Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view MasterLoop.c @ 187:fe5ad5726e36
counters working ...sort of
| author | Nina Engelhardt <nengel@mailbox.tu-berlin.de> |
|---|---|
| date | Fri, 03 Feb 2012 17:32:48 +0100 |
| parents | 69eb54ce9c4b |
| children | 20358f56e498 |
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 //First animation of each MasterVP will in turn animate this part
94 // of setup code.. (VP creator sets up the stack as if this function
95 // was called normally, but actually get here by jmp)
96 //So, setup values about stack ptr, jmp pt and all that
97 //masterPr->nextInstrPt = &&masterLoopStartPt;
100 //Note, got rid of writing the stack and frame ptr up here, because
101 // only one
102 // core can ever animate a given MasterVP, so don't need to communicate
103 // new frame and stack ptr to the MasterVP storage before a second
104 // version of that MasterVP can get animated on a different core.
105 //Also got rid of the busy-wait.
108 //masterLoopStartPt:
109 while(1){
111 //============================= MEASUREMENT STUFF ========================
112 #ifdef MEAS__TIME_MASTER
113 //Total Master time includes one coreloop time -- just assume the core
114 // loop time is same for Master as for AppVPs, even though it may be
115 // smaller due to higher predictability of the fixed jmp.
116 saveLowTimeStampCountInto( masterPr->startMasterTSCLow );
117 #endif
119 //========================================================================
121 masterEnv = (MasterEnv*)_VMSMasterEnv;
123 //GCC may optimize so doesn't always re-define from frame-storage
124 masterPr = (VirtProcr*)volatileMasterPr; //just to make sure after jmp
125 thisCoresIdx = masterPr->coreAnimatedBy;
126 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
127 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
129 requestHandler = masterEnv->requestHandler;
130 slaveScheduler = masterEnv->slaveScheduler;
131 semanticEnv = masterEnv->semanticEnv;
133 #ifdef MEAS__PERF_COUNTERS
134 CounterHandler counterHandler = masterEnv->counterHandler;
135 #endif
137 //Poll each slot's Done flag
138 numSlotsFilled = 0;
139 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
140 {
141 currSlot = schedSlots[ slotIdx ];
143 if( currSlot->workIsDone )
144 {
145 currSlot->workIsDone = FALSE;
146 currSlot->needsProcrAssigned = TRUE;
148 //process requests from slave to master
149 //====================== MEASUREMENT STUFF ===================
150 #ifdef MEAS__TIME_PLUGIN
151 int32 startStamp1, endStamp1;
152 saveLowTimeStampCountInto( startStamp1 );
153 #endif
154 #ifdef MEAS__PERF_COUNTERS
155 VirtProcr p_sav = *(currSlot->procrAssignedToSlot);
156 uint64 cycles, instrs;
157 saveCyclesAndInstrs(thisCoresIdx,cycles, instrs);
158 (*counterHandler)(MasterLoop_beforeReqHdlr,&p_sav,cycles,instrs);
159 #endif
160 //============================================================
161 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
162 //====================== MEASUREMENT STUFF ===================
163 #ifdef MEAS__TIME_PLUGIN
164 saveLowTimeStampCountInto( endStamp1 );
165 addIntervalToHist( startStamp1, endStamp1,
166 _VMSMasterEnv->reqHdlrLowTimeHist );
167 addIntervalToHist( startStamp1, endStamp1,
168 _VMSMasterEnv->reqHdlrHighTimeHist );
169 #endif
170 #ifdef MEAS__PERF_COUNTERS
171 //done with constraints check
172 uint64 cycles2,instrs2;
173 saveCyclesAndInstrs(thisCoresIdx,cycles2, instrs2);
174 (*counterHandler)(MasterLoop_afterReqHdlr,&p_sav,cycles2,instrs2);
175 #endif
176 //============================================================
177 }
178 if( currSlot->needsProcrAssigned )
179 { //give slot a new virt procr
180 #ifdef MEAS__PERF_COUNTERS
181 //start assigner
182 uint64 tmp_cycles;
183 uint64 tmp_instrs;
184 saveCyclesAndInstrs(thisCoresIdx,tmp_cycles,tmp_instrs);
185 //FIXME WTF AM I DOING WHY DOES THIS EVEN WORK
186 //(*counterHandler)(MasterLoop_beforeNextAssign,schedVirtPr,tmp_cycles,tmp_instrs);
187 #endif
188 schedVirtPr =
189 (*slaveScheduler)( semanticEnv, thisCoresIdx, slotIdx );
191 if( schedVirtPr != NULL )
192 { currSlot->procrAssignedToSlot = schedVirtPr;
193 schedVirtPr->schedSlot = currSlot;
194 currSlot->needsProcrAssigned = FALSE;
195 numSlotsFilled += 1;
197 writeVMSQ( schedVirtPr, readyToAnimateQ );
199 #ifdef MEAS__PERF_COUNTERS
200 uint64 cycles;
201 uint64 instrs;
202 saveCyclesAndInstrs(thisCoresIdx,cycles,instrs);
203 (*counterHandler)(MasterLoop_beforeAssign,schedVirtPr,tmp_cycles,tmp_instrs);
204 (*counterHandler)(MasterLoop_afterAssign,schedVirtPr,cycles,instrs);
205 #endif
206 }
207 }
209 }
212 #ifdef USE_WORK_STEALING
213 //If no slots filled, means no more work, look for work to steal.
214 if( numSlotsFilled == 0 )
215 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
216 }
217 #endif
220 #ifdef MEAS__TIME_MASTER
221 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
222 #endif
224 masterSwitchToCoreLoop(animatingPr);
225 flushRegisters();
226 }//MasterLoop
229 }
233 /*This has a race condition -- the coreloops are accessing their own queues
234 * at the same time that this work-stealer on a different core is trying to
235 */
236 void inline
237 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
238 VirtProcr *masterPr )
239 {
240 VirtProcr *stolenPr;
241 int32 coreIdx, i;
242 VMSQueueStruc *currQ;
244 stolenPr = NULL;
245 coreIdx = masterPr->coreAnimatedBy;
246 for( i = 0; i < NUM_CORES -1; i++ )
247 {
248 if( coreIdx >= NUM_CORES -1 )
249 { coreIdx = 0;
250 }
251 else
252 { coreIdx++;
253 }
254 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
255 if( numInVMSQ( currQ ) > 0 )
256 { stolenPr = readVMSQ (currQ );
257 break;
258 }
259 }
261 if( stolenPr != NULL )
262 { currSlot->procrAssignedToSlot = stolenPr;
263 stolenPr->schedSlot = currSlot;
264 currSlot->needsProcrAssigned = FALSE;
266 writeVMSQ( stolenPr, readyToAnimateQ );
267 }
268 }
270 /*This algorithm makes the common case fast. Make the coreloop passive,
271 * and show its progress. Make the stealer control a gate that coreloop
272 * has to pass.
273 *To avoid interference, only one stealer at a time. Use a global
274 * stealer-lock.
275 *
276 *The pattern is based on a gate -- stealer shuts the gate, then monitors
277 * to be sure any already past make it all the way out, before starting.
278 *So, have a "progress" measure just before the gate, then have two after it,
279 * one is in a "waiting room" outside the gate, the other is at the exit.
280 *Then, the stealer first shuts the gate, then checks the progress measure
281 * outside it, then looks to see if the progress measure at the exit is the
282 * same. If yes, it knows the protected area is empty 'cause no other way
283 * to get in and the last to get in also exited.
284 *If the progress measure at the exit is not the same, then the stealer goes
285 * into a loop checking both the waiting-area and the exit progress-measures
286 * until one of them shows the same as the measure outside the gate. Might
287 * as well re-read the measure outside the gate each go around, just to be
288 * sure. It is guaranteed that one of the two will eventually match the one
289 * outside the gate.
290 *
291 *Here's an informal proof of correctness:
292 *The gate can be closed at any point, and have only four cases:
293 * 1) coreloop made it past the gate-closing but not yet past the exit
294 * 2) coreloop made it past the pre-gate progress update but not yet past
295 * the gate,
296 * 3) coreloop is right before the pre-gate update
297 * 4) coreloop is past the exit and far from the pre-gate update.
298 *
299 * Covering the cases in reverse order,
300 * 4) is not a problem -- stealer will read pre-gate progress, see that it
301 * matches exit progress, and the gate is closed, so stealer can proceed.
302 * 3) stealer will read pre-gate progress just after coreloop updates it..
303 * so stealer goes into a loop until the coreloop causes wait-progress
304 * to match pre-gate progress, so then stealer can proceed
305 * 2) same as 3..
306 * 1) stealer reads pre-gate progress, sees that it's different than exit,
307 * so goes into loop until exit matches pre-gate, now it knows coreloop
308 * is not in protected and cannot get back in, so can proceed.
309 *
310 *Implementation for the stealer:
311 *
312 *First, acquire the stealer lock -- only cores with no work to do will
313 * compete to steal, so not a big performance penalty having only one --
314 * will rarely have multiple stealers in a system with plenty of work -- and
315 * in a system with little work, it doesn't matter.
316 *
317 *Note, have single-reader, single-writer pattern for all variables used to
318 * communicate between stealer and victims
319 *
320 *So, scan the queues of the core loops, until find non-empty. Each core
321 * has its own list that it scans. The list goes in order from closest to
322 * furthest core, so it steals first from close cores. Later can add
323 * taking info from the app about overlapping footprints, and scan all the
324 * others then choose work with the most footprint overlap with the contents
325 * of this core's cache.
326 *
327 *Now, have a victim want to take work from. So, shut the gate in that
328 * coreloop, by setting the "gate closed" var on its stack to TRUE.
329 *Then, read the core's pre-gate progress and compare to the core's exit
330 * progress.
331 *If same, can proceed to take work from the coreloop's queue. When done,
332 * write FALSE to gate closed var.
333 *If different, then enter a loop that reads the pre-gate progress, then
334 * compares to exit progress then to wait progress. When one of two
335 * matches, proceed. Take work from the coreloop's queue. When done,
336 * write FALSE to the gate closed var.
337 *
338 */
339 void inline
340 gateProtected_stealWorkInto( SchedSlot *currSlot,
341 VMSQueueStruc *myReadyToAnimateQ,
342 VirtProcr *masterPr )
343 {
344 VirtProcr *stolenPr;
345 int32 coreIdx, i, haveAVictim, gotLock;
346 VMSQueueStruc *victimsQ;
348 volatile GateStruc *vicGate;
349 int32 coreMightBeInProtected;
353 //see if any other cores have work available to steal
354 haveAVictim = FALSE;
355 coreIdx = masterPr->coreAnimatedBy;
356 for( i = 0; i < NUM_CORES -1; i++ )
357 {
358 if( coreIdx >= NUM_CORES -1 )
359 { coreIdx = 0;
360 }
361 else
362 { coreIdx++;
363 }
364 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
365 if( numInVMSQ( victimsQ ) > 0 )
366 { haveAVictim = TRUE;
367 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
368 break;
369 }
370 }
371 if( !haveAVictim ) return; //no work to steal, exit
373 //have a victim core, now get the stealer-lock
374 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
375 UNLOCKED, LOCKED );
376 if( !gotLock ) return; //go back to core loop, which will re-start master
379 //====== Start Gate-protection =======
380 vicGate->gateClosed = TRUE;
381 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
382 while( coreMightBeInProtected )
383 { //wait until sure
384 if( vicGate->preGateProgress == vicGate->waitProgress )
385 coreMightBeInProtected = FALSE;
386 if( vicGate->preGateProgress == vicGate->exitProgress )
387 coreMightBeInProtected = FALSE;
388 }
390 stolenPr = readVMSQ ( victimsQ );
392 vicGate->gateClosed = FALSE;
393 //======= End Gate-protection =======
396 if( stolenPr != NULL ) //victim could have been in protected and taken
397 { currSlot->procrAssignedToSlot = stolenPr;
398 stolenPr->schedSlot = currSlot;
399 currSlot->needsProcrAssigned = FALSE;
401 writeVMSQ( stolenPr, myReadyToAnimateQ );
402 }
404 //unlock the work stealing lock
405 _VMSMasterEnv->workStealingLock = UNLOCKED;
406 }