Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view MasterLoop.c @ 185:28cc465f7eb7
new counter interface works now
| author | Nina Engelhardt <nengel@mailbox.tu-berlin.de> |
|---|---|
| date | Fri, 13 Jan 2012 18:34:31 +0100 |
| parents | 50b29548d4f0 |
| children | 69eb54ce9c4b |
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
118 //========================================================================
120 masterEnv = (MasterEnv*)_VMSMasterEnv;
122 //GCC may optimize so doesn't always re-define from frame-storage
123 masterPr = (VirtProcr*)volatileMasterPr; //just to make sure after jmp
124 thisCoresIdx = masterPr->coreAnimatedBy;
125 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
126 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
128 requestHandler = masterEnv->requestHandler;
129 slaveScheduler = masterEnv->slaveScheduler;
130 semanticEnv = masterEnv->semanticEnv;
132 #ifdef MEAS__PERF_COUNTERS
133 CounterHandler counterHandler = masterEnv->counterHandler;
134 #endif
136 //Poll each slot's Done flag
137 numSlotsFilled = 0;
138 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
139 {
140 currSlot = schedSlots[ slotIdx ];
142 if( currSlot->workIsDone )
143 {
144 currSlot->workIsDone = FALSE;
145 currSlot->needsProcrAssigned = TRUE;
147 //process requests from slave to master
148 //====================== MEASUREMENT STUFF ===================
149 #ifdef MEAS__TIME_PLUGIN
150 int32 startStamp1, endStamp1;
151 saveLowTimeStampCountInto( startStamp1 );
152 #endif
153 #ifdef MEAS__PERF_COUNTERS
154 uint64 cycles, instrs;
155 saveCyclesAndInstrs(thisCoresIdx,cycles, instrs);
156 (*counterHandler)(MasterLoop_beforeReqHdlr,currSlot->procrAssignedToSlot,cycles,instrs);
157 #endif
158 //============================================================
159 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
160 //====================== MEASUREMENT STUFF ===================
161 #ifdef MEAS__TIME_PLUGIN
162 saveLowTimeStampCountInto( endStamp1 );
163 addIntervalToHist( startStamp1, endStamp1,
164 _VMSMasterEnv->reqHdlrLowTimeHist );
165 addIntervalToHist( startStamp1, endStamp1,
166 _VMSMasterEnv->reqHdlrHighTimeHist );
167 #endif
168 #ifdef MEAS__PERF_COUNTERS
169 //done with constraints check
170 uint64 cycles2,instrs2;
171 saveCyclesAndInstrs(thisCoresIdx,cycles2, instrs2);
172 (*counterHandler)(MasterLoop_afterReqHdlr,currSlot->procrAssignedToSlot,cycles2,instrs2);
173 #endif
174 //============================================================
175 }
176 if( currSlot->needsProcrAssigned )
177 { //give slot a new virt procr
178 #ifdef MEAS__PERF_COUNTERS
179 //start assigner
180 uint64 tmp_cycles;
181 uint64 tmp_instrs;
182 saveCyclesAndInstrs(thisCoresIdx,tmp_cycles,tmp_instrs);
183 #endif
184 schedVirtPr =
185 (*slaveScheduler)( semanticEnv, thisCoresIdx, slotIdx );
187 if( schedVirtPr != NULL )
188 { currSlot->procrAssignedToSlot = schedVirtPr;
189 schedVirtPr->schedSlot = currSlot;
190 currSlot->needsProcrAssigned = FALSE;
191 numSlotsFilled += 1;
193 writeVMSQ( schedVirtPr, readyToAnimateQ );
195 #ifdef MEAS__PERF_COUNTERS
196 uint64 cycles;
197 uint64 instrs;
198 saveCyclesAndInstrs(thisCoresIdx,cycles,instrs);
199 (*counterHandler)(MasterLoop_beforeAssign,schedVirtPr,tmp_cycles,tmp_instrs);
200 (*counterHandler)(MasterLoop_afterAssign,schedVirtPr,cycles,instrs);
201 #endif
202 }
203 }
205 }
208 #ifdef USE_WORK_STEALING
209 //If no slots filled, means no more work, look for work to steal.
210 if( numSlotsFilled == 0 )
211 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
212 }
213 #endif
216 #ifdef MEAS__TIME_MASTER
217 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
218 #endif
220 masterSwitchToCoreLoop(animatingPr);
221 flushRegisters();
222 }//MasterLoop
225 }
229 /*This has a race condition -- the coreloops are accessing their own queues
230 * at the same time that this work-stealer on a different core is trying to
231 */
232 void inline
233 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
234 VirtProcr *masterPr )
235 {
236 VirtProcr *stolenPr;
237 int32 coreIdx, i;
238 VMSQueueStruc *currQ;
240 stolenPr = NULL;
241 coreIdx = masterPr->coreAnimatedBy;
242 for( i = 0; i < NUM_CORES -1; i++ )
243 {
244 if( coreIdx >= NUM_CORES -1 )
245 { coreIdx = 0;
246 }
247 else
248 { coreIdx++;
249 }
250 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
251 if( numInVMSQ( currQ ) > 0 )
252 { stolenPr = readVMSQ (currQ );
253 break;
254 }
255 }
257 if( stolenPr != NULL )
258 { currSlot->procrAssignedToSlot = stolenPr;
259 stolenPr->schedSlot = currSlot;
260 currSlot->needsProcrAssigned = FALSE;
262 writeVMSQ( stolenPr, readyToAnimateQ );
263 }
264 }
266 /*This algorithm makes the common case fast. Make the coreloop passive,
267 * and show its progress. Make the stealer control a gate that coreloop
268 * has to pass.
269 *To avoid interference, only one stealer at a time. Use a global
270 * stealer-lock.
271 *
272 *The pattern is based on a gate -- stealer shuts the gate, then monitors
273 * to be sure any already past make it all the way out, before starting.
274 *So, have a "progress" measure just before the gate, then have two after it,
275 * one is in a "waiting room" outside the gate, the other is at the exit.
276 *Then, the stealer first shuts the gate, then checks the progress measure
277 * outside it, then looks to see if the progress measure at the exit is the
278 * same. If yes, it knows the protected area is empty 'cause no other way
279 * to get in and the last to get in also exited.
280 *If the progress measure at the exit is not the same, then the stealer goes
281 * into a loop checking both the waiting-area and the exit progress-measures
282 * until one of them shows the same as the measure outside the gate. Might
283 * as well re-read the measure outside the gate each go around, just to be
284 * sure. It is guaranteed that one of the two will eventually match the one
285 * outside the gate.
286 *
287 *Here's an informal proof of correctness:
288 *The gate can be closed at any point, and have only four cases:
289 * 1) coreloop made it past the gate-closing but not yet past the exit
290 * 2) coreloop made it past the pre-gate progress update but not yet past
291 * the gate,
292 * 3) coreloop is right before the pre-gate update
293 * 4) coreloop is past the exit and far from the pre-gate update.
294 *
295 * Covering the cases in reverse order,
296 * 4) is not a problem -- stealer will read pre-gate progress, see that it
297 * matches exit progress, and the gate is closed, so stealer can proceed.
298 * 3) stealer will read pre-gate progress just after coreloop updates it..
299 * so stealer goes into a loop until the coreloop causes wait-progress
300 * to match pre-gate progress, so then stealer can proceed
301 * 2) same as 3..
302 * 1) stealer reads pre-gate progress, sees that it's different than exit,
303 * so goes into loop until exit matches pre-gate, now it knows coreloop
304 * is not in protected and cannot get back in, so can proceed.
305 *
306 *Implementation for the stealer:
307 *
308 *First, acquire the stealer lock -- only cores with no work to do will
309 * compete to steal, so not a big performance penalty having only one --
310 * will rarely have multiple stealers in a system with plenty of work -- and
311 * in a system with little work, it doesn't matter.
312 *
313 *Note, have single-reader, single-writer pattern for all variables used to
314 * communicate between stealer and victims
315 *
316 *So, scan the queues of the core loops, until find non-empty. Each core
317 * has its own list that it scans. The list goes in order from closest to
318 * furthest core, so it steals first from close cores. Later can add
319 * taking info from the app about overlapping footprints, and scan all the
320 * others then choose work with the most footprint overlap with the contents
321 * of this core's cache.
322 *
323 *Now, have a victim want to take work from. So, shut the gate in that
324 * coreloop, by setting the "gate closed" var on its stack to TRUE.
325 *Then, read the core's pre-gate progress and compare to the core's exit
326 * progress.
327 *If same, can proceed to take work from the coreloop's queue. When done,
328 * write FALSE to gate closed var.
329 *If different, then enter a loop that reads the pre-gate progress, then
330 * compares to exit progress then to wait progress. When one of two
331 * matches, proceed. Take work from the coreloop's queue. When done,
332 * write FALSE to the gate closed var.
333 *
334 */
335 void inline
336 gateProtected_stealWorkInto( SchedSlot *currSlot,
337 VMSQueueStruc *myReadyToAnimateQ,
338 VirtProcr *masterPr )
339 {
340 VirtProcr *stolenPr;
341 int32 coreIdx, i, haveAVictim, gotLock;
342 VMSQueueStruc *victimsQ;
344 volatile GateStruc *vicGate;
345 int32 coreMightBeInProtected;
349 //see if any other cores have work available to steal
350 haveAVictim = FALSE;
351 coreIdx = masterPr->coreAnimatedBy;
352 for( i = 0; i < NUM_CORES -1; i++ )
353 {
354 if( coreIdx >= NUM_CORES -1 )
355 { coreIdx = 0;
356 }
357 else
358 { coreIdx++;
359 }
360 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
361 if( numInVMSQ( victimsQ ) > 0 )
362 { haveAVictim = TRUE;
363 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
364 break;
365 }
366 }
367 if( !haveAVictim ) return; //no work to steal, exit
369 //have a victim core, now get the stealer-lock
370 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
371 UNLOCKED, LOCKED );
372 if( !gotLock ) return; //go back to core loop, which will re-start master
375 //====== Start Gate-protection =======
376 vicGate->gateClosed = TRUE;
377 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
378 while( coreMightBeInProtected )
379 { //wait until sure
380 if( vicGate->preGateProgress == vicGate->waitProgress )
381 coreMightBeInProtected = FALSE;
382 if( vicGate->preGateProgress == vicGate->exitProgress )
383 coreMightBeInProtected = FALSE;
384 }
386 stolenPr = readVMSQ ( victimsQ );
388 vicGate->gateClosed = FALSE;
389 //======= End Gate-protection =======
392 if( stolenPr != NULL ) //victim could have been in protected and taken
393 { currSlot->procrAssignedToSlot = stolenPr;
394 stolenPr->schedSlot = currSlot;
395 currSlot->needsProcrAssigned = FALSE;
397 writeVMSQ( stolenPr, myReadyToAnimateQ );
398 }
400 //unlock the work stealing lock
401 _VMSMasterEnv->workStealingLock = UNLOCKED;
402 }