| rev |
line source |
|
Me@0
|
1 /*
|
|
Me@38
|
2 * Copyright 2010 OpenSourceStewardshipFoundation
|
|
Me@0
|
3 *
|
|
Me@0
|
4 * Licensed under BSD
|
|
Me@0
|
5 */
|
|
Me@0
|
6
|
|
Me@0
|
7 #include <stdio.h>
|
|
Me@0
|
8 #include <stdlib.h>
|
|
Me@0
|
9 #include <malloc.h>
|
|
Me@0
|
10
|
|
Me@0
|
11 #include "VMS.h"
|
|
Me@0
|
12 #include "Queue_impl/BlockingQueue.h"
|
|
Me@38
|
13 #include "Histogram/Histogram.h"
|
|
Me@0
|
14
|
|
Me@0
|
15
|
|
Me@26
|
16 #define thdAttrs NULL
|
|
Me@26
|
17
|
|
Me@22
|
18 //===========================================================================
|
|
Me@22
|
19 void
|
|
Me@22
|
20 shutdownFn( void *dummy, VirtProcr *dummy2 );
|
|
Me@22
|
21
|
|
Me@31
|
22 SchedSlot **
|
|
Me@31
|
23 create_sched_slots();
|
|
Me@22
|
24
|
|
Me@28
|
25 void
|
|
Me@28
|
26 create_masterEnv();
|
|
Me@28
|
27
|
|
Me@28
|
28 void
|
|
Me@28
|
29 create_the_coreLoop_OS_threads();
|
|
Me@28
|
30
|
|
Me@26
|
31 pthread_mutex_t suspendLock = PTHREAD_MUTEX_INITIALIZER;
|
|
Me@26
|
32 pthread_cond_t suspend_cond = PTHREAD_COND_INITIALIZER;
|
|
Me@26
|
33
|
|
Me@22
|
34 //===========================================================================
|
|
Me@22
|
35
|
|
Me@0
|
36 /*Setup has two phases:
|
|
Me@0
|
37 * 1) Semantic layer first calls init_VMS, which creates masterEnv, and puts
|
|
Me@8
|
38 * the master virt procr into the work-queue, ready for first "call"
|
|
Me@8
|
39 * 2) Semantic layer then does its own init, which creates the seed virt
|
|
Me@8
|
40 * procr inside the semantic layer, ready to schedule it when
|
|
Me@0
|
41 * asked by the first run of the masterLoop.
|
|
Me@0
|
42 *
|
|
Me@0
|
43 *This part is bit weird because VMS really wants to be "always there", and
|
|
Me@0
|
44 * have applications attach and detach.. for now, this VMS is part of
|
|
Me@0
|
45 * the app, so the VMS system starts up as part of running the app.
|
|
Me@0
|
46 *
|
|
Me@8
|
47 *The semantic layer is isolated from the VMS internals by making the
|
|
Me@8
|
48 * semantic layer do setup to a state that it's ready with its
|
|
Me@8
|
49 * initial virt procrs, ready to schedule them to slots when the masterLoop
|
|
Me@0
|
50 * asks. Without this pattern, the semantic layer's setup would
|
|
Me@8
|
51 * have to modify slots directly to assign the initial virt-procrs, and put
|
|
Me@31
|
52 * them into the readyToAnimateQ itself, breaking the isolation completely.
|
|
Me@0
|
53 *
|
|
Me@0
|
54 *
|
|
Me@8
|
55 *The semantic layer creates the initial virt procr(s), and adds its
|
|
Me@8
|
56 * own environment to masterEnv, and fills in the pointers to
|
|
Me@0
|
57 * the requestHandler and slaveScheduler plug-in functions
|
|
Me@8
|
58 */
|
|
Me@8
|
59
|
|
Me@8
|
60 /*This allocates VMS data structures, populates the master VMSProc,
|
|
Me@0
|
61 * and master environment, and returns the master environment to the semantic
|
|
Me@0
|
62 * layer.
|
|
Me@0
|
63 */
|
|
Me@8
|
64 void
|
|
Me@8
|
65 VMS__init()
|
|
Me@28
|
66 {
|
|
Me@28
|
67 create_masterEnv();
|
|
Me@28
|
68 create_the_coreLoop_OS_threads();
|
|
Me@28
|
69 }
|
|
Me@28
|
70
|
|
Me@28
|
71 /*To initialize the sequential version, just don't create the threads
|
|
Me@28
|
72 */
|
|
Me@28
|
73 void
|
|
Me@28
|
74 VMS__init_Seq()
|
|
Me@28
|
75 {
|
|
Me@28
|
76 create_masterEnv();
|
|
Me@28
|
77 }
|
|
Me@28
|
78
|
|
Me@28
|
79 void
|
|
Me@28
|
80 create_masterEnv()
|
|
Me@31
|
81 { MasterEnv *masterEnv;
|
|
Me@47
|
82 VMSQueueStruc **readyToAnimateQs;
|
|
Me@31
|
83 int coreIdx;
|
|
Me@31
|
84 VirtProcr **masterVPs;
|
|
Me@31
|
85 SchedSlot ***allSchedSlots; //ptr to array of ptrs
|
|
Me@31
|
86
|
|
Me@31
|
87 //Make the master env, which holds everything else
|
|
Me@1
|
88 _VMSMasterEnv = malloc( sizeof(MasterEnv) );
|
|
Me@1
|
89 masterEnv = _VMSMasterEnv;
|
|
Me@31
|
90 //Need to set start pt here 'cause used by seed procr, which is created
|
|
Me@31
|
91 // before the first core loop starts up. -- not sure how yet..
|
|
Me@31
|
92 // masterEnv->coreLoopStartPt = ;
|
|
Me@31
|
93 // masterEnv->coreLoopEndPt = ;
|
|
Me@31
|
94
|
|
Me@31
|
95 //Make a readyToAnimateQ for each core loop
|
|
Me@47
|
96 readyToAnimateQs = malloc( NUM_CORES * sizeof(VMSQueueStruc *) );
|
|
Me@31
|
97 masterVPs = malloc( NUM_CORES * sizeof(VirtProcr *) );
|
|
Me@0
|
98
|
|
Me@31
|
99 //One array for each core, 3 in array, core's masterVP scheds all
|
|
Me@31
|
100 allSchedSlots = malloc( NUM_CORES * sizeof(SchedSlot *) );
|
|
Me@0
|
101
|
|
Me@31
|
102 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@31
|
103 {
|
|
Me@31
|
104 readyToAnimateQs[ coreIdx ] = makeSRSWQ();
|
|
Me@31
|
105
|
|
Me@31
|
106 //Q: should give masterVP core-specific into as its init data?
|
|
Me@31
|
107 masterVPs[ coreIdx ] = VMS__create_procr( &masterLoop, masterEnv );
|
|
Me@31
|
108 masterVPs[ coreIdx ]->coreAnimatedBy = coreIdx;
|
|
Me@31
|
109 allSchedSlots[ coreIdx ] = create_sched_slots(); //makes for one core
|
|
Me@31
|
110 }
|
|
Me@31
|
111 _VMSMasterEnv->readyToAnimateQs = readyToAnimateQs;
|
|
Me@31
|
112 _VMSMasterEnv->masterVPs = masterVPs;
|
|
Me@31
|
113 _VMSMasterEnv->allSchedSlots = allSchedSlots;
|
|
Me@0
|
114
|
|
Me@48
|
115 //============================= MEASUREMENT STUFF ========================
|
|
Me@48
|
116 #ifdef MEAS__TIME_MASTER
|
|
Me@28
|
117
|
|
Me@48
|
118 _VMSMasterEnv->stats->masterTimeHist = makeHistogram( 25, 500, 800 );
|
|
Me@48
|
119 _VMSMasterEnv->stats->masterLockHist = makeHistogram( 25, 0, 100000 );
|
|
Me@48
|
120 _VMSMasterEnv->stats->createHist = makeHistogram( 25, 0, 5000 );
|
|
Me@48
|
121 #endif
|
|
Me@48
|
122 //========================================================================
|
|
Me@12
|
123
|
|
Me@31
|
124 //Aug 19, 2010: no longer need to place initial masterVP into queue
|
|
Me@31
|
125 // because coreLoop now controls -- animates its masterVP when no work
|
|
Me@31
|
126
|
|
Me@30
|
127
|
|
Me@30
|
128 //==================== malloc substitute ========================
|
|
Me@30
|
129 //
|
|
Me@30
|
130 //Testing whether malloc is using thread-local storage and therefore
|
|
Me@30
|
131 // causing unreliable behavior.
|
|
Me@30
|
132 //Just allocate a massive chunk of memory and roll own malloc/free and
|
|
Me@30
|
133 // make app use VMS__malloc_to, which will suspend and perform malloc
|
|
Me@30
|
134 // in the master, taking from this massive chunk.
|
|
Me@30
|
135
|
|
Me@30
|
136 // initFreeList();
|
|
Me@38
|
137
|
|
Me@0
|
138 }
|
|
Me@0
|
139
|
|
Me@30
|
140 /*
|
|
Me@30
|
141 void
|
|
Me@30
|
142 initMasterMalloc()
|
|
Me@30
|
143 {
|
|
Me@30
|
144 _VMSMasterEnv->mallocChunk = malloc( MASSIVE_MALLOC_SIZE );
|
|
Me@30
|
145
|
|
Me@30
|
146 //The free-list element is the first several locations of an
|
|
Me@30
|
147 // allocated chunk -- the address given to the application is pre-
|
|
Me@30
|
148 // pended with both the ownership structure and the free-list struc.
|
|
Me@30
|
149 //So, write the values of these into the first locations of
|
|
Me@30
|
150 // mallocChunk -- which marks it as free & puts in its size.
|
|
Me@30
|
151 listElem = (FreeListElem *)_VMSMasterEnv->mallocChunk;
|
|
Me@30
|
152 listElem->size = MASSIVE_MALLOC_SIZE - NUM_PREPEND_BYTES
|
|
Me@30
|
153 listElem->next = NULL;
|
|
Me@30
|
154 }
|
|
Me@30
|
155
|
|
Me@30
|
156 void
|
|
Me@30
|
157 dissipateMasterMalloc()
|
|
Me@30
|
158 {
|
|
Me@30
|
159 //Just foo code -- to get going -- doing as if free list were link-list
|
|
Me@30
|
160 currElem = _VMSMasterEnv->freeList;
|
|
Me@30
|
161 while( currElem != NULL )
|
|
Me@30
|
162 {
|
|
Me@30
|
163 nextElem = currElem->next;
|
|
Me@30
|
164 masterFree( currElem );
|
|
Me@30
|
165 currElem = nextElem;
|
|
Me@30
|
166 }
|
|
Me@30
|
167 free( _VMSMasterEnv->freeList );
|
|
Me@30
|
168 }
|
|
Me@30
|
169 */
|
|
Me@30
|
170
|
|
Me@31
|
171 SchedSlot **
|
|
Me@31
|
172 create_sched_slots()
|
|
Me@31
|
173 { SchedSlot **schedSlots;
|
|
Me@0
|
174 int i;
|
|
Me@0
|
175
|
|
Me@8
|
176 schedSlots = malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
|
|
Me@8
|
177
|
|
Me@1
|
178 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
|
|
Me@0
|
179 {
|
|
Me@8
|
180 schedSlots[i] = malloc( sizeof(SchedSlot) );
|
|
Me@8
|
181
|
|
Me@1
|
182 //Set state to mean "handling requests done, slot needs filling"
|
|
Me@8
|
183 schedSlots[i]->workIsDone = FALSE;
|
|
Me@8
|
184 schedSlots[i]->needsProcrAssigned = TRUE;
|
|
Me@0
|
185 }
|
|
Me@31
|
186 return schedSlots;
|
|
Me@31
|
187 }
|
|
Me@31
|
188
|
|
Me@31
|
189
|
|
Me@31
|
190 void
|
|
Me@31
|
191 freeSchedSlots( SchedSlot **schedSlots )
|
|
Me@31
|
192 { int i;
|
|
Me@31
|
193 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
|
|
Me@31
|
194 {
|
|
Me@31
|
195 free( schedSlots[i] );
|
|
Me@31
|
196 }
|
|
Me@31
|
197 free( schedSlots );
|
|
Me@0
|
198 }
|
|
Me@0
|
199
|
|
Me@8
|
200
|
|
Me@28
|
201 void
|
|
Me@28
|
202 create_the_coreLoop_OS_threads()
|
|
Me@28
|
203 {
|
|
Me@28
|
204 //========================================================================
|
|
Me@28
|
205 // Create the Threads
|
|
Me@47
|
206 int coreIdx, retCode, i;
|
|
Me@47
|
207
|
|
Me@47
|
208 //create the arrays used to measure TSC offsets between cores
|
|
Me@47
|
209 pongNums = malloc( NUM_CORES * sizeof( int ) );
|
|
Me@47
|
210 pingTimes = malloc( NUM_CORES * NUM_TSC_ROUND_TRIPS * sizeof( TSCount ) );
|
|
Me@47
|
211 pongTimes = malloc( NUM_CORES * NUM_TSC_ROUND_TRIPS * sizeof( TSCount ) );
|
|
Me@47
|
212
|
|
Me@47
|
213 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@47
|
214 {
|
|
Me@47
|
215 pongNums[ coreIdx ] = 0;
|
|
Me@47
|
216 for( i = 0; i < NUM_TSC_ROUND_TRIPS; i++ )
|
|
Me@47
|
217 {
|
|
Me@47
|
218 pingTimes[ coreIdx * NUM_TSC_ROUND_TRIPS + i ] = (TSCount) 0;
|
|
Me@47
|
219 pingTimes[ coreIdx * NUM_TSC_ROUND_TRIPS + i ] = (TSCount) 0;
|
|
Me@47
|
220 }
|
|
Me@47
|
221 }
|
|
Me@28
|
222
|
|
Me@28
|
223 //Need the threads to be created suspended, and wait for a signal
|
|
Me@28
|
224 // before proceeding -- gives time after creating to initialize other
|
|
Me@28
|
225 // stuff before the coreLoops set off.
|
|
Me@28
|
226 _VMSMasterEnv->setupComplete = 0;
|
|
Me@28
|
227
|
|
Me@28
|
228 //Make the threads that animate the core loops
|
|
Me@28
|
229 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@28
|
230 { coreLoopThdParams[coreIdx] = malloc( sizeof(ThdParams) );
|
|
Me@28
|
231 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
|
|
Me@28
|
232
|
|
Me@28
|
233 retCode =
|
|
Me@28
|
234 pthread_create( &(coreLoopThdHandles[coreIdx]),
|
|
Me@28
|
235 thdAttrs,
|
|
Me@28
|
236 &coreLoop,
|
|
Me@28
|
237 (void *)(coreLoopThdParams[coreIdx]) );
|
|
Me@28
|
238 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(0);}
|
|
Me@28
|
239 }
|
|
Me@28
|
240 }
|
|
Me@28
|
241
|
|
Me@0
|
242 /*Semantic layer calls this when it want the system to start running..
|
|
Me@0
|
243 *
|
|
Me@24
|
244 *This starts the core loops running then waits for them to exit.
|
|
Me@0
|
245 */
|
|
Me@12
|
246 void
|
|
Me@24
|
247 VMS__start_the_work_then_wait_until_done()
|
|
Me@12
|
248 { int coreIdx;
|
|
Me@24
|
249 //Start the core loops running
|
|
Me@24
|
250 //===========================================================================
|
|
Me@25
|
251 TSCount startCount, endCount;
|
|
Me@24
|
252 unsigned long long count = 0, freq = 0;
|
|
Me@25
|
253 double runTime;
|
|
Me@0
|
254
|
|
Me@47
|
255 startCount = getTSC();
|
|
Me@25
|
256
|
|
Me@25
|
257 //tell the core loop threads that setup is complete
|
|
Me@25
|
258 //get lock, to lock out any threads still starting up -- they'll see
|
|
Me@25
|
259 // that setupComplete is true before entering while loop, and so never
|
|
Me@25
|
260 // wait on the condition
|
|
Me@26
|
261 pthread_mutex_lock( &suspendLock );
|
|
Me@25
|
262 _VMSMasterEnv->setupComplete = 1;
|
|
Me@26
|
263 pthread_mutex_unlock( &suspendLock );
|
|
Me@26
|
264 pthread_cond_broadcast( &suspend_cond );
|
|
Me@25
|
265
|
|
Me@25
|
266
|
|
Me@24
|
267 //wait for all to complete
|
|
Me@8
|
268 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@8
|
269 {
|
|
Me@25
|
270 pthread_join( coreLoopThdHandles[coreIdx], NULL );
|
|
Me@24
|
271 }
|
|
Me@25
|
272
|
|
Me@24
|
273 //NOTE: do not clean up VMS env here -- semantic layer has to have
|
|
Me@24
|
274 // a chance to clean up its environment first, then do a call to free
|
|
Me@24
|
275 // the Master env and rest of VMS locations
|
|
Me@24
|
276
|
|
Me@24
|
277
|
|
Me@47
|
278 endCount = getTSC();
|
|
Me@25
|
279 count = endCount - startCount;
|
|
Me@24
|
280
|
|
Me@25
|
281 runTime = (double)count / (double)TSCOUNT_FREQ;
|
|
Me@25
|
282
|
|
Me@25
|
283 printf("\n Time startup to shutdown: %f\n", runTime); fflush( stdin );
|
|
Me@8
|
284 }
|
|
Me@0
|
285
|
|
Me@28
|
286 /*Only difference between version with an OS thread pinned to each core and
|
|
Me@28
|
287 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
|
|
Me@28
|
288 */
|
|
Me@28
|
289 void
|
|
Me@28
|
290 VMS__start_the_work_then_wait_until_done_Seq()
|
|
Me@28
|
291 {
|
|
Me@28
|
292 //Instead of un-suspending threads, just call the one and only
|
|
Me@28
|
293 // core loop (sequential version), in the main thread.
|
|
Me@28
|
294 coreLoop_Seq( NULL );
|
|
Me@28
|
295
|
|
Me@28
|
296 }
|
|
Me@28
|
297
|
|
Me@0
|
298
|
|
Me@0
|
299
|
|
Me@8
|
300 /*Create stack, then create __cdecl structure on it and put initialData and
|
|
Me@8
|
301 * pointer to the new structure instance into the parameter positions on
|
|
Me@8
|
302 * the stack
|
|
Me@8
|
303 *Then put function pointer into nextInstrPt -- the stack is setup in std
|
|
Me@8
|
304 * call structure, so jumping to function ptr is same as a GCC generated
|
|
Me@8
|
305 * function call
|
|
Me@8
|
306 *No need to save registers on old stack frame, because there's no old
|
|
Me@8
|
307 * animator state to return to --
|
|
Me@8
|
308 *
|
|
Me@8
|
309 */
|
|
Me@8
|
310 VirtProcr *
|
|
Me@8
|
311 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
|
|
Me@8
|
312 { VirtProcr *newPr;
|
|
Me@8
|
313 char *stackLocs, *stackPtr;
|
|
Me@8
|
314
|
|
Me@48
|
315 //============================= MEASUREMENT STUFF ========================
|
|
Me@48
|
316 #ifdef MEAS__TIME_MASTER
|
|
Me@48
|
317 int32 startStamp;
|
|
Me@48
|
318 saveLowTimeStampCountInto( startStamp );
|
|
Me@48
|
319 #endif
|
|
Me@48
|
320 //========================================================================
|
|
Me@48
|
321
|
|
Me@8
|
322 newPr = malloc( sizeof(VirtProcr) );
|
|
Me@12
|
323 newPr->procrID = numProcrsCreated++;
|
|
Me@8
|
324 newPr->nextInstrPt = fnPtr;
|
|
Me@8
|
325 newPr->initialData = initialData;
|
|
Me@31
|
326 newPr->requests = NULL;
|
|
Me@38
|
327 newPr->schedSlot = NULL;
|
|
Me@31
|
328 // newPr->coreLoopStartPt = _VMSMasterEnv->coreLoopStartPt;
|
|
Me@8
|
329
|
|
Me@14
|
330 //fnPtr takes two params -- void *initData & void *animProcr
|
|
Me@8
|
331 //alloc stack locations, make stackPtr be the highest addr minus room
|
|
Me@14
|
332 // for 2 params + return addr. Return addr (NULL) is in loc pointed to
|
|
Me@14
|
333 // by stackPtr, initData at stackPtr + 4 bytes, animatingPr just above
|
|
Me@22
|
334 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
|
|
Me@47
|
335 if(stackLocs == 0) {perror("error: malloc stack"); exit(1);}
|
|
Me@22
|
336 newPr->startOfStack = stackLocs;
|
|
Me@22
|
337 stackPtr = ( (char *)stackLocs + VIRT_PROCR_STACK_SIZE - 0x10 );
|
|
Me@8
|
338 //setup __cdecl on stack -- coreloop will switch to stackPtr before jmp
|
|
Me@22
|
339 *( (int *)stackPtr + 2 ) = (int) newPr; //rightmost param -- 32bit pointer
|
|
Me@14
|
340 *( (int *)stackPtr + 1 ) = (int) initialData; //next param to left
|
|
Me@8
|
341 newPr->stackPtr = stackPtr; //core loop will switch to this, then
|
|
Me@8
|
342 newPr->framePtr = stackPtr; //suspend loop will save new stack & frame ptr
|
|
Me@8
|
343
|
|
Me@48
|
344 //============================= MEASUREMENT STUFF ========================
|
|
Me@48
|
345 #ifdef MEAS__TIME_MASTER
|
|
Me@48
|
346 int32 endStamp;
|
|
Me@48
|
347 saveLowTimeStampCountInto( endStamp );
|
|
Me@48
|
348 addIntervalToHist( startStamp, endStamp,
|
|
Me@48
|
349 _VMSMasterEnv->stats->createHist );
|
|
Me@48
|
350 #endif
|
|
Me@48
|
351 //========================================================================
|
|
Me@48
|
352
|
|
Me@8
|
353 return newPr;
|
|
Me@8
|
354 }
|
|
Me@8
|
355
|
|
Me@8
|
356
|
|
Me@26
|
357 /*there is a label inside this function -- save the addr of this label in
|
|
Me@0
|
358 * the callingPr struc, as the pick-up point from which to start the next
|
|
Me@0
|
359 * work-unit for that procr. If turns out have to save registers, then
|
|
Me@0
|
360 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
|
|
Me@0
|
361 * "done with work-unit" label. The procr struc is in the request in the
|
|
Me@0
|
362 * slave that animated the just-ended work-unit, so all the state is saved
|
|
Me@0
|
363 * there, and will get passed along, inside the request handler, to the
|
|
Me@0
|
364 * next work-unit for that procr.
|
|
Me@0
|
365 */
|
|
Me@8
|
366 void
|
|
Me@38
|
367 VMS__suspend_procr( VirtProcr *animatingPr )
|
|
Me@14
|
368 { void *jmpPt, *stackPtrAddr, *framePtrAddr, *coreLoopStackPtr;
|
|
Me@14
|
369 void *coreLoopFramePtr;
|
|
Me@0
|
370
|
|
Me@14
|
371 //The request to master will cause this suspended virt procr to get
|
|
Me@14
|
372 // scheduled again at some future point -- to resume, core loop jumps
|
|
Me@14
|
373 // to the resume point (below), which causes restore of saved regs and
|
|
Me@14
|
374 // "return" from this call.
|
|
Me@38
|
375 animatingPr->nextInstrPt = &&ResumePt;
|
|
Me@1
|
376
|
|
Me@1
|
377 //return ownership of the virt procr and sched slot to Master virt pr
|
|
Me@38
|
378 animatingPr->schedSlot->workIsDone = TRUE;
|
|
Me@14
|
379 // coreIdx = callingPr->coreAnimatedBy;
|
|
Me@1
|
380
|
|
Me@38
|
381 stackPtrAddr = &(animatingPr->stackPtr);
|
|
Me@38
|
382 framePtrAddr = &(animatingPr->framePtr);
|
|
Me@26
|
383
|
|
Me@31
|
384 jmpPt = _VMSMasterEnv->coreLoopStartPt;
|
|
Me@38
|
385 coreLoopFramePtr = animatingPr->coreLoopFramePtr;//need this only
|
|
Me@38
|
386 coreLoopStackPtr = animatingPr->coreLoopStackPtr;//safety
|
|
Me@1
|
387
|
|
Me@26
|
388 //Save the virt procr's stack and frame ptrs,
|
|
Me@18
|
389 asm volatile("movl %0, %%eax; \
|
|
Me@18
|
390 movl %%esp, (%%eax); \
|
|
Me@18
|
391 movl %1, %%eax; \
|
|
Me@26
|
392 movl %%ebp, (%%eax) "\
|
|
Me@26
|
393 /* outputs */ : "=g" (stackPtrAddr), "=g" (framePtrAddr) \
|
|
Me@26
|
394 /* inputs */ : \
|
|
Me@26
|
395 /* clobber */ : "%eax" \
|
|
Me@26
|
396 );
|
|
Me@26
|
397
|
|
Me@41
|
398 //=========================== Measurement stuff ========================
|
|
Me@38
|
399 #ifdef MEAS__TIME_STAMP_SUSP
|
|
Me@41
|
400 //record time stamp: compare to time-stamp recorded below
|
|
Me@38
|
401 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
|
|
Me@38
|
402 #endif
|
|
Me@41
|
403 //=======================================================================
|
|
Me@38
|
404
|
|
Me@26
|
405 //restore coreloop's frame ptr, then jump back to "start" of core loop
|
|
Me@26
|
406 //Note, GCC compiles to assembly that saves esp and ebp in the stack
|
|
Me@26
|
407 // frame -- so have to explicitly do assembly that saves to memory
|
|
Me@26
|
408 asm volatile("movl %0, %%eax; \
|
|
Me@26
|
409 movl %1, %%esp; \
|
|
Me@26
|
410 movl %2, %%ebp; \
|
|
Me@18
|
411 jmp %%eax " \
|
|
Me@26
|
412 /* outputs */ : \
|
|
Me@26
|
413 /* inputs */ : "m" (jmpPt), "m"(coreLoopStackPtr), "m"(coreLoopFramePtr)\
|
|
Me@18
|
414 /* clobber */ : "memory", "%eax", "%ebx", "%ecx", "%edx", "%edi","%esi" \
|
|
Me@12
|
415 ); //list everything as clobbered to force GCC to save all
|
|
Me@12
|
416 // live vars that are in regs on stack before this
|
|
Me@12
|
417 // assembly, so that stack pointer is correct, before jmp
|
|
Me@1
|
418
|
|
Me@1
|
419 ResumePt:
|
|
Me@38
|
420 #ifdef MEAS__TIME_STAMP_SUSP
|
|
Me@41
|
421 //NOTE: only take low part of count -- do sanity check when take diff
|
|
Me@38
|
422 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
|
|
Me@38
|
423 #endif
|
|
Me@38
|
424
|
|
Me@0
|
425 return;
|
|
Me@0
|
426 }
|
|
Me@0
|
427
|
|
Me@22
|
428
|
|
Me@22
|
429
|
|
Me@22
|
430
|
|
Me@38
|
431 /*
|
|
Me@22
|
432 *This adds a request to dissipate, then suspends the processor so that the
|
|
Me@22
|
433 * request handler will receive the request. The request handler is what
|
|
Me@22
|
434 * does the work of freeing memory and removing the processor from the
|
|
Me@22
|
435 * semantic environment's data structures.
|
|
Me@22
|
436 *The request handler also is what figures out when to shutdown the VMS
|
|
Me@22
|
437 * system -- which causes all the core loop threads to die, and returns from
|
|
Me@22
|
438 * the call that started up VMS to perform the work.
|
|
Me@22
|
439 *
|
|
Me@22
|
440 *This form is a bit misleading to understand if one is trying to figure out
|
|
Me@22
|
441 * how VMS works -- it looks like a normal function call, but inside it
|
|
Me@22
|
442 * sends a request to the request handler and suspends the processor, which
|
|
Me@22
|
443 * jumps out of the VMS__dissipate_procr function, and out of all nestings
|
|
Me@22
|
444 * above it, transferring the work of dissipating to the request handler,
|
|
Me@22
|
445 * which then does the actual work -- causing the processor that animated
|
|
Me@22
|
446 * the call of this function to disappear and the "hanging" state of this
|
|
Me@22
|
447 * function to just poof into thin air -- the virtual processor's trace
|
|
Me@22
|
448 * never returns from this call, but instead the virtual processor's trace
|
|
Me@22
|
449 * gets suspended in this call and all the virt processor's state disap-
|
|
Me@22
|
450 * pears -- making that suspend the last thing in the virt procr's trace.
|
|
Me@8
|
451 */
|
|
Me@8
|
452 void
|
|
Me@22
|
453 VMS__dissipate_procr( VirtProcr *procrToDissipate )
|
|
Me@22
|
454 { VMSReqst *req;
|
|
Me@22
|
455
|
|
Me@22
|
456 req = malloc( sizeof(VMSReqst) );
|
|
Me@22
|
457 // req->virtProcrFrom = callingPr;
|
|
Me@22
|
458 req->reqType = dissipate;
|
|
Me@22
|
459 req->nextReqst = procrToDissipate->requests;
|
|
Me@22
|
460 procrToDissipate->requests = req;
|
|
Me@22
|
461
|
|
Me@22
|
462 VMS__suspend_procr( procrToDissipate );
|
|
Me@22
|
463 }
|
|
Me@22
|
464
|
|
Me@22
|
465
|
|
Me@22
|
466 /*This inserts the semantic-layer's request data into standard VMS carrier
|
|
Me@22
|
467 */
|
|
Me@22
|
468 inline void
|
|
Me@24
|
469 VMS__add_sem_request( void *semReqData, VirtProcr *callingPr )
|
|
Me@22
|
470 { VMSReqst *req;
|
|
Me@22
|
471
|
|
Me@22
|
472 req = malloc( sizeof(VMSReqst) );
|
|
Me@22
|
473 // req->virtProcrFrom = callingPr;
|
|
Me@22
|
474 req->reqType = semantic;
|
|
Me@22
|
475 req->semReqData = semReqData;
|
|
Me@22
|
476 req->nextReqst = callingPr->requests;
|
|
Me@22
|
477 callingPr->requests = req;
|
|
Me@22
|
478 }
|
|
Me@22
|
479
|
|
Me@22
|
480
|
|
Me@38
|
481 /*Use this to get first request before starting request handler's loop
|
|
Me@38
|
482 */
|
|
Me@24
|
483 VMSReqst *
|
|
Me@24
|
484 VMS__take_top_request_from( VirtProcr *procrWithReq )
|
|
Me@24
|
485 { VMSReqst *req;
|
|
Me@24
|
486
|
|
Me@24
|
487 req = procrWithReq->requests;
|
|
Me@24
|
488 if( req == NULL ) return req;
|
|
Me@31
|
489
|
|
Me@24
|
490 procrWithReq->requests = procrWithReq->requests->nextReqst;
|
|
Me@24
|
491 return req;
|
|
Me@24
|
492 }
|
|
Me@24
|
493
|
|
Me@38
|
494 /*A subtle bug due to freeing then accessing "next" after freed caused this
|
|
Me@38
|
495 * form of call to be put in -- so call this at end of request handler loop
|
|
Me@38
|
496 * that iterates through the requests.
|
|
Me@38
|
497 */
|
|
Me@31
|
498 VMSReqst *
|
|
Me@31
|
499 VMS__free_top_and_give_next_request_from( VirtProcr *procrWithReq )
|
|
Me@31
|
500 { VMSReqst *req;
|
|
Me@31
|
501
|
|
Me@31
|
502 req = procrWithReq->requests;
|
|
Me@38
|
503 if( req == NULL ) return NULL;
|
|
Me@31
|
504
|
|
Me@31
|
505 procrWithReq->requests = procrWithReq->requests->nextReqst;
|
|
Me@31
|
506 VMS__free_request( req );
|
|
Me@31
|
507 return procrWithReq->requests;
|
|
Me@31
|
508 }
|
|
Me@31
|
509
|
|
Me@38
|
510
|
|
Me@38
|
511 //TODO: add a semantic-layer supplied "freer" for the semantic-data portion
|
|
Me@38
|
512 // of a request -- IE call with both a virt procr and a fn-ptr to request
|
|
Me@38
|
513 // freer (also maybe put sem request freer as a field in virt procr?)
|
|
Me@38
|
514 //MeasVMS relies right now on this only freeing VMS layer of request -- the
|
|
Me@38
|
515 // semantic portion of request is alloc'd and freed by request handler
|
|
Me@38
|
516 void
|
|
Me@38
|
517 VMS__free_request( VMSReqst *req )
|
|
Me@38
|
518 {
|
|
Me@38
|
519 free( req );
|
|
Me@38
|
520 }
|
|
Me@38
|
521
|
|
Me@38
|
522
|
|
Me@38
|
523
|
|
Me@24
|
524 inline int
|
|
Me@24
|
525 VMS__isSemanticReqst( VMSReqst *req )
|
|
Me@22
|
526 {
|
|
Me@24
|
527 return ( req->reqType == semantic );
|
|
Me@24
|
528 }
|
|
Me@22
|
529
|
|
Me@24
|
530
|
|
Me@24
|
531 inline void *
|
|
Me@24
|
532 VMS__take_sem_reqst_from( VMSReqst *req )
|
|
Me@24
|
533 {
|
|
Me@24
|
534 return req->semReqData;
|
|
Me@24
|
535 }
|
|
Me@24
|
536
|
|
Me@24
|
537 inline int
|
|
Me@24
|
538 VMS__isDissipateReqst( VMSReqst *req )
|
|
Me@24
|
539 {
|
|
Me@24
|
540 return ( req->reqType == dissipate );
|
|
Me@24
|
541 }
|
|
Me@24
|
542
|
|
Me@24
|
543 inline int
|
|
Me@24
|
544 VMS__isCreateReqst( VMSReqst *req )
|
|
Me@24
|
545 {
|
|
Me@24
|
546 return ( req->reqType == regCreated );
|
|
Me@24
|
547 }
|
|
Me@24
|
548
|
|
Me@24
|
549 void
|
|
Me@38
|
550 VMS__send_req_to_register_new_procr(VirtProcr *newPr, VirtProcr *reqstingPr)
|
|
Me@24
|
551 { VMSReqst *req;
|
|
Me@24
|
552
|
|
Me@24
|
553 req = malloc( sizeof(VMSReqst) );
|
|
Me@24
|
554 req->reqType = regCreated;
|
|
Me@24
|
555 req->semReqData = newPr;
|
|
Me@24
|
556 req->nextReqst = reqstingPr->requests;
|
|
Me@24
|
557 reqstingPr->requests = req;
|
|
Me@24
|
558
|
|
Me@24
|
559 VMS__suspend_procr( reqstingPr );
|
|
Me@22
|
560 }
|
|
Me@22
|
561
|
|
Me@22
|
562
|
|
Me@22
|
563
|
|
Me@24
|
564 /*This must be called by the request handler plugin -- it cannot be called
|
|
Me@24
|
565 * from the semantic library "dissipate processor" function -- instead, the
|
|
Me@24
|
566 * semantic layer has to generate a request for the plug-in to call this
|
|
Me@24
|
567 * function.
|
|
Me@24
|
568 *The reason is that this frees the virtual processor's stack -- which is
|
|
Me@24
|
569 * still in use inside semantic library calls!
|
|
Me@24
|
570 *
|
|
Me@24
|
571 *This frees or recycles all the state owned by and comprising the VMS
|
|
Me@24
|
572 * portion of the animating virtual procr. The request handler must first
|
|
Me@24
|
573 * free any semantic data created for the processor that didn't use the
|
|
Me@24
|
574 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
|
|
Me@24
|
575 * system to disown any state that did use VMS_malloc, and then frees the
|
|
Me@24
|
576 * statck and the processor-struct itself.
|
|
Me@24
|
577 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
|
|
Me@24
|
578 * state, then that state gets freed (or sent to recycling) as a side-effect
|
|
Me@24
|
579 * of dis-owning it.
|
|
Me@24
|
580 */
|
|
Me@24
|
581 void
|
|
Me@29
|
582 VMS__handle_dissipate_reqst( VirtProcr *animatingPr )
|
|
Me@24
|
583 {
|
|
Me@24
|
584 //dis-own all locations owned by this processor, causing to be freed
|
|
Me@24
|
585 // any locations that it is (was) sole owner of
|
|
Me@29
|
586 //TODO: implement VMS__malloc system, including "give up ownership"
|
|
Me@24
|
587
|
|
Me@24
|
588 //The dissipate request might still be attached, so remove and free it
|
|
Me@38
|
589 VMS__free_top_and_give_next_request_from( animatingPr );
|
|
Me@24
|
590
|
|
Me@24
|
591 //NOTE: initialData was given to the processor, so should either have
|
|
Me@24
|
592 // been alloc'd with VMS__malloc, or freed by the level above animPr.
|
|
Me@24
|
593 //So, all that's left to free here is the stack and the VirtProcr struc
|
|
Me@24
|
594 // itself
|
|
Me@24
|
595 free( animatingPr->startOfStack );
|
|
Me@24
|
596 free( animatingPr );
|
|
Me@24
|
597 }
|
|
Me@24
|
598
|
|
Me@24
|
599
|
|
Me@29
|
600 //TODO: re-architect so that have clean separation between request handler
|
|
Me@29
|
601 // and master loop, for dissipate, create, shutdown, and other non-semantic
|
|
Me@29
|
602 // requests. Issue is chain: one removes requests from AppVP, one dispatches
|
|
Me@29
|
603 // on type of request, and one handles each type.. but some types require
|
|
Me@29
|
604 // action from both request handler and master loop -- maybe just give the
|
|
Me@29
|
605 // request handler calls like: VMS__handle_X_request_type
|
|
Me@24
|
606
|
|
Me@29
|
607 void
|
|
Me@29
|
608 endOSThreadFn( void *initData, VirtProcr *animatingPr );
|
|
Me@29
|
609
|
|
Me@29
|
610 /*This is called by the semantic layer's request handler when it decides its
|
|
Me@29
|
611 * time to shut down the VMS system. Calling this causes the core loop OS
|
|
Me@29
|
612 * threads to exit, which unblocks the entry-point function that started up
|
|
Me@29
|
613 * VMS, and allows it to grab the result and return to the original single-
|
|
Me@29
|
614 * threaded application.
|
|
Me@22
|
615 *
|
|
Me@29
|
616 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
|
|
Me@29
|
617 * and-wait function has to free a bunch of stuff after it detects the
|
|
Me@29
|
618 * threads have all died: the masterEnv, the thread-related locations,
|
|
Me@29
|
619 * masterVP any AppVPs that might still be allocated and sitting in the
|
|
Me@29
|
620 * semantic environment, or have been orphaned in the _VMSWorkQ.
|
|
Me@29
|
621 *
|
|
Me@29
|
622 *NOTE: the semantic plug-in is expected to use VMS__malloc_to to get all the
|
|
Me@29
|
623 * locations it needs, and give ownership to masterVP. Then, they will be
|
|
Me@29
|
624 * automatically freed when the masterVP is dissipated. (This happens after
|
|
Me@29
|
625 * the core loop threads have all exited)
|
|
Me@22
|
626 *
|
|
Me@29
|
627 *In here,create one core-loop shut-down processor for each core loop and put
|
|
Me@31
|
628 * them all directly into the readyToAnimateQ.
|
|
Me@29
|
629 *Note, this function can ONLY be called after the semantic environment no
|
|
Me@29
|
630 * longer cares if AppVPs get animated after the point this is called. In
|
|
Me@29
|
631 * other words, this can be used as an abort, or else it should only be
|
|
Me@29
|
632 * called when all AppVPs have finished dissipate requests -- only at that
|
|
Me@29
|
633 * point is it sure that all results have completed.
|
|
Me@22
|
634 */
|
|
Me@22
|
635 void
|
|
Me@29
|
636 VMS__handle_shutdown_reqst( void *dummy, VirtProcr *animatingPr )
|
|
Me@8
|
637 { int coreIdx;
|
|
Me@14
|
638 VirtProcr *shutDownPr;
|
|
Me@22
|
639
|
|
Me@29
|
640 //create the shutdown processors, one for each core loop -- put them
|
|
Me@31
|
641 // directly into the Q -- each core will die when gets one
|
|
Me@8
|
642 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@8
|
643 {
|
|
Me@29
|
644 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
|
|
Me@31
|
645 writeSRSWQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
|
|
Me@8
|
646 }
|
|
Me@22
|
647
|
|
Me@12
|
648 }
|
|
Me@12
|
649
|
|
Me@12
|
650
|
|
Me@29
|
651 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
|
|
Me@29
|
652 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
|
|
Me@29
|
653 *This function has the sole purpose of setting the stack and framePtr
|
|
Me@29
|
654 * to the coreLoop's stack and framePtr.. it does that then jumps to the
|
|
Me@29
|
655 * core loop's shutdown point -- might be able to just call Pthread_exit
|
|
Me@30
|
656 * from here, but am going back to the pthread's stack and setting everything
|
|
Me@29
|
657 * up just as if it never jumped out, before calling pthread_exit.
|
|
Me@29
|
658 *The end-point of core loop will free the stack and so forth of the
|
|
Me@29
|
659 * processor that animates this function, (this fn is transfering the
|
|
Me@29
|
660 * animator of the AppVP that is in turn animating this function over
|
|
Me@29
|
661 * to core loop function -- note that this slices out a level of virtual
|
|
Me@29
|
662 * processors).
|
|
Me@29
|
663 */
|
|
Me@29
|
664 void
|
|
Me@29
|
665 endOSThreadFn( void *initData, VirtProcr *animatingPr )
|
|
Me@29
|
666 { void *jmpPt, *coreLoopStackPtr, *coreLoopFramePtr;
|
|
Me@29
|
667
|
|
Me@29
|
668 jmpPt = _VMSMasterEnv->coreLoopEndPt;
|
|
Me@29
|
669 coreLoopStackPtr = animatingPr->coreLoopStackPtr;
|
|
Me@29
|
670 coreLoopFramePtr = animatingPr->coreLoopFramePtr;
|
|
Me@29
|
671
|
|
Me@29
|
672
|
|
Me@29
|
673 asm volatile("movl %0, %%eax; \
|
|
Me@29
|
674 movl %1, %%esp; \
|
|
Me@29
|
675 movl %2, %%ebp; \
|
|
Me@29
|
676 jmp %%eax " \
|
|
Me@29
|
677 /* outputs */ : \
|
|
Me@29
|
678 /* inputs */ : "m" (jmpPt), "m"(coreLoopStackPtr), "m"(coreLoopFramePtr)\
|
|
Me@29
|
679 /* clobber */ : "memory", "%eax", "%ebx", "%ecx", "%edx", "%edi","%esi" \
|
|
Me@29
|
680 );
|
|
Me@29
|
681 }
|
|
Me@29
|
682
|
|
Me@29
|
683
|
|
Me@31
|
684 /*This is called after the threads have shut down and control has returned
|
|
Me@30
|
685 * to the semantic layer, in the entry point function in the main thread.
|
|
Me@30
|
686 * It has to free anything allocated during VMS_init, and any other alloc'd
|
|
Me@24
|
687 * locations that might be left over.
|
|
Me@24
|
688 */
|
|
Me@24
|
689 void
|
|
Me@29
|
690 VMS__cleanup_after_shutdown()
|
|
Me@31
|
691 {
|
|
Me@47
|
692 VMSQueueStruc **readyToAnimateQs;
|
|
Me@31
|
693 int coreIdx;
|
|
Me@31
|
694 VirtProcr **masterVPs;
|
|
Me@31
|
695 SchedSlot ***allSchedSlots; //ptr to array of ptrs
|
|
Me@31
|
696
|
|
Me@31
|
697 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
|
|
Me@31
|
698 masterVPs = _VMSMasterEnv->masterVPs;
|
|
Me@31
|
699 allSchedSlots = _VMSMasterEnv->allSchedSlots;
|
|
Me@31
|
700
|
|
Me@31
|
701 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@24
|
702 {
|
|
Me@31
|
703 freeSRSWQ( readyToAnimateQs[ coreIdx ] );
|
|
Me@31
|
704
|
|
Me@31
|
705 VMS__handle_dissipate_reqst( masterVPs[ coreIdx ] );
|
|
Me@31
|
706
|
|
Me@31
|
707 freeSchedSlots( allSchedSlots[ coreIdx ] );
|
|
Me@24
|
708 }
|
|
Me@31
|
709
|
|
Me@31
|
710 free( _VMSMasterEnv->readyToAnimateQs );
|
|
Me@31
|
711 free( _VMSMasterEnv->masterVPs );
|
|
Me@31
|
712 free( _VMSMasterEnv->allSchedSlots );
|
|
Me@24
|
713
|
|
Me@24
|
714 free( _VMSMasterEnv );
|
|
Me@24
|
715 }
|
|
Me@24
|
716
|
|
Me@24
|
717
|
|
Me@24
|
718 //===========================================================================
|
|
Me@12
|
719
|
|
Me@47
|
720 inline TSCount getTSC()
|
|
Me@12
|
721 { unsigned int low, high;
|
|
Me@12
|
722 TSCount out;
|
|
Me@12
|
723
|
|
Me@12
|
724 saveTimeStampCountInto( low, high );
|
|
Me@12
|
725 out = high;
|
|
Me@12
|
726 out = (out << 32) + low;
|
|
Me@12
|
727 return out;
|
|
Me@12
|
728 }
|
|
Me@12
|
729
|
