Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view MasterLoop.c @ 136:f1374d5dbb98
set currentMasterProcrID to 0 for corrent free on shutdown
| author | Merten Sach <msach@mailbox.tu-berlin.de> |
|---|---|
| date | Fri, 16 Sep 2011 20:13:33 +0200 |
| parents | a9b72021f053 |
| children | 99343ffe1918 |
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"
14 #include "scheduling.h"
15 #include "inter_VMS_requests.h"
16 #include "inter_VMS_requests_handler.h"
18 //===========================================================================
19 void inline
20 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
21 VirtProcr *masterPr);
23 void inline
24 handleInterMasterReq( InterMasterReqst *currReq, void *_semEnv,
25 VirtProcr *masterPr);
27 void inline
28 handleInterVMSCoreReq( InterVMSCoreReqst *currReq, VirtProcr *masterPr);
30 //===========================================================================
34 /*This code is animated by the virtual Master processor.
35 *
36 *Polls each sched slot exactly once, hands any requests made by a newly
37 * done slave to the "request handler" plug-in function
38 *
39 *Any slots that need a virt procr assigned are given to the "schedule"
40 * plug-in function, which tries to assign a virt procr (slave) to it.
41 *
42 *When all slots needing a processor have been given to the schedule plug-in,
43 * a fraction of the procrs successfully scheduled are put into the
44 * work queue, then a continuation of this function is put in, then the rest
45 * of the virt procrs that were successfully scheduled.
46 *
47 *The first thing the continuation does is busy-wait until the previous
48 * animation completes. This is because an (unlikely) continuation may
49 * sneak through queue before previous continuation is done putting second
50 * part of scheduled slaves in, which is the only race condition.
51 *
52 */
54 /*May 29, 2010 -- birth a Master during init so that first core loop to
55 * start running gets it and does all the stuff for a newly born --
56 * from then on, will be doing continuation, but do suspension self
57 * directly at end of master loop
58 *So VMS__init just births the master virtual processor same way it births
59 * all the others -- then does any extra setup needed and puts it into the
60 * work queue.
61 *However means have to make masterEnv a global static volatile.
62 *
63 *
64 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
65 * avoids the suspected bug in the system stack that causes bizarre faults
66 * at random places in the system code.
67 *
68 *So, this function is coupled to each of the MasterVPs, -- meaning this
69 * function can't rely on a particular stack and frame -- each MasterVP that
70 * animates this function has a different stack.
71 *
72 *At this point, the masterLoop does not write itself into the queue anymore,
73 * instead, the coreLoop acquires the masterLock when it has nothing to
74 * animate, and then animates its own masterLoop. However, still try to put
75 * several AppVPs into the queue to amortize the startup cost of switching
76 * to the MasterVP. Note, don't have to worry about latency of requests much
77 * because most requests generate work for same core -- only latency issue
78 * is case when other cores starved and one core's requests generate work
79 * for them -- so keep max in queue to 3 or 4..
80 */
81 void masterLoop( void *initData, VirtProcr *animatingPr )
82 {
83 int32 slotIdx, numSlotsFilled;
84 VirtProcr *schedVirtPr;
85 SchedSlot *currSlot, **schedSlots;
86 MasterEnv *masterEnv;
87 VMSQueueStruc *readyToAnimateQ;
89 SlaveScheduler slaveScheduler;
90 RequestHandler requestHandler;
91 void *semanticEnv;
93 int32 thisCoresIdx;
94 VirtProcr *masterPr;
95 volatile VirtProcr *volatileMasterPr;
97 volatileMasterPr = animatingPr;
98 masterPr = (VirtProcr*)volatileMasterPr; //used to force re-define after jmp
99 masterEnv = (MasterEnv*)_VMSMasterEnv;
101 //First animation of each MasterVP will in turn animate this part
102 // of setup code.. (VP creator sets up the stack as if this function
103 // was called normally, but actually get here by jmp)
105 //Sept 2011
106 //Old code jumped directly to this point, but doesn't work on x64
107 // So, just make this an endless loop, and do assembly function at end
108 // that saves its own return addr, then jumps to core_loop.
109 while(1)
110 {
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 is 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 //GCC may optimize so doesn't always re-define from frame-storage
121 thisCoresIdx = masterPr->coreAnimatedBy;
122 masterEnv->currentMasterProcrID = thisCoresIdx;
123 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
124 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
126 requestHandler = masterEnv->requestHandler;
127 slaveScheduler = masterEnv->slaveScheduler;
128 semanticEnv = masterEnv->semanticEnv;
130 //First, check for requests from other MasterVPs, and handle them
131 InterMasterReqst* currReqst = masterEnv->interMasterRequestsFor[thisCoresIdx];
132 while(currReqst)
133 {
134 handleInterMasterReq( currReqst, semanticEnv, masterPr );
135 currReqst = currReqst->nextReqst;
136 }
137 masterEnv->interMasterRequestsFor[thisCoresIdx] = NULL;
139 //Second, check for own request that were handled for other MasterVPs
140 currReqst = masterEnv->interMasterRequestsSentBy[thisCoresIdx];
141 while(currReqst && currReqst->obsolete)
142 {
143 InterMasterReqst *nextReqst = currReqst->nextSentReqst;
144 VMS__free(currReqst);
145 currReqst = nextReqst;
146 }
147 masterEnv->interMasterRequestsSentBy[thisCoresIdx] = currReqst;
149 //Now, take care of the SlaveVPs
150 //Go through the slots -- if Slave there newly suspended, handle its request
151 // then, either way, ask assigner to fill each slot
152 numSlotsFilled = 0;
153 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
154 {
155 currSlot = schedSlots[ slotIdx ];
157 if( currSlot->workIsDone )
158 {
159 currSlot->workIsDone = FALSE;
160 currSlot->needsProcrAssigned = TRUE;
162 //process requests from slave to master
163 //====================== MEASUREMENT STUFF ===================
164 #ifdef MEAS__TIME_PLUGIN
165 int32 startStamp1, endStamp1;
166 saveLowTimeStampCountInto( startStamp1 );
167 #endif
168 //============================================================
169 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
170 //====================== MEASUREMENT STUFF ===================
171 #ifdef MEAS__TIME_PLUGIN
172 saveLowTimeStampCountInto( endStamp1 );
173 addIntervalToHist( startStamp1, endStamp1,
174 _VMSMasterEnv->reqHdlrLowTimeHist );
175 addIntervalToHist( startStamp1, endStamp1,
176 _VMSMasterEnv->reqHdlrHighTimeHist );
177 #endif
178 //============================================================
179 }
180 if( currSlot->needsProcrAssigned )
181 { //give slot a new virt procr
182 schedVirtPr =
183 (*slaveScheduler)( semanticEnv, thisCoresIdx );
185 if( schedVirtPr != NULL )
186 { currSlot->procrAssignedToSlot = schedVirtPr;
187 schedVirtPr->schedSlot = currSlot;
188 schedVirtPr->coreAnimatedBy = thisCoresIdx;
189 currSlot->needsProcrAssigned = FALSE;
190 numSlotsFilled += 1;
192 writeVMSQ( schedVirtPr, readyToAnimateQ );
193 }
194 }
195 }
198 #ifdef USE_WORK_STEALING
199 //If no slots filled, means no more work, look for work to steal.
200 if( numSlotsFilled == 0 )
201 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
202 }
203 #endif
206 #ifdef MEAS__TIME_MASTER
207 saveLowTimeStampCountInto( masterPr->endMasterTSCLow );
208 #endif
210 masterSwitchToCoreLoop(animatingPr);
211 flushRegisters();
212 }//while(1) MasterLoop
213 }
215 /*This is for inter-master communication. Either the master itself or
216 * the plugin sends one of these requests. Some are handled here, by the
217 * master_loop, others are handed off to the plugin.
218 */
219 void inline
220 handleInterMasterReq( InterMasterReqst *currReq, void *_semEnv,
221 VirtProcr *masterPr )
222 {
224 switch( currReq->reqType )
225 {
226 case destVMSCore:
227 handleInterVMSCoreReq( (InterVMSCoreReqst *)currReq, masterPr);
228 break;
229 case destPlugin:
230 _VMSMasterEnv->interPluginReqHdlr( ((InterPluginReqst *)currReq)->pluginReq,
231 _semEnv );
232 break;
233 default:
234 break;
235 }
236 }
238 void inline
239 handleInterVMSCoreReq( InterVMSCoreReqst *currReq, VirtProcr *masterPr )
240 {
241 switch( currReq->reqType )
242 {
243 case transfer_free_ptr:
244 handleTransferFree( currReq, masterPr );
245 currReq->obsolete = 1; //now the sender can free the structure
246 break;
247 default:
248 break;
249 }
250 }
252 /*Work Stealing Alg -- racy one
253 *This algorithm has a race condition -- the coreloops are accessing their
254 * own queues at the same time that this work-stealer on a different core
255 * is trying to.
256 *The second stealing alg, below, protects against this.
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 /*Work Stealing alg -- protected one
293 *This algorithm makes the common case fast. Make the coreloop passive,
294 * and show its progress. Make the stealer control a gate that coreloop
295 * has to pass.
296 *To avoid interference, only one stealer at a time. Use a global
297 * stealer-lock.
298 *
299 *The pattern is based on a gate -- stealer shuts the gate, then monitors
300 * to be sure any already past make it all the way out, before starting.
301 *So, have a "progress" measure just before the gate, then have two after it,
302 * one is in a "waiting room" outside the gate, the other is at the exit.
303 *Then, the stealer first shuts the gate, then checks the progress measure
304 * outside it, then looks to see if the progress measure at the exit is the
305 * same. If yes, it knows the protected area is empty 'cause no other way
306 * to get in and the last to get in also exited.
307 *If the progress measure at the exit is not the same, then the stealer goes
308 * into a loop checking both the waiting-area and the exit progress-measures
309 * until one of them shows the same as the measure outside the gate. Might
310 * as well re-read the measure outside the gate each go around, just to be
311 * sure. It is guaranteed that one of the two will eventually match the one
312 * outside the gate.
313 *
314 *Here's an informal proof of correctness:
315 *The gate can be closed at any point, and have only four cases:
316 * 1) coreloop made it past the gate-closing but not yet past the exit
317 * 2) coreloop made it past the pre-gate progress update but not yet past
318 * the gate,
319 * 3) coreloop is right before the pre-gate update
320 * 4) coreloop is past the exit and far from the pre-gate update.
321 *
322 * Covering the cases in reverse order,
323 * 4) is not a problem -- stealer will read pre-gate progress, see that it
324 * matches exit progress, and the gate is closed, so stealer can proceed.
325 * 3) stealer will read pre-gate progress just after coreloop updates it..
326 * so stealer goes into a loop until the coreloop causes wait-progress
327 * to match pre-gate progress, so then stealer can proceed
328 * 2) same as 3..
329 * 1) stealer reads pre-gate progress, sees that it's different than exit,
330 * so goes into loop until exit matches pre-gate, now it knows coreloop
331 * is not in protected and cannot get back in, so can proceed.
332 *
333 *Implementation for the stealer:
334 *
335 *First, acquire the stealer lock -- only cores with no work to do will
336 * compete to steal, so not a big performance penalty having only one --
337 * will rarely have multiple stealers in a system with plenty of work -- and
338 * in a system with little work, it doesn't matter.
339 *
340 *Note, have single-reader, single-writer pattern for all variables used to
341 * communicate between stealer and victims
342 *
343 *So, scan the queues of the core loops, until find non-empty. Each core
344 * has its own list that it scans. The list goes in order from closest to
345 * furthest core, so it steals first from close cores. Later can add
346 * taking info from the app about overlapping footprints, and scan all the
347 * others then choose work with the most footprint overlap with the contents
348 * of this core's cache.
349 *
350 *Now, have a victim want to take work from. So, shut the gate in that
351 * coreloop, by setting the "gate closed" var on its stack to TRUE.
352 *Then, read the core's pre-gate progress and compare to the core's exit
353 * progress.
354 *If same, can proceed to take work from the coreloop's queue. When done,
355 * write FALSE to gate closed var.
356 *If different, then enter a loop that reads the pre-gate progress, then
357 * compares to exit progress then to wait progress. When one of two
358 * matches, proceed. Take work from the coreloop's queue. When done,
359 * write FALSE to the gate closed var.
360 *
361 */
362 void inline
363 gateProtected_stealWorkInto( SchedSlot *currSlot,
364 VMSQueueStruc *myReadyToAnimateQ,
365 VirtProcr *masterPr )
366 {
367 VirtProcr *stolenPr;
368 int32 coreIdx, i, haveAVictim, gotLock;
369 VMSQueueStruc *victimsQ;
371 volatile GateStruc *vicGate;
372 int32 coreMightBeInProtected;
376 //see if any other cores have work available to steal
377 haveAVictim = FALSE;
378 coreIdx = masterPr->coreAnimatedBy;
379 for( i = 0; i < NUM_CORES -1; i++ )
380 {
381 if( coreIdx >= NUM_CORES -1 )
382 { coreIdx = 0;
383 }
384 else
385 { coreIdx++;
386 }
387 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
388 if( numInVMSQ( victimsQ ) > 0 )
389 { haveAVictim = TRUE;
390 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
391 break;
392 }
393 }
394 if( !haveAVictim ) return; //no work to steal, exit
396 //have a victim core, now get the stealer-lock
397 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
398 UNLOCKED, LOCKED );
399 if( !gotLock ) return; //go back to core loop, which will re-start master
402 //====== Start Gate-protection =======
403 vicGate->gateClosed = TRUE;
404 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
405 while( coreMightBeInProtected )
406 { //wait until sure
407 if( vicGate->preGateProgress == vicGate->waitProgress )
408 coreMightBeInProtected = FALSE;
409 if( vicGate->preGateProgress == vicGate->exitProgress )
410 coreMightBeInProtected = FALSE;
411 }
413 stolenPr = readVMSQ ( victimsQ );
415 vicGate->gateClosed = FALSE;
416 //======= End Gate-protection =======
419 if( stolenPr != NULL ) //victim could have been in protected and took it
420 { currSlot->procrAssignedToSlot = stolenPr;
421 stolenPr->schedSlot = currSlot;
422 currSlot->needsProcrAssigned = FALSE;
424 writeVMSQ( stolenPr, myReadyToAnimateQ );
425 }
427 //unlock the work stealing lock
428 _VMSMasterEnv->workStealingLock = UNLOCKED;
429 }