Fri, 16 Jul 2010 21:33:21 -0700
6962947: shared TaskQueue statistics
Reviewed-by: tonyp, ysr
1 /*
2 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_psPromotionManager.cpp.incl"
28 PSPromotionManager** PSPromotionManager::_manager_array = NULL;
29 OopStarTaskQueueSet* PSPromotionManager::_stack_array_depth = NULL;
30 OopTaskQueueSet* PSPromotionManager::_stack_array_breadth = NULL;
31 PSOldGen* PSPromotionManager::_old_gen = NULL;
32 MutableSpace* PSPromotionManager::_young_space = NULL;
34 void PSPromotionManager::initialize() {
35 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
36 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
38 _old_gen = heap->old_gen();
39 _young_space = heap->young_gen()->to_space();
41 assert(_manager_array == NULL, "Attempt to initialize twice");
42 _manager_array = NEW_C_HEAP_ARRAY(PSPromotionManager*, ParallelGCThreads+1 );
43 guarantee(_manager_array != NULL, "Could not initialize promotion manager");
45 if (UseDepthFirstScavengeOrder) {
46 _stack_array_depth = new OopStarTaskQueueSet(ParallelGCThreads);
47 guarantee(_stack_array_depth != NULL, "Count not initialize promotion manager");
48 } else {
49 _stack_array_breadth = new OopTaskQueueSet(ParallelGCThreads);
50 guarantee(_stack_array_breadth != NULL, "Count not initialize promotion manager");
51 }
53 // Create and register the PSPromotionManager(s) for the worker threads.
54 for(uint i=0; i<ParallelGCThreads; i++) {
55 _manager_array[i] = new PSPromotionManager();
56 guarantee(_manager_array[i] != NULL, "Could not create PSPromotionManager");
57 if (UseDepthFirstScavengeOrder) {
58 stack_array_depth()->register_queue(i, _manager_array[i]->claimed_stack_depth());
59 } else {
60 stack_array_breadth()->register_queue(i, _manager_array[i]->claimed_stack_breadth());
61 }
62 }
64 // The VMThread gets its own PSPromotionManager, which is not available
65 // for work stealing.
66 _manager_array[ParallelGCThreads] = new PSPromotionManager();
67 guarantee(_manager_array[ParallelGCThreads] != NULL, "Could not create PSPromotionManager");
68 }
70 PSPromotionManager* PSPromotionManager::gc_thread_promotion_manager(int index) {
71 assert(index >= 0 && index < (int)ParallelGCThreads, "index out of range");
72 assert(_manager_array != NULL, "Sanity");
73 return _manager_array[index];
74 }
76 PSPromotionManager* PSPromotionManager::vm_thread_promotion_manager() {
77 assert(_manager_array != NULL, "Sanity");
78 return _manager_array[ParallelGCThreads];
79 }
81 void PSPromotionManager::pre_scavenge() {
82 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
83 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
85 _young_space = heap->young_gen()->to_space();
87 for(uint i=0; i<ParallelGCThreads+1; i++) {
88 manager_array(i)->reset();
89 }
90 }
92 void PSPromotionManager::post_scavenge() {
93 TASKQUEUE_STATS_ONLY(if (PrintGCDetails && ParallelGCVerbose) print_stats());
94 for (uint i = 0; i < ParallelGCThreads + 1; i++) {
95 PSPromotionManager* manager = manager_array(i);
96 if (UseDepthFirstScavengeOrder) {
97 assert(manager->claimed_stack_depth()->is_empty(), "should be empty");
98 } else {
99 assert(manager->claimed_stack_breadth()->is_empty(), "should be empty");
100 }
101 manager->flush_labs();
102 }
103 }
105 #if TASKQUEUE_STATS
106 void
107 PSPromotionManager::print_taskqueue_stats(uint i) const {
108 const TaskQueueStats& stats = depth_first() ?
109 _claimed_stack_depth.stats : _claimed_stack_breadth.stats;
110 tty->print("%3u ", i);
111 stats.print();
112 tty->cr();
113 }
115 void
116 PSPromotionManager::print_local_stats(uint i) const {
117 #define FMT " " SIZE_FORMAT_W(10)
118 tty->print_cr("%3u" FMT FMT FMT FMT, i, _masked_pushes, _masked_steals,
119 _arrays_chunked, _array_chunks_processed);
120 #undef FMT
121 }
123 static const char* const pm_stats_hdr[] = {
124 " --------masked------- arrays array",
125 "thr push steal chunked chunks",
126 "--- ---------- ---------- ---------- ----------"
127 };
129 void
130 PSPromotionManager::print_stats() {
131 const bool df = UseDepthFirstScavengeOrder;
132 tty->print_cr("== GC Task Stats (%s-First), GC %3d", df ? "Depth" : "Breadth",
133 Universe::heap()->total_collections());
135 tty->print("thr "); TaskQueueStats::print_header(1); tty->cr();
136 tty->print("--- "); TaskQueueStats::print_header(2); tty->cr();
137 for (uint i = 0; i < ParallelGCThreads + 1; ++i) {
138 manager_array(i)->print_taskqueue_stats(i);
139 }
141 const uint hlines = sizeof(pm_stats_hdr) / sizeof(pm_stats_hdr[0]);
142 for (uint i = 0; i < hlines; ++i) tty->print_cr(pm_stats_hdr[i]);
143 for (uint i = 0; i < ParallelGCThreads + 1; ++i) {
144 manager_array(i)->print_local_stats(i);
145 }
146 }
148 void
149 PSPromotionManager::reset_stats() {
150 TaskQueueStats& stats = depth_first() ?
151 claimed_stack_depth()->stats : claimed_stack_breadth()->stats;
152 stats.reset();
153 _masked_pushes = _masked_steals = 0;
154 _arrays_chunked = _array_chunks_processed = 0;
155 }
156 #endif // TASKQUEUE_STATS
158 PSPromotionManager::PSPromotionManager() {
159 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
160 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
161 _depth_first = UseDepthFirstScavengeOrder;
163 // We set the old lab's start array.
164 _old_lab.set_start_array(old_gen()->start_array());
166 uint queue_size;
167 if (depth_first()) {
168 claimed_stack_depth()->initialize();
169 queue_size = claimed_stack_depth()->max_elems();
170 } else {
171 claimed_stack_breadth()->initialize();
172 queue_size = claimed_stack_breadth()->max_elems();
173 }
175 _totally_drain = (ParallelGCThreads == 1) || (GCDrainStackTargetSize == 0);
176 if (_totally_drain) {
177 _target_stack_size = 0;
178 } else {
179 // don't let the target stack size to be more than 1/4 of the entries
180 _target_stack_size = (uint) MIN2((uint) GCDrainStackTargetSize,
181 (uint) (queue_size / 4));
182 }
184 _array_chunk_size = ParGCArrayScanChunk;
185 // let's choose 1.5x the chunk size
186 _min_array_size_for_chunking = 3 * _array_chunk_size / 2;
188 reset();
189 }
191 void PSPromotionManager::reset() {
192 assert(stacks_empty(), "reset of non-empty stack");
194 // We need to get an assert in here to make sure the labs are always flushed.
196 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
197 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
199 // Do not prefill the LAB's, save heap wastage!
200 HeapWord* lab_base = young_space()->top();
201 _young_lab.initialize(MemRegion(lab_base, (size_t)0));
202 _young_gen_is_full = false;
204 lab_base = old_gen()->object_space()->top();
205 _old_lab.initialize(MemRegion(lab_base, (size_t)0));
206 _old_gen_is_full = false;
208 _prefetch_queue.clear();
210 TASKQUEUE_STATS_ONLY(reset_stats());
211 }
214 void PSPromotionManager::drain_stacks_depth(bool totally_drain) {
215 assert(depth_first(), "invariant");
216 assert(claimed_stack_depth()->overflow_stack() != NULL, "invariant");
217 totally_drain = totally_drain || _totally_drain;
219 #ifdef ASSERT
220 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
221 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
222 MutableSpace* to_space = heap->young_gen()->to_space();
223 MutableSpace* old_space = heap->old_gen()->object_space();
224 MutableSpace* perm_space = heap->perm_gen()->object_space();
225 #endif /* ASSERT */
227 OopStarTaskQueue* const tq = claimed_stack_depth();
228 do {
229 StarTask p;
231 // Drain overflow stack first, so other threads can steal from
232 // claimed stack while we work.
233 while (tq->pop_overflow(p)) {
234 process_popped_location_depth(p);
235 }
237 if (totally_drain) {
238 while (tq->pop_local(p)) {
239 process_popped_location_depth(p);
240 }
241 } else {
242 while (tq->size() > _target_stack_size && tq->pop_local(p)) {
243 process_popped_location_depth(p);
244 }
245 }
246 } while (totally_drain && !tq->taskqueue_empty() || !tq->overflow_empty());
248 assert(!totally_drain || tq->taskqueue_empty(), "Sanity");
249 assert(totally_drain || tq->size() <= _target_stack_size, "Sanity");
250 assert(tq->overflow_empty(), "Sanity");
251 }
253 void PSPromotionManager::drain_stacks_breadth(bool totally_drain) {
254 assert(!depth_first(), "invariant");
255 assert(claimed_stack_breadth()->overflow_stack() != NULL, "invariant");
256 totally_drain = totally_drain || _totally_drain;
258 #ifdef ASSERT
259 ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
260 assert(heap->kind() == CollectedHeap::ParallelScavengeHeap, "Sanity");
261 MutableSpace* to_space = heap->young_gen()->to_space();
262 MutableSpace* old_space = heap->old_gen()->object_space();
263 MutableSpace* perm_space = heap->perm_gen()->object_space();
264 #endif /* ASSERT */
266 OverflowTaskQueue<oop>* const tq = claimed_stack_breadth();
267 do {
268 oop obj;
270 // Drain overflow stack first, so other threads can steal from
271 // claimed stack while we work.
272 while (tq->pop_overflow(obj)) {
273 obj->copy_contents(this);
274 }
276 if (totally_drain) {
277 while (tq->pop_local(obj)) {
278 obj->copy_contents(this);
279 }
280 } else {
281 while (tq->size() > _target_stack_size && tq->pop_local(obj)) {
282 obj->copy_contents(this);
283 }
284 }
286 // If we could not find any other work, flush the prefetch queue
287 if (tq->is_empty()) {
288 flush_prefetch_queue();
289 }
290 } while (totally_drain && !tq->taskqueue_empty() || !tq->overflow_empty());
292 assert(!totally_drain || tq->taskqueue_empty(), "Sanity");
293 assert(totally_drain || tq->size() <= _target_stack_size, "Sanity");
294 assert(tq->overflow_empty(), "Sanity");
295 }
297 void PSPromotionManager::flush_labs() {
298 assert(stacks_empty(), "Attempt to flush lab with live stack");
300 // If either promotion lab fills up, we can flush the
301 // lab but not refill it, so check first.
302 assert(!_young_lab.is_flushed() || _young_gen_is_full, "Sanity");
303 if (!_young_lab.is_flushed())
304 _young_lab.flush();
306 assert(!_old_lab.is_flushed() || _old_gen_is_full, "Sanity");
307 if (!_old_lab.is_flushed())
308 _old_lab.flush();
310 // Let PSScavenge know if we overflowed
311 if (_young_gen_is_full) {
312 PSScavenge::set_survivor_overflow(true);
313 }
314 }
316 //
317 // This method is pretty bulky. It would be nice to split it up
318 // into smaller submethods, but we need to be careful not to hurt
319 // performance.
320 //
322 oop PSPromotionManager::copy_to_survivor_space(oop o, bool depth_first) {
323 assert(PSScavenge::should_scavenge(&o), "Sanity");
325 oop new_obj = NULL;
327 // NOTE! We must be very careful with any methods that access the mark
328 // in o. There may be multiple threads racing on it, and it may be forwarded
329 // at any time. Do not use oop methods for accessing the mark!
330 markOop test_mark = o->mark();
332 // The same test as "o->is_forwarded()"
333 if (!test_mark->is_marked()) {
334 bool new_obj_is_tenured = false;
335 size_t new_obj_size = o->size();
337 // Find the objects age, MT safe.
338 int age = (test_mark->has_displaced_mark_helper() /* o->has_displaced_mark() */) ?
339 test_mark->displaced_mark_helper()->age() : test_mark->age();
341 // Try allocating obj in to-space (unless too old)
342 if (age < PSScavenge::tenuring_threshold()) {
343 new_obj = (oop) _young_lab.allocate(new_obj_size);
344 if (new_obj == NULL && !_young_gen_is_full) {
345 // Do we allocate directly, or flush and refill?
346 if (new_obj_size > (YoungPLABSize / 2)) {
347 // Allocate this object directly
348 new_obj = (oop)young_space()->cas_allocate(new_obj_size);
349 } else {
350 // Flush and fill
351 _young_lab.flush();
353 HeapWord* lab_base = young_space()->cas_allocate(YoungPLABSize);
354 if (lab_base != NULL) {
355 _young_lab.initialize(MemRegion(lab_base, YoungPLABSize));
356 // Try the young lab allocation again.
357 new_obj = (oop) _young_lab.allocate(new_obj_size);
358 } else {
359 _young_gen_is_full = true;
360 }
361 }
362 }
363 }
365 // Otherwise try allocating obj tenured
366 if (new_obj == NULL) {
367 #ifndef PRODUCT
368 if (Universe::heap()->promotion_should_fail()) {
369 return oop_promotion_failed(o, test_mark);
370 }
371 #endif // #ifndef PRODUCT
373 new_obj = (oop) _old_lab.allocate(new_obj_size);
374 new_obj_is_tenured = true;
376 if (new_obj == NULL) {
377 if (!_old_gen_is_full) {
378 // Do we allocate directly, or flush and refill?
379 if (new_obj_size > (OldPLABSize / 2)) {
380 // Allocate this object directly
381 new_obj = (oop)old_gen()->cas_allocate(new_obj_size);
382 } else {
383 // Flush and fill
384 _old_lab.flush();
386 HeapWord* lab_base = old_gen()->cas_allocate(OldPLABSize);
387 if(lab_base != NULL) {
388 _old_lab.initialize(MemRegion(lab_base, OldPLABSize));
389 // Try the old lab allocation again.
390 new_obj = (oop) _old_lab.allocate(new_obj_size);
391 }
392 }
393 }
395 // This is the promotion failed test, and code handling.
396 // The code belongs here for two reasons. It is slightly
397 // different thatn the code below, and cannot share the
398 // CAS testing code. Keeping the code here also minimizes
399 // the impact on the common case fast path code.
401 if (new_obj == NULL) {
402 _old_gen_is_full = true;
403 return oop_promotion_failed(o, test_mark);
404 }
405 }
406 }
408 assert(new_obj != NULL, "allocation should have succeeded");
410 // Copy obj
411 Copy::aligned_disjoint_words((HeapWord*)o, (HeapWord*)new_obj, new_obj_size);
413 // Now we have to CAS in the header.
414 if (o->cas_forward_to(new_obj, test_mark)) {
415 // We won any races, we "own" this object.
416 assert(new_obj == o->forwardee(), "Sanity");
418 // Increment age if obj still in new generation. Now that
419 // we're dealing with a markOop that cannot change, it is
420 // okay to use the non mt safe oop methods.
421 if (!new_obj_is_tenured) {
422 new_obj->incr_age();
423 assert(young_space()->contains(new_obj), "Attempt to push non-promoted obj");
424 }
426 if (depth_first) {
427 // Do the size comparison first with new_obj_size, which we
428 // already have. Hopefully, only a few objects are larger than
429 // _min_array_size_for_chunking, and most of them will be arrays.
430 // So, the is->objArray() test would be very infrequent.
431 if (new_obj_size > _min_array_size_for_chunking &&
432 new_obj->is_objArray() &&
433 PSChunkLargeArrays) {
434 // we'll chunk it
435 oop* const masked_o = mask_chunked_array_oop(o);
436 push_depth(masked_o);
437 TASKQUEUE_STATS_ONLY(++_arrays_chunked; ++_masked_pushes);
438 } else {
439 // we'll just push its contents
440 new_obj->push_contents(this);
441 }
442 } else {
443 push_breadth(new_obj);
444 }
445 } else {
446 // We lost, someone else "owns" this object
447 guarantee(o->is_forwarded(), "Object must be forwarded if the cas failed.");
449 // Try to deallocate the space. If it was directly allocated we cannot
450 // deallocate it, so we have to test. If the deallocation fails,
451 // overwrite with a filler object.
452 if (new_obj_is_tenured) {
453 if (!_old_lab.unallocate_object(new_obj)) {
454 CollectedHeap::fill_with_object((HeapWord*) new_obj, new_obj_size);
455 }
456 } else if (!_young_lab.unallocate_object(new_obj)) {
457 CollectedHeap::fill_with_object((HeapWord*) new_obj, new_obj_size);
458 }
460 // don't update this before the unallocation!
461 new_obj = o->forwardee();
462 }
463 } else {
464 assert(o->is_forwarded(), "Sanity");
465 new_obj = o->forwardee();
466 }
468 #ifdef DEBUG
469 // This code must come after the CAS test, or it will print incorrect
470 // information.
471 if (TraceScavenge) {
472 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (" SIZE_FORMAT ")}",
473 PSScavenge::should_scavenge(&new_obj) ? "copying" : "tenuring",
474 new_obj->blueprint()->internal_name(), o, new_obj, new_obj->size());
475 }
476 #endif
478 return new_obj;
479 }
481 template <class T> void PSPromotionManager::process_array_chunk_work(
482 oop obj,
483 int start, int end) {
484 assert(start < end, "invariant");
485 T* const base = (T*)objArrayOop(obj)->base();
486 T* p = base + start;
487 T* const chunk_end = base + end;
488 while (p < chunk_end) {
489 if (PSScavenge::should_scavenge(p)) {
490 claim_or_forward_depth(p);
491 }
492 ++p;
493 }
494 }
496 void PSPromotionManager::process_array_chunk(oop old) {
497 assert(PSChunkLargeArrays, "invariant");
498 assert(old->is_objArray(), "invariant");
499 assert(old->is_forwarded(), "invariant");
501 TASKQUEUE_STATS_ONLY(++_array_chunks_processed);
503 oop const obj = old->forwardee();
505 int start;
506 int const end = arrayOop(old)->length();
507 if (end > (int) _min_array_size_for_chunking) {
508 // we'll chunk more
509 start = end - _array_chunk_size;
510 assert(start > 0, "invariant");
511 arrayOop(old)->set_length(start);
512 push_depth(mask_chunked_array_oop(old));
513 TASKQUEUE_STATS_ONLY(++_masked_pushes);
514 } else {
515 // this is the final chunk for this array
516 start = 0;
517 int const actual_length = arrayOop(obj)->length();
518 arrayOop(old)->set_length(actual_length);
519 }
521 if (UseCompressedOops) {
522 process_array_chunk_work<narrowOop>(obj, start, end);
523 } else {
524 process_array_chunk_work<oop>(obj, start, end);
525 }
526 }
528 oop PSPromotionManager::oop_promotion_failed(oop obj, markOop obj_mark) {
529 assert(_old_gen_is_full || PromotionFailureALot, "Sanity");
531 // Attempt to CAS in the header.
532 // This tests if the header is still the same as when
533 // this started. If it is the same (i.e., no forwarding
534 // pointer has been installed), then this thread owns
535 // it.
536 if (obj->cas_forward_to(obj, obj_mark)) {
537 // We won any races, we "own" this object.
538 assert(obj == obj->forwardee(), "Sanity");
540 if (depth_first()) {
541 obj->push_contents(this);
542 } else {
543 // Don't bother incrementing the age, just push
544 // onto the claimed_stack..
545 push_breadth(obj);
546 }
548 // Save the mark if needed
549 PSScavenge::oop_promotion_failed(obj, obj_mark);
550 } else {
551 // We lost, someone else "owns" this object
552 guarantee(obj->is_forwarded(), "Object must be forwarded if the cas failed.");
554 // No unallocation to worry about.
555 obj = obj->forwardee();
556 }
558 #ifdef DEBUG
559 if (TraceScavenge) {
560 gclog_or_tty->print_cr("{%s %s 0x%x (%d)}",
561 "promotion-failure",
562 obj->blueprint()->internal_name(),
563 obj, obj->size());
565 }
566 #endif
568 return obj;
569 }