2 #include "structdefs.h"
3 #include "multicoregccompact.h"
4 #include "runtime_arch.h"
5 #include "multicoreruntime.h"
6 #include "multicoregarbage.h"
8 #include "multicoremem_helper.h"
11 int gc_countRunningCores() {
13 for(int i = 0; i < NUMCORES4GC; i++) {
21 void initOrig_Dst(struct moveHelper * orig,struct moveHelper * to) {
23 to->localblocknum = 0;
24 BASEPTR(to->base, BAMBOO_NUM_OF_CORE, to->localblocknum);
26 to->bound=to->base+BLOCKSIZE(to->localblocknum);
29 orig->localblocknum = 0;
30 orig->ptr=orig->base = to->base;
31 orig->bound=orig->base+BLOCKSIZE(orig->localblocknum);
33 to->pagebound=to->base+BAMBOO_PAGE_SIZE;
34 orig->pagebound=orig->base+BAMBOO_PAGE_SIZE;
38 void getSpaceLocally(struct moveHelper *to) {
39 //we have space on our core...just keep going
41 BASEPTR(to->base,BAMBOO_NUM_OF_CORE, to->localblocknum);
43 to->bound=to->base+BLOCKSIZE(to->localblocknum);
45 to->pagebound=to->base+BAMBOO_PAGE_SIZE;
49 //This function is called on the master core only...and typically by
50 //the message interrupt handler
52 void handleReturnMem_I(unsigned int cnum, void *heaptop) {
53 unsigned int blockindex;
54 BLOCKINDEX(blockindex, heaptop);
55 unsigned INTPTR localblocknum=GLOBALBLOCK2LOCAL(blockindex);
56 //this core is done as far as memory usage is concerned
59 struct blockrecord * blockrecord=&allocationinfo.blocktable[blockindex];
61 blockrecord->status=BS_FREE;
62 blockrecord->usedspace=(unsigned INTPTR)(heaptop-OFFSET2BASEVA(blockindex)-gcbaseva);
63 blockrecord->freespace=BLOCKSIZE(localblocknum)-blockrecord->usedspace;
64 /* Update the lowest free block */
65 if (blockindex < allocationinfo.lowestfreeblock) {
66 allocationinfo.lowestfreeblock=blockindex;
69 /* This is our own block...means we should mark other blocks above us as free*/
71 if (cnum==blockrecord->corenum) {
72 unsigned INTPTR nextlocalblocknum=localblocknum+1;
73 for(;nextlocalblocknum<numblockspercore;nextlocalblocknum++) {
74 unsigned INTPTR blocknum=BLOCKINDEX2(cnum, nextlocalblocknum);
75 struct blockrecord * nextblockrecord=&allocationinfo.blocktable[blocknum];
76 nextblockrecord->status=BS_FREE;
77 nextblockrecord->usedspace=0;
78 //this is true because this cannot be the lowest block
79 nextblockrecord->freespace=BLOCKSIZE(1);
83 //this could be the last one....
84 int count=gc_countRunningCores();
85 if (gcmovepending==count) {
86 // All cores have stopped...hand out memory as necessary to handle all requests
87 handleMemoryRequests_I();
89 //see if returned memory blocks let us resolve requests
90 useReturnedMem(cnum, allocationinfo.lowestfreeblock);
94 void useReturnedMem(unsigned int retcorenum, block_t localblockindex) {
95 for(int i=0;i<NUMCORES4GC;i++) {
96 unsigned INTPTR requiredmem=gcrequiredmems[i];
98 unsigned INTPTR desiredmem=maxusefulmems[i];
99 unsigned INTPTR threshold=(desiredmem<MINMEMORYCHUNKSIZE)? desiredmem: MINMEMORYCHUNKSIZE;
100 unsigned INTPTR memcheck=requiredmem>threshold?requiredmem:threshold;
103 for(block_t nextlocalblocknum=localblockindex;nextlocalblocknum<numblockspercore;nextlocalblocknum++) {
104 unsigned INTPTR blocknum=BLOCKINDEX2(retcorenum, nextlocalblocknum);
105 struct blockrecord * nextblockrecord=&allocationinfo.blocktable[blocknum];
106 if (nextblockrecord->status==BS_FREE) {
107 unsigned INTPTR freespace=nextblockrecord->freespace&~BAMBOO_CACHE_LINE_MASK;
108 if (freespace>=memcheck) {
109 nextblockrecord->status=BS_USED;
110 void *blockptr=OFFSET2BASEVA(blocknum)+gcbaseva;
111 unsigned INTPTR usedspace=((nextblockrecord->usedspace-1)&~BAMBOO_CACHE_LINE_MASK)+BAMBOO_CACHE_LINE_SIZE;
112 //taken care of one block
114 void *startaddr=blockptr+usedspace;
117 if (i==STARTUPCORE) {
119 gcmovestartaddr = startaddr;
120 } else if(BAMBOO_CHECK_SEND_MODE()) {
121 cache_msg_2_I(i,GCMOVESTART,startaddr);
123 send_msg_2_I(i,GCMOVESTART,startaddr);
132 void handleReturnMem(unsigned int cnum, void *heaptop) {
133 BAMBOO_ENTER_RUNTIME_MODE_FROM_CLIENT();
134 handleReturnMem_I(cnum, heaptop);
135 BAMBOO_ENTER_CLIENT_MODE_FROM_RUNTIME();
138 void getSpaceRemotely(struct moveHelper *to, unsigned int minimumbytes) {
139 //need to get another block from elsewhere
140 //set flag to wait for memory
142 if (BAMBOO_NUM_OF_CORE==STARTUPCORE) {
144 BAMBOO_ENTER_RUNTIME_MODE_FROM_CLIENT();
145 void *startaddr=handlegcfinishcompact_I(BAMBOO_NUM_OF_CORE, minimumbytes, gccurr_heaptop);
146 BAMBOO_ENTER_CLIENT_MODE_FROM_RUNTIME();
149 gcmovestartaddr=startaddr;
155 //send request for memory
156 send_msg_4(STARTUPCORE,GCFINISHCOMPACT,BAMBOO_NUM_OF_CORE, minimumbytes, gccurr_heaptop);
157 //wait for flag to be set that we received message
163 to->ptr = gcmovestartaddr;
165 //set localblock number to high number to indicate this block isn't local
166 to->localblocknum = MAXBLOCK;
167 unsigned int globalblocknum;
168 BLOCKINDEX(globalblocknum, to->ptr);
169 to->base = gcbaseva + OFFSET2BASEVA(globalblocknum);
170 to->bound=gcbaseva+BOUNDPTR(globalblocknum);
171 #ifdef GC_CACHE_ADAPT
172 to->pagebound=(void *)((int)((int)(to->ptr)&(~(BAMBOO_PAGE_SIZE-1)))+BAMBOO_PAGE_SIZE);
176 void getSpace(struct moveHelper *to, unsigned int minimumbytes) {
177 //need more space to compact into
178 if ((to->localblocknum+1) < gcblock2fill) {
181 getSpaceRemotely(to, minimumbytes);
185 void compacthelper(struct moveHelper * orig,struct moveHelper * to) {
186 bool senttopmessage=false;
188 if ((gccurr_heaptop <= ((unsigned INTPTR)(to->bound-to->ptr)))&&!senttopmessage) {
189 //This block is the last for this core...let the startup know
190 if (BAMBOO_NUM_OF_CORE==STARTUPCORE) {
191 handleReturnMem(BAMBOO_NUM_OF_CORE, to->ptr+gccurr_heaptop);
193 send_msg_3(STARTUPCORE, GCRETURNMEM, BAMBOO_NUM_OF_CORE, to->ptr+gccurr_heaptop);
195 //Only send the message once
198 unsigned int minimumbytes=COMPACTUNITS(orig, to);
199 if (orig->ptr==orig->bound) {
200 //need more data to compact
202 orig->localblocknum++;
203 BASEPTR(orig->base,BAMBOO_NUM_OF_CORE, orig->localblocknum);
204 orig->ptr=orig->base;
205 orig->bound=orig->base+BLOCKSIZE(orig->localblocknum);
206 #ifdef GC_CACHE_ADAPT
207 orig->pagebound=orig->base+BAMBOO_PAGE_SIZE;
209 if (orig->base >= gcbaseva+BAMBOO_SHARED_MEM_SIZE) {
210 CACHEADAPT_FINISH_COMPACT(to->ptr);
214 if (minimumbytes!=0) {
215 getSpace(to, minimumbytes);
218 if (BAMBOO_NUM_OF_CORE==STARTUPCORE) {
219 BAMBOO_ENTER_RUNTIME_MODE_FROM_CLIENT();
220 handlegcfinishcompact_I(BAMBOO_NUM_OF_CORE, 0, 0);
221 BAMBOO_ENTER_CLIENT_MODE_FROM_RUNTIME();
223 send_msg_4(STARTUPCORE,GCFINISHCOMPACT,BAMBOO_NUM_OF_CORE, 0, 0);
227 void * checkNeighbors_I(int ncorenum, unsigned INTPTR requiredmem, unsigned INTPTR desiredmem) {
228 int minblockindex=allocationinfo.lowestfreeblock/NUMCORES4GC;
229 unsigned INTPTR threshold=(desiredmem<MINMEMORYCHUNKSIZE)? desiredmem: MINMEMORYCHUNKSIZE;
230 unsigned INTPTR memcheck=requiredmem>threshold?requiredmem:threshold;
232 for(block_t lblock=minblockindex;lblock<numblockspercore;lblock++) {
233 for(int i=0;i<NUM_CORES2TEST;i++) {
234 int neighborcore=core2test[ncorenum][i];
235 if (neighborcore!=-1) {
236 block_t globalblockindex=BLOCKINDEX2(neighborcore, lblock);
237 struct blockrecord * block=&allocationinfo.blocktable[globalblockindex];
238 if (block->status==BS_FREE) {
239 unsigned INTPTR freespace=block->freespace&~BAMBOO_CACHE_LINE_MASK;
240 if (memcheck<=freespace) {
243 block->status=BS_USED;
244 void *blockptr=OFFSET2BASEVA(globalblockindex)+gcbaseva;
245 unsigned INTPTR usedspace=((block->usedspace-1)&~BAMBOO_CACHE_LINE_MASK)+BAMBOO_CACHE_LINE_SIZE;
246 return blockptr+usedspace;
255 void * globalSearch_I(unsigned int topblock, unsigned INTPTR requiredmem, unsigned INTPTR desiredmem) {
256 unsigned int firstfree=NOFREEBLOCK;
257 unsigned INTPTR threshold=(desiredmem<MINMEMORYCHUNKSIZE)? desiredmem: MINMEMORYCHUNKSIZE;
258 unsigned INTPTR memcheck=requiredmem>threshold?requiredmem:threshold;
260 for(block_t i=allocationinfo.lowestfreeblock;i<topblock;i++) {
261 struct blockrecord * block=&allocationinfo.blocktable[i];
262 if (block->status==BS_FREE) {
263 if(firstfree==NOFREEBLOCK)
265 unsigned INTPTR freespace=block->freespace&~BAMBOO_CACHE_LINE_MASK;
266 if (memcheck<=freespace) {
269 block->status=BS_USED;
270 void *blockptr=OFFSET2BASEVA(i)+gcbaseva;
271 unsigned INTPTR usedspace=((block->usedspace-1)&~BAMBOO_CACHE_LINE_MASK)+BAMBOO_CACHE_LINE_SIZE;
272 allocationinfo.lowestfreeblock=firstfree;
273 return blockptr+usedspace;
277 allocationinfo.lowestfreeblock=firstfree;
281 void handleOneMemoryRequest(int core, unsigned int lowestblock) {
282 unsigned INTPTR requiredmem=gcrequiredmems[core];
283 unsigned INTPTR desiredmem=maxusefulmems[core];
284 block_t firstfree=NOFREEBLOCK;
285 unsigned INTPTR threshold=(desiredmem<MINMEMORYCHUNKSIZE)? desiredmem: MINMEMORYCHUNKSIZE;
286 unsigned INTPTR memcheck=requiredmem>threshold?requiredmem:threshold;
288 for(block_t searchblock=lowestblock;searchblock<GCNUMBLOCK;searchblock++) {
289 struct blockrecord * block=&allocationinfo.blocktable[searchblock];
290 if (block->status==BS_FREE) {
291 if(firstfree==NOFREEBLOCK)
292 firstfree=searchblock;
293 //don't take a block from another core that hasn't returned its memory yet
294 if (block->corenum!=core&&returnedmem[block->corenum])
297 unsigned INTPTR freespace=block->freespace&~BAMBOO_CACHE_LINE_MASK;
298 if (freespace>=memcheck) {
299 //TODO: should check memory block at same level on our own core...if that works, use it to preserve locality
303 block->status=BS_USED;
304 void *blockptr=OFFSET2BASEVA(searchblock)+gcbaseva;
305 unsigned INTPTR usedspace=((block->usedspace-1)&~BAMBOO_CACHE_LINE_MASK)+BAMBOO_CACHE_LINE_SIZE;
306 allocationinfo.lowestfreeblock=firstfree;
307 //taken care of one block
309 void *startaddr=blockptr+usedspace;
310 if (core==STARTUPCORE) {
312 gcmovestartaddr=startaddr;
313 } else if(BAMBOO_CHECK_SEND_MODE()) {
314 cache_msg_2_I(core,GCMOVESTART,startaddr);
316 send_msg_2_I(core,GCMOVESTART,startaddr);
322 //this is bad...ran out of memory
323 printf("Out of memory. Was trying for %u bytes\n", threshold);
327 void handleMemoryRequests_I() {
328 unsigned int lowestblock=allocationinfo.lowestfreeblock;
329 if (lowestblock==NOFREEBLOCK) {
330 lowestblock=numblockspercore*NUMCORES4GC;
333 for(int i=0;i < NUMCORES4GC; i++) {
334 if (gcrequiredmems[i]) {
335 handleOneMemoryRequest(i, lowestblock);
336 lowestblock=allocationinfo.lowestfreeblock;
341 /* should be invoked with interrupt turned off */
343 void * gcfindSpareMem_I(unsigned INTPTR requiredmem, unsigned INTPTR desiredmem,unsigned int requiredcore) {
344 if (allocationinfo.lowestfreeblock!=NOFREEBLOCK) {
345 //There are spare blocks
346 unsigned int topblock=numblockspercore*NUMCORES4GC;
349 if (memblock=checkNeighbors_I(requiredcore, requiredmem, desiredmem)) {
351 } else if (memblock=globalSearch_I(topblock, requiredmem, desiredmem)) {
356 // If we cannot find spare mem right now, hold the request
357 gcrequiredmems[requiredcore] = requiredmem;
358 maxusefulmems[requiredcore]=desiredmem;
361 int count=gc_countRunningCores();
362 if (gcmovepending==count) {
363 // All cores have stopped...hand out memory as necessary to handle all requests
364 handleMemoryRequests_I();
370 #ifdef GC_CACHE_ADAPT
371 unsigned int compactblockshelper(struct moveHelper * orig, struct moveHelper * to) {
372 unsigned int minimumbytes=0;
373 void *origptr=orig->ptr;
374 void *origbound=orig->bound;
375 void * tmporig=orig->ptr;
376 void * tmpto=to->ptr;
379 //call compactblocks using the page boundaries at the current bounds
380 minimumbytes=compactblocks(orig, to);
381 if(minimumbytes == 0) {
382 //bump the orig page bound...
383 //use old orig pointer to make sure we get correct block
384 CACHEADAPT_FINISH_SRC_PAGE(tmporig, tmpto, to->ptr);
385 if (orig->ptr<origbound) {
388 orig->pagebound=orig->pagebound+BAMBOO_PAGE_SIZE;
393 // require more memory
394 void *endtoptr=to->ptr+minimumbytes;
395 if (endtoptr>to->bound) {
396 CACHEADAPT_FINISH_DST_PAGE(orig->ptr, tmpto, to->ptr, 0);
399 CACHEADAPT_FINISH_DST_PAGE(orig->ptr, tmpto, to->ptr, minimumbytes);
400 to->pagebound=((((unsigned INTPTR)endtoptr)-1)&~(BAMBOO_PAGE_SIZE-1))+BAMBOO_PAGE_SIZE;
401 //update pointers to avoid double counting the stuff we already added in
402 tmporig=orig->ptr+minimumbytes;
403 tmpto=to->ptr+minimumbytes;
410 /* This function is performance critical... spend more time optimizing it */
412 unsigned int compactblocks(struct moveHelper * orig, struct moveHelper * to) {
413 void *toptrinit=to->ptr;
414 void *toptr=toptrinit;
415 void *origptr=orig->ptr;
416 #ifdef GC_CACHE_ADAPT
417 void *origbound=orig->pagebound;
418 void *tobound=to->pagebound;
419 //set to the first line so we don't need conditions
420 void *lastflush=(void *)(((unsigned INTPTR)&gcmappingtbl[OBJMAPPINGINDEX(origptr)])&~(BAMBOO_CACHE_LINE_MASK));
422 void *origbound=orig->bound;
423 void *tobound=to->bound;
425 unsigned INTPTR origendoffset=ALIGNTOTABLEINDEX((unsigned INTPTR)(origbound-gcbaseva));
426 unsigned int objlength;
428 while(origptr<origbound) {
429 //Try to skip over stuff fast first
430 unsigned INTPTR offset=(unsigned INTPTR) (origptr-gcbaseva);
431 unsigned INTPTR arrayoffset=ALIGNTOTABLEINDEX(offset);
432 if (!gcmarktbl[arrayoffset]) {
435 if (arrayoffset>=origendoffset) {
436 //finished with block(a page in CACHE_ADAPT version)...
439 gccurr_heaptop-=(unsigned INTPTR)(toptr-toptrinit);
440 #ifdef GC_CACHE_ADAPT
441 BAMBOO_CACHE_FLUSH_LINE(lastflush);
445 } while(!gcmarktbl[arrayoffset]);
446 origptr=CONVERTTABLEINDEXTOPTR(arrayoffset);
449 //Scan more carefully next
450 objlength=getMarkedLength(origptr);
452 if (objlength!=NOTMARKED) {
453 unsigned int length=ALIGNSIZETOBYTES(objlength);
455 //code between this and next comment should be removed
459 gettype_size(origptr, &type, &size);
460 size=((size-1)&(~(ALIGNMENTSIZE-1)))+ALIGNMENTSIZE;
463 tprintf("BAD SIZE IN BITMAP: type=%u object=%x size=%u length=%u\n", type, origptr, size, length);
464 unsigned INTPTR alignsize=ALIGNOBJSIZE((unsigned INTPTR)(origptr-gcbaseva));
465 unsigned INTPTR hibits=alignsize>>4;
466 unsigned INTPTR lobits=(alignsize&15)<<1;
467 tprintf("hibits=%x lobits=%x\n", hibits, lobits);
468 tprintf("hi=%x lo=%x\n", gcmarktbl[hibits], gcmarktbl[hibits+1]);
471 //end of code to remove
473 void *endtoptr=toptr+length;
474 if (endtoptr>tobound) {
475 gccurr_heaptop-=(unsigned INTPTR)(toptr-toptrinit);
478 #ifdef GC_CACHE_ADAPT
479 BAMBOO_CACHE_FLUSH_LINE(lastflush);
483 //good to move objects and update pointers
485 void ** mapptr=&gcmappingtbl[OBJMAPPINGINDEX(origptr)];
488 #ifdef GC_CACHE_ADAPT
489 void *maskmapptr=(void *)(((unsigned INTPTR)mapptr)&~(BAMBOO_CACHE_LINE_MASK));
491 if (lastflush!=maskmapptr) {
492 BAMBOO_CACHE_FLUSH_LINE(lastflush);
493 lastflush=maskmapptr;
500 origptr+=ALIGNMENTSIZE;
504 gccurr_heaptop-=(unsigned INTPTR)(toptr-toptrinit);
505 #ifdef GC_CACHE_ADAPT
506 BAMBOO_CACHE_FLUSH_LINE(lastflush);
512 BAMBOO_ASSERT(COMPACTPHASE == gc_status_info.gcphase);
514 // initialize structs for compacting
515 struct moveHelper orig;
516 struct moveHelper to;
517 initOrig_Dst(&orig, &to);
519 compacthelper(&orig, &to);
520 #ifdef GC_CACHE_ADAPT
525 void master_compact() {
526 // predict number of blocks to fill for each core
527 numblockspercore = loadbalance()+1;
529 GC_PRINTF("mark phase finished \n");
531 gc_resetCoreStatus();
532 //initialize local data structures first....we don't want remote requests messing data up
533 unsigned int initblocks=numblockspercore*NUMCORES4GC;
534 allocationinfo.lowestfreeblock=NOFREEBLOCK;
537 for(int i=0;i<initblocks;i++) {
538 allocationinfo.blocktable[i].status=BS_USED;
542 for(int i=initblocks;i<GCNUMBLOCK;i++) {
543 allocationinfo.blocktable[i].status=BS_FREE;
544 allocationinfo.blocktable[i].usedspace=0;
545 //this is true because all cores have at least one block already...
546 allocationinfo.blocktable[i].freespace=BLOCKSIZE(1);
549 //start all of the cores
550 for(int i = 0; i < NUMCORES4GC; i++) {
551 // init some data strutures for compact phase
552 gcrequiredmems[i] = 0;
555 //send start compact messages to all cores
556 if(i != STARTUPCORE) {
557 send_msg_2(i, GCSTARTCOMPACT, numblockspercore);
559 gcblock2fill = numblockspercore;
562 GCPROFILE_ITEM_MASTER();
565 /* wait for all cores to finish compacting */
566 GC_PRINTF("master finished\n");
568 while(!gc_checkCoreStatus())
572 void *nextvalid=gcbaseva;
573 for(void *tmp=gcbaseva; tmp<gcbaseva+BAMBOO_SHARED_MEM_SIZE;tmp+=ALIGNMENTSIZE) {
574 unsigned int objlength=getMarkedLength(tmp);
575 void *forwarding=gcmappingtbl[OBJMAPPINGINDEX(tmp)];
576 if (tmp>=nextvalid&&((objlength!=0)!=(forwarding!=NULL))) {
577 tprintf("Maps disagree tmp=%x olength=%u forwarding=%x\n",tmp, objlength, forwarding);
579 if (tmp<nextvalid&&forwarding!=NULL) {
580 tprintf("Weird forwarding pointer\n");
582 if (tmp>=nextvalid&&(objlength!=0||forwarding!=NULL)) {
583 unsigned int length=ALIGNSIZETOBYTES(objlength);
586 nextvalid=tmp+length;
587 gettype_size(tmp, &type, &size);
588 size=((size-1)&(~(ALIGNMENTSIZE-1)))+ALIGNMENTSIZE;
590 tprintf("Bad size in bitmap: tmp=%x length=%u size=%u type=%u\n", tmp, length, size, type);
593 BLOCKINDEX(blockindex, forwarding);
594 struct blockrecord * block=&allocationinfo.blocktable[blockindex];
595 void *blockptr=OFFSET2BASEVA(blockindex)+gcbaseva;
597 if (block->status==BS_FREE) {
598 if (forwarding>(blockptr+block->usedspace)) {
599 tprintf("Pointer references free space forwarding=%x tmp=%x length=%u type=%u blockindex=%u, baseptr=%x, usedspace=%u, status=%u\n", forwarding, tmp, length, type,blockindex, blockptr, block->usedspace, block->status);
606 GCPROFILE_ITEM_MASTER();
608 //just in case we didn't get blocks back...
609 if (allocationinfo.lowestfreeblock==NOFREEBLOCK)
610 allocationinfo.lowestfreeblock=numblockspercore*NUMCORES4GC;
612 // compute live object space
613 GCPROFILE_RECORD_SPACE_MASTER();
614 GC_PRINTF("compact phase finished \n");
617 #endif // MULTICORE_GC