Mercurial > cgi-bin > hgwebdir.cgi > VMS > VMS_Implementations > VMS_impls > VMS__MC_shared_impl
view MasterLoop.c @ 209:0c83ea8adefc
Close to compilable version of common_ancestor -- still includes HW dep stuff
| author | Some Random Person <seanhalle@yahoo.com> |
|---|---|
| date | Sun, 04 Mar 2012 14:26:35 -0800 |
| parents | eaf7e4c58c9e |
| children | a18539c0bc37 |
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 SlaveVP *masterVP );
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 Slv assigned are given to the "schedule"
30 * plug-in function, which tries to assign a Slv (slave) to it.
31 *
32 *When all slots needing a processor have been given to the schedule plug-in,
33 * a fraction of the slaves successfully scheduled are put into the
34 * work queue, then a continuation of this function is put in, then the rest
35 * of the Slvs 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_WL__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 AppSlvs 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, SlaveVP *animatingSlv )
74 {
75 int32 slotIdx, numSlotsFilled;
76 SlaveVP *schedSlaveVP;
77 SchedSlot *currSlot, **schedSlots;
78 MasterEnv *masterEnv;
79 VMSQueueStruc *readyToAnimateQ;
81 Sched_Assigner slaveAssigner;
82 RequestHandler requestHandler;
83 void *semanticEnv;
85 int32 thisCoresIdx;
86 SlaveVP *masterVP;
87 volatile SlaveVP *volatileMasterVP;
89 volatileMasterVP = animatingSlv;
90 masterVP = (SlaveVP*)volatileMasterVP; //used to force re-define after jmp
92 //First animation of each MasterVP will in turn animate this part
93 // of setup code.. (Slv 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 //masterVP->resumeInstrPtr = &&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 while(1){
110 MEAS__Capture_Pre_Master_Point
112 masterEnv = (MasterEnv*)_VMSMasterEnv;
114 //GCC may optimize so doesn't always re-define from frame-storage
115 masterVP = (SlaveVP*)volatileMasterVP; //just to make sure after jmp
116 thisCoresIdx = masterVP->coreAnimatedBy;
117 readyToAnimateQ = masterEnv->readyToAnimateQs[thisCoresIdx];
118 schedSlots = masterEnv->allSchedSlots[thisCoresIdx];
120 requestHandler = masterEnv->requestHandler;
121 slaveAssigner = masterEnv->slaveAssigner;
122 semanticEnv = masterEnv->semanticEnv;
125 //Poll each slot's Done flag
126 numSlotsFilled = 0;
127 for( slotIdx = 0; slotIdx < NUM_SCHED_SLOTS; slotIdx++)
128 {
129 currSlot = schedSlots[ slotIdx ];
131 if( currSlot->workIsDone )
132 {
133 currSlot->workIsDone = FALSE;
134 currSlot->needsSlaveAssigned = TRUE;
136 //process requests from slave to master
137 //====================== MEASUREMENT STUFF ===================
138 #ifdef MEAS__TURN_ON_PLUGIN_MEAS
139 int32 startStamp1, endStamp1;
140 saveLowTimeStampCountInto( startStamp1 );
141 #endif
142 //============================================================
143 (*requestHandler)( currSlot->slaveAssignedToSlot, semanticEnv );
144 //====================== MEASUREMENT STUFF ===================
145 #ifdef MEAS__TURN_ON_PLUGIN_MEAS
146 saveLowTimeStampCountInto( endStamp1 );
147 addIntervalToHist( startStamp1, endStamp1,
148 _VMSMasterEnv->reqHdlrLowTimeHist );
149 addIntervalToHist( startStamp1, endStamp1,
150 _VMSMasterEnv->reqHdlrHighTimeHist );
151 #endif
152 //============================================================
153 }
154 if( currSlot->needsSlaveAssigned )
155 { //give slot a new Slv
156 schedSlaveVP =
157 (*slaveAssigner)( semanticEnv, thisCoresIdx );
159 if( schedSlaveVP != NULL )
160 { currSlot->slaveAssignedToSlot = schedSlaveVP;
161 schedSlaveVP->schedSlot = currSlot;
162 currSlot->needsSlaveAssigned = FALSE;
163 numSlotsFilled += 1;
165 writeVMSQ( schedSlaveVP, readyToAnimateQ );
166 }
167 }
168 }
171 #ifdef USE_WORK_STEALING
172 //If no slots filled, means no more work, look for work to steal.
173 if( numSlotsFilled == 0 )
174 { gateProtected_stealWorkInto( currSlot, readyToAnimateQ, masterVP );
175 }
176 #endif
178 MEAS__Capture_Post_Master_Point;
180 masterSwitchToCoreLoop(animatingSlv);
181 flushRegisters();
182 }//MasterLoop
185 }
189 /*This has a race condition -- the coreloops are accessing their own queues
190 * at the same time that this work-stealer on a different core is trying to
191 */
192 void inline
193 stealWorkInto( SchedSlot *currSlot, VMSQueueStruc *readyToAnimateQ,
194 SlaveVP *masterVP )
195 {
196 SlaveVP *stolenSlv;
197 int32 coreIdx, i;
198 VMSQueueStruc *currQ;
200 stolenSlv = NULL;
201 coreIdx = masterVP->coreAnimatedBy;
202 for( i = 0; i < NUM_CORES -1; i++ )
203 {
204 if( coreIdx >= NUM_CORES -1 )
205 { coreIdx = 0;
206 }
207 else
208 { coreIdx++;
209 }
210 currQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
211 if( numInVMSQ( currQ ) > 0 )
212 { stolenSlv = readVMSQ (currQ );
213 break;
214 }
215 }
217 if( stolenSlv != NULL )
218 { currSlot->slaveAssignedToSlot = stolenSlv;
219 stolenSlv->schedSlot = currSlot;
220 currSlot->needsSlaveAssigned = FALSE;
222 writeVMSQ( stolenSlv, readyToAnimateQ );
223 }
224 }
226 /*This algorithm makes the common case fast. Make the coreloop passive,
227 * and show its progress. Make the stealer control a gate that coreloop
228 * has to pass.
229 *To avoid interference, only one stealer at a time. Use a global
230 * stealer-lock.
231 *
232 *The pattern is based on a gate -- stealer shuts the gate, then monitors
233 * to be sure any already past make it all the way out, before starting.
234 *So, have a "progress" measure just before the gate, then have two after it,
235 * one is in a "waiting room" outside the gate, the other is at the exit.
236 *Then, the stealer first shuts the gate, then checks the progress measure
237 * outside it, then looks to see if the progress measure at the exit is the
238 * same. If yes, it knows the protected area is empty 'cause no other way
239 * to get in and the last to get in also exited.
240 *If the progress measure at the exit is not the same, then the stealer goes
241 * into a loop checking both the waiting-area and the exit progress-measures
242 * until one of them shows the same as the measure outside the gate. Might
243 * as well re-read the measure outside the gate each go around, just to be
244 * sure. It is guaranteed that one of the two will eventually match the one
245 * outside the gate.
246 *
247 *Here's an informal proof of correctness:
248 *The gate can be closed at any point, and have only four cases:
249 * 1) coreloop made it past the gate-closing but not yet past the exit
250 * 2) coreloop made it past the pre-gate progress update but not yet past
251 * the gate,
252 * 3) coreloop is right before the pre-gate update
253 * 4) coreloop is past the exit and far from the pre-gate update.
254 *
255 * Covering the cases in reverse order,
256 * 4) is not a problem -- stealer will read pre-gate progress, see that it
257 * matches exit progress, and the gate is closed, so stealer can proceed.
258 * 3) stealer will read pre-gate progress just after coreloop updates it..
259 * so stealer goes into a loop until the coreloop causes wait-progress
260 * to match pre-gate progress, so then stealer can proceed
261 * 2) same as 3..
262 * 1) stealer reads pre-gate progress, sees that it's different than exit,
263 * so goes into loop until exit matches pre-gate, now it knows coreloop
264 * is not in protected and cannot get back in, so can proceed.
265 *
266 *Implementation for the stealer:
267 *
268 *First, acquire the stealer lock -- only cores with no work to do will
269 * compete to steal, so not a big performance penalty having only one --
270 * will rarely have multiple stealers in a system with plenty of work -- and
271 * in a system with little work, it doesn't matter.
272 *
273 *Note, have single-reader, single-writer pattern for all variables used to
274 * communicate between stealer and victims
275 *
276 *So, scan the queues of the core loops, until find non-empty. Each core
277 * has its own list that it scans. The list goes in order from closest to
278 * furthest core, so it steals first from close cores. Later can add
279 * taking info from the app about overlapping footprints, and scan all the
280 * others then choose work with the most footprint overlap with the contents
281 * of this core's cache.
282 *
283 *Now, have a victim want to take work from. So, shut the gate in that
284 * coreloop, by setting the "gate closed" var on its stack to TRUE.
285 *Then, read the core's pre-gate progress and compare to the core's exit
286 * progress.
287 *If same, can proceed to take work from the coreloop's queue. When done,
288 * write FALSE to gate closed var.
289 *If different, then enter a loop that reads the pre-gate progress, then
290 * compares to exit progress then to wait progress. When one of two
291 * matches, proceed. Take work from the coreloop's queue. When done,
292 * write FALSE to the gate closed var.
293 *
294 */
295 void inline
296 gateProtected_stealWorkInto( SchedSlot *currSlot,
297 VMSQueueStruc *myReadyToAnimateQ,
298 SlaveVP *masterVP )
299 {
300 SlaveVP *stolenSlv;
301 int32 coreIdx, i, haveAVictim, gotLock;
302 VMSQueueStruc *victimsQ;
304 volatile GateStruc *vicGate;
305 int32 coreMightBeInProtected;
309 //see if any other cores have work available to steal
310 haveAVictim = FALSE;
311 coreIdx = masterVP->coreAnimatedBy;
312 for( i = 0; i < NUM_CORES -1; i++ )
313 {
314 if( coreIdx >= NUM_CORES -1 )
315 { coreIdx = 0;
316 }
317 else
318 { coreIdx++;
319 }
320 victimsQ = _VMSMasterEnv->readyToAnimateQs[coreIdx];
321 if( numInVMSQ( victimsQ ) > 0 )
322 { haveAVictim = TRUE;
323 vicGate = _VMSMasterEnv->workStealingGates[ coreIdx ];
324 break;
325 }
326 }
327 if( !haveAVictim ) return; //no work to steal, exit
329 //have a victim core, now get the stealer-lock
330 gotLock =__sync_bool_compare_and_swap( &(_VMSMasterEnv->workStealingLock),
331 UNLOCKED, LOCKED );
332 if( !gotLock ) return; //go back to core loop, which will re-start master
335 //====== Start Gate-protection =======
336 vicGate->gateClosed = TRUE;
337 coreMightBeInProtected= vicGate->preGateProgress != vicGate->exitProgress;
338 while( coreMightBeInProtected )
339 { //wait until sure
340 if( vicGate->preGateProgress == vicGate->waitProgress )
341 coreMightBeInProtected = FALSE;
342 if( vicGate->preGateProgress == vicGate->exitProgress )
343 coreMightBeInProtected = FALSE;
344 }
346 stolenSlv = readVMSQ ( victimsQ );
348 vicGate->gateClosed = FALSE;
349 //======= End Gate-protection =======
352 if( stolenSlv != NULL ) //victim could have been in protected and taken
353 { currSlot->slaveAssignedToSlot = stolenSlv;
354 stolenSlv->schedSlot = currSlot;
355 currSlot->needsSlaveAssigned = FALSE;
357 writeVMSQ( stolenSlv, myReadyToAnimateQ );
358 }
360 //unlock the work stealing lock
361 _VMSMasterEnv->workStealingLock = UNLOCKED;
362 }