Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view MasterLoop.c @ 61:984f7d78bfdf
Merge See what happens -- merged test stuff into Nov 8 VMS version
| author | SeanHalle |
|---|---|
| date | Thu, 11 Nov 2010 06:19:51 -0800 |
| parents | 054006c26b92 3bac84e4e56e |
| children |
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"
15 //===========================================================================
16 void inline
17 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
18 VirtProcr *masterPr );
20 //===========================================================================
24 /*This code is animated by the virtual Master processor.
25 *
26 *Polls each sched slot exactly once, hands any requests made by a newly
27 * done slave to the "request handler" plug-in function
28 *
29 *Any slots that need a virt procr assigned are given to the "schedule"
30 * plug-in function, which tries to assign a virt procr (slave) to it.
31 *
32 *When all slots needing a processor have been given to the schedule plug-in,
33 * a fraction of the procrs successfully scheduled are put into the
34 * work queue, then a continuation of this function is put in, then the rest
35 * of the virt procrs that were successfully scheduled.
36 *
37 *The first thing the continuation does is busy-wait until the previous
38 * animation completes. This is because an (unlikely) continuation may
39 * sneak through queue before previous continuation is done putting second
40 * part of scheduled slaves in, which is the only race condition.
41 *
42 */
44 /*May 29, 2010 -- birth a Master during init so that first core loop to
45 * start running gets it and does all the stuff for a newly born --
46 * from then on, will be doing continuation, but do suspension self
47 * directly at end of master loop
48 *So VMS__init just births the master virtual processor same way it births
49 * all the others -- then does any extra setup needed and puts it into the
50 * work queue.
51 *However means have to make masterEnv a global static volatile the same way
52 * did with readyToAnimateQ in core loop. -- for performance, put the
53 * jump to the core loop directly in here, and have it directly jump back.
54 *
55 *
56 *Aug 18, 2010 -- Going to a separate MasterVP for each core, to see if this
57 * avoids the suspected bug in the system stack that causes bizarre faults
58 * at random places in the system code.
59 *
60 *So, this function is coupled to each of the MasterVPs, -- meaning this
61 * function can't rely on a particular stack and frame -- each MasterVP that
62 * animates this function has a different one.
63 *
64 *At this point, the masterLoop does not write itself into the queue anymore,
65 * instead, the coreLoop acquires the masterLock when it has nothing to
66 * animate, and then animates its own masterLoop. However, still try to put
67 * several AppVPs into the queue to amortize the startup cost of switching
68 * to the MasterVP. Note, don't have to worry about latency of requests much
69 * because most requests generate work for same core -- only latency issue
70 * is case when other cores starved and one core's requests generate work
71 * for them -- so keep max in queue to 3 or 4..
72 */
73 void masterLoop( void *initData, VirtProcr *animatingPr )
74 {
75 int32 slotIdx, numSlotsFilled;
76 VirtProcr *schedVirtPr;
77 SchedSlot *currSlot, **schedSlots;
78 MasterEnv *masterEnv;
79 VMSQueueStruc *readyToAnimateQ;
81 SlaveScheduler slaveScheduler;
82 RequestHandler requestHandler;
83 void *semanticEnv;
85 int32 thisCoresIdx;
86 VirtProcr *masterPr;
87 volatile VirtProcr *volatileMasterPr;
89 volatileMasterPr = animatingPr;
90 masterPr = volatileMasterPr; //used to force re-define after jmp
92 //First animation of each MasterVP will in turn animate this part
93 // of setup code.. (VP creator sets up the stack as if this function
94 // was called normally, but actually get here by jmp)
95 //So, setup values about stack ptr, jmp pt and all that
96 masterPr->nextInstrPt = &&masterLoopStartPt;
99 //Note, got rid of writing the stack and frame ptr up here, because
100 // only one
101 // core can ever animate a given MasterVP, so don't need to communicate
102 // new frame and stack ptr to the MasterVP storage before a second
103 // version of that MasterVP can get animated on a different core.
104 //Also got rid of the busy-wait.
107 masterLoopStartPt:
108 //============================= MEASUREMENT STUFF ========================
109 #ifdef MEAS__TIME_MASTER
110 int startStamp, endStamp;
111 //Total Master time includes one coreloop time -- just assume the core
112 // loop time is same for Master as for AppVPs, even though it will be
113 // smaller due to high predictability of the fixed jmp.
114 saveLowTimeStampCountInto( startStamp );
115 #endif
116 //========================================================================
118 masterEnv = _VMSMasterEnv;
120 //GCC may optimize so doesn't always re-define from frame-storage
121 masterPr = volatileMasterPr; //just to make sure after jmp
122 thisCoresIdx = masterPr->coreAnimatedBy;
123 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
124 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
126 requestHandler = masterEnv->requestHandler;
127 slaveScheduler = masterEnv->slaveScheduler;
128 semanticEnv = masterEnv->semanticEnv;
131 //Poll each slot's Done flag
132 numSlotsFilled = 0;
133 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
134 {
135 currSlot = schedSlots[ slotIdx ];
137 if( currSlot->workIsDone )
138 {
139 currSlot->workIsDone = FALSE;
140 currSlot->needsProcrAssigned = TRUE;
142 //process requests from slave to master
143 (*requestHandler)( currSlot->procrAssignedToSlot, semanticEnv );
144 }
145 if( currSlot->needsProcrAssigned )
146 { //give slot a new virt procr
147 schedVirtPr =
148 (*slaveScheduler)( semanticEnv, thisCoresIdx );
150 if( schedVirtPr != NULL )
151 { currSlot->procrAssignedToSlot = schedVirtPr;
152 schedVirtPr->schedSlot = currSlot;
153 currSlot->needsProcrAssigned = FALSE;
154 numSlotsFilled += 1;
156 writeVMSQ( schedVirtPr, readyToAnimateQ );
157 }
158 }
159 }
162 #ifdef USE_WORK_STEALING
163 //If no slots filled, means no more work, look for work to steal.
164 if( numSlotsFilled == 0 )
165 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterPr );
166 }
167 #endif
170 //============================= MEASUREMENT STUFF ========================
171 #ifdef MEAS__TIME_MASTER
172 saveLowTimeStampCountInto( endStamp );
173 addIntervalToHist(startStamp,endStamp,_VMSMasterEnv->stats->masterTimeHist);
174 #endif
175 //========================================================================
178 masterSwitchToCoreLoop( masterPr )
179 }
183 /*This has a race condition -- the coreloops are accessing their own queues
184 * at the same time that this work-stealer on a different core is trying to
185 */
186 void inline
187 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
188 VirtProcr *masterPr )
189 {
190 VirtProcr *stolenPr;
191 int32 coreIdx, i;
192 VMSQueueStruc *currQ;
194 stolenPr = NULL;
195 coreIdx = masterPr->coreAnimatedBy;
196 for( i = 0; i < NUM_CORES -1; i++ )
197 {
198 if( coreIdx >= NUM_CORES -1 )
199 { coreIdx = 0;
200 }
201 else
202 { coreIdx++;
203 }
204 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
205 if( numInVMSQ( currQ ) > 0 )
206 { stolenPr = readVMSQ (currQ );
207 break;
208 }
209 }
211 if( stolenPr != NULL )
212 { currSlot->procrAssignedToSlot = stolenPr;
213 stolenPr->schedSlot = currSlot;
214 currSlot->needsProcrAssigned = FALSE;
216 writeVMSQ( stolenPr, readyToAnimateQ );
217 }
218 }
220 /*This algorithm makes the common case fast. Make the coreloop passive,
221 * and show its progress. Make the stealer control a gate that coreloop
222 * has to pass.
223 *To avoid interference, only one stealer at a time. Use a global
224 * stealer-lock.
225 *
226 *The pattern is based on a gate -- stealer shuts the gate, then monitors
227 * to be sure any already past make it all the way out, before starting.
228 *So, have a "progress" measure just before the gate, then have two after it,
229 * one is in a "waiting room" outside the gate, the other is at the exit.
230 *Then, the stealer first shuts the gate, then checks the progress measure
231 * outside it, then looks to see if the progress measure at the exit is the
232 * same. If yes, it knows the protected area is empty 'cause no other way
233 * to get in and the last to get in also exited.
234 *If the progress measure at the exit is not the same, then the stealer goes
235 * into a loop checking both the waiting-area and the exit progress-measures
236 * until one of them shows the same as the measure outside the gate. Might
237 * as well re-read the measure outside the gate each go around, just to be
238 * sure. It is guaranteed that one of the two will eventually match the one
239 * outside the gate.
240 *
241 *Here's an informal proof of correctness:
242 *The gate can be closed at any point, and have only four cases:
243 * 1) coreloop made it past the gate-closing but not yet past the exit
244 * 2) coreloop made it past the pre-gate progress update but not yet past
245 * the gate,
246 * 3) coreloop is right before the pre-gate update
247 * 4) coreloop is past the exit and far from the pre-gate update.
248 *
249 * Covering the cases in reverse order,
250 * 4) is not a problem -- stealer will read pre-gate progress, see that it
251 * matches exit progress, and the gate is closed, so stealer can proceed.
252 * 3) stealer will read pre-gate progress just after coreloop updates it..
253 * so stealer goes into a loop until the coreloop causes wait-progress
254 * to match pre-gate progress, so then stealer can proceed
255 * 2) same as 3..
256 * 1) stealer reads pre-gate progress, sees that it's different than exit,
257 * so goes into loop until exit matches pre-gate, now it knows coreloop
258 * is not in protected and cannot get back in, so can proceed.
259 *
260 *Implementation for the stealer:
261 *
262 *First, acquire the stealer lock -- only cores with no work to do will
263 * compete to steal, so not a big performance penalty having only one --
264 * will rarely have multiple stealers in a system with plenty of work -- and
265 * in a system with little work, it doesn't matter.
266 *
267 *Note, have single-reader, single-writer pattern for all variables used to
268 * communicate between stealer and victims
269 *
270 *So, scan the queues of the core loops, until find non-empty. Each core
271 * has its own list that it scans. The list goes in order from closest to
272 * furthest core, so it steals first from close cores. Later can add
273 * taking info from the app about overlapping footprints, and scan all the
274 * others then choose work with the most footprint overlap with the contents
275 * of this core's cache.
276 *
277 *Now, have a victim want to take work from. So, shut the gate in that
278 * coreloop, by setting the "gate closed" var on its stack to TRUE.
279 *Then, read the core's pre-gate progress and compare to the core's exit
280 * progress.
281 *If same, can proceed to take work from the coreloop's queue. When done,
282 * write FALSE to gate closed var.
283 *If different, then enter a loop that reads the pre-gate progress, then
284 * compares to exit progress then to wait progress. When one of two
285 * matches, proceed. Take work from the coreloop's queue. When done,
286 * write FALSE to the gate closed var.
287 *
288 */
289 void inline
290 gateProtected_stealWorkInto( SchedSlot *currSlot,
291 VMSQueueStruc *myReadyToAnimateQ,
292 VirtProcr *masterPr )
293 {
294 VirtProcr *stolenPr;
295 int32 coreIdx, i, haveAVictim, gotLock;
296 VMSQueueStruc *victimsQ;
298 volatile GateStruc *vicGate;
299 int32 coreMightBeInProtected;
303 //see if any other cores have work available to steal
304 haveAVictim = FALSE;
305 coreIdx = masterPr->coreAnimatedBy;
306 for( i = 0; i < NUM_CORES -1; i++ )
307 {
308 if( coreIdx >= NUM_CORES -1 )
309 { coreIdx = 0;
310 }
311 else
312 { coreIdx++;
313 }
314 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
315 if( numInVMSQ( victimsQ ) > 0 )
316 { haveAVictim = TRUE;
317 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
318 break;
319 }
320 }
321 if( !haveAVictim ) return; //no work to steal, exit
323 //have a victim core, now get the stealer-lock
324 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
325 UNLOCKED, LOCKED );
326 if( !gotLock ) return; //go back to core loop, which will re-start master
329 //====== Start Gate-protection =======
330 vicGate->gateClosed = TRUE;
331 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
332 while( coreMightBeInProtected )
333 { //wait until sure
334 if( vicGate->preGateProgress == vicGate->waitProgress )
335 coreMightBeInProtected = FALSE;
336 if( vicGate->preGateProgress == vicGate->exitProgress )
337 coreMightBeInProtected = FALSE;
338 }
340 stolenPr = readVMSQ ( victimsQ );
342 vicGate->gateClosed = FALSE;
343 //======= End Gate-protection =======
346 if( stolenPr != NULL ) //victim could have been in protected and taken
347 { currSlot->procrAssignedToSlot = stolenPr;
348 stolenPr->schedSlot = currSlot;
349 currSlot->needsProcrAssigned = FALSE;
351 writeVMSQ( stolenPr, myReadyToAnimateQ );
352 }
354 //unlock the work stealing lock
355 _VMSMasterEnv->workStealingLock = UNLOCKED;
356 }