| 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@31
|
82 SRSWQueueStruc **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@31
|
96 readyToAnimateQs = malloc( NUM_CORES * sizeof(SRSWQueueStruc *) );
|
|
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@28
|
115
|
|
Me@12
|
116
|
|
Me@31
|
117 //Aug 19, 2010: no longer need to place initial masterVP into queue
|
|
Me@31
|
118 // because coreLoop now controls -- animates its masterVP when no work
|
|
Me@31
|
119
|
|
Me@30
|
120
|
|
Me@30
|
121 //==================== malloc substitute ========================
|
|
Me@30
|
122 //
|
|
Me@30
|
123 //Testing whether malloc is using thread-local storage and therefore
|
|
Me@30
|
124 // causing unreliable behavior.
|
|
Me@30
|
125 //Just allocate a massive chunk of memory and roll own malloc/free and
|
|
Me@30
|
126 // make app use VMS__malloc_to, which will suspend and perform malloc
|
|
Me@30
|
127 // in the master, taking from this massive chunk.
|
|
Me@30
|
128
|
|
Me@30
|
129 // initFreeList();
|
|
Me@38
|
130
|
|
Me@38
|
131 //============================= MEASUREMENT STUFF ========================
|
|
Me@38
|
132 #ifdef MEAS__TIME_STAMP_SUSP
|
|
Me@40
|
133 //RDTSC may run out of order, and so measure a time-span different
|
|
Me@40
|
134 // from the desired time-span -- got some weird changes in suspend
|
|
Me@40
|
135 // hist when added Master hist
|
|
Me@40
|
136 _VMSMasterEnv->measSuspHist = makeHistogram( 25, 110, 1300 );
|
|
Me@38
|
137 #endif
|
|
Me@38
|
138
|
|
Me@38
|
139 #ifdef MEAS__TIME_MASTER
|
|
Me@38
|
140 _VMSMasterEnv->measMasterHist = makeHistogram( 25, 500, 800 );
|
|
Me@38
|
141 #endif
|
|
Me@38
|
142 //========================================================================
|
|
Me@38
|
143
|
|
Me@0
|
144 }
|
|
Me@0
|
145
|
|
Me@30
|
146 /*
|
|
Me@30
|
147 void
|
|
Me@30
|
148 initMasterMalloc()
|
|
Me@30
|
149 {
|
|
Me@30
|
150 _VMSMasterEnv->mallocChunk = malloc( MASSIVE_MALLOC_SIZE );
|
|
Me@30
|
151
|
|
Me@30
|
152 //The free-list element is the first several locations of an
|
|
Me@30
|
153 // allocated chunk -- the address given to the application is pre-
|
|
Me@30
|
154 // pended with both the ownership structure and the free-list struc.
|
|
Me@30
|
155 //So, write the values of these into the first locations of
|
|
Me@30
|
156 // mallocChunk -- which marks it as free & puts in its size.
|
|
Me@30
|
157 listElem = (FreeListElem *)_VMSMasterEnv->mallocChunk;
|
|
Me@30
|
158 listElem->size = MASSIVE_MALLOC_SIZE - NUM_PREPEND_BYTES
|
|
Me@30
|
159 listElem->next = NULL;
|
|
Me@30
|
160 }
|
|
Me@30
|
161
|
|
Me@30
|
162 void
|
|
Me@30
|
163 dissipateMasterMalloc()
|
|
Me@30
|
164 {
|
|
Me@30
|
165 //Just foo code -- to get going -- doing as if free list were link-list
|
|
Me@30
|
166 currElem = _VMSMasterEnv->freeList;
|
|
Me@30
|
167 while( currElem != NULL )
|
|
Me@30
|
168 {
|
|
Me@30
|
169 nextElem = currElem->next;
|
|
Me@30
|
170 masterFree( currElem );
|
|
Me@30
|
171 currElem = nextElem;
|
|
Me@30
|
172 }
|
|
Me@30
|
173 free( _VMSMasterEnv->freeList );
|
|
Me@30
|
174 }
|
|
Me@30
|
175 */
|
|
Me@30
|
176
|
|
Me@31
|
177 SchedSlot **
|
|
Me@31
|
178 create_sched_slots()
|
|
Me@31
|
179 { SchedSlot **schedSlots;
|
|
Me@0
|
180 int i;
|
|
Me@0
|
181
|
|
Me@8
|
182 schedSlots = malloc( NUM_SCHED_SLOTS * sizeof(SchedSlot *) );
|
|
Me@8
|
183
|
|
Me@1
|
184 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
|
|
Me@0
|
185 {
|
|
Me@8
|
186 schedSlots[i] = malloc( sizeof(SchedSlot) );
|
|
Me@8
|
187
|
|
Me@1
|
188 //Set state to mean "handling requests done, slot needs filling"
|
|
Me@8
|
189 schedSlots[i]->workIsDone = FALSE;
|
|
Me@8
|
190 schedSlots[i]->needsProcrAssigned = TRUE;
|
|
Me@0
|
191 }
|
|
Me@31
|
192 return schedSlots;
|
|
Me@31
|
193 }
|
|
Me@31
|
194
|
|
Me@31
|
195
|
|
Me@31
|
196 void
|
|
Me@31
|
197 freeSchedSlots( SchedSlot **schedSlots )
|
|
Me@31
|
198 { int i;
|
|
Me@31
|
199 for( i = 0; i < NUM_SCHED_SLOTS; i++ )
|
|
Me@31
|
200 {
|
|
Me@31
|
201 free( schedSlots[i] );
|
|
Me@31
|
202 }
|
|
Me@31
|
203 free( schedSlots );
|
|
Me@0
|
204 }
|
|
Me@0
|
205
|
|
Me@8
|
206
|
|
Me@28
|
207 void
|
|
Me@28
|
208 create_the_coreLoop_OS_threads()
|
|
Me@28
|
209 {
|
|
Me@28
|
210 //========================================================================
|
|
Me@28
|
211 // Create the Threads
|
|
Me@28
|
212 int coreIdx, retCode;
|
|
Me@28
|
213
|
|
Me@28
|
214 //Need the threads to be created suspended, and wait for a signal
|
|
Me@28
|
215 // before proceeding -- gives time after creating to initialize other
|
|
Me@28
|
216 // stuff before the coreLoops set off.
|
|
Me@28
|
217 _VMSMasterEnv->setupComplete = 0;
|
|
Me@28
|
218
|
|
Me@28
|
219 //Make the threads that animate the core loops
|
|
Me@28
|
220 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@28
|
221 { coreLoopThdParams[coreIdx] = malloc( sizeof(ThdParams) );
|
|
Me@28
|
222 coreLoopThdParams[coreIdx]->coreNum = coreIdx;
|
|
Me@28
|
223
|
|
Me@28
|
224 retCode =
|
|
Me@28
|
225 pthread_create( &(coreLoopThdHandles[coreIdx]),
|
|
Me@28
|
226 thdAttrs,
|
|
Me@28
|
227 &coreLoop,
|
|
Me@28
|
228 (void *)(coreLoopThdParams[coreIdx]) );
|
|
Me@28
|
229 if(retCode){printf("ERROR creating thread: %d\n", retCode); exit(0);}
|
|
Me@28
|
230 }
|
|
Me@28
|
231 }
|
|
Me@28
|
232
|
|
Me@0
|
233 /*Semantic layer calls this when it want the system to start running..
|
|
Me@0
|
234 *
|
|
Me@24
|
235 *This starts the core loops running then waits for them to exit.
|
|
Me@0
|
236 */
|
|
Me@12
|
237 void
|
|
Me@24
|
238 VMS__start_the_work_then_wait_until_done()
|
|
Me@12
|
239 { int coreIdx;
|
|
Me@24
|
240 //Start the core loops running
|
|
Me@24
|
241 //===========================================================================
|
|
Me@25
|
242 TSCount startCount, endCount;
|
|
Me@24
|
243 unsigned long long count = 0, freq = 0;
|
|
Me@25
|
244 double runTime;
|
|
Me@0
|
245
|
|
Me@25
|
246 startCount = getTSCount();
|
|
Me@25
|
247
|
|
Me@25
|
248 //tell the core loop threads that setup is complete
|
|
Me@25
|
249 //get lock, to lock out any threads still starting up -- they'll see
|
|
Me@25
|
250 // that setupComplete is true before entering while loop, and so never
|
|
Me@25
|
251 // wait on the condition
|
|
Me@26
|
252 pthread_mutex_lock( &suspendLock );
|
|
Me@25
|
253 _VMSMasterEnv->setupComplete = 1;
|
|
Me@26
|
254 pthread_mutex_unlock( &suspendLock );
|
|
Me@26
|
255 pthread_cond_broadcast( &suspend_cond );
|
|
Me@25
|
256
|
|
Me@25
|
257
|
|
Me@24
|
258 //wait for all to complete
|
|
Me@8
|
259 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@8
|
260 {
|
|
Me@25
|
261 pthread_join( coreLoopThdHandles[coreIdx], NULL );
|
|
Me@24
|
262 }
|
|
Me@25
|
263
|
|
Me@24
|
264 //NOTE: do not clean up VMS env here -- semantic layer has to have
|
|
Me@24
|
265 // a chance to clean up its environment first, then do a call to free
|
|
Me@24
|
266 // the Master env and rest of VMS locations
|
|
Me@24
|
267
|
|
Me@24
|
268
|
|
Me@25
|
269 endCount = getTSCount();
|
|
Me@25
|
270 count = endCount - startCount;
|
|
Me@24
|
271
|
|
Me@25
|
272 runTime = (double)count / (double)TSCOUNT_FREQ;
|
|
Me@25
|
273
|
|
Me@25
|
274 printf("\n Time startup to shutdown: %f\n", runTime); fflush( stdin );
|
|
Me@8
|
275 }
|
|
Me@0
|
276
|
|
Me@28
|
277 /*Only difference between version with an OS thread pinned to each core and
|
|
Me@28
|
278 * the sequential version of VMS is VMS__init_Seq, this, and coreLoop_Seq.
|
|
Me@28
|
279 */
|
|
Me@28
|
280 void
|
|
Me@28
|
281 VMS__start_the_work_then_wait_until_done_Seq()
|
|
Me@28
|
282 {
|
|
Me@28
|
283 //Instead of un-suspending threads, just call the one and only
|
|
Me@28
|
284 // core loop (sequential version), in the main thread.
|
|
Me@28
|
285 coreLoop_Seq( NULL );
|
|
Me@28
|
286
|
|
Me@28
|
287 }
|
|
Me@28
|
288
|
|
Me@0
|
289
|
|
Me@0
|
290
|
|
Me@8
|
291 /*Create stack, then create __cdecl structure on it and put initialData and
|
|
Me@8
|
292 * pointer to the new structure instance into the parameter positions on
|
|
Me@8
|
293 * the stack
|
|
Me@8
|
294 *Then put function pointer into nextInstrPt -- the stack is setup in std
|
|
Me@8
|
295 * call structure, so jumping to function ptr is same as a GCC generated
|
|
Me@8
|
296 * function call
|
|
Me@8
|
297 *No need to save registers on old stack frame, because there's no old
|
|
Me@8
|
298 * animator state to return to --
|
|
Me@8
|
299 *
|
|
Me@8
|
300 */
|
|
Me@8
|
301 VirtProcr *
|
|
Me@8
|
302 VMS__create_procr( VirtProcrFnPtr fnPtr, void *initialData )
|
|
Me@8
|
303 { VirtProcr *newPr;
|
|
Me@8
|
304 char *stackLocs, *stackPtr;
|
|
Me@8
|
305
|
|
Me@8
|
306 newPr = malloc( sizeof(VirtProcr) );
|
|
Me@12
|
307 newPr->procrID = numProcrsCreated++;
|
|
Me@8
|
308 newPr->nextInstrPt = fnPtr;
|
|
Me@8
|
309 newPr->initialData = initialData;
|
|
Me@31
|
310 newPr->requests = NULL;
|
|
Me@38
|
311 newPr->schedSlot = NULL;
|
|
Me@31
|
312 // newPr->coreLoopStartPt = _VMSMasterEnv->coreLoopStartPt;
|
|
Me@8
|
313
|
|
Me@14
|
314 //fnPtr takes two params -- void *initData & void *animProcr
|
|
Me@8
|
315 //alloc stack locations, make stackPtr be the highest addr minus room
|
|
Me@14
|
316 // for 2 params + return addr. Return addr (NULL) is in loc pointed to
|
|
Me@14
|
317 // by stackPtr, initData at stackPtr + 4 bytes, animatingPr just above
|
|
Me@22
|
318 stackLocs = malloc( VIRT_PROCR_STACK_SIZE );
|
|
Me@26
|
319 if(stackLocs == 0)
|
|
Me@26
|
320 {perror("malloc stack"); exit(1);}
|
|
Me@22
|
321 newPr->startOfStack = stackLocs;
|
|
Me@22
|
322 stackPtr = ( (char *)stackLocs + VIRT_PROCR_STACK_SIZE - 0x10 );
|
|
Me@8
|
323 //setup __cdecl on stack -- coreloop will switch to stackPtr before jmp
|
|
Me@22
|
324 *( (int *)stackPtr + 2 ) = (int) newPr; //rightmost param -- 32bit pointer
|
|
Me@14
|
325 *( (int *)stackPtr + 1 ) = (int) initialData; //next param to left
|
|
Me@8
|
326 newPr->stackPtr = stackPtr; //core loop will switch to this, then
|
|
Me@8
|
327 newPr->framePtr = stackPtr; //suspend loop will save new stack & frame ptr
|
|
Me@8
|
328
|
|
Me@8
|
329 return newPr;
|
|
Me@8
|
330 }
|
|
Me@8
|
331
|
|
Me@8
|
332
|
|
Me@26
|
333 /*there is a label inside this function -- save the addr of this label in
|
|
Me@0
|
334 * the callingPr struc, as the pick-up point from which to start the next
|
|
Me@0
|
335 * work-unit for that procr. If turns out have to save registers, then
|
|
Me@0
|
336 * save them in the procr struc too. Then do assembly jump to the CoreLoop's
|
|
Me@0
|
337 * "done with work-unit" label. The procr struc is in the request in the
|
|
Me@0
|
338 * slave that animated the just-ended work-unit, so all the state is saved
|
|
Me@0
|
339 * there, and will get passed along, inside the request handler, to the
|
|
Me@0
|
340 * next work-unit for that procr.
|
|
Me@0
|
341 */
|
|
Me@8
|
342 void
|
|
Me@38
|
343 VMS__suspend_procr( VirtProcr *animatingPr )
|
|
Me@14
|
344 { void *jmpPt, *stackPtrAddr, *framePtrAddr, *coreLoopStackPtr;
|
|
Me@14
|
345 void *coreLoopFramePtr;
|
|
Me@0
|
346
|
|
Me@14
|
347 //The request to master will cause this suspended virt procr to get
|
|
Me@14
|
348 // scheduled again at some future point -- to resume, core loop jumps
|
|
Me@14
|
349 // to the resume point (below), which causes restore of saved regs and
|
|
Me@14
|
350 // "return" from this call.
|
|
Me@38
|
351 animatingPr->nextInstrPt = &&ResumePt;
|
|
Me@1
|
352
|
|
Me@1
|
353 //return ownership of the virt procr and sched slot to Master virt pr
|
|
Me@38
|
354 animatingPr->schedSlot->workIsDone = TRUE;
|
|
Me@14
|
355 // coreIdx = callingPr->coreAnimatedBy;
|
|
Me@1
|
356
|
|
Me@38
|
357 stackPtrAddr = &(animatingPr->stackPtr);
|
|
Me@38
|
358 framePtrAddr = &(animatingPr->framePtr);
|
|
Me@26
|
359
|
|
Me@31
|
360 jmpPt = _VMSMasterEnv->coreLoopStartPt;
|
|
Me@38
|
361 coreLoopFramePtr = animatingPr->coreLoopFramePtr;//need this only
|
|
Me@38
|
362 coreLoopStackPtr = animatingPr->coreLoopStackPtr;//safety
|
|
Me@1
|
363
|
|
Me@26
|
364 //Save the virt procr's stack and frame ptrs,
|
|
Me@18
|
365 asm volatile("movl %0, %%eax; \
|
|
Me@18
|
366 movl %%esp, (%%eax); \
|
|
Me@18
|
367 movl %1, %%eax; \
|
|
Me@26
|
368 movl %%ebp, (%%eax) "\
|
|
Me@26
|
369 /* outputs */ : "=g" (stackPtrAddr), "=g" (framePtrAddr) \
|
|
Me@26
|
370 /* inputs */ : \
|
|
Me@26
|
371 /* clobber */ : "%eax" \
|
|
Me@26
|
372 );
|
|
Me@26
|
373
|
|
Me@38
|
374 #ifdef MEAS__TIME_STAMP_SUSP
|
|
Me@40
|
375 //record time stamp: compare to time-stamp recorded below, at resume
|
|
Me@38
|
376 saveLowTimeStampCountInto( animatingPr->preSuspTSCLow );
|
|
Me@38
|
377 #endif
|
|
Me@38
|
378
|
|
Me@26
|
379 //restore coreloop's frame ptr, then jump back to "start" of core loop
|
|
Me@26
|
380 //Note, GCC compiles to assembly that saves esp and ebp in the stack
|
|
Me@26
|
381 // frame -- so have to explicitly do assembly that saves to memory
|
|
Me@26
|
382 asm volatile("movl %0, %%eax; \
|
|
Me@26
|
383 movl %1, %%esp; \
|
|
Me@26
|
384 movl %2, %%ebp; \
|
|
Me@18
|
385 jmp %%eax " \
|
|
Me@26
|
386 /* outputs */ : \
|
|
Me@26
|
387 /* inputs */ : "m" (jmpPt), "m"(coreLoopStackPtr), "m"(coreLoopFramePtr)\
|
|
Me@18
|
388 /* clobber */ : "memory", "%eax", "%ebx", "%ecx", "%edx", "%edi","%esi" \
|
|
Me@12
|
389 ); //list everything as clobbered to force GCC to save all
|
|
Me@12
|
390 // live vars that are in regs on stack before this
|
|
Me@12
|
391 // assembly, so that stack pointer is correct, before jmp
|
|
Me@1
|
392
|
|
Me@1
|
393 ResumePt:
|
|
Me@38
|
394 #ifdef MEAS__TIME_STAMP_SUSP
|
|
Me@38
|
395 saveLowTimeStampCountInto( animatingPr->postSuspTSCLow );
|
|
Me@38
|
396 //Take difference between the pre-suspend and post-suspend times
|
|
Me@38
|
397 // and do sanity check to see if rollover happened between
|
|
Me@38
|
398 int diff = animatingPr->postSuspTSCLow - animatingPr->preSuspTSCLow;
|
|
Me@38
|
399 if( diff > 1000000 ) diff = 0;
|
|
Me@38
|
400 addToHist( diff, _VMSMasterEnv->measSuspHist );
|
|
Me@38
|
401
|
|
Me@38
|
402 #endif
|
|
Me@38
|
403
|
|
Me@0
|
404 return;
|
|
Me@0
|
405 }
|
|
Me@0
|
406
|
|
Me@22
|
407
|
|
Me@22
|
408
|
|
Me@22
|
409
|
|
Me@38
|
410 /*
|
|
Me@22
|
411 *This adds a request to dissipate, then suspends the processor so that the
|
|
Me@22
|
412 * request handler will receive the request. The request handler is what
|
|
Me@22
|
413 * does the work of freeing memory and removing the processor from the
|
|
Me@22
|
414 * semantic environment's data structures.
|
|
Me@22
|
415 *The request handler also is what figures out when to shutdown the VMS
|
|
Me@22
|
416 * system -- which causes all the core loop threads to die, and returns from
|
|
Me@22
|
417 * the call that started up VMS to perform the work.
|
|
Me@22
|
418 *
|
|
Me@22
|
419 *This form is a bit misleading to understand if one is trying to figure out
|
|
Me@22
|
420 * how VMS works -- it looks like a normal function call, but inside it
|
|
Me@22
|
421 * sends a request to the request handler and suspends the processor, which
|
|
Me@22
|
422 * jumps out of the VMS__dissipate_procr function, and out of all nestings
|
|
Me@22
|
423 * above it, transferring the work of dissipating to the request handler,
|
|
Me@22
|
424 * which then does the actual work -- causing the processor that animated
|
|
Me@22
|
425 * the call of this function to disappear and the "hanging" state of this
|
|
Me@22
|
426 * function to just poof into thin air -- the virtual processor's trace
|
|
Me@22
|
427 * never returns from this call, but instead the virtual processor's trace
|
|
Me@22
|
428 * gets suspended in this call and all the virt processor's state disap-
|
|
Me@22
|
429 * pears -- making that suspend the last thing in the virt procr's trace.
|
|
Me@8
|
430 */
|
|
Me@8
|
431 void
|
|
Me@22
|
432 VMS__dissipate_procr( VirtProcr *procrToDissipate )
|
|
Me@22
|
433 { VMSReqst *req;
|
|
Me@22
|
434
|
|
Me@22
|
435 req = malloc( sizeof(VMSReqst) );
|
|
Me@22
|
436 // req->virtProcrFrom = callingPr;
|
|
Me@22
|
437 req->reqType = dissipate;
|
|
Me@22
|
438 req->nextReqst = procrToDissipate->requests;
|
|
Me@22
|
439 procrToDissipate->requests = req;
|
|
Me@22
|
440
|
|
Me@22
|
441 VMS__suspend_procr( procrToDissipate );
|
|
Me@22
|
442 }
|
|
Me@22
|
443
|
|
Me@22
|
444
|
|
Me@22
|
445 /*This inserts the semantic-layer's request data into standard VMS carrier
|
|
Me@22
|
446 */
|
|
Me@22
|
447 inline void
|
|
Me@24
|
448 VMS__add_sem_request( void *semReqData, VirtProcr *callingPr )
|
|
Me@22
|
449 { VMSReqst *req;
|
|
Me@22
|
450
|
|
Me@22
|
451 req = malloc( sizeof(VMSReqst) );
|
|
Me@22
|
452 // req->virtProcrFrom = callingPr;
|
|
Me@22
|
453 req->reqType = semantic;
|
|
Me@22
|
454 req->semReqData = semReqData;
|
|
Me@22
|
455 req->nextReqst = callingPr->requests;
|
|
Me@22
|
456 callingPr->requests = req;
|
|
Me@22
|
457 }
|
|
Me@22
|
458
|
|
Me@22
|
459
|
|
Me@38
|
460 /*Use this to get first request before starting request handler's loop
|
|
Me@38
|
461 */
|
|
Me@24
|
462 VMSReqst *
|
|
Me@24
|
463 VMS__take_top_request_from( VirtProcr *procrWithReq )
|
|
Me@24
|
464 { VMSReqst *req;
|
|
Me@24
|
465
|
|
Me@24
|
466 req = procrWithReq->requests;
|
|
Me@24
|
467 if( req == NULL ) return req;
|
|
Me@31
|
468
|
|
Me@24
|
469 procrWithReq->requests = procrWithReq->requests->nextReqst;
|
|
Me@24
|
470 return req;
|
|
Me@24
|
471 }
|
|
Me@24
|
472
|
|
Me@38
|
473 /*A subtle bug due to freeing then accessing "next" after freed caused this
|
|
Me@38
|
474 * form of call to be put in -- so call this at end of request handler loop
|
|
Me@38
|
475 * that iterates through the requests.
|
|
Me@38
|
476 */
|
|
Me@31
|
477 VMSReqst *
|
|
Me@31
|
478 VMS__free_top_and_give_next_request_from( VirtProcr *procrWithReq )
|
|
Me@31
|
479 { VMSReqst *req;
|
|
Me@31
|
480
|
|
Me@31
|
481 req = procrWithReq->requests;
|
|
Me@38
|
482 if( req == NULL ) return NULL;
|
|
Me@31
|
483
|
|
Me@31
|
484 procrWithReq->requests = procrWithReq->requests->nextReqst;
|
|
Me@31
|
485 VMS__free_request( req );
|
|
Me@31
|
486 return procrWithReq->requests;
|
|
Me@31
|
487 }
|
|
Me@31
|
488
|
|
Me@38
|
489
|
|
Me@38
|
490 //TODO: add a semantic-layer supplied "freer" for the semantic-data portion
|
|
Me@38
|
491 // of a request -- IE call with both a virt procr and a fn-ptr to request
|
|
Me@38
|
492 // freer (also maybe put sem request freer as a field in virt procr?)
|
|
Me@38
|
493 //MeasVMS relies right now on this only freeing VMS layer of request -- the
|
|
Me@38
|
494 // semantic portion of request is alloc'd and freed by request handler
|
|
Me@38
|
495 void
|
|
Me@38
|
496 VMS__free_request( VMSReqst *req )
|
|
Me@38
|
497 {
|
|
Me@38
|
498 free( req );
|
|
Me@38
|
499 }
|
|
Me@38
|
500
|
|
Me@38
|
501
|
|
Me@38
|
502
|
|
Me@24
|
503 inline int
|
|
Me@24
|
504 VMS__isSemanticReqst( VMSReqst *req )
|
|
Me@22
|
505 {
|
|
Me@24
|
506 return ( req->reqType == semantic );
|
|
Me@24
|
507 }
|
|
Me@22
|
508
|
|
Me@24
|
509
|
|
Me@24
|
510 inline void *
|
|
Me@24
|
511 VMS__take_sem_reqst_from( VMSReqst *req )
|
|
Me@24
|
512 {
|
|
Me@24
|
513 return req->semReqData;
|
|
Me@24
|
514 }
|
|
Me@24
|
515
|
|
Me@24
|
516 inline int
|
|
Me@24
|
517 VMS__isDissipateReqst( VMSReqst *req )
|
|
Me@24
|
518 {
|
|
Me@24
|
519 return ( req->reqType == dissipate );
|
|
Me@24
|
520 }
|
|
Me@24
|
521
|
|
Me@24
|
522 inline int
|
|
Me@24
|
523 VMS__isCreateReqst( VMSReqst *req )
|
|
Me@24
|
524 {
|
|
Me@24
|
525 return ( req->reqType == regCreated );
|
|
Me@24
|
526 }
|
|
Me@24
|
527
|
|
Me@24
|
528 void
|
|
Me@38
|
529 VMS__send_req_to_register_new_procr(VirtProcr *newPr, VirtProcr *reqstingPr)
|
|
Me@24
|
530 { VMSReqst *req;
|
|
Me@24
|
531
|
|
Me@24
|
532 req = malloc( sizeof(VMSReqst) );
|
|
Me@24
|
533 req->reqType = regCreated;
|
|
Me@24
|
534 req->semReqData = newPr;
|
|
Me@24
|
535 req->nextReqst = reqstingPr->requests;
|
|
Me@24
|
536 reqstingPr->requests = req;
|
|
Me@24
|
537
|
|
Me@24
|
538 VMS__suspend_procr( reqstingPr );
|
|
Me@22
|
539 }
|
|
Me@22
|
540
|
|
Me@22
|
541
|
|
Me@22
|
542
|
|
Me@24
|
543 /*This must be called by the request handler plugin -- it cannot be called
|
|
Me@24
|
544 * from the semantic library "dissipate processor" function -- instead, the
|
|
Me@24
|
545 * semantic layer has to generate a request for the plug-in to call this
|
|
Me@24
|
546 * function.
|
|
Me@24
|
547 *The reason is that this frees the virtual processor's stack -- which is
|
|
Me@24
|
548 * still in use inside semantic library calls!
|
|
Me@24
|
549 *
|
|
Me@24
|
550 *This frees or recycles all the state owned by and comprising the VMS
|
|
Me@24
|
551 * portion of the animating virtual procr. The request handler must first
|
|
Me@24
|
552 * free any semantic data created for the processor that didn't use the
|
|
Me@24
|
553 * VMS_malloc mechanism. Then it calls this, which first asks the malloc
|
|
Me@24
|
554 * system to disown any state that did use VMS_malloc, and then frees the
|
|
Me@24
|
555 * statck and the processor-struct itself.
|
|
Me@24
|
556 *If the dissipated processor is the sole (remaining) owner of VMS__malloc'd
|
|
Me@24
|
557 * state, then that state gets freed (or sent to recycling) as a side-effect
|
|
Me@24
|
558 * of dis-owning it.
|
|
Me@24
|
559 */
|
|
Me@24
|
560 void
|
|
Me@29
|
561 VMS__handle_dissipate_reqst( VirtProcr *animatingPr )
|
|
Me@24
|
562 {
|
|
Me@24
|
563 //dis-own all locations owned by this processor, causing to be freed
|
|
Me@24
|
564 // any locations that it is (was) sole owner of
|
|
Me@29
|
565 //TODO: implement VMS__malloc system, including "give up ownership"
|
|
Me@24
|
566
|
|
Me@24
|
567 //The dissipate request might still be attached, so remove and free it
|
|
Me@38
|
568 VMS__free_top_and_give_next_request_from( animatingPr );
|
|
Me@24
|
569
|
|
Me@24
|
570 //NOTE: initialData was given to the processor, so should either have
|
|
Me@24
|
571 // been alloc'd with VMS__malloc, or freed by the level above animPr.
|
|
Me@24
|
572 //So, all that's left to free here is the stack and the VirtProcr struc
|
|
Me@24
|
573 // itself
|
|
Me@24
|
574 free( animatingPr->startOfStack );
|
|
Me@24
|
575 free( animatingPr );
|
|
Me@24
|
576 }
|
|
Me@24
|
577
|
|
Me@24
|
578
|
|
Me@29
|
579 //TODO: re-architect so that have clean separation between request handler
|
|
Me@29
|
580 // and master loop, for dissipate, create, shutdown, and other non-semantic
|
|
Me@29
|
581 // requests. Issue is chain: one removes requests from AppVP, one dispatches
|
|
Me@29
|
582 // on type of request, and one handles each type.. but some types require
|
|
Me@29
|
583 // action from both request handler and master loop -- maybe just give the
|
|
Me@29
|
584 // request handler calls like: VMS__handle_X_request_type
|
|
Me@24
|
585
|
|
Me@29
|
586 void
|
|
Me@29
|
587 endOSThreadFn( void *initData, VirtProcr *animatingPr );
|
|
Me@29
|
588
|
|
Me@29
|
589 /*This is called by the semantic layer's request handler when it decides its
|
|
Me@29
|
590 * time to shut down the VMS system. Calling this causes the core loop OS
|
|
Me@29
|
591 * threads to exit, which unblocks the entry-point function that started up
|
|
Me@29
|
592 * VMS, and allows it to grab the result and return to the original single-
|
|
Me@29
|
593 * threaded application.
|
|
Me@22
|
594 *
|
|
Me@29
|
595 *The _VMSMasterEnv is needed by this shut down function, so the create-seed-
|
|
Me@29
|
596 * and-wait function has to free a bunch of stuff after it detects the
|
|
Me@29
|
597 * threads have all died: the masterEnv, the thread-related locations,
|
|
Me@29
|
598 * masterVP any AppVPs that might still be allocated and sitting in the
|
|
Me@29
|
599 * semantic environment, or have been orphaned in the _VMSWorkQ.
|
|
Me@29
|
600 *
|
|
Me@29
|
601 *NOTE: the semantic plug-in is expected to use VMS__malloc_to to get all the
|
|
Me@29
|
602 * locations it needs, and give ownership to masterVP. Then, they will be
|
|
Me@29
|
603 * automatically freed when the masterVP is dissipated. (This happens after
|
|
Me@29
|
604 * the core loop threads have all exited)
|
|
Me@22
|
605 *
|
|
Me@29
|
606 *In here,create one core-loop shut-down processor for each core loop and put
|
|
Me@31
|
607 * them all directly into the readyToAnimateQ.
|
|
Me@29
|
608 *Note, this function can ONLY be called after the semantic environment no
|
|
Me@29
|
609 * longer cares if AppVPs get animated after the point this is called. In
|
|
Me@29
|
610 * other words, this can be used as an abort, or else it should only be
|
|
Me@29
|
611 * called when all AppVPs have finished dissipate requests -- only at that
|
|
Me@29
|
612 * point is it sure that all results have completed.
|
|
Me@22
|
613 */
|
|
Me@22
|
614 void
|
|
Me@29
|
615 VMS__handle_shutdown_reqst( void *dummy, VirtProcr *animatingPr )
|
|
Me@8
|
616 { int coreIdx;
|
|
Me@14
|
617 VirtProcr *shutDownPr;
|
|
Me@22
|
618
|
|
Me@29
|
619 //create the shutdown processors, one for each core loop -- put them
|
|
Me@31
|
620 // directly into the Q -- each core will die when gets one
|
|
Me@8
|
621 for( coreIdx=0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@8
|
622 {
|
|
Me@29
|
623 shutDownPr = VMS__create_procr( &endOSThreadFn, NULL );
|
|
Me@31
|
624 writeSRSWQ( shutDownPr, _VMSMasterEnv->readyToAnimateQs[coreIdx] );
|
|
Me@8
|
625 }
|
|
Me@22
|
626
|
|
Me@12
|
627 }
|
|
Me@12
|
628
|
|
Me@12
|
629
|
|
Me@29
|
630 /*Am trying to be cute, avoiding IF statement in coreLoop that checks for
|
|
Me@29
|
631 * a special shutdown procr. Ended up with extra-complex shutdown sequence.
|
|
Me@29
|
632 *This function has the sole purpose of setting the stack and framePtr
|
|
Me@29
|
633 * to the coreLoop's stack and framePtr.. it does that then jumps to the
|
|
Me@29
|
634 * core loop's shutdown point -- might be able to just call Pthread_exit
|
|
Me@30
|
635 * from here, but am going back to the pthread's stack and setting everything
|
|
Me@29
|
636 * up just as if it never jumped out, before calling pthread_exit.
|
|
Me@29
|
637 *The end-point of core loop will free the stack and so forth of the
|
|
Me@29
|
638 * processor that animates this function, (this fn is transfering the
|
|
Me@29
|
639 * animator of the AppVP that is in turn animating this function over
|
|
Me@29
|
640 * to core loop function -- note that this slices out a level of virtual
|
|
Me@29
|
641 * processors).
|
|
Me@29
|
642 */
|
|
Me@29
|
643 void
|
|
Me@29
|
644 endOSThreadFn( void *initData, VirtProcr *animatingPr )
|
|
Me@29
|
645 { void *jmpPt, *coreLoopStackPtr, *coreLoopFramePtr;
|
|
Me@29
|
646
|
|
Me@29
|
647 jmpPt = _VMSMasterEnv->coreLoopEndPt;
|
|
Me@29
|
648 coreLoopStackPtr = animatingPr->coreLoopStackPtr;
|
|
Me@29
|
649 coreLoopFramePtr = animatingPr->coreLoopFramePtr;
|
|
Me@29
|
650
|
|
Me@29
|
651
|
|
Me@29
|
652 asm volatile("movl %0, %%eax; \
|
|
Me@29
|
653 movl %1, %%esp; \
|
|
Me@29
|
654 movl %2, %%ebp; \
|
|
Me@29
|
655 jmp %%eax " \
|
|
Me@29
|
656 /* outputs */ : \
|
|
Me@29
|
657 /* inputs */ : "m" (jmpPt), "m"(coreLoopStackPtr), "m"(coreLoopFramePtr)\
|
|
Me@29
|
658 /* clobber */ : "memory", "%eax", "%ebx", "%ecx", "%edx", "%edi","%esi" \
|
|
Me@29
|
659 );
|
|
Me@29
|
660 }
|
|
Me@29
|
661
|
|
Me@29
|
662
|
|
Me@31
|
663 /*This is called after the threads have shut down and control has returned
|
|
Me@30
|
664 * to the semantic layer, in the entry point function in the main thread.
|
|
Me@30
|
665 * It has to free anything allocated during VMS_init, and any other alloc'd
|
|
Me@24
|
666 * locations that might be left over.
|
|
Me@24
|
667 */
|
|
Me@24
|
668 void
|
|
Me@29
|
669 VMS__cleanup_after_shutdown()
|
|
Me@31
|
670 {
|
|
Me@31
|
671 SRSWQueueStruc **readyToAnimateQs;
|
|
Me@31
|
672 int coreIdx;
|
|
Me@31
|
673 VirtProcr **masterVPs;
|
|
Me@31
|
674 SchedSlot ***allSchedSlots; //ptr to array of ptrs
|
|
Me@31
|
675
|
|
Me@31
|
676 readyToAnimateQs = _VMSMasterEnv->readyToAnimateQs;
|
|
Me@31
|
677 masterVPs = _VMSMasterEnv->masterVPs;
|
|
Me@31
|
678 allSchedSlots = _VMSMasterEnv->allSchedSlots;
|
|
Me@31
|
679
|
|
Me@31
|
680 for( coreIdx = 0; coreIdx < NUM_CORES; coreIdx++ )
|
|
Me@24
|
681 {
|
|
Me@31
|
682 freeSRSWQ( readyToAnimateQs[ coreIdx ] );
|
|
Me@31
|
683
|
|
Me@31
|
684 VMS__handle_dissipate_reqst( masterVPs[ coreIdx ] );
|
|
Me@31
|
685
|
|
Me@31
|
686 freeSchedSlots( allSchedSlots[ coreIdx ] );
|
|
Me@24
|
687 }
|
|
Me@31
|
688
|
|
Me@31
|
689 free( _VMSMasterEnv->readyToAnimateQs );
|
|
Me@31
|
690 free( _VMSMasterEnv->masterVPs );
|
|
Me@31
|
691 free( _VMSMasterEnv->allSchedSlots );
|
|
Me@24
|
692
|
|
Me@24
|
693 free( _VMSMasterEnv );
|
|
Me@24
|
694 }
|
|
Me@24
|
695
|
|
Me@24
|
696
|
|
Me@24
|
697 //===========================================================================
|
|
Me@12
|
698
|
|
Me@12
|
699 inline TSCount getTSCount()
|
|
Me@12
|
700 { unsigned int low, high;
|
|
Me@12
|
701 TSCount out;
|
|
Me@12
|
702
|
|
Me@12
|
703 saveTimeStampCountInto( low, high );
|
|
Me@12
|
704 out = high;
|
|
Me@12
|
705 out = (out << 32) + low;
|
|
Me@12
|
706 return out;
|
|
Me@12
|
707 }
|
|
Me@12
|
708
|
