Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view MasterLoop.c @ 177:3bd35fc83c61
move loopgraph recording to plugin (\! changed scheduler fn prototype \!)
| author | Nina Engelhardt <nengel@mailbox.tu-berlin.de> |
|---|---|
| date | Wed, 04 Jan 2012 16:40:10 +0100 |
| parents | 981acd1db6af |
| children | 7523ee70d66c |
line source
1 /*
2 * Copyright 2010 OpenSourceStewardshipFoundation
3 *
4 * Licensed under BSD
5 */
9 #include <stdio.h>
10 #include <stddef.h>
12 #include "VMS.h"
13 #include "ProcrContext.h"
16 //===========================================================================
17 void inline
18 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
19 VirtProcr *masterPr );
21 //===========================================================================
25 /*This code is animated by the virtual Master processor.
26 *
27 *Polls each sched slot exactly once, hands any requests made by a newly
28 * done slave to the "request handler" plug-in function
29 *
30 *Any slots that need a virt procr assigned are given to the "schedule"
31 * plug-in function, which tries to assign a virt procr (slave) to it.
32 *
33 *When all slots needing a processor have been given to the schedule plug-in,
34 * a fraction of the procrs successfully scheduled are put into the
35 * work queue, then a continuation of this function is put in, then the rest
36 * of the virt procrs that were successfully scheduled.
37 *
38 *The first thing the continuation does is busy-wait until the previous
39 * animation completes. This is because an (unlikely) continuation may
40 * sneak through queue before previous continuation is done putting second
41 * part of scheduled slaves in, which is the only race condition.
42 *
43 */
45 /*May 29, 2010 -- birth a Master during init so that first core loop to
46 * start running gets it and does all the stuff for a newly born --
47 * from then on, will be doing continuation, but do suspension self
48 * directly at end of master loop
49 *So VMS__init just births the master virtual processor same way it births
50 * all the others -- then does any extra setup needed and puts it into the
51 * work queue.
52 *However means have to make masterEnv a global static volatile the same way
53 * did with readyToAnimateQ in core loop. -- for performance, put the
54 * jump to the core loop directly in here, and have it directly jump back.
55 *
56 *
57 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
58 * avoids the suspected bug in the system stack that causes bizarre faults
59 * at random places in the system code.
60 *
61 *So, this function is coupled to each of the MasterVPs, -- meaning this
62 * function can't rely on a particular stack and frame -- each MasterVP that
63 * animates this function has a different one.
64 *
65 *At this point, the masterLoop does not write itself into the queue anymore,
66 * instead, the coreLoop acquires the masterLock when it has nothing to
67 * animate, and then animates its own masterLoop. However, still try to put
68 * several AppVPs into the queue to amortize the startup cost of switching
69 * to the MasterVP. Note, don't have to worry about latency of requests much
70 * because most requests generate work for same core -- only latency issue
71 * is case when other cores starved and one core's requests generate work
72 * for them -- so keep max in queue to 3 or 4..
73 */
74 void masterLoop( void *initData, VirtProcr *animatingPr )
75 {
76 int32 slotIdx, numSlotsFilled;
77 VirtProcr *schedVirtPr;
78 SchedSlot *currSlot, **schedSlots;
79 MasterEnv *masterEnv;
80 VMSQueueStruc *readyToAnimateQ;
82 SlaveScheduler slaveScheduler;
83 RequestHandler requestHandler;
84 void *semanticEnv;
86 int32 thisCoresIdx;
87 VirtProcr *masterPr;
88 volatile VirtProcr *volatileMasterPr;
90 volatileMasterPr = animatingPr;
91 masterPr = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
93 bulb b = new_bulb();
94 numSlotsFilled=1;
95 //First animation of each MasterVP will in turn animate this part
96 // of setup code.. (VP creator sets up the stack as if this function
97 // was called normally, but actually get here by jmp)
98 //So, setup values about stack ptr, jmp pt and all that
99 //masterPr->nextInstrPt = &&masterLoopStartPt;
102 //Note, got rid of writing the stack and frame ptr up here, because
103 // only one
104 // core can ever animate a given MasterVP, so don't need to communicate
105 // new frame and stack ptr to the MasterVP storage before a second
106 // version of that MasterVP can get animated on a different core.
107 //Also got rid of the busy-wait.
110 //masterLoopStartPt:
111 while(1){
113 //============================= MEASUREMENT STUFF ========================
114 #ifdef MEAS__TIME_MASTER
115 //Total Master time includes one coreloop time -- just assume the core
116 // loop time is same for Master as for AppVPs, even though it may be
117 // smaller due to higher predictability of the fixed jmp.
118 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
119 #endif
120 //========================================================================
122 masterEnv = (MasterEnv*)_VMSMasterEnv;
124 //GCC may optimize so doesn't always re-define from frame-storage
125 masterPr = (VirtProcr*)volatileMasterPr; //just to make sure after jmp
126 thisCoresIdx = masterPr->coreAnimatedBy;
127 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
128 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
130 requestHandler = masterEnv->requestHandler;
131 slaveScheduler = masterEnv->slaveScheduler;
132 semanticEnv = masterEnv->semanticEnv;
134 #ifdef DETECT_LOOP_GRAPH
136 #endif
138 //Poll each slot's Done flag
139 numSlotsFilled = 0;
140 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
141 {
142 currSlot = schedSlots[ slotIdx ];
144 if( currSlot->workIsDone )
145 {
146 currSlot->workIsDone = FALSE;
147 currSlot->needsProcrAssigned = TRUE;
149 //process requests from slave to master
150 //====================== MEASUREMENT STUFF ===================
151 #ifdef MEAS__TIME_PLUGIN
152 int32 startStamp1, endStamp1;
153 saveLowTimeStampCountInto( startStamp1 );
154 #endif
155 #ifdef MEAS__PERF_COUNTERS
156 int lastRecordIdx = currSlot->procrAssignedToSlot->counter_history_array_info->numInArray -1;
157 CounterRecord* lastRecord = currSlot->procrAssignedToSlot->counter_history[lastRecordIdx];
158 lastRecord->req_core = thisCoresIdx;
159 saveCyclesAndInstrs(thisCoresIdx,lastRecord->next_task_req_cycles,lastRecord->next_task_req_instrs);
160 //End of task, start of next task
161 //print counters from last run
162 addToDynArray((void*)lastRecord,masterEnv->counter_history_array_info);
163 print_record_csv_to_file(lastRecord,_VMSMasterEnv->counteroutput);
169 //print_record_human_readable(lastRecord);
170 //create new entry in record array
171 CounterRecord* newRecord = VMS__malloc(sizeof(CounterRecord));
172 newRecord->req_core = thisCoresIdx;
173 newRecord->vp_id = currSlot->procrAssignedToSlot->procrID;
174 newRecord->task_position = lastRecord->task_position + 1;
175 newRecord->req_cycles = lastRecord->next_task_req_cycles;
176 newRecord->req_instrs = lastRecord->next_task_req_instrs;
177 getReturnAddressBeforeLibraryCall(currSlot->procrAssignedToSlot, &(newRecord->addr_of_libcall_for_req));
178 addToDynArray( (void*) newRecord, currSlot->procrAssignedToSlot->counter_history_array_info);
179 lastRecord = newRecord;
180 #endif
181 //============================================================
182 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
183 //====================== MEASUREMENT STUFF ===================
184 #ifdef MEAS__TIME_PLUGIN
185 saveLowTimeStampCountInto( endStamp1 );
186 addIntervalToHist( startStamp1, endStamp1,
187 _VMSMasterEnv->reqHdlrLowTimeHist );
188 addIntervalToHist( startStamp1, endStamp1,
189 _VMSMasterEnv->reqHdlrHighTimeHist );
190 #endif
191 #ifdef MEAS__PERF_COUNTERS
192 //done with constraints check
193 saveCyclesAndInstrs(thisCoresIdx,lastRecord->sc_done_cycles,lastRecord->sc_done_instrs);
194 saveLowTimeStampCountInto(lastRecord->blocked_timestamp);
195 #endif
196 //============================================================
197 }
198 if( currSlot->needsProcrAssigned )
199 { //give slot a new virt procr
200 #ifdef MEAS__PERF_COUNTERS
201 //start assigner
202 uint64 tmp_cycles;
203 uint64 tmp_instrs;
204 saveCyclesAndInstrs(thisCoresIdx,tmp_cycles,tmp_instrs);
205 #endif
206 schedVirtPr =
207 (*slaveScheduler)( semanticEnv, thisCoresIdx, slotIdx );
209 if( schedVirtPr != NULL )
210 { currSlot->procrAssignedToSlot = schedVirtPr;
211 schedVirtPr->schedSlot = currSlot;
212 currSlot->needsProcrAssigned = FALSE;
213 numSlotsFilled += 1;
215 #ifdef MEAS__PERF_COUNTERS
216 //end assigner
217 int lastRecordIdx = currSlot->procrAssignedToSlot->counter_history_array_info->numInArray -1;
218 CounterRecord* lastRecord = currSlot->procrAssignedToSlot->counter_history[lastRecordIdx];
219 lastRecord->assigning_core = thisCoresIdx;
220 lastRecord->start_assign_cycles = tmp_cycles;
221 lastRecord->start_assign_instrs = tmp_instrs;
222 saveCyclesAndInstrs(thisCoresIdx,lastRecord->end_assign_cycles,lastRecord->end_assign_instrs);
223 #endif
224 #ifdef DETECT_LOOP_GRAPH
226 #endif
227 writeVMSQ( schedVirtPr, readyToAnimateQ );
228 }
229 }
231 }
234 #ifdef USE_WORK_STEALING
235 //If no slots filled, means no more work, look for work to steal.
236 if( numSlotsFilled == 0 )
237 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
238 }
239 #endif
242 #ifdef MEAS__TIME_MASTER
243 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
244 #endif
246 masterSwitchToCoreLoop(animatingPr);
247 flushRegisters();
248 }//MasterLoop
251 }
255 /*This has a race condition -- the coreloops are accessing their own queues
256 * at the same time that this work-stealer on a different core is trying to
257 */
258 void inline
259 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
260 VirtProcr *masterPr )
261 {
262 VirtProcr *stolenPr;
263 int32 coreIdx, i;
264 VMSQueueStruc *currQ;
266 stolenPr = NULL;
267 coreIdx = masterPr->coreAnimatedBy;
268 for( i = 0; i < NUM_CORES -1; i++ )
269 {
270 if( coreIdx >= NUM_CORES -1 )
271 { coreIdx = 0;
272 }
273 else
274 { coreIdx++;
275 }
276 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
277 if( numInVMSQ( currQ ) > 0 )
278 { stolenPr = readVMSQ (currQ );
279 break;
280 }
281 }
283 if( stolenPr != NULL )
284 { currSlot->procrAssignedToSlot = stolenPr;
285 stolenPr->schedSlot = currSlot;
286 currSlot->needsProcrAssigned = FALSE;
288 writeVMSQ( stolenPr, readyToAnimateQ );
289 }
290 }
292 /*This algorithm makes the common case fast. Make the coreloop passive,
293 * and show its progress. Make the stealer control a gate that coreloop
294 * has to pass.
295 *To avoid interference, only one stealer at a time. Use a global
296 * stealer-lock.
297 *
298 *The pattern is based on a gate -- stealer shuts the gate, then monitors
299 * to be sure any already past make it all the way out, before starting.
300 *So, have a "progress" measure just before the gate, then have two after it,
301 * one is in a "waiting room" outside the gate, the other is at the exit.
302 *Then, the stealer first shuts the gate, then checks the progress measure
303 * outside it, then looks to see if the progress measure at the exit is the
304 * same. If yes, it knows the protected area is empty 'cause no other way
305 * to get in and the last to get in also exited.
306 *If the progress measure at the exit is not the same, then the stealer goes
307 * into a loop checking both the waiting-area and the exit progress-measures
308 * until one of them shows the same as the measure outside the gate. Might
309 * as well re-read the measure outside the gate each go around, just to be
310 * sure. It is guaranteed that one of the two will eventually match the one
311 * outside the gate.
312 *
313 *Here's an informal proof of correctness:
314 *The gate can be closed at any point, and have only four cases:
315 * 1) coreloop made it past the gate-closing but not yet past the exit
316 * 2) coreloop made it past the pre-gate progress update but not yet past
317 * the gate,
318 * 3) coreloop is right before the pre-gate update
319 * 4) coreloop is past the exit and far from the pre-gate update.
320 *
321 * Covering the cases in reverse order,
322 * 4) is not a problem -- stealer will read pre-gate progress, see that it
323 * matches exit progress, and the gate is closed, so stealer can proceed.
324 * 3) stealer will read pre-gate progress just after coreloop updates it..
325 * so stealer goes into a loop until the coreloop causes wait-progress
326 * to match pre-gate progress, so then stealer can proceed
327 * 2) same as 3..
328 * 1) stealer reads pre-gate progress, sees that it's different than exit,
329 * so goes into loop until exit matches pre-gate, now it knows coreloop
330 * is not in protected and cannot get back in, so can proceed.
331 *
332 *Implementation for the stealer:
333 *
334 *First, acquire the stealer lock -- only cores with no work to do will
335 * compete to steal, so not a big performance penalty having only one --
336 * will rarely have multiple stealers in a system with plenty of work -- and
337 * in a system with little work, it doesn't matter.
338 *
339 *Note, have single-reader, single-writer pattern for all variables used to
340 * communicate between stealer and victims
341 *
342 *So, scan the queues of the core loops, until find non-empty. Each core
343 * has its own list that it scans. The list goes in order from closest to
344 * furthest core, so it steals first from close cores. Later can add
345 * taking info from the app about overlapping footprints, and scan all the
346 * others then choose work with the most footprint overlap with the contents
347 * of this core's cache.
348 *
349 *Now, have a victim want to take work from. So, shut the gate in that
350 * coreloop, by setting the "gate closed" var on its stack to TRUE.
351 *Then, read the core's pre-gate progress and compare to the core's exit
352 * progress.
353 *If same, can proceed to take work from the coreloop's queue. When done,
354 * write FALSE to gate closed var.
355 *If different, then enter a loop that reads the pre-gate progress, then
356 * compares to exit progress then to wait progress. When one of two
357 * matches, proceed. Take work from the coreloop's queue. When done,
358 * write FALSE to the gate closed var.
359 *
360 */
361 void inline
362 gateProtected_stealWorkInto( SchedSlot *currSlot,
363 VMSQueueStruc *myReadyToAnimateQ,
364 VirtProcr *masterPr )
365 {
366 VirtProcr *stolenPr;
367 int32 coreIdx, i, haveAVictim, gotLock;
368 VMSQueueStruc *victimsQ;
370 volatile GateStruc *vicGate;
371 int32 coreMightBeInProtected;
375 //see if any other cores have work available to steal
376 haveAVictim = FALSE;
377 coreIdx = masterPr->coreAnimatedBy;
378 for( i = 0; i < NUM_CORES -1; i++ )
379 {
380 if( coreIdx >= NUM_CORES -1 )
381 { coreIdx = 0;
382 }
383 else
384 { coreIdx++;
385 }
386 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
387 if( numInVMSQ( victimsQ ) > 0 )
388 { haveAVictim = TRUE;
389 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
390 break;
391 }
392 }
393 if( !haveAVictim ) return; //no work to steal, exit
395 //have a victim core, now get the stealer-lock
396 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
397 UNLOCKED, LOCKED );
398 if( !gotLock ) return; //go back to core loop, which will re-start master
401 //====== Start Gate-protection =======
402 vicGate->gateClosed = TRUE;
403 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
404 while( coreMightBeInProtected )
405 { //wait until sure
406 if( vicGate->preGateProgress == vicGate->waitProgress )
407 coreMightBeInProtected = FALSE;
408 if( vicGate->preGateProgress == vicGate->exitProgress )
409 coreMightBeInProtected = FALSE;
410 }
412 stolenPr = readVMSQ ( victimsQ );
414 vicGate->gateClosed = FALSE;
415 //======= End Gate-protection =======
418 if( stolenPr != NULL ) //victim could have been in protected and taken
419 { currSlot->procrAssignedToSlot = stolenPr;
420 stolenPr->schedSlot = currSlot;
421 currSlot->needsProcrAssigned = FALSE;
423 writeVMSQ( stolenPr, myReadyToAnimateQ );
424 }
426 //unlock the work stealing lock
427 _VMSMasterEnv->workStealingLock = UNLOCKED;
428 }