = 0; /* only group on PMU */
229 hw_event.exclude_user = 0; /* don't count user */
230 hw_event.exclude_kernel = 1; /* ditto kernel */
231 hw_event.exclude_hv = 1; /* ditto hypervisor */
232 hw_event.exclude_idle = 0; /* don't count when idle */
233 hw_event.mmap = 0; /* include mmap data */
234 hw_event.comm = 0; /* include comm data */
237 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
238 {
239 hw_event.config = 0x0000000000000000; //cycles
240 _VMSMasterEnv->cycles_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
241 0,//pid_t pid,
242 coreIdx,//int cpu,
243 -1,//int group_fd,
244 0//unsigned long flags
245 );
246 if (_VMSMasterEnv->cycles_counter_fd[coreIdx]<0){
247 fprintf(stderr,"On core %d: ",coreIdx);
248 perror("Failed to open cycles counter");
249 }
250 hw_event.config = 0x0000000000000001; //instrs
251 _VMSMasterEnv->instrs_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
252 0,//pid_t pid,
253 coreIdx,//int cpu,
254 -1,//int group_fd,
255 0//unsigned long flags
256 );
257 if (_VMSMasterEnv->instrs_counter_fd[coreIdx]<0){
258 fprintf(stderr,"On core %d: ",coreIdx);
259 perror("Failed to open instrs counter");
260 }
261 }
262 prctl(PR_TASK_PERF_EVENTS_ENABLE);
263 uint64 tmpc,tmpi;
264 saveCyclesAndInstrs(0,tmpc,tmpi);
265 printf("Start: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
266 #endif
268 //========================================================================
270 }
272 SchedSlot **
273 create_sched_slots()
274 { SchedSlot **schedSlots;
275 int i;
277 schedSlots = VMS__malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
279 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
280 {
281 schedSlots[i] = VMS__malloc( sizeof(SchedSlot) );
283 //Set state to mean "handling requests done, slot needs filling"
284 schedSlots[i]->workIsDone = FALSE;
285 schedSlots[i]->needsProcrAssigned = TRUE;
286 }
287 return schedSlots;
288 }
291 void
292 freeSchedSlots( SchedSlot **schedSlots )
293 { int i;
294 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
295 {
296 VMS__free( schedSlots[i] );
297 }
298 VMS__free( schedSlots );
299 }
302 void
303 create_the_coreLoop_OS_threads()
304 {
305 //========================================================================
306 // Create the Threads
307 int coreIdx, retCode;
309 //Need the threads to be created suspended, and wait for a signal
310 // before proceeding -- gives time after creating to initialize other
311 // stuff before the coreLoops set off.
312 _VMSMasterEnv->setupComplete = 0;
314 //Make the threads that animate the core loops
315 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
316 { coreLoopThdParams[coreIdx] = VMS__malloc( sizeof(ThdParams) );
317 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
319 retCode =
320 pthread_create( &(coreLoopThdHandles[coreIdx]),
321 thdAttrs,
322 &coreLoop,
323 (void *)(coreLoopThdParams[coreIdx]) );
324 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
325 }
326 }
328 /*Semantic layer calls this when it want the system to start running..
329 *
330 *This starts the core loops running then waits for them to exit.
331 */
332 void
333 VMS__start_the_work_then_wait_until_done()
334 { int coreIdx;
335 //Start the core loops running
337 //tell the core loop threads that setup is complete
338 //get lock, to lock out any threads still starting up -- they'll see
339 // that setupComplete is true before entering while loop, and so never
340 // wait on the condition
341 pthread_mutex_lock( &suspendLock );
342 _VMSMasterEnv->setupComplete = 1;
343 pthread_mutex_unlock( &suspendLock );
344 pthread_cond_broadcast( &suspend_cond );
347 //wait for all to complete
348 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
349 {
350 pthread_join( coreLoopThdHandles[coreIdx], NULL );
351 }
353 //NOTE: do not clean up VMS env here -- semantic layer has to have
354 // a chance to clean up its environment first, then do a call to free
355 // the Master env and rest of VMS locations
356 }
358 #ifdef SEQUENTIAL
359 /*Only difference between version with an OS thread pinned to each core and
360 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
361 */
362 void
363 VMS__start_the_work_then_wait_until_done_Seq()
364 {
365 //Instead of un-suspending threads, just call the one and only
366 // core loop (sequential version), in the main thread.
367 coreLoop_Seq( NULL );
368 flushRegisters();
370 }
371 #endif
373 inline VirtProcr *
374 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
375 { VirtProcr *newPr;
376 void *stackLocs;
378 newPr = VMS__malloc( sizeof(VirtProcr) );
379 stackLocs = VMS__malloc( VIRT_PROCR_STACK_SIZE );
380 if( stackLocs == 0 )
381 { perror("VMS__malloc stack"); exit(1); }
383 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
384 }
386 /* "ext" designates that it's for use outside the VMS system -- should only
387 * be called from main thread or other thread -- never from code animated by
388 * a VMS virtual processor.
389 */
390 inline VirtProcr *
391 VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
392 { VirtProcr *newPr;
393 char *stackLocs;
395 newPr = malloc( sizeof(VirtProcr) );
396 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
397 if( stackLocs == 0 )
398 { perror("malloc stack"); exit(1); }
400 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
401 }
404 /*Anticipating multi-tasking
405 */
406 void *
407 VMS__give_sem_env_for( VirtProcr *animPr )
408 {
409 return _VMSMasterEnv->semanticEnv;
410 }
411 //===========================================================================
412 /*there is a label inside this function -- save the addr of this label in
413 * the callingPr struc, as the pick-up point from which to start the next
414 * work-unit for that procr. If turns out have to save registers, then
415 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
416 * "done with work-unit" label. The procr struc is in the request in the
417 * slave that animated the just-ended work-unit, so all the state is saved
418 * there, and will get passed along, inside the request handler, to the
419 * next work-unit for that procr.
420 */
421 void
422 VMS__suspend_procr( VirtProcr *animatingPr )
423 {
425 //The request to master will cause this suspended virt procr to get
426 // scheduled again at some future point -- to resume, core loop jumps
427 // to the resume point (below), which causes restore of saved regs and
428 // "return" from this call.
429 //animatingPr->nextInstrPt = &&ResumePt;
431 //return ownership of the virt procr and sched slot to Master virt pr
432 animatingPr->schedSlot->workIsDone = TRUE;
434 //=========================== Measurement stuff ========================
435 #ifdef MEAS__TIME_STAMP_SUSP
436 //record time stamp: compare to time-stamp recorded below
437 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
438 #endif
439 //=======================================================================
441 switchToCoreLoop(animatingPr);
442 flushRegisters();
444 //=======================================================================
446 #ifdef MEAS__TIME_STAMP_SUSP
447 //NOTE: only take low part of count -- do sanity check when take diff
448 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
449 #endif
451 return;
452 }
456 /*For this implementation of VMS, it may not make much sense to have the
457 * system of requests for creating a new processor done this way.. but over
458 * the scope of single-master, multi-master, mult-tasking, OS-implementing,
459 * distributed-memory, and so on, this gives VMS implementation a chance to
460 * do stuff before suspend, in the AppVP, and in the Master before the plugin
461 * is called, as well as in the lang-lib before this is called, and in the
462 * plugin. So, this gives both VMS and language implementations a chance to
463 * intercept at various points and do order-dependent stuff.
464 *Having a standard VMSNewPrReqData struc allows the language to create and
465 * free the struc, while VMS knows how to get the newPr if it wants it, and
466 * it lets the lang have lang-specific data related to creation transported
467 * to the plugin.
468 */
469 void
470 VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr )
471 { VMSReqst req;
473 req.reqType = createReq;
474 req.semReqData = semReqData;
475 req.nextReqst = reqstingPr->requests;
476 reqstingPr->requests = &req;
478 VMS__suspend_procr( reqstingPr );
479 }
482 /*
483 *This adds a request to dissipate, then suspends the processor so that the
484 * request handler will receive the request. The request handler is what
485 * does the work of freeing memory and removing the processor from the
486 * semantic environment's data structures.
487 *The request handler also is what figures out when to shutdown the VMS
488 * system -- which causes all the core loop threads to die, and returns from
489 * the call that started up VMS to perform the work.
490 *
491 *This form is a bit misleading to understand if one is trying to figure out
492 * how VMS works -- it looks like a normal function call, but inside it
493 * sends a request to the request handler and suspends the processor, which
494 * jumps out of the VMS__dissipate_procr function, and out of all nestings
495 * above it, transferring the work of dissipating to the request handler,
496 * which then does the actual work -- causing the processor that animated
497 * the call of this function to disappear and the "hanging" state of this
498 * function to just poof into thin air -- the virtual processor's trace
499 * never returns from this call, but instead the virtual processor's trace
500 * gets suspended in this call and all the virt processor's state disap-
501 * pears -- making that suspend the last thing in the virt procr's trace.
502 */
503 void
504 VMS__send_dissipate_req( VirtProcr *procrToDissipate )
505 { VMSReqst req;
507 req.reqType = dissipate;
508 req.nextReqst = procrToDissipate->requests;
509 procrToDissipate->requests = &req;
511 VMS__suspend_procr( procrToDissipate );
512 }
515 /* "ext" designates that it's for use outside the VMS system -- should only
516 * be called from main thread or other thread -- never from code animated by
517 * a VMS virtual processor.
518 *
519 *Use this version to dissipate VPs created outside the VMS system.
520 */
521 void
522 VMS_ext__dissipate_procr( VirtProcr *procrToDissipate )
523 {
524 //NOTE: initialData was given to the processor, so should either have
525 // been alloc'd with VMS__malloc, or freed by the level above animPr.
526 //So, all that's left to free here is the stack and the VirtProcr struc
527 // itself
528 //Note, should not stack-allocate initial data -- no guarantee, in
529 // general that creating processor will outlive ones it creates.
530 free( procrToDissipate->startOfStack );
531 free( procrToDissipate );
532 }
536 /*This call's name indicates that request is malloc'd -- so req handler
537 * has to free any extra requests tacked on before a send, using this.
538 *
539 * This inserts the semantic-layer's request data into standard VMS carrier
540 * request data-struct that is mallocd. The sem request doesn't need to
541 * be malloc'd if this is called inside the same call chain before the
542 * send of the last request is called.
543 *
544 *The request handler has to call VMS__free_VMSReq for any of these
545 */
546 inline void
547 VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData,
548 VirtProcr *callingPr )
549 { VMSReqst *req;
551 req = VMS__malloc( sizeof(VMSReqst) );
552 req->reqType = semantic;
553 req->semReqData = semReqData;
554 req->nextReqst = callingPr->requests;
555 callingPr->requests = req;
556 }
558 /*This inserts the semantic-layer's request data into standard VMS carrier
559 * request data-struct is allocated on stack of this call & ptr to it sent
560 * to plugin
561 *Then it does suspend, to cause request to be sent.
562 */
563 inline void
564 VMS__send_sem_request( void *semReqData, VirtProcr *callingPr )
565 { VMSReqst req;
567 req.reqType = semantic;
568 req.semReqData = semReqData;
569 req.nextReqst = callingPr->requests;
570 callingPr->requests = &req;
572 VMS__suspend_procr( callingPr );
573 }
576 inline void
577 VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr )
578 { VMSReqst req;
580 req.reqType = VMSSemantic;
581 req.semReqData = semReqData;
582 req.nextReqst = callingPr->requests; //gab any other preceeding
583 callingPr->requests = &req;
585 VMS__suspend_procr( callingPr );
586 }
589 /*
590 */
591 VMSReqst *
592 VMS__take_next_request_out_of( VirtProcr *procrWithReq )
593 { VMSReqst *req;
595 req = procrWithReq->requests;
596 if( req == NULL ) return NULL;
598 procrWithReq->requests = procrWithReq->requests->nextReqst;
599 return req;
600 }
603 inline void *
604 VMS__take_sem_reqst_from( VMSReqst *req )
605 {
606 return req->semReqData;
607 }
611 /* This is for OS requests and VMS infrastructure requests, such as to create
612 * a probe -- a probe is inside the heart of VMS-core, it's not part of any
613 * language -- but it's also a semantic thing that's triggered from and used
614 * in the application.. so it crosses abstractions.. so, need some special
615 * pattern here for handling such requests.
616 * Doing this just like it were a second language sharing VMS-core.
617 *
618 * This is called from the language's request handler when it sees a request
619 * of type VMSSemReq
620 *
621 * TODO: Later change this, to give probes their own separate plugin & have
622 * VMS-core steer the request to appropriate plugin
623 * Do the same for OS calls -- look later at it..
624 */
625 void inline
626 VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv,
627 ResumePrFnPtr resumePrFnPtr )
628 { VMSSemReq *semReq;
629 IntervalProbe *newProbe;
631 semReq = req->semReqData;
633 newProbe = VMS__malloc( sizeof(IntervalProbe) );
634 newProbe->nameStr = VMS__strDup( semReq->nameStr );
635 newProbe->hist = NULL;
636 newProbe->schedChoiceWasRecorded = FALSE;
638 //This runs in masterVP, so no race-condition worries
639 newProbe->probeID =
640 addToDynArray( newProbe, _VMSMasterEnv->dynIntervalProbesInfo );
642 requestingPr->dataRetFromReq = newProbe;
644 (*resumePrFnPtr)( requestingPr, semEnv );
645 }
649 /*This must be called by the request handler plugin -- it cannot be called
650 * from the semantic library "dissipate processor" function -- instead, the
651 * semantic layer has to generate a request, and the plug-in calls this
652 * function.
653 *The reason is that this frees the virtual processor's stack -- which is
654 * still in use inside semantic library calls!
655 *
656 *This frees or recycles all the state owned by and comprising the VMS
657 * portion of the animating virtual procr. The request handler must first
658 * free any semantic data created for the processor that didn't use the
659 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
660 * system to disown any state that did use VMS_malloc, and then frees the
661 * statck and the processor-struct itself.
662 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
663 * state, then that state gets freed (or sent to recycling) as a side-effect
664 * of dis-owning it.
665 */
666 void
667 VMS__dissipate_procr( VirtProcr *animatingPr )
668 {
669 //dis-own all locations owned by this processor, causing to be freed
670 // any locations that it is (was) sole owner of
671 //TODO: implement VMS__malloc system, including "give up ownership"
674 //NOTE: initialData was given to the processor, so should either have
675 // been alloc'd with VMS__malloc, or freed by the level above animPr.
676 //So, all that's left to free here is the stack and the VirtProcr struc
677 // itself
678 //Note, should not stack-allocate initial data -- no guarantee, in
679 // general that creating processor will outlive ones it creates.
680 VMS__free( animatingPr->startOfStack );
681 VMS__free( animatingPr );
682 }
685 //TODO: look at architecting cleanest separation between request handler
686 // and master loop, for dissipate, create, shutdown, and other non-semantic
687 // requests. Issue is chain: one removes requests from AppVP, one dispatches
688 // on type of request, and one handles each type.. but some types require
689 // action from both request handler and master loop -- maybe just give the
690 // request handler calls like: VMS__handle_X_request_type
693 /*This is called by the semantic layer's request handler when it decides its
694 * time to shut down the VMS system. Calling this causes the core loop OS
695 * threads to exit, which unblocks the entry-point function that started up
696 * VMS, and allows it to grab the result and return to the original single-
697 * threaded application.
698 *
699 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
700 * and-wait function has to free a bunch of stuff after it detects the
701 * threads have all died: the masterEnv, the thread-related locations,
702 * masterVP any AppVPs that might still be allocated and sitting in the
703 * semantic environment, or have been orphaned in the _VMSWorkQ.
704 *
705 *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the
706 * locations it needs, and give ownership to masterVP. Then, they will be
707 * automatically freed.
708 *
709 *In here,create one core-loop shut-down processor for each core loop and put
710 * them all directly into the readyToAnimateQ.
711 *Note, this function can ONLY be called after the semantic environment no
712 * longer cares if AppVPs get animated after the point this is called. In
713 * other words, this can be used as an abort, or else it should only be
714 * called when all AppVPs have finished dissipate requests -- only at that
715 * point is it sure that all results have completed.
716 */
717 void
718 VMS__shutdown()
719 { int coreIdx;
720 VirtProcr *shutDownPr;
722 //create the shutdown processors, one for each core loop -- put them
723 // directly into the Q -- each core will die when gets one
724 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
725 { //Note, this is running in the master
726 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
727 writeVMSQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
728 }
729 #ifdef MEAS__PERF_COUNTERS
730 uint64 tmpc,tmpi;
731 saveCyclesAndInstrs(0,tmpc,tmpi);
732 printf("End: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
733 prctl(PR_TASK_PERF_EVENTS_DISABLE);
734 /*
735 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ){
736 close(_VMSMasterEnv->cycles_counter_fd[coreIdx]);
737 close(_VMSMasterEnv->instrs_counter_fd[coreIdx]);
738 }
739 */
740 #endif
741 }
744 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
745 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
746 *This function has the sole purpose of setting the stack and framePtr
747 * to the coreLoop's stack and framePtr.. it does that then jumps to the
748 * core loop's shutdown point -- might be able to just call Pthread_exit
749 * from here, but am going back to the pthread's stack and setting everything
750 * up just as if it never jumped out, before calling pthread_exit.
751 *The end-point of core loop will free the stack and so forth of the
752 * processor that animates this function, (this fn is transfering the
753 * animator of the AppVP that is in turn animating this function over
754 * to core loop function -- note that this slices out a level of virtual
755 * processors).
756 */
757 void
758 endOSThreadFn( void *initData, VirtProcr *animatingPr )
759 {
760 #ifdef SEQUENTIAL
761 asmTerminateCoreLoopSeq(animatingPr);
762 #else
763 asmTerminateCoreLoop(animatingPr);
764 #endif
765 }
768 /*This is called from the startup & shutdown
769 */
770 void
771 VMS__cleanup_at_end_of_shutdown()
772 {
773 //unused
774 //VMSQueueStruc **readyToAnimateQs;
775 //int coreIdx;
776 //VirtProcr **masterVPs;
777 //SchedSlot ***allSchedSlots; //ptr to array of ptrs
779 //Before getting rid of everything, print out any measurements made
780 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&printHist );
781 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&saveHistToFile);
782 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, &freeHist );
785 #ifdef MEAS__TIME_PLUGIN
786 printHist( _VMSMasterEnv->reqHdlrLowTimeHist );
787 saveHistToFile( _VMSMasterEnv->reqHdlrLowTimeHist );
788 printHist( _VMSMasterEnv->reqHdlrHighTimeHist );
789 saveHistToFile( _VMSMasterEnv->reqHdlrHighTimeHist );
790 freeHistExt( _VMSMasterEnv->reqHdlrLowTimeHist );
791 freeHistExt( _VMSMasterEnv->reqHdlrHighTimeHist );
792 #endif
793 #ifdef MEAS__TIME_MALLOC
794 printHist( _VMSMasterEnv->mallocTimeHist );
795 saveHistToFile( _VMSMasterEnv->mallocTimeHist );
796 printHist( _VMSMasterEnv->freeTimeHist );
797 saveHistToFile( _VMSMasterEnv->freeTimeHist );
798 freeHistExt( _VMSMasterEnv->mallocTimeHist );
799 freeHistExt( _VMSMasterEnv->freeTimeHist );
800 #endif
801 #ifdef MEAS__TIME_MASTER_LOCK
802 printHist( _VMSMasterEnv->masterLockLowTimeHist );
803 printHist( _VMSMasterEnv->masterLockHighTimeHist );
804 #endif
805 #ifdef MEAS__TIME_MASTER
806 printHist( _VMSMasterEnv->pluginTimeHist );
807 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
808 {
809 freeVMSQ( readyToAnimateQs[ coreIdx ] );
810 //master VPs were created external to VMS, so use external free
811 VMS__dissipate_procr( masterVPs[ coreIdx ] );
813 freeSchedSlots( allSchedSlots[ coreIdx ] );
814 }
815 #endif
816 #ifdef MEAS__TIME_STAMP_SUSP
817 printHist( _VMSMasterEnv->pluginTimeHist );
818 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
819 {
820 freeVMSQ( readyToAnimateQs[ coreIdx ] );
821 //master VPs were created external to VMS, so use external free
822 VMS__dissipate_procr( masterVPs[ coreIdx ] );
824 freeSchedSlots( allSchedSlots[ coreIdx ] );
825 }
826 #endif
828 //All the environment data has been allocated with VMS__malloc, so just
829 // free its internal big-chunk and all inside it disappear.
830 /*
831 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
832 masterVPs = _VMSMasterEnv->masterVPs;
833 allSchedSlots = _VMSMasterEnv->allSchedSlots;
835 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
836 {
837 freeVMSQ( readyToAnimateQs[ coreIdx ] );
838 //master VPs were created external to VMS, so use external free
839 VMS__dissipate_procr( masterVPs[ coreIdx ] );
841 freeSchedSlots( allSchedSlots[ coreIdx ] );
842 }
844 VMS__free( _VMSMasterEnv->readyToAnimateQs );
845 VMS__free( _VMSMasterEnv->masterVPs );
846 VMS__free( _VMSMasterEnv->allSchedSlots );
848 //============================= MEASUREMENT STUFF ========================
849 #ifdef STATS__TURN_ON_PROBES
850 freeDynArrayDeep( _VMSMasterEnv->dynIntervalProbesInfo, &VMS__free_probe);
851 #endif
852 //========================================================================
853 */
854 //These are the only two that use system free
855 VMS_ext__free_free_list( _VMSMasterEnv->freeListHead );
856 free( (void *)_VMSMasterEnv );
857 }
860 //================================
863 /*Later, improve this -- for now, just exits the application after printing
864 * the error message.
865 */
866 void
867 VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData )
868 {
869 printf("%s",msgStr);
870 fflush(stdin);
871 exit(1);
872 }