Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view VMS.c @ 166:aefd87f9d12f
loop graph
| author | Nina Engelhardt |
|---|---|
| date | Tue, 18 Oct 2011 15:53:04 +0200 |
| parents | 395f58384a5c |
| children | 981acd1db6af |
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, 100,
122 "malloc_time_hist");
123 _VMSMasterEnv->freeTimeHist = makeFixedBinHistExt( 80, 0, 100,
124 "free_time_hist");
125 #endif
126 #ifdef MEAS__TIME_PLUGIN
127 _VMSMasterEnv->reqHdlrLowTimeHist = makeFixedBinHistExt( 1000, 0, 100,
128 "plugin_low_time_hist");
129 _VMSMasterEnv->reqHdlrHighTimeHist = makeFixedBinHistExt( 1000, 0, 100,
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();
189 #ifdef DETECT_DEPENDENCIES
190 _VMSMasterEnv->dependencies = VMS__malloc(10*sizeof(void*));
191 _VMSMasterEnv->dependenciesInfo = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->dependencies),10);
192 #endif
194 #ifdef DETECT_LOOP_GRAPH
195 _VMSMasterEnv->loop_graph = VMS__malloc(10*sizeof(void*));
196 _VMSMasterEnv->loop_graph_array_info = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->loop_graph),10);
197 int loop_i;
198 for(loop_i=0;loop_i<NUM_CORES;loop_i++){
199 _VMSMasterEnv->loop_counter[loop_i]=0;
200 }
202 #endif
204 #ifdef MEAS__PERF_COUNTERS
205 _VMSMasterEnv->counter_history = VMS__malloc(10*sizeof(void*));
206 _VMSMasterEnv->counter_history_array_info = makePrivDynArrayInfoFrom((void***)&(_VMSMasterEnv->counter_history),10);
207 //printf("Creating HW counters...");
208 FILE* output;
209 int n;
210 char filename[255];
211 for(n=0;n<255;n++)
212 {
213 sprintf(filename, "./counters/Counters.%d.csv",n);
214 output = fopen(filename,"r");
215 if(output)
216 {
217 fclose(output);
218 }else{
219 break;
220 }
221 }
222 printf("Saving Counter measurements to File: %s ...\n", filename);
223 output = fopen(filename,"w+");
224 _VMSMasterEnv->counteroutput = output;
226 struct perf_event_attr hw_event;
227 memset(&hw_event,0,sizeof(hw_event));
228 hw_event.type = PERF_TYPE_HARDWARE;
229 hw_event.size = sizeof(hw_event);
230 hw_event.disabled = 1;
231 hw_event.freq = 0;
232 hw_event.inherit = 1; /* children inherit it */
233 hw_event.pinned = 1; /* must always be on PMU */
234 hw_event.exclusive = 0; /* only group on PMU */
235 hw_event.exclude_user = 0; /* don't count user */
236 hw_event.exclude_kernel = 1; /* ditto kernel */
237 hw_event.exclude_hv = 1; /* ditto hypervisor */
238 hw_event.exclude_idle = 0; /* don't count when idle */
239 hw_event.mmap = 0; /* include mmap data */
240 hw_event.comm = 0; /* include comm data */
243 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
244 {
245 hw_event.config = 0x0000000000000000; //cycles
246 _VMSMasterEnv->cycles_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
247 0,//pid_t pid,
248 -1,//int cpu,
249 -1,//int group_fd,
250 0//unsigned long flags
251 );
252 if (_VMSMasterEnv->cycles_counter_fd[coreIdx]<0){
253 fprintf(stderr,"On core %d: ",coreIdx);
254 perror("Failed to open cycles counter");
255 }
256 hw_event.config = 0x0000000000000001; //instrs
257 _VMSMasterEnv->instrs_counter_fd[coreIdx] = syscall(__NR_perf_event_open, &hw_event,
258 0,//pid_t pid,
259 -1,//int cpu,
260 -1,//int group_fd,
261 0//unsigned long flags
262 );
263 if (_VMSMasterEnv->instrs_counter_fd[coreIdx]<0){
264 fprintf(stderr,"On core %d: ",coreIdx);
265 perror("Failed to open instrs counter");
266 }
267 }
268 prctl(PR_TASK_PERF_EVENTS_ENABLE);
269 uint64 tmpc,tmpi;
270 saveCyclesAndInstrs(0,tmpc,tmpi);
271 printf("Start: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
272 #endif
274 //========================================================================
276 }
278 SchedSlot **
279 create_sched_slots()
280 { SchedSlot **schedSlots;
281 int i;
283 schedSlots = VMS__malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
285 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
286 {
287 schedSlots[i] = VMS__malloc( sizeof(SchedSlot) );
289 //Set state to mean "handling requests done, slot needs filling"
290 schedSlots[i]->workIsDone = FALSE;
291 schedSlots[i]->needsProcrAssigned = TRUE;
292 }
293 return schedSlots;
294 }
297 void
298 freeSchedSlots( SchedSlot **schedSlots )
299 { int i;
300 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
301 {
302 VMS__free( schedSlots[i] );
303 }
304 VMS__free( schedSlots );
305 }
308 void
309 create_the_coreLoop_OS_threads()
310 {
311 //========================================================================
312 // Create the Threads
313 int coreIdx, retCode;
315 //Need the threads to be created suspended, and wait for a signal
316 // before proceeding -- gives time after creating to initialize other
317 // stuff before the coreLoops set off.
318 _VMSMasterEnv->setupComplete = 0;
320 //Make the threads that animate the core loops
321 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
322 { coreLoopThdParams[coreIdx] = VMS__malloc( sizeof(ThdParams) );
323 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
325 retCode =
326 pthread_create( &(coreLoopThdHandles[coreIdx]),
327 thdAttrs,
328 &coreLoop,
329 (void *)(coreLoopThdParams[coreIdx]) );
330 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(1);}
331 }
332 }
334 /*Semantic layer calls this when it want the system to start running..
335 *
336 *This starts the core loops running then waits for them to exit.
337 */
338 void
339 VMS__start_the_work_then_wait_until_done()
340 { int coreIdx;
341 //Start the core loops running
343 //tell the core loop threads that setup is complete
344 //get lock, to lock out any threads still starting up -- they'll see
345 // that setupComplete is true before entering while loop, and so never
346 // wait on the condition
347 pthread_mutex_lock( &suspendLock );
348 _VMSMasterEnv->setupComplete = 1;
349 pthread_mutex_unlock( &suspendLock );
350 pthread_cond_broadcast( &suspend_cond );
353 //wait for all to complete
354 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
355 {
356 pthread_join( coreLoopThdHandles[coreIdx], NULL );
357 }
359 //NOTE: do not clean up VMS env here -- semantic layer has to have
360 // a chance to clean up its environment first, then do a call to free
361 // the Master env and rest of VMS locations
362 }
364 #ifdef SEQUENTIAL
365 /*Only difference between version with an OS thread pinned to each core and
366 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
367 */
368 void
369 VMS__start_the_work_then_wait_until_done_Seq()
370 {
371 //Instead of un-suspending threads, just call the one and only
372 // core loop (sequential version), in the main thread.
373 coreLoop_Seq( NULL );
374 flushRegisters();
376 }
377 #endif
379 inline VirtProcr *
380 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
381 { VirtProcr *newPr;
382 void *stackLocs;
384 newPr = VMS__malloc( sizeof(VirtProcr) );
385 stackLocs = VMS__malloc( VIRT_PROCR_STACK_SIZE );
386 if( stackLocs == 0 )
387 { perror("VMS__malloc stack"); exit(1); }
389 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
390 }
392 /* "ext" designates that it's for use outside the VMS system -- should only
393 * be called from main thread or other thread -- never from code animated by
394 * a VMS virtual processor.
395 */
396 inline VirtProcr *
397 VMS_ext__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
398 { VirtProcr *newPr;
399 char *stackLocs;
401 newPr = malloc( sizeof(VirtProcr) );
402 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
403 if( stackLocs == 0 )
404 { perror("malloc stack"); exit(1); }
406 return create_procr_helper( newPr, fnPtr, initialData, stackLocs );
407 }
410 /*Anticipating multi-tasking
411 */
412 void *
413 VMS__give_sem_env_for( VirtProcr *animPr )
414 {
415 return _VMSMasterEnv->semanticEnv;
416 }
417 //===========================================================================
418 /*there is a label inside this function -- save the addr of this label in
419 * the callingPr struc, as the pick-up point from which to start the next
420 * work-unit for that procr. If turns out have to save registers, then
421 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
422 * "done with work-unit" label. The procr struc is in the request in the
423 * slave that animated the just-ended work-unit, so all the state is saved
424 * there, and will get passed along, inside the request handler, to the
425 * next work-unit for that procr.
426 */
427 void
428 VMS__suspend_procr( VirtProcr *animatingPr )
429 {
431 //The request to master will cause this suspended virt procr to get
432 // scheduled again at some future point -- to resume, core loop jumps
433 // to the resume point (below), which causes restore of saved regs and
434 // "return" from this call.
435 //animatingPr->nextInstrPt = &&ResumePt;
437 //return ownership of the virt procr and sched slot to Master virt pr
438 animatingPr->schedSlot->workIsDone = TRUE;
440 //=========================== Measurement stuff ========================
441 #ifdef MEAS__TIME_STAMP_SUSP
442 //record time stamp: compare to time-stamp recorded below
443 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
444 #endif
445 //=======================================================================
447 switchToCoreLoop(animatingPr);
448 flushRegisters();
450 //=======================================================================
452 #ifdef MEAS__TIME_STAMP_SUSP
453 //NOTE: only take low part of count -- do sanity check when take diff
454 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
455 #endif
457 return;
458 }
462 /*For this implementation of VMS, it may not make much sense to have the
463 * system of requests for creating a new processor done this way.. but over
464 * the scope of single-master, multi-master, mult-tasking, OS-implementing,
465 * distributed-memory, and so on, this gives VMS implementation a chance to
466 * do stuff before suspend, in the AppVP, and in the Master before the plugin
467 * is called, as well as in the lang-lib before this is called, and in the
468 * plugin. So, this gives both VMS and language implementations a chance to
469 * intercept at various points and do order-dependent stuff.
470 *Having a standard VMSNewPrReqData struc allows the language to create and
471 * free the struc, while VMS knows how to get the newPr if it wants it, and
472 * it lets the lang have lang-specific data related to creation transported
473 * to the plugin.
474 */
475 __attribute__ ((noinline)) void
476 VMS__send_create_procr_req( void *semReqData, VirtProcr *reqstingPr )
478 { VMSReqst req;
480 req.reqType = createReq;
481 req.semReqData = semReqData;
482 req.nextReqst = reqstingPr->requests;
483 reqstingPr->requests = &req;
485 VMS__suspend_procr( reqstingPr );
486 }
489 /*
490 *This adds a request to dissipate, then suspends the processor so that the
491 * request handler will receive the request. The request handler is what
492 * does the work of freeing memory and removing the processor from the
493 * semantic environment's data structures.
494 *The request handler also is what figures out when to shutdown the VMS
495 * system -- which causes all the core loop threads to die, and returns from
496 * the call that started up VMS to perform the work.
497 *
498 *This form is a bit misleading to understand if one is trying to figure out
499 * how VMS works -- it looks like a normal function call, but inside it
500 * sends a request to the request handler and suspends the processor, which
501 * jumps out of the VMS__dissipate_procr function, and out of all nestings
502 * above it, transferring the work of dissipating to the request handler,
503 * which then does the actual work -- causing the processor that animated
504 * the call of this function to disappear and the "hanging" state of this
505 * function to just poof into thin air -- the virtual processor's trace
506 * never returns from this call, but instead the virtual processor's trace
507 * gets suspended in this call and all the virt processor's state disap-
508 * pears -- making that suspend the last thing in the virt procr's trace.
509 */
510 __attribute__ ((noinline)) void
511 VMS__send_dissipate_req( VirtProcr *procrToDissipate )
512 { VMSReqst req;
514 req.reqType = dissipate;
515 req.nextReqst = procrToDissipate->requests;
516 procrToDissipate->requests = &req;
518 VMS__suspend_procr( procrToDissipate );
519 }
522 /* "ext" designates that it's for use outside the VMS system -- should only
523 * be called from main thread or other thread -- never from code animated by
524 * a VMS virtual processor.
525 *
526 *Use this version to dissipate VPs created outside the VMS system.
527 */
528 void
529 VMS_ext__dissipate_procr( VirtProcr *procrToDissipate )
530 {
531 //NOTE: initialData was given to the processor, so should either have
532 // been alloc'd with VMS__malloc, or freed by the level above animPr.
533 //So, all that's left to free here is the stack and the VirtProcr struc
534 // itself
535 //Note, should not stack-allocate initial data -- no guarantee, in
536 // general that creating processor will outlive ones it creates.
537 free( procrToDissipate->startOfStack );
538 free( procrToDissipate );
539 }
543 /*This call's name indicates that request is malloc'd -- so req handler
544 * has to free any extra requests tacked on before a send, using this.
545 *
546 * This inserts the semantic-layer's request data into standard VMS carrier
547 * request data-struct that is mallocd. The sem request doesn't need to
548 * be malloc'd if this is called inside the same call chain before the
549 * send of the last request is called.
550 *
551 *The request handler has to call VMS__free_VMSReq for any of these
552 */
553 inline void
554 VMS__add_sem_request_in_mallocd_VMSReqst( void *semReqData,
555 VirtProcr *callingPr )
556 { VMSReqst *req;
558 req = VMS__malloc( sizeof(VMSReqst) );
559 req->reqType = semantic;
560 req->semReqData = semReqData;
561 req->nextReqst = callingPr->requests;
562 callingPr->requests = req;
563 }
565 /*This inserts the semantic-layer's request data into standard VMS carrier
566 * request data-struct is allocated on stack of this call & ptr to it sent
567 * to plugin
568 *Then it does suspend, to cause request to be sent.
569 */
570 /*inline*/__attribute__ ((noinline)) void
571 VMS__send_sem_request( void *semReqData, VirtProcr *callingPr )
572 { VMSReqst req;
574 req.reqType = semantic;
575 req.semReqData = semReqData;
576 req.nextReqst = callingPr->requests;
577 callingPr->requests = &req;
579 VMS__suspend_procr( callingPr );
580 }
583 /*inline*/ __attribute__ ((noinline)) void
584 VMS__send_VMSSem_request( void *semReqData, VirtProcr *callingPr )
586 { VMSReqst req;
588 req.reqType = VMSSemantic;
589 req.semReqData = semReqData;
590 req.nextReqst = callingPr->requests; //gab any other preceeding
591 callingPr->requests = &req;
593 VMS__suspend_procr( callingPr );
594 }
597 /*
598 */
599 VMSReqst *
600 VMS__take_next_request_out_of( VirtProcr *procrWithReq )
601 { VMSReqst *req;
603 req = procrWithReq->requests;
604 if( req == NULL ) return NULL;
606 procrWithReq->requests = procrWithReq->requests->nextReqst;
607 return req;
608 }
611 inline void *
612 VMS__take_sem_reqst_from( VMSReqst *req )
613 {
614 return req->semReqData;
615 }
619 /* This is for OS requests and VMS infrastructure requests, such as to create
620 * a probe -- a probe is inside the heart of VMS-core, it's not part of any
621 * language -- but it's also a semantic thing that's triggered from and used
622 * in the application.. so it crosses abstractions.. so, need some special
623 * pattern here for handling such requests.
624 * Doing this just like it were a second language sharing VMS-core.
625 *
626 * This is called from the language's request handler when it sees a request
627 * of type VMSSemReq
628 *
629 * TODO: Later change this, to give probes their own separate plugin & have
630 * VMS-core steer the request to appropriate plugin
631 * Do the same for OS calls -- look later at it..
632 */
633 void inline
634 VMS__handle_VMSSemReq( VMSReqst *req, VirtProcr *requestingPr, void *semEnv,
635 ResumePrFnPtr resumePrFnPtr )
636 { VMSSemReq *semReq;
637 IntervalProbe *newProbe;
639 semReq = req->semReqData;
641 newProbe = VMS__malloc( sizeof(IntervalProbe) );
642 newProbe->nameStr = VMS__strDup( semReq->nameStr );
643 newProbe->hist = NULL;
644 newProbe->schedChoiceWasRecorded = FALSE;
646 //This runs in masterVP, so no race-condition worries
647 newProbe->probeID =
648 addToDynArray( newProbe, _VMSMasterEnv->dynIntervalProbesInfo );
650 requestingPr->dataRetFromReq = newProbe;
652 (*resumePrFnPtr)( requestingPr, semEnv );
653 }
657 /*This must be called by the request handler plugin -- it cannot be called
658 * from the semantic library "dissipate processor" function -- instead, the
659 * semantic layer has to generate a request, and the plug-in calls this
660 * function.
661 *The reason is that this frees the virtual processor's stack -- which is
662 * still in use inside semantic library calls!
663 *
664 *This frees or recycles all the state owned by and comprising the VMS
665 * portion of the animating virtual procr. The request handler must first
666 * free any semantic data created for the processor that didn't use the
667 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
668 * system to disown any state that did use VMS_malloc, and then frees the
669 * statck and the processor-struct itself.
670 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
671 * state, then that state gets freed (or sent to recycling) as a side-effect
672 * of dis-owning it.
673 */
674 void
675 VMS__dissipate_procr( VirtProcr *animatingPr )
676 {
677 //dis-own all locations owned by this processor, causing to be freed
678 // any locations that it is (was) sole owner of
679 //TODO: implement VMS__malloc system, including "give up ownership"
682 //NOTE: initialData was given to the processor, so should either have
683 // been alloc'd with VMS__malloc, or freed by the level above animPr.
684 //So, all that's left to free here is the stack and the VirtProcr struc
685 // itself
686 //Note, should not stack-allocate initial data -- no guarantee, in
687 // general that creating processor will outlive ones it creates.
688 VMS__free( animatingPr->startOfStack );
689 VMS__free( animatingPr );
690 }
693 //TODO: look at architecting cleanest separation between request handler
694 // and master loop, for dissipate, create, shutdown, and other non-semantic
695 // requests. Issue is chain: one removes requests from AppVP, one dispatches
696 // on type of request, and one handles each type.. but some types require
697 // action from both request handler and master loop -- maybe just give the
698 // request handler calls like: VMS__handle_X_request_type
701 /*This is called by the semantic layer's request handler when it decides its
702 * time to shut down the VMS system. Calling this causes the core loop OS
703 * threads to exit, which unblocks the entry-point function that started up
704 * VMS, and allows it to grab the result and return to the original single-
705 * threaded application.
706 *
707 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
708 * and-wait function has to free a bunch of stuff after it detects the
709 * threads have all died: the masterEnv, the thread-related locations,
710 * masterVP any AppVPs that might still be allocated and sitting in the
711 * semantic environment, or have been orphaned in the _VMSWorkQ.
712 *
713 *NOTE: the semantic plug-in is expected to use VMS__malloc to get all the
714 * locations it needs, and give ownership to masterVP. Then, they will be
715 * automatically freed.
716 *
717 *In here,create one core-loop shut-down processor for each core loop and put
718 * them all directly into the readyToAnimateQ.
719 *Note, this function can ONLY be called after the semantic environment no
720 * longer cares if AppVPs get animated after the point this is called. In
721 * other words, this can be used as an abort, or else it should only be
722 * called when all AppVPs have finished dissipate requests -- only at that
723 * point is it sure that all results have completed.
724 */
725 void
726 VMS__shutdown()
727 { int coreIdx;
728 VirtProcr *shutDownPr;
730 //create the shutdown processors, one for each core loop -- put them
731 // directly into the Q -- each core will die when gets one
732 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
733 { //Note, this is running in the master
734 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
735 writeVMSQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
736 }
737 #ifdef MEAS__PERF_COUNTERS
738 uint64 tmpc,tmpi;
739 saveCyclesAndInstrs(0,tmpc,tmpi);
740 printf("End: cycles = %llu, instrs = %llu\n",tmpc,tmpi);
741 prctl(PR_TASK_PERF_EVENTS_DISABLE);
742 /*
743 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ ){
744 close(_VMSMasterEnv->cycles_counter_fd[coreIdx]);
745 close(_VMSMasterEnv->instrs_counter_fd[coreIdx]);
746 }
747 */
748 #endif
749 }
752 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
753 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
754 *This function has the sole purpose of setting the stack and framePtr
755 * to the coreLoop's stack and framePtr.. it does that then jumps to the
756 * core loop's shutdown point -- might be able to just call Pthread_exit
757 * from here, but am going back to the pthread's stack and setting everything
758 * up just as if it never jumped out, before calling pthread_exit.
759 *The end-point of core loop will free the stack and so forth of the
760 * processor that animates this function, (this fn is transfering the
761 * animator of the AppVP that is in turn animating this function over
762 * to core loop function -- note that this slices out a level of virtual
763 * processors).
764 */
765 void
766 endOSThreadFn( void *initData, VirtProcr *animatingPr )
767 {
768 #ifdef SEQUENTIAL
769 asmTerminateCoreLoopSeq(animatingPr);
770 #else
771 asmTerminateCoreLoop(animatingPr);
772 #endif
773 }
776 /*This is called from the startup & shutdown
777 */
778 void
779 VMS__cleanup_at_end_of_shutdown()
780 {
781 //unused
782 //VMSQueueStruc **readyToAnimateQs;
783 //int coreIdx;
784 //VirtProcr **masterVPs;
785 //SchedSlot ***allSchedSlots; //ptr to array of ptrs
787 //Before getting rid of everything, print out any measurements made
788 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&printHist );
789 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, (DynArrayFnPtr)&saveHistToFile);
790 //forAllInDynArrayDo( _VMSMasterEnv->measHistsInfo, &freeHistExt );
791 #ifdef DETECT_DEPENDENCIES
792 FILE* output;
793 int n;
794 char filename[255];
795 for(n=0;n<255;n++)
796 {
797 sprintf(filename, "./counters/Dependencies.%d.dot",n);
798 output = fopen(filename,"r");
799 if(output)
800 {
801 fclose(output);
802 }else{
803 break;
804 }
805 }
806 if(n<255){
807 printf("Saving Dependencies to File: %s ...\n", filename);
808 output = fopen(filename,"w+");
809 if(output!=NULL){
810 set_dependency_file(output);
811 fprintf(output,"digraph Dependencies {\n");
812 set_dot_file(output);
813 forAllInDynArrayDo(_VMSMasterEnv->counter_history_array_info, &print_dot_node_info );
814 forAllInDynArrayDo( _VMSMasterEnv->dependenciesInfo, &print_dependency_to_file );
815 fprintf(output,"}\n");
816 } else
817 printf("Opening Dependencies file failed. Please check that folder \"counters\" exists in run directory.\n");
818 } else {
819 printf("Could not open Dependencies file, please clean \"counters\" folder. (Must contain less than 255 files.)\n");
820 }
821 #endif
822 #ifdef DETECT_LOOP_GRAPH
823 FILE* loop_output;
824 int loop_n;
825 char loop_filename[255];
826 for(loop_n=0;loop_n<255;loop_n++)
827 {
828 sprintf(loop_filename, "./counters/LoopGraph.%d.dot",loop_n);
829 loop_output = fopen(loop_filename,"r");
830 if(loop_output)
831 {
832 fclose(loop_output);
833 }else{
834 break;
835 }
836 }
837 if(loop_n<255){
838 printf("Saving Loop Graph to File: %s ...\n", loop_filename);
839 loop_output = fopen(loop_filename,"w+");
840 if(loop_output!=NULL){
841 set_dependency_file(loop_output);
842 fprintf(loop_output,"digraph Loop {\n");
843 set_loop_file(loop_output);
844 forAllInDynArrayDo(_VMSMasterEnv->counter_history_array_info, &print_dot_node_info );
845 forAllInDynArrayDo( _VMSMasterEnv->loop_graph_array_info, &print_per_slot_to_file );
846 //int coreIdx;
847 //for(coreIdx=0;coreIdx<NUM_CORES;coreIdx++){
848 // fprintf(loop_output,"sync%d_%d [shape=rect,label=\"Sync\"];\n",coreIdx,_VMSMasterEnv->loop_counter[coreIdx]);
849 //}
850 fprintf(loop_output,"}\n");
851 } else
852 printf("Opening Loop Graph file failed. Please check that folder \"counters\" exists in run directory.\n");
853 } else {
854 printf("Could not open Loop Graph file, please clean \"counters\" folder. (Must contain less than 255 files.)\n");
855 }
856 #endif
857 #ifdef MEAS__TIME_PLUGIN
858 printHist( _VMSMasterEnv->reqHdlrLowTimeHist );
859 saveHistToFile( _VMSMasterEnv->reqHdlrLowTimeHist );
860 printHist( _VMSMasterEnv->reqHdlrHighTimeHist );
861 saveHistToFile( _VMSMasterEnv->reqHdlrHighTimeHist );
862 freeHistExt( _VMSMasterEnv->reqHdlrLowTimeHist );
863 freeHistExt( _VMSMasterEnv->reqHdlrHighTimeHist );
864 #endif
865 #ifdef MEAS__TIME_MALLOC
866 printHist( _VMSMasterEnv->mallocTimeHist );
867 saveHistToFile( _VMSMasterEnv->mallocTimeHist );
868 printHist( _VMSMasterEnv->freeTimeHist );
869 saveHistToFile( _VMSMasterEnv->freeTimeHist );
870 freeHistExt( _VMSMasterEnv->mallocTimeHist );
871 freeHistExt( _VMSMasterEnv->freeTimeHist );
872 #endif
873 #ifdef MEAS__TIME_MASTER_LOCK
874 printHist( _VMSMasterEnv->masterLockLowTimeHist );
875 printHist( _VMSMasterEnv->masterLockHighTimeHist );
876 #endif
877 #ifdef MEAS__TIME_MASTER
878 printHist( _VMSMasterEnv->pluginTimeHist );
879 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
880 {
881 freeVMSQ( readyToAnimateQs[ coreIdx ] );
882 //master VPs were created external to VMS, so use external free
883 VMS__dissipate_procr( masterVPs[ coreIdx ] );
885 freeSchedSlots( allSchedSlots[ coreIdx ] );
886 }
887 #endif
888 #ifdef MEAS__TIME_STAMP_SUSP
889 printHist( _VMSMasterEnv->pluginTimeHist );
890 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
891 {
892 freeVMSQ( readyToAnimateQs[ coreIdx ] );
893 //master VPs were created external to VMS, so use external free
894 VMS__dissipate_procr( masterVPs[ coreIdx ] );
896 freeSchedSlots( allSchedSlots[ coreIdx ] );
897 }
898 #endif
900 //All the environment data has been allocated with VMS__malloc, so just
901 // free its internal big-chunk and all inside it disappear.
902 /*
903 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
904 masterVPs = _VMSMasterEnv->masterVPs;
905 allSchedSlots = _VMSMasterEnv->allSchedSlots;
907 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
908 {
909 freeVMSQ( readyToAnimateQs[ coreIdx ] );
910 //master VPs were created external to VMS, so use external free
911 VMS__dissipate_procr( masterVPs[ coreIdx ] );
913 freeSchedSlots( allSchedSlots[ coreIdx ] );
914 }
916 VMS__free( _VMSMasterEnv->readyToAnimateQs );
917 VMS__free( _VMSMasterEnv->masterVPs );
918 VMS__free( _VMSMasterEnv->allSchedSlots );
920 //============================= MEASUREMENT STUFF ========================
921 #ifdef STATS__TURN_ON_PROBES
922 freeDynArrayDeep( _VMSMasterEnv->dynIntervalProbesInfo, &VMS__free_probe);
923 #endif
924 //========================================================================
925 */
926 //These are the only two that use system free
927 VMS_ext__free_free_list( _VMSMasterEnv->freeListHead );
928 free( (void *)_VMSMasterEnv );
929 }
932 //================================
935 /*Later, improve this -- for now, just exits the application after printing
936 * the error message.
937 */
938 void
939 VMS__throw_exception( char *msgStr, VirtProcr *reqstPr, VMSExcp *excpData )
940 {
941 printf("%s",msgStr);
942 fflush(stdin);
943 exit(1);
944 }