Tue, 04 Aug 2009 16:00:17 -0700
6819077: G1: first GC thread coming late into the GC.
Summary: The first worker thread is delayed when entering the GC because it clears the card count table that is used in identifying hot cards. Replace the card count table with a dynamically sized evicting hash table that includes an epoch based counter.
Reviewed-by: iveresov, tonyp
1 /*
2 * Copyright 2001-2009 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_g1RemSet.cpp.incl"
28 #define CARD_REPEAT_HISTO 0
30 #if CARD_REPEAT_HISTO
31 static size_t ct_freq_sz;
32 static jbyte* ct_freq = NULL;
34 void init_ct_freq_table(size_t heap_sz_bytes) {
35 if (ct_freq == NULL) {
36 ct_freq_sz = heap_sz_bytes/CardTableModRefBS::card_size;
37 ct_freq = new jbyte[ct_freq_sz];
38 for (size_t j = 0; j < ct_freq_sz; j++) ct_freq[j] = 0;
39 }
40 }
42 void ct_freq_note_card(size_t index) {
43 assert(0 <= index && index < ct_freq_sz, "Bounds error.");
44 if (ct_freq[index] < 100) { ct_freq[index]++; }
45 }
47 static IntHistogram card_repeat_count(10, 10);
49 void ct_freq_update_histo_and_reset() {
50 for (size_t j = 0; j < ct_freq_sz; j++) {
51 card_repeat_count.add_entry(ct_freq[j]);
52 ct_freq[j] = 0;
53 }
55 }
56 #endif
59 class IntoCSOopClosure: public OopsInHeapRegionClosure {
60 OopsInHeapRegionClosure* _blk;
61 G1CollectedHeap* _g1;
62 public:
63 IntoCSOopClosure(G1CollectedHeap* g1, OopsInHeapRegionClosure* blk) :
64 _g1(g1), _blk(blk) {}
65 void set_region(HeapRegion* from) {
66 _blk->set_region(from);
67 }
68 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
69 virtual void do_oop( oop* p) { do_oop_work(p); }
70 template <class T> void do_oop_work(T* p) {
71 oop obj = oopDesc::load_decode_heap_oop(p);
72 if (_g1->obj_in_cs(obj)) _blk->do_oop(p);
73 }
74 bool apply_to_weak_ref_discovered_field() { return true; }
75 bool idempotent() { return true; }
76 };
78 class IntoCSRegionClosure: public HeapRegionClosure {
79 IntoCSOopClosure _blk;
80 G1CollectedHeap* _g1;
81 public:
82 IntoCSRegionClosure(G1CollectedHeap* g1, OopsInHeapRegionClosure* blk) :
83 _g1(g1), _blk(g1, blk) {}
84 bool doHeapRegion(HeapRegion* r) {
85 if (!r->in_collection_set()) {
86 _blk.set_region(r);
87 if (r->isHumongous()) {
88 if (r->startsHumongous()) {
89 oop obj = oop(r->bottom());
90 obj->oop_iterate(&_blk);
91 }
92 } else {
93 r->oop_before_save_marks_iterate(&_blk);
94 }
95 }
96 return false;
97 }
98 };
100 void
101 StupidG1RemSet::oops_into_collection_set_do(OopsInHeapRegionClosure* oc,
102 int worker_i) {
103 IntoCSRegionClosure rc(_g1, oc);
104 _g1->heap_region_iterate(&rc);
105 }
107 class VerifyRSCleanCardOopClosure: public OopClosure {
108 G1CollectedHeap* _g1;
109 public:
110 VerifyRSCleanCardOopClosure(G1CollectedHeap* g1) : _g1(g1) {}
112 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
113 virtual void do_oop( oop* p) { do_oop_work(p); }
114 template <class T> void do_oop_work(T* p) {
115 oop obj = oopDesc::load_decode_heap_oop(p);
116 HeapRegion* to = _g1->heap_region_containing(obj);
117 guarantee(to == NULL || !to->in_collection_set(),
118 "Missed a rem set member.");
119 }
120 };
122 HRInto_G1RemSet::HRInto_G1RemSet(G1CollectedHeap* g1, CardTableModRefBS* ct_bs)
123 : G1RemSet(g1), _ct_bs(ct_bs), _g1p(_g1->g1_policy()),
124 _cg1r(g1->concurrent_g1_refine()),
125 _par_traversal_in_progress(false), _new_refs(NULL),
126 _cards_scanned(NULL), _total_cards_scanned(0)
127 {
128 _seq_task = new SubTasksDone(NumSeqTasks);
129 guarantee(n_workers() > 0, "There should be some workers");
130 _new_refs = NEW_C_HEAP_ARRAY(GrowableArray<OopOrNarrowOopStar>*, n_workers());
131 for (uint i = 0; i < n_workers(); i++) {
132 _new_refs[i] = new (ResourceObj::C_HEAP) GrowableArray<OopOrNarrowOopStar>(8192,true);
133 }
134 }
136 HRInto_G1RemSet::~HRInto_G1RemSet() {
137 delete _seq_task;
138 for (uint i = 0; i < n_workers(); i++) {
139 delete _new_refs[i];
140 }
141 FREE_C_HEAP_ARRAY(GrowableArray<OopOrNarrowOopStar>*, _new_refs);
142 }
144 void CountNonCleanMemRegionClosure::do_MemRegion(MemRegion mr) {
145 if (_g1->is_in_g1_reserved(mr.start())) {
146 _n += (int) ((mr.byte_size() / CardTableModRefBS::card_size));
147 if (_start_first == NULL) _start_first = mr.start();
148 }
149 }
151 class ScanRSClosure : public HeapRegionClosure {
152 size_t _cards_done, _cards;
153 G1CollectedHeap* _g1h;
154 OopsInHeapRegionClosure* _oc;
155 G1BlockOffsetSharedArray* _bot_shared;
156 CardTableModRefBS *_ct_bs;
157 int _worker_i;
158 bool _try_claimed;
159 size_t _min_skip_distance, _max_skip_distance;
160 public:
161 ScanRSClosure(OopsInHeapRegionClosure* oc, int worker_i) :
162 _oc(oc),
163 _cards(0),
164 _cards_done(0),
165 _worker_i(worker_i),
166 _try_claimed(false)
167 {
168 _g1h = G1CollectedHeap::heap();
169 _bot_shared = _g1h->bot_shared();
170 _ct_bs = (CardTableModRefBS*) (_g1h->barrier_set());
171 _min_skip_distance = 16;
172 _max_skip_distance = 2 * _g1h->n_par_threads() * _min_skip_distance;
173 }
175 void set_try_claimed() { _try_claimed = true; }
177 void scanCard(size_t index, HeapRegion *r) {
178 _cards_done++;
179 DirtyCardToOopClosure* cl =
180 r->new_dcto_closure(_oc,
181 CardTableModRefBS::Precise,
182 HeapRegionDCTOC::IntoCSFilterKind);
184 // Set the "from" region in the closure.
185 _oc->set_region(r);
186 HeapWord* card_start = _bot_shared->address_for_index(index);
187 HeapWord* card_end = card_start + G1BlockOffsetSharedArray::N_words;
188 Space *sp = SharedHeap::heap()->space_containing(card_start);
189 MemRegion sm_region;
190 if (ParallelGCThreads > 0) {
191 // first find the used area
192 sm_region = sp->used_region_at_save_marks();
193 } else {
194 // The closure is not idempotent. We shouldn't look at objects
195 // allocated during the GC.
196 sm_region = sp->used_region_at_save_marks();
197 }
198 MemRegion mr = sm_region.intersection(MemRegion(card_start,card_end));
199 if (!mr.is_empty()) {
200 cl->do_MemRegion(mr);
201 }
202 }
204 void printCard(HeapRegion* card_region, size_t card_index,
205 HeapWord* card_start) {
206 gclog_or_tty->print_cr("T %d Region [" PTR_FORMAT ", " PTR_FORMAT ") "
207 "RS names card %p: "
208 "[" PTR_FORMAT ", " PTR_FORMAT ")",
209 _worker_i,
210 card_region->bottom(), card_region->end(),
211 card_index,
212 card_start, card_start + G1BlockOffsetSharedArray::N_words);
213 }
215 bool doHeapRegion(HeapRegion* r) {
216 assert(r->in_collection_set(), "should only be called on elements of CS.");
217 HeapRegionRemSet* hrrs = r->rem_set();
218 if (hrrs->iter_is_complete()) return false; // All done.
219 if (!_try_claimed && !hrrs->claim_iter()) return false;
220 _g1h->push_dirty_cards_region(r);
221 // If we didn't return above, then
222 // _try_claimed || r->claim_iter()
223 // is true: either we're supposed to work on claimed-but-not-complete
224 // regions, or we successfully claimed the region.
225 HeapRegionRemSetIterator* iter = _g1h->rem_set_iterator(_worker_i);
226 hrrs->init_iterator(iter);
227 size_t card_index;
228 size_t skip_distance = 0, current_card = 0, jump_to_card = 0;
229 while (iter->has_next(card_index)) {
230 if (current_card < jump_to_card) {
231 ++current_card;
232 continue;
233 }
234 HeapWord* card_start = _g1h->bot_shared()->address_for_index(card_index);
235 #if 0
236 gclog_or_tty->print("Rem set iteration yielded card [" PTR_FORMAT ", " PTR_FORMAT ").\n",
237 card_start, card_start + CardTableModRefBS::card_size_in_words);
238 #endif
240 HeapRegion* card_region = _g1h->heap_region_containing(card_start);
241 assert(card_region != NULL, "Yielding cards not in the heap?");
242 _cards++;
244 if (!card_region->is_on_dirty_cards_region_list()) {
245 _g1h->push_dirty_cards_region(card_region);
246 }
248 // If the card is dirty, then we will scan it during updateRS.
249 if (!card_region->in_collection_set() && !_ct_bs->is_card_dirty(card_index)) {
250 if (!_ct_bs->is_card_claimed(card_index) && _ct_bs->claim_card(card_index)) {
251 scanCard(card_index, card_region);
252 } else if (_try_claimed) {
253 if (jump_to_card == 0 || jump_to_card != current_card) {
254 // We did some useful work in the previous iteration.
255 // Decrease the distance.
256 skip_distance = MAX2(skip_distance >> 1, _min_skip_distance);
257 } else {
258 // Previous iteration resulted in a claim failure.
259 // Increase the distance.
260 skip_distance = MIN2(skip_distance << 1, _max_skip_distance);
261 }
262 jump_to_card = current_card + skip_distance;
263 }
264 }
265 ++current_card;
266 }
267 if (!_try_claimed) {
268 hrrs->set_iter_complete();
269 }
270 return false;
271 }
272 // Set all cards back to clean.
273 void cleanup() {_g1h->cleanUpCardTable();}
274 size_t cards_done() { return _cards_done;}
275 size_t cards_looked_up() { return _cards;}
276 };
278 // We want the parallel threads to start their scanning at
279 // different collection set regions to avoid contention.
280 // If we have:
281 // n collection set regions
282 // p threads
283 // Then thread t will start at region t * floor (n/p)
285 HeapRegion* HRInto_G1RemSet::calculateStartRegion(int worker_i) {
286 HeapRegion* result = _g1p->collection_set();
287 if (ParallelGCThreads > 0) {
288 size_t cs_size = _g1p->collection_set_size();
289 int n_workers = _g1->workers()->total_workers();
290 size_t cs_spans = cs_size / n_workers;
291 size_t ind = cs_spans * worker_i;
292 for (size_t i = 0; i < ind; i++)
293 result = result->next_in_collection_set();
294 }
295 return result;
296 }
298 void HRInto_G1RemSet::scanRS(OopsInHeapRegionClosure* oc, int worker_i) {
299 double rs_time_start = os::elapsedTime();
300 HeapRegion *startRegion = calculateStartRegion(worker_i);
302 BufferingOopsInHeapRegionClosure boc(oc);
303 ScanRSClosure scanRScl(&boc, worker_i);
304 _g1->collection_set_iterate_from(startRegion, &scanRScl);
305 scanRScl.set_try_claimed();
306 _g1->collection_set_iterate_from(startRegion, &scanRScl);
308 boc.done();
309 double closure_app_time_sec = boc.closure_app_seconds();
310 double scan_rs_time_sec = (os::elapsedTime() - rs_time_start) -
311 closure_app_time_sec;
312 double closure_app_time_ms = closure_app_time_sec * 1000.0;
314 assert( _cards_scanned != NULL, "invariant" );
315 _cards_scanned[worker_i] = scanRScl.cards_done();
317 _g1p->record_scan_rs_start_time(worker_i, rs_time_start * 1000.0);
318 _g1p->record_scan_rs_time(worker_i, scan_rs_time_sec * 1000.0);
320 double scan_new_refs_time_ms = _g1p->get_scan_new_refs_time(worker_i);
321 if (scan_new_refs_time_ms > 0.0) {
322 closure_app_time_ms += scan_new_refs_time_ms;
323 }
325 _g1p->record_obj_copy_time(worker_i, closure_app_time_ms);
326 }
328 void HRInto_G1RemSet::updateRS(int worker_i) {
329 ConcurrentG1Refine* cg1r = _g1->concurrent_g1_refine();
331 double start = os::elapsedTime();
332 _g1p->record_update_rs_start_time(worker_i, start * 1000.0);
334 // Apply the appropriate closure to all remaining log entries.
335 _g1->iterate_dirty_card_closure(false, worker_i);
336 // Now there should be no dirty cards.
337 if (G1RSLogCheckCardTable) {
338 CountNonCleanMemRegionClosure cl(_g1);
339 _ct_bs->mod_card_iterate(&cl);
340 // XXX This isn't true any more: keeping cards of young regions
341 // marked dirty broke it. Need some reasonable fix.
342 guarantee(cl.n() == 0, "Card table should be clean.");
343 }
345 _g1p->record_update_rs_time(worker_i, (os::elapsedTime() - start) * 1000.0);
346 }
348 #ifndef PRODUCT
349 class PrintRSClosure : public HeapRegionClosure {
350 int _count;
351 public:
352 PrintRSClosure() : _count(0) {}
353 bool doHeapRegion(HeapRegion* r) {
354 HeapRegionRemSet* hrrs = r->rem_set();
355 _count += (int) hrrs->occupied();
356 if (hrrs->occupied() == 0) {
357 gclog_or_tty->print("Heap Region [" PTR_FORMAT ", " PTR_FORMAT ") "
358 "has no remset entries\n",
359 r->bottom(), r->end());
360 } else {
361 gclog_or_tty->print("Printing rem set for heap region [" PTR_FORMAT ", " PTR_FORMAT ")\n",
362 r->bottom(), r->end());
363 r->print();
364 hrrs->print();
365 gclog_or_tty->print("\nDone printing rem set\n");
366 }
367 return false;
368 }
369 int occupied() {return _count;}
370 };
371 #endif
373 class CountRSSizeClosure: public HeapRegionClosure {
374 size_t _n;
375 size_t _tot;
376 size_t _max;
377 HeapRegion* _max_r;
378 enum {
379 N = 20,
380 MIN = 6
381 };
382 int _histo[N];
383 public:
384 CountRSSizeClosure() : _n(0), _tot(0), _max(0), _max_r(NULL) {
385 for (int i = 0; i < N; i++) _histo[i] = 0;
386 }
387 bool doHeapRegion(HeapRegion* r) {
388 if (!r->continuesHumongous()) {
389 size_t occ = r->rem_set()->occupied();
390 _n++;
391 _tot += occ;
392 if (occ > _max) {
393 _max = occ;
394 _max_r = r;
395 }
396 // Fit it into a histo bin.
397 int s = 1 << MIN;
398 int i = 0;
399 while (occ > (size_t) s && i < (N-1)) {
400 s = s << 1;
401 i++;
402 }
403 _histo[i]++;
404 }
405 return false;
406 }
407 size_t n() { return _n; }
408 size_t tot() { return _tot; }
409 size_t mx() { return _max; }
410 HeapRegion* mxr() { return _max_r; }
411 void print_histo() {
412 int mx = N;
413 while (mx >= 0) {
414 if (_histo[mx-1] > 0) break;
415 mx--;
416 }
417 gclog_or_tty->print_cr("Number of regions with given RS sizes:");
418 gclog_or_tty->print_cr(" <= %8d %8d", 1 << MIN, _histo[0]);
419 for (int i = 1; i < mx-1; i++) {
420 gclog_or_tty->print_cr(" %8d - %8d %8d",
421 (1 << (MIN + i - 1)) + 1,
422 1 << (MIN + i),
423 _histo[i]);
424 }
425 gclog_or_tty->print_cr(" > %8d %8d", (1 << (MIN+mx-2))+1, _histo[mx-1]);
426 }
427 };
429 template <class T> void
430 HRInto_G1RemSet::scanNewRefsRS_work(OopsInHeapRegionClosure* oc,
431 int worker_i) {
432 double scan_new_refs_start_sec = os::elapsedTime();
433 G1CollectedHeap* g1h = G1CollectedHeap::heap();
434 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (g1h->barrier_set());
435 for (int i = 0; i < _new_refs[worker_i]->length(); i++) {
436 T* p = (T*) _new_refs[worker_i]->at(i);
437 oop obj = oopDesc::load_decode_heap_oop(p);
438 // *p was in the collection set when p was pushed on "_new_refs", but
439 // another thread may have processed this location from an RS, so it
440 // might not point into the CS any longer. If so, it's obviously been
441 // processed, and we don't need to do anything further.
442 if (g1h->obj_in_cs(obj)) {
443 HeapRegion* r = g1h->heap_region_containing(p);
445 DEBUG_ONLY(HeapRegion* to = g1h->heap_region_containing(obj));
446 oc->set_region(r);
447 // If "p" has already been processed concurrently, this is
448 // idempotent.
449 oc->do_oop(p);
450 }
451 }
452 _g1p->record_scan_new_refs_time(worker_i,
453 (os::elapsedTime() - scan_new_refs_start_sec)
454 * 1000.0);
455 }
457 void HRInto_G1RemSet::cleanupHRRS() {
458 HeapRegionRemSet::cleanup();
459 }
461 void
462 HRInto_G1RemSet::oops_into_collection_set_do(OopsInHeapRegionClosure* oc,
463 int worker_i) {
464 #if CARD_REPEAT_HISTO
465 ct_freq_update_histo_and_reset();
466 #endif
467 if (worker_i == 0) {
468 _cg1r->clear_and_record_card_counts();
469 }
471 // Make this into a command-line flag...
472 if (G1RSCountHisto && (ParallelGCThreads == 0 || worker_i == 0)) {
473 CountRSSizeClosure count_cl;
474 _g1->heap_region_iterate(&count_cl);
475 gclog_or_tty->print_cr("Avg of %d RS counts is %f, max is %d, "
476 "max region is " PTR_FORMAT,
477 count_cl.n(), (float)count_cl.tot()/(float)count_cl.n(),
478 count_cl.mx(), count_cl.mxr());
479 count_cl.print_histo();
480 }
482 if (ParallelGCThreads > 0) {
483 // The two flags below were introduced temporarily to serialize
484 // the updating and scanning of remembered sets. There are some
485 // race conditions when these two operations are done in parallel
486 // and they are causing failures. When we resolve said race
487 // conditions, we'll revert back to parallel remembered set
488 // updating and scanning. See CRs 6677707 and 6677708.
489 if (G1ParallelRSetUpdatingEnabled || (worker_i == 0)) {
490 updateRS(worker_i);
491 scanNewRefsRS(oc, worker_i);
492 } else {
493 _g1p->record_update_rs_start_time(worker_i, os::elapsedTime() * 1000.0);
494 _g1p->record_update_rs_processed_buffers(worker_i, 0.0);
495 _g1p->record_update_rs_time(worker_i, 0.0);
496 _g1p->record_scan_new_refs_time(worker_i, 0.0);
497 }
498 if (G1ParallelRSetScanningEnabled || (worker_i == 0)) {
499 scanRS(oc, worker_i);
500 } else {
501 _g1p->record_scan_rs_start_time(worker_i, os::elapsedTime() * 1000.0);
502 _g1p->record_scan_rs_time(worker_i, 0.0);
503 }
504 } else {
505 assert(worker_i == 0, "invariant");
506 updateRS(0);
507 scanNewRefsRS(oc, 0);
508 scanRS(oc, 0);
509 }
510 }
512 void HRInto_G1RemSet::
513 prepare_for_oops_into_collection_set_do() {
514 #if G1_REM_SET_LOGGING
515 PrintRSClosure cl;
516 _g1->collection_set_iterate(&cl);
517 #endif
518 cleanupHRRS();
519 ConcurrentG1Refine* cg1r = _g1->concurrent_g1_refine();
520 _g1->set_refine_cte_cl_concurrency(false);
521 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
522 dcqs.concatenate_logs();
524 assert(!_par_traversal_in_progress, "Invariant between iterations.");
525 if (ParallelGCThreads > 0) {
526 set_par_traversal(true);
527 _seq_task->set_par_threads((int)n_workers());
528 }
529 guarantee( _cards_scanned == NULL, "invariant" );
530 _cards_scanned = NEW_C_HEAP_ARRAY(size_t, n_workers());
531 for (uint i = 0; i < n_workers(); ++i) {
532 _cards_scanned[i] = 0;
533 }
534 _total_cards_scanned = 0;
535 }
538 class cleanUpIteratorsClosure : public HeapRegionClosure {
539 bool doHeapRegion(HeapRegion *r) {
540 HeapRegionRemSet* hrrs = r->rem_set();
541 hrrs->init_for_par_iteration();
542 return false;
543 }
544 };
546 class UpdateRSetOopsIntoCSImmediate : public OopClosure {
547 G1CollectedHeap* _g1;
548 public:
549 UpdateRSetOopsIntoCSImmediate(G1CollectedHeap* g1) : _g1(g1) { }
550 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
551 virtual void do_oop( oop* p) { do_oop_work(p); }
552 template <class T> void do_oop_work(T* p) {
553 HeapRegion* to = _g1->heap_region_containing(oopDesc::load_decode_heap_oop(p));
554 if (to->in_collection_set()) {
555 to->rem_set()->add_reference(p, 0);
556 }
557 }
558 };
560 class UpdateRSetOopsIntoCSDeferred : public OopClosure {
561 G1CollectedHeap* _g1;
562 CardTableModRefBS* _ct_bs;
563 DirtyCardQueue* _dcq;
564 public:
565 UpdateRSetOopsIntoCSDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
566 _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) { }
567 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
568 virtual void do_oop( oop* p) { do_oop_work(p); }
569 template <class T> void do_oop_work(T* p) {
570 oop obj = oopDesc::load_decode_heap_oop(p);
571 if (_g1->obj_in_cs(obj)) {
572 size_t card_index = _ct_bs->index_for(p);
573 if (_ct_bs->mark_card_deferred(card_index)) {
574 _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
575 }
576 }
577 }
578 };
580 template <class T> void HRInto_G1RemSet::new_refs_iterate_work(OopClosure* cl) {
581 for (size_t i = 0; i < n_workers(); i++) {
582 for (int j = 0; j < _new_refs[i]->length(); j++) {
583 T* p = (T*) _new_refs[i]->at(j);
584 cl->do_oop(p);
585 }
586 }
587 }
589 void HRInto_G1RemSet::cleanup_after_oops_into_collection_set_do() {
590 guarantee( _cards_scanned != NULL, "invariant" );
591 _total_cards_scanned = 0;
592 for (uint i = 0; i < n_workers(); ++i)
593 _total_cards_scanned += _cards_scanned[i];
594 FREE_C_HEAP_ARRAY(size_t, _cards_scanned);
595 _cards_scanned = NULL;
596 // Cleanup after copy
597 #if G1_REM_SET_LOGGING
598 PrintRSClosure cl;
599 _g1->heap_region_iterate(&cl);
600 #endif
601 _g1->set_refine_cte_cl_concurrency(true);
602 cleanUpIteratorsClosure iterClosure;
603 _g1->collection_set_iterate(&iterClosure);
604 // Set all cards back to clean.
605 _g1->cleanUpCardTable();
607 if (ParallelGCThreads > 0) {
608 set_par_traversal(false);
609 }
611 if (_g1->evacuation_failed()) {
612 // Restore remembered sets for the regions pointing into
613 // the collection set.
614 if (G1DeferredRSUpdate) {
615 DirtyCardQueue dcq(&_g1->dirty_card_queue_set());
616 UpdateRSetOopsIntoCSDeferred deferred_update(_g1, &dcq);
617 new_refs_iterate(&deferred_update);
618 } else {
619 UpdateRSetOopsIntoCSImmediate immediate_update(_g1);
620 new_refs_iterate(&immediate_update);
621 }
622 }
623 for (uint i = 0; i < n_workers(); i++) {
624 _new_refs[i]->clear();
625 }
627 assert(!_par_traversal_in_progress, "Invariant between iterations.");
628 }
630 class UpdateRSObjectClosure: public ObjectClosure {
631 UpdateRSOopClosure* _update_rs_oop_cl;
632 public:
633 UpdateRSObjectClosure(UpdateRSOopClosure* update_rs_oop_cl) :
634 _update_rs_oop_cl(update_rs_oop_cl) {}
635 void do_object(oop obj) {
636 obj->oop_iterate(_update_rs_oop_cl);
637 }
639 };
641 class ScrubRSClosure: public HeapRegionClosure {
642 G1CollectedHeap* _g1h;
643 BitMap* _region_bm;
644 BitMap* _card_bm;
645 CardTableModRefBS* _ctbs;
646 public:
647 ScrubRSClosure(BitMap* region_bm, BitMap* card_bm) :
648 _g1h(G1CollectedHeap::heap()),
649 _region_bm(region_bm), _card_bm(card_bm),
650 _ctbs(NULL)
651 {
652 ModRefBarrierSet* bs = _g1h->mr_bs();
653 guarantee(bs->is_a(BarrierSet::CardTableModRef), "Precondition");
654 _ctbs = (CardTableModRefBS*)bs;
655 }
657 bool doHeapRegion(HeapRegion* r) {
658 if (!r->continuesHumongous()) {
659 r->rem_set()->scrub(_ctbs, _region_bm, _card_bm);
660 }
661 return false;
662 }
663 };
665 void HRInto_G1RemSet::scrub(BitMap* region_bm, BitMap* card_bm) {
666 ScrubRSClosure scrub_cl(region_bm, card_bm);
667 _g1->heap_region_iterate(&scrub_cl);
668 }
670 void HRInto_G1RemSet::scrub_par(BitMap* region_bm, BitMap* card_bm,
671 int worker_num, int claim_val) {
672 ScrubRSClosure scrub_cl(region_bm, card_bm);
673 _g1->heap_region_par_iterate_chunked(&scrub_cl, worker_num, claim_val);
674 }
677 static IntHistogram out_of_histo(50, 50);
679 void HRInto_G1RemSet::concurrentRefineOneCard_impl(jbyte* card_ptr, int worker_i) {
680 // Construct the region representing the card.
681 HeapWord* start = _ct_bs->addr_for(card_ptr);
682 // And find the region containing it.
683 HeapRegion* r = _g1->heap_region_containing(start);
684 assert(r != NULL, "unexpected null");
686 HeapWord* end = _ct_bs->addr_for(card_ptr + 1);
687 MemRegion dirtyRegion(start, end);
689 #if CARD_REPEAT_HISTO
690 init_ct_freq_table(_g1->g1_reserved_obj_bytes());
691 ct_freq_note_card(_ct_bs->index_for(start));
692 #endif
694 UpdateRSOopClosure update_rs_oop_cl(this, worker_i);
695 update_rs_oop_cl.set_from(r);
696 FilterOutOfRegionClosure filter_then_update_rs_oop_cl(r, &update_rs_oop_cl);
698 // Undirty the card.
699 *card_ptr = CardTableModRefBS::clean_card_val();
700 // We must complete this write before we do any of the reads below.
701 OrderAccess::storeload();
702 // And process it, being careful of unallocated portions of TLAB's.
703 HeapWord* stop_point =
704 r->oops_on_card_seq_iterate_careful(dirtyRegion,
705 &filter_then_update_rs_oop_cl);
706 // If stop_point is non-null, then we encountered an unallocated region
707 // (perhaps the unfilled portion of a TLAB.) For now, we'll dirty the
708 // card and re-enqueue: if we put off the card until a GC pause, then the
709 // unallocated portion will be filled in. Alternatively, we might try
710 // the full complexity of the technique used in "regular" precleaning.
711 if (stop_point != NULL) {
712 // The card might have gotten re-dirtied and re-enqueued while we
713 // worked. (In fact, it's pretty likely.)
714 if (*card_ptr != CardTableModRefBS::dirty_card_val()) {
715 *card_ptr = CardTableModRefBS::dirty_card_val();
716 MutexLockerEx x(Shared_DirtyCardQ_lock,
717 Mutex::_no_safepoint_check_flag);
718 DirtyCardQueue* sdcq =
719 JavaThread::dirty_card_queue_set().shared_dirty_card_queue();
720 sdcq->enqueue(card_ptr);
721 }
722 } else {
723 out_of_histo.add_entry(filter_then_update_rs_oop_cl.out_of_region());
724 _conc_refine_cards++;
725 }
726 }
728 void HRInto_G1RemSet::concurrentRefineOneCard(jbyte* card_ptr, int worker_i) {
729 // If the card is no longer dirty, nothing to do.
730 if (*card_ptr != CardTableModRefBS::dirty_card_val()) return;
732 // Construct the region representing the card.
733 HeapWord* start = _ct_bs->addr_for(card_ptr);
734 // And find the region containing it.
735 HeapRegion* r = _g1->heap_region_containing(start);
736 if (r == NULL) {
737 guarantee(_g1->is_in_permanent(start), "Or else where?");
738 return; // Not in the G1 heap (might be in perm, for example.)
739 }
740 // Why do we have to check here whether a card is on a young region,
741 // given that we dirty young regions and, as a result, the
742 // post-barrier is supposed to filter them out and never to enqueue
743 // them? When we allocate a new region as the "allocation region" we
744 // actually dirty its cards after we release the lock, since card
745 // dirtying while holding the lock was a performance bottleneck. So,
746 // as a result, it is possible for other threads to actually
747 // allocate objects in the region (after the acquire the lock)
748 // before all the cards on the region are dirtied. This is unlikely,
749 // and it doesn't happen often, but it can happen. So, the extra
750 // check below filters out those cards.
751 if (r->is_young()) {
752 return;
753 }
754 // While we are processing RSet buffers during the collection, we
755 // actually don't want to scan any cards on the collection set,
756 // since we don't want to update remebered sets with entries that
757 // point into the collection set, given that live objects from the
758 // collection set are about to move and such entries will be stale
759 // very soon. This change also deals with a reliability issue which
760 // involves scanning a card in the collection set and coming across
761 // an array that was being chunked and looking malformed. Note,
762 // however, that if evacuation fails, we have to scan any objects
763 // that were not moved and create any missing entries.
764 if (r->in_collection_set()) {
765 return;
766 }
768 // Should we defer processing the card?
769 //
770 // Previously the result from the insert_cache call would be
771 // either card_ptr (implying that card_ptr was currently "cold"),
772 // null (meaning we had inserted the card ptr into the "hot"
773 // cache, which had some headroom), or a "hot" card ptr
774 // extracted from the "hot" cache.
775 //
776 // Now that the _card_counts cache in the ConcurrentG1Refine
777 // instance is an evicting hash table, the result we get back
778 // could be from evicting the card ptr in an already occupied
779 // bucket (in which case we have replaced the card ptr in the
780 // bucket with card_ptr and "defer" is set to false). To avoid
781 // having a data structure (updates to which would need a lock)
782 // to hold these unprocessed dirty cards, we need to immediately
783 // process card_ptr. The actions needed to be taken on return
784 // from cache_insert are summarized in the following table:
785 //
786 // res defer action
787 // --------------------------------------------------------------
788 // null false card evicted from _card_counts & replaced with
789 // card_ptr; evicted ptr added to hot cache.
790 // No need to process res; immediately process card_ptr
791 //
792 // null true card not evicted from _card_counts; card_ptr added
793 // to hot cache.
794 // Nothing to do.
795 //
796 // non-null false card evicted from _card_counts & replaced with
797 // card_ptr; evicted ptr is currently "cold" or
798 // caused an eviction from the hot cache.
799 // Immediately process res; process card_ptr.
800 //
801 // non-null true card not evicted from _card_counts; card_ptr is
802 // currently cold, or caused an eviction from hot
803 // cache.
804 // Immediately process res; no need to process card_ptr.
806 jbyte* res = card_ptr;
807 bool defer = false;
808 if (_cg1r->use_cache()) {
809 jbyte* res = _cg1r->cache_insert(card_ptr, &defer);
810 if (res != NULL && (res != card_ptr || defer)) {
811 start = _ct_bs->addr_for(res);
812 r = _g1->heap_region_containing(start);
813 if (r == NULL) {
814 assert(_g1->is_in_permanent(start), "Or else where?");
815 } else {
816 guarantee(!r->is_young(), "It was evicted in the current minor cycle.");
817 // Process card pointer we get back from the hot card cache
818 concurrentRefineOneCard_impl(res, worker_i);
819 }
820 }
821 }
823 if (!defer) {
824 concurrentRefineOneCard_impl(card_ptr, worker_i);
825 }
826 }
828 class HRRSStatsIter: public HeapRegionClosure {
829 size_t _occupied;
830 size_t _total_mem_sz;
831 size_t _max_mem_sz;
832 HeapRegion* _max_mem_sz_region;
833 public:
834 HRRSStatsIter() :
835 _occupied(0),
836 _total_mem_sz(0),
837 _max_mem_sz(0),
838 _max_mem_sz_region(NULL)
839 {}
841 bool doHeapRegion(HeapRegion* r) {
842 if (r->continuesHumongous()) return false;
843 size_t mem_sz = r->rem_set()->mem_size();
844 if (mem_sz > _max_mem_sz) {
845 _max_mem_sz = mem_sz;
846 _max_mem_sz_region = r;
847 }
848 _total_mem_sz += mem_sz;
849 size_t occ = r->rem_set()->occupied();
850 _occupied += occ;
851 return false;
852 }
853 size_t total_mem_sz() { return _total_mem_sz; }
854 size_t max_mem_sz() { return _max_mem_sz; }
855 size_t occupied() { return _occupied; }
856 HeapRegion* max_mem_sz_region() { return _max_mem_sz_region; }
857 };
859 class PrintRSThreadVTimeClosure : public ThreadClosure {
860 public:
861 virtual void do_thread(Thread *t) {
862 ConcurrentG1RefineThread* crt = (ConcurrentG1RefineThread*) t;
863 gclog_or_tty->print(" %5.2f", crt->vtime_accum());
864 }
865 };
867 void HRInto_G1RemSet::print_summary_info() {
868 G1CollectedHeap* g1 = G1CollectedHeap::heap();
870 #if CARD_REPEAT_HISTO
871 gclog_or_tty->print_cr("\nG1 card_repeat count histogram: ");
872 gclog_or_tty->print_cr(" # of repeats --> # of cards with that number.");
873 card_repeat_count.print_on(gclog_or_tty);
874 #endif
876 if (FILTEROUTOFREGIONCLOSURE_DOHISTOGRAMCOUNT) {
877 gclog_or_tty->print_cr("\nG1 rem-set out-of-region histogram: ");
878 gclog_or_tty->print_cr(" # of CS ptrs --> # of cards with that number.");
879 out_of_histo.print_on(gclog_or_tty);
880 }
881 gclog_or_tty->print_cr("\n Concurrent RS processed %d cards",
882 _conc_refine_cards);
883 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
884 jint tot_processed_buffers =
885 dcqs.processed_buffers_mut() + dcqs.processed_buffers_rs_thread();
886 gclog_or_tty->print_cr(" Of %d completed buffers:", tot_processed_buffers);
887 gclog_or_tty->print_cr(" %8d (%5.1f%%) by conc RS threads.",
888 dcqs.processed_buffers_rs_thread(),
889 100.0*(float)dcqs.processed_buffers_rs_thread()/
890 (float)tot_processed_buffers);
891 gclog_or_tty->print_cr(" %8d (%5.1f%%) by mutator threads.",
892 dcqs.processed_buffers_mut(),
893 100.0*(float)dcqs.processed_buffers_mut()/
894 (float)tot_processed_buffers);
895 gclog_or_tty->print_cr(" Conc RS threads times(s)");
896 PrintRSThreadVTimeClosure p;
897 gclog_or_tty->print(" ");
898 g1->concurrent_g1_refine()->threads_do(&p);
899 gclog_or_tty->print_cr("");
901 if (G1UseHRIntoRS) {
902 HRRSStatsIter blk;
903 g1->heap_region_iterate(&blk);
904 gclog_or_tty->print_cr(" Total heap region rem set sizes = " SIZE_FORMAT "K."
905 " Max = " SIZE_FORMAT "K.",
906 blk.total_mem_sz()/K, blk.max_mem_sz()/K);
907 gclog_or_tty->print_cr(" Static structures = " SIZE_FORMAT "K,"
908 " free_lists = " SIZE_FORMAT "K.",
909 HeapRegionRemSet::static_mem_size()/K,
910 HeapRegionRemSet::fl_mem_size()/K);
911 gclog_or_tty->print_cr(" %d occupied cards represented.",
912 blk.occupied());
913 gclog_or_tty->print_cr(" Max sz region = [" PTR_FORMAT ", " PTR_FORMAT " )"
914 ", cap = " SIZE_FORMAT "K, occ = " SIZE_FORMAT "K.",
915 blk.max_mem_sz_region()->bottom(), blk.max_mem_sz_region()->end(),
916 (blk.max_mem_sz_region()->rem_set()->mem_size() + K - 1)/K,
917 (blk.max_mem_sz_region()->rem_set()->occupied() + K - 1)/K);
918 gclog_or_tty->print_cr(" Did %d coarsenings.",
919 HeapRegionRemSet::n_coarsenings());
921 }
922 }
923 void HRInto_G1RemSet::prepare_for_verify() {
924 if (G1HRRSFlushLogBuffersOnVerify &&
925 (VerifyBeforeGC || VerifyAfterGC)
926 && !_g1->full_collection()) {
927 cleanupHRRS();
928 _g1->set_refine_cte_cl_concurrency(false);
929 if (SafepointSynchronize::is_at_safepoint()) {
930 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
931 dcqs.concatenate_logs();
932 }
933 bool cg1r_use_cache = _cg1r->use_cache();
934 _cg1r->set_use_cache(false);
935 updateRS(0);
936 _cg1r->set_use_cache(cg1r_use_cache);
938 assert(JavaThread::dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
939 }
940 }