代码位置:
gc_interface 接口和gc_implementation
这是jdk1.7 几种垃圾回收器
先看公共接口类中的 代码:collectedHeap.cpp
/*
* Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "classfile/systemDictionary.hpp"
#include "gc_implementation/shared/vmGCOperations.hpp"
#include "gc_interface/collectedHeap.hpp"
#include "gc_interface/collectedHeap.inline.hpp"
#include "oops/oop.inline.hpp"
#include "runtime/init.hpp"
#include "services/heapDumper.hpp"
#ifdef TARGET_OS_FAMILY_linux
# include "thread_linux.inline.hpp"
#endif
#ifdef TARGET_OS_FAMILY_solaris
# include "thread_solaris.inline.hpp"
#endif
#ifdef TARGET_OS_FAMILY_windows
# include "thread_windows.inline.hpp"
#endif
#ifdef ASSERT
int CollectedHeap::_fire_out_of_memory_count = 0;
#endif
size_t CollectedHeap::_filler_array_max_size = 0;
// Memory state functions.
CollectedHeap::CollectedHeap() : _n_par_threads(0)
{
const size_t max_len = size_t(arrayOopDesc::max_array_length(T_INT));
const size_t elements_per_word = HeapWordSize / sizeof(jint);
_filler_array_max_size = align_object_size(filler_array_hdr_size() +
max_len * elements_per_word);
_barrier_set = NULL;
_is_gc_active = false;
_total_collections = _total_full_collections = 0;
_gc_cause = _gc_lastcause = GCCause::_no_gc;
NOT_PRODUCT(_promotion_failure_alot_count = 0;)
NOT_PRODUCT(_promotion_failure_alot_gc_number = 0;)
if (UsePerfData) {
EXCEPTION_MARK;
// create the gc cause jvmstat counters
_perf_gc_cause = PerfDataManager::create_string_variable(SUN_GC, "cause",
80, GCCause::to_string(_gc_cause), CHECK);
_perf_gc_lastcause =
PerfDataManager::create_string_variable(SUN_GC, "lastCause",
80, GCCause::to_string(_gc_lastcause), CHECK);
}
_defer_initial_card_mark = false; // strengthened by subclass in pre_initialize() below.
}
void CollectedHeap::pre_initialize() {
// Used for ReduceInitialCardMarks (when COMPILER2 is used);
// otherwise remains unused.
#ifdef COMPILER2
_defer_initial_card_mark = ReduceInitialCardMarks && can_elide_tlab_store_barriers()
&& (DeferInitialCardMark || card_mark_must_follow_store());
#else
assert(_defer_initial_card_mark == false, "Who would set it?");
#endif
}
#ifndef PRODUCT
void CollectedHeap::check_for_bad_heap_word_value(HeapWord* addr, size_t size) {
if (CheckMemoryInitialization && ZapUnusedHeapArea) {
for (size_t slot = 0; slot < size; slot += 1) {
assert((*(intptr_t*) (addr + slot)) != ((intptr_t) badHeapWordVal),
"Found badHeapWordValue in post-allocation check");
}
}
}
void CollectedHeap::check_for_non_bad_heap_word_value(HeapWord* addr, size_t size) {
if (CheckMemoryInitialization && ZapUnusedHeapArea) {
for (size_t slot = 0; slot < size; slot += 1) {
assert((*(intptr_t*) (addr + slot)) == ((intptr_t) badHeapWordVal),
"Found non badHeapWordValue in pre-allocation check");
}
}
}
#endif // PRODUCT
#ifdef ASSERT
void CollectedHeap::check_for_valid_allocation_state() {
Thread *thread = Thread::current();
// How to choose between a pending exception and a potential
// OutOfMemoryError? Don't allow pending exceptions.
// This is a VM policy failure, so how do we exhaustively test it?
assert(!thread->has_pending_exception(),
"shouldn't be allocating with pending exception");
if (StrictSafepointChecks) {
assert(thread->allow_allocation(),
"Allocation done by thread for which allocation is blocked "
"by No_Allocation_Verifier!");
// Allocation of an oop can always invoke a safepoint,
// hence, the true argument
thread->check_for_valid_safepoint_state(true);
}
}
#endif
HeapWord* CollectedHeap::allocate_from_tlab_slow(Thread* thread, size_t size) {
// Retain tlab and allocate object in shared space if
// the amount free in the tlab is too large to discard.
if (thread->tlab().free() > thread->tlab().refill_waste_limit()) {
thread->tlab().record_slow_allocation(size);
return NULL;
}
// Discard tlab and allocate a new one.
// To minimize fragmentation, the last TLAB may be smaller than the rest.
size_t new_tlab_size = thread->tlab().compute_size(size);
thread->tlab().clear_before_allocation();
if (new_tlab_size == 0) {
return NULL;
}
// Allocate a new TLAB...
HeapWord* obj = Universe::heap()->allocate_new_tlab(new_tlab_size);
if (obj == NULL) {
return NULL;
}
if (ZeroTLAB) {
// ..and clear it.
Copy::zero_to_words(obj, new_tlab_size);
} else {
// ...and clear just the allocated object.
Copy::zero_to_words(obj, size);
}
thread->tlab().fill(obj, obj + size, new_tlab_size);
return obj;
}
void CollectedHeap::flush_deferred_store_barrier(JavaThread* thread) {
MemRegion deferred = thread->deferred_card_mark();
if (!deferred.is_empty()) {
assert(_defer_initial_card_mark, "Otherwise should be empty");
{
// Verify that the storage points to a parsable object in heap
DEBUG_ONLY(oop old_obj = oop(deferred.start());)
assert(is_in(old_obj), "Not in allocated heap");
assert(!can_elide_initializing_store_barrier(old_obj),
"Else should have been filtered in new_store_pre_barrier()");
assert(!is_in_permanent(old_obj), "Sanity: not expected");
assert(old_obj->is_oop(true), "Not an oop");
assert(old_obj->is_parsable(), "Will not be concurrently parsable");
assert(deferred.word_size() == (size_t)(old_obj->size()),
"Mismatch: multiple objects?");
}
BarrierSet* bs = barrier_set();
assert(bs->has_write_region_opt(), "No write_region() on BarrierSet");
bs->write_region(deferred);
// "Clear" the deferred_card_mark field
thread->set_deferred_card_mark(MemRegion());
}
assert(thread->deferred_card_mark().is_empty(), "invariant");
}
// Helper for ReduceInitialCardMarks. For performance,
// compiled code may elide card-marks for initializing stores
// to a newly allocated object along the fast-path. We
// compensate for such elided card-marks as follows:
// (a) Generational, non-concurrent collectors, such as
// GenCollectedHeap(ParNew,DefNew,Tenured) and
// ParallelScavengeHeap(ParallelGC, ParallelOldGC)
// need the card-mark if and only if the region is
// in the old gen, and do not care if the card-mark
// succeeds or precedes the initializing stores themselves,
// so long as the card-mark is completed before the next
// scavenge. For all these cases, we can do a card mark
// at the point at which we do a slow path allocation
// in the old gen, i.e. in this call.
// (b) GenCollectedHeap(ConcurrentMarkSweepGeneration) requires
// in addition that the card-mark for an old gen allocated
// object strictly follow any associated initializing stores.
// In these cases, the memRegion remembered below is
// used to card-mark the entire region either just before the next
// slow-path allocation by this thread or just before the next scavenge or
// CMS-associated safepoint, whichever of these events happens first.
// (The implicit assumption is that the object has been fully
// initialized by this point, a fact that we assert when doing the
// card-mark.)
// (c) G1CollectedHeap(G1) uses two kinds of write barriers. When a
// G1 concurrent marking is in progress an SATB (pre-write-)barrier is
// is used to remember the pre-value of any store. Initializing
// stores will not need this barrier, so we need not worry about
// compensating for the missing pre-barrier here. Turning now
// to the post-barrier, we note that G1 needs a RS update barrier
// which simply enqueues a (sequence of) dirty cards which may
// optionally be refined by the concurrent update threads. Note
// that this barrier need only be applied to a non-young write,
// but, like in CMS, because of the presence of concurrent refinement
// (much like CMS' precleaning), must strictly follow the oop-store.
// Thus, using the same protocol for maintaining the intended
// invariants turns out, serendepitously, to be the same for both
// G1 and CMS.
//
// For any future collector, this code should be reexamined with
// that specific collector in mind, and the documentation above suitably
// extended and updated.
oop CollectedHeap::new_store_pre_barrier(JavaThread* thread, oop new_obj) {
// If a previous card-mark was deferred, flush it now.
flush_deferred_store_barrier(thread);
if (can_elide_initializing_store_barrier(new_obj)) {
// The deferred_card_mark region should be empty
// following the flush above.
assert(thread->deferred_card_mark().is_empty(), "Error");
} else {
MemRegion mr((HeapWord*)new_obj, new_obj->size());
assert(!mr.is_empty(), "Error");
if (_defer_initial_card_mark) {
// Defer the card mark
thread->set_deferred_card_mark(mr);
} else {
// Do the card mark
BarrierSet* bs = barrier_set();
assert(bs->has_write_region_opt(), "No write_region() on BarrierSet");
bs->write_region(mr);
}
}
return new_obj;
}
size_t CollectedHeap::filler_array_hdr_size() {
return size_t(align_object_offset(arrayOopDesc::header_size(T_INT))); // align to Long
}
size_t CollectedHeap::filler_array_min_size() {
return align_object_size(filler_array_hdr_size()); // align to MinObjAlignment
}
size_t CollectedHeap::filler_array_max_size() {
return _filler_array_max_size;
}
#ifdef ASSERT
void CollectedHeap::fill_args_check(HeapWord* start, size_t words)
{
assert(words >= min_fill_size(), "too small to fill");
assert(words % MinObjAlignment == 0, "unaligned size");
assert(Universe::heap()->is_in_reserved(start), "not in heap");
assert(Universe::heap()->is_in_reserved(start + words - 1), "not in heap");
}
void CollectedHeap::zap_filler_array(HeapWord* start, size_t words, bool zap)
{
if (ZapFillerObjects && zap) {
Copy::fill_to_words(start + filler_array_hdr_size(),
words - filler_array_hdr_size(), 0XDEAFBABE);
}
}
#endif // ASSERT
void
CollectedHeap::fill_with_array(HeapWord* start, size_t words, bool zap)
{
assert(words >= filler_array_min_size(), "too small for an array");
assert(words <= filler_array_max_size(), "too big for a single object");
const size_t payload_size = words - filler_array_hdr_size();
const size_t len = payload_size * HeapWordSize / sizeof(jint);
// Set the length first for concurrent GC.
((arrayOop)start)->set_length((int)len);
post_allocation_setup_common(Universe::intArrayKlassObj(), start, words);
DEBUG_ONLY(zap_filler_array(start, words, zap);)
}
void
CollectedHeap::fill_with_object_impl(HeapWord* start, size_t words, bool zap)
{
assert(words <= filler_array_max_size(), "too big for a single object");
if (words >= filler_array_min_size()) {
fill_with_array(start, words, zap);
} else if (words > 0) {
assert(words == min_fill_size(), "unaligned size");
post_allocation_setup_common(SystemDictionary::Object_klass(), start,
words);
}
}
void CollectedHeap::fill_with_object(HeapWord* start, size_t words, bool zap)
{
DEBUG_ONLY(fill_args_check(start, words);)
HandleMark hm; // Free handles before leaving.
fill_with_object_impl(start, words, zap);
}
void CollectedHeap::fill_with_objects(HeapWord* start, size_t words, bool zap)
{
DEBUG_ONLY(fill_args_check(start, words);)
HandleMark hm; // Free handles before leaving.
#ifdef _LP64
// A single array can fill ~8G, so multiple objects are needed only in 64-bit.
// First fill with arrays, ensuring that any remaining space is big enough to
// fill. The remainder is filled with a single object.
const size_t min = min_fill_size();
const size_t max = filler_array_max_size();
while (words > max) {
const size_t cur = words - max >= min ? max : max - min;
fill_with_array(start, cur, zap);
start += cur;
words -= cur;
}
#endif
fill_with_object_impl(start, words, zap);
}
HeapWord* CollectedHeap::allocate_new_tlab(size_t size) {
guarantee(false, "thread-local allocation buffers not supported");
return NULL;
}
void CollectedHeap::ensure_parsability(bool retire_tlabs) {
// The second disjunct in the assertion below makes a concession
// for the start-up verification done while the VM is being
// created. Callers be careful that you know that mutators
// aren't going to interfere -- for instance, this is permissible
// if we are still single-threaded and have either not yet
// started allocating (nothing much to verify) or we have
// started allocating but are now a full-fledged JavaThread
// (and have thus made our TLAB's) available for filling.
assert(SafepointSynchronize::is_at_safepoint() ||
!is_init_completed(),
"Should only be called at a safepoint or at start-up"
" otherwise concurrent mutator activity may make heap "
" unparsable again");
const bool use_tlab = UseTLAB;
const bool deferred = _defer_initial_card_mark;
// The main thread starts allocating via a TLAB even before it
// has added itself to the threads list at vm boot-up.
assert(!use_tlab || Threads::first() != NULL,
"Attempt to fill tlabs before main thread has been added"
" to threads list is doomed to failure!");
for (JavaThread *thread = Threads::first(); thread; thread = thread->next()) {
if (use_tlab) thread->tlab().make_parsable(retire_tlabs);
#ifdef COMPILER2
// The deferred store barriers must all have been flushed to the
// card-table (or other remembered set structure) before GC starts
// processing the card-table (or other remembered set).
if (deferred) flush_deferred_store_barrier(thread);
#else
assert(!deferred, "Should be false");
assert(thread->deferred_card_mark().is_empty(), "Should be empty");
#endif
}
}
void CollectedHeap::accumulate_statistics_all_tlabs() {
if (UseTLAB) {
assert(SafepointSynchronize::is_at_safepoint() ||
!is_init_completed(),
"should only accumulate statistics on tlabs at safepoint");
ThreadLocalAllocBuffer::accumulate_statistics_before_gc();
}
}
void CollectedHeap::resize_all_tlabs() {
if (UseTLAB) {
assert(SafepointSynchronize::is_at_safepoint() ||
!is_init_completed(),
"should only resize tlabs at safepoint");
ThreadLocalAllocBuffer::resize_all_tlabs();
}
}
void CollectedHeap::pre_full_gc_dump() {
if (HeapDumpBeforeFullGC) {
TraceTime tt("Heap Dump: ", PrintGCDetails, false, gclog_or_tty);
// We are doing a "major" collection and a heap dump before
// major collection has been requested.
HeapDumper::dump_heap();
}
if (PrintClassHistogramBeforeFullGC) {
TraceTime tt("Class Histogram: ", PrintGCDetails, true, gclog_or_tty);
VM_GC_HeapInspection inspector(gclog_or_tty, false /* ! full gc */, false /* ! prologue */);
inspector.doit();
}
}
void CollectedHeap::post_full_gc_dump() {
if (HeapDumpAfterFullGC) {
TraceTime tt("Heap Dump", PrintGCDetails, false, gclog_or_tty);
HeapDumper::dump_heap();
}
if (PrintClassHistogramAfterFullGC) {
TraceTime tt("Class Histogram", PrintGCDetails, true, gclog_or_tty);
VM_GC_HeapInspection inspector(gclog_or_tty, false /* ! full gc */, false /* ! prologue */);
inspector.doit();
}
}
gcCause.cpp
Copyright (c) 2002, 2011, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "gc_interface/gcCause.hpp"
const char* GCCause::to_string(GCCause::Cause cause) {
switch (cause) {
case _java_lang_system_gc:
return "System.gc()";
case _full_gc_alot:
return "FullGCAlot";
case _scavenge_alot:
return "ScavengeAlot";
case _allocation_profiler:
return "Allocation Profiler";
case _jvmti_force_gc:
return "JvmtiEnv ForceGarbageCollection";
case _no_gc:
return "No GC";
case _allocation_failure:
return "Allocation Failure";
case _gc_locker:
return "GCLocker Initiated GC";
case _heap_inspection:
return "Heap Inspection Initiated GC";
case _heap_dump:
return "Heap Dump Initiated GC";
case _tenured_generation_full:
return "Tenured Generation Full";
case _permanent_generation_full:
return "Permanent Generation Full";
case _cms_generation_full:
return "CMS Generation Full";
case _cms_initial_mark:
return "CMS Initial Mark";
case _cms_final_remark:
return "CMS Final Remark";
case _old_generation_expanded_on_last_scavenge:
return "Old Generation Expanded On Last Scavenge";
case _old_generation_too_full_to_scavenge:
return "Old Generation Too Full To Scavenge";
case _g1_inc_collection_pause:
return "G1 Evacuation Pause";
case _last_ditch_collection:
return "Last ditch collection";
case _last_gc_cause:
return "ILLEGAL VALUE - last gc cause - ILLEGAL VALUE";
default:
return "unknown GCCause";
}
ShouldNotReachHere();
}
parNew 垃圾回收器:
asParNewGeneration.cpp 代码:
/*
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp"
#include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp"
#include "gc_implementation/parNew/asParNewGeneration.hpp"
#include "gc_implementation/parNew/parNewGeneration.hpp"
#include "gc_implementation/shared/markSweep.inline.hpp"
#include "gc_implementation/shared/spaceDecorator.hpp"
#include "memory/defNewGeneration.inline.hpp"
#include "memory/referencePolicy.hpp"
#include "oops/markOop.inline.hpp"
#include "oops/oop.pcgc.inline.hpp"
ASParNewGeneration::ASParNewGeneration(ReservedSpace rs,
size_t initial_byte_size,
size_t min_byte_size,
int level) :
ParNewGeneration(rs, initial_byte_size, level),
_min_gen_size(min_byte_size) {}
const char* ASParNewGeneration::name() const {
return "adaptive size par new generation";
}
void ASParNewGeneration::adjust_desired_tenuring_threshold() {
assert(UseAdaptiveSizePolicy,
"Should only be used with UseAdaptiveSizePolicy");
}
void ASParNewGeneration::resize(size_t eden_size, size_t survivor_size) {
// Resize the generation if needed. If the generation resize
// reports false, do not attempt to resize the spaces.
if (resize_generation(eden_size, survivor_size)) {
// Then we lay out the spaces inside the generation
resize_spaces(eden_size, survivor_size);
space_invariants();
if (PrintAdaptiveSizePolicy && Verbose) {
gclog_or_tty->print_cr("Young generation size: "
"desired eden: " SIZE_FORMAT " survivor: " SIZE_FORMAT
" used: " SIZE_FORMAT " capacity: " SIZE_FORMAT
" gen limits: " SIZE_FORMAT " / " SIZE_FORMAT,
eden_size, survivor_size, used(), capacity(),
max_gen_size(), min_gen_size());
}
}
}
size_t ASParNewGeneration::available_to_min_gen() {
assert(virtual_space()->committed_size() >= min_gen_size(), "Invariant");
return virtual_space()->committed_size() - min_gen_size();
}
// This method assumes that from-space has live data and that
// any shrinkage of the young gen is limited by location of
// from-space.
size_t ASParNewGeneration::available_to_live() const {
#undef SHRINKS_AT_END_OF_EDEN
#ifdef SHRINKS_AT_END_OF_EDEN
size_t delta_in_survivor = 0;
ParallelScavengeHeap* heap = (ParallelScavengeHeap*)Universe::heap();
const size_t space_alignment = heap->intra_heap_alignment();
const size_t gen_alignment = heap->object_heap_alignment();
MutableSpace* space_shrinking = NULL;
if (from_space()->end() > to_space()->end()) {
space_shrinking = from_space();
} else {
space_shrinking = to_space();
}
// Include any space that is committed but not included in
// the survivor spaces.
assert(((HeapWord*)virtual_space()->high()) >= space_shrinking->end(),
"Survivor space beyond high end");
size_t unused_committed = pointer_delta(virtual_space()->high(),
space_shrinking->end(), sizeof(char));
if (space_shrinking->is_empty()) {
// Don't let the space shrink to 0
assert(space_shrinking->capacity_in_bytes() >= space_alignment,
"Space is too small");
delta_in_survivor = space_shrinking->capacity_in_bytes() - space_alignment;
} else {
delta_in_survivor = pointer_delta(space_shrinking->end(),
space_shrinking->top(),
sizeof(char));
}
size_t delta_in_bytes = unused_committed + delta_in_survivor;
delta_in_bytes = align_size_down(delta_in_bytes, gen_alignment);
return delta_in_bytes;
#else
// The only space available for shrinking is in to-space if it
// is above from-space.
if (to()->bottom() > from()->bottom()) {
const size_t alignment = os::vm_page_size();
if (to()->capacity() < alignment) {
return 0;
} else {
return to()->capacity() - alignment;
}
} else {
return 0;
}
#endif
}
// Return the number of bytes available for resizing down the young
// generation. This is the minimum of
// input "bytes"
// bytes to the minimum young gen size
// bytes to the size currently being used + some small extra
size_t ASParNewGeneration::limit_gen_shrink (size_t bytes) {
// Allow shrinkage into the current eden but keep eden large enough
// to maintain the minimum young gen size
bytes = MIN3(bytes, available_to_min_gen(), available_to_live());
return align_size_down(bytes, os::vm_page_size());
}
// Note that the the alignment used is the OS page size as
// opposed to an alignment associated with the virtual space
// (as is done in the ASPSYoungGen/ASPSOldGen)
bool ASParNewGeneration::resize_generation(size_t eden_size,
size_t survivor_size) {
const size_t alignment = os::vm_page_size();
size_t orig_size = virtual_space()->committed_size();
bool size_changed = false;
// There used to be this guarantee there.
// guarantee ((eden_size + 2*survivor_size) <= _max_gen_size, "incorrect input arguments");
// Code below forces this requirement. In addition the desired eden
// size and disired survivor sizes are desired goals and may
// exceed the total generation size.
assert(min_gen_size() <= orig_size && orig_size <= max_gen_size(),
"just checking");
// Adjust new generation size
const size_t eden_plus_survivors =
align_size_up(eden_size + 2 * survivor_size, alignment);
size_t desired_size = MAX2(MIN2(eden_plus_survivors, max_gen_size()),
min_gen_size());
assert(desired_size <= max_gen_size(), "just checking");
if (desired_size > orig_size) {
// Grow the generation
size_t change = desired_size - orig_size;
assert(change % alignment == 0, "just checking");
if (expand(change)) {
return false; // Error if we fail to resize!
}
size_changed = true;
} else if (desired_size < orig_size) {
size_t desired_change = orig_size - desired_size;
assert(desired_change % alignment == 0, "just checking");
desired_change = limit_gen_shrink(desired_change);
if (desired_change > 0) {
virtual_space()->shrink_by(desired_change);
reset_survivors_after_shrink();
size_changed = true;
}
} else {
if (Verbose && PrintGC) {
if (orig_size == max_gen_size()) {
gclog_or_tty->print_cr("ASParNew generation size at maximum: "
SIZE_FORMAT "K", orig_size/K);
} else if (orig_size == min_gen_size()) {
gclog_or_tty->print_cr("ASParNew generation size at minium: "
SIZE_FORMAT "K", orig_size/K);
}
}
}
if (size_changed) {
MemRegion cmr((HeapWord*)virtual_space()->low(),
(HeapWord*)virtual_space()->high());
GenCollectedHeap::heap()->barrier_set()->resize_covered_region(cmr);
if (Verbose && PrintGC) {
size_t current_size = virtual_space()->committed_size();
gclog_or_tty->print_cr("ASParNew generation size changed: "
SIZE_FORMAT "K->" SIZE_FORMAT "K",
orig_size/K, current_size/K);
}
}
guarantee(eden_plus_survivors <= virtual_space()->committed_size() ||
virtual_space()->committed_size() == max_gen_size(), "Sanity");
return true;
}
void ASParNewGeneration::reset_survivors_after_shrink() {
GenCollectedHeap* gch = GenCollectedHeap::heap();
HeapWord* new_end = (HeapWord*)virtual_space()->high();
if (from()->end() > to()->end()) {
assert(new_end >= from()->end(), "Shrinking past from-space");
} else {
assert(new_end >= to()->bottom(), "Shrink was too large");
// Was there a shrink of the survivor space?
if (new_end < to()->end()) {
MemRegion mr(to()->bottom(), new_end);
to()->initialize(mr,
SpaceDecorator::DontClear,
SpaceDecorator::DontMangle);
}
}
}
void ASParNewGeneration::resize_spaces(size_t requested_eden_size,
size_t requested_survivor_size) {
assert(UseAdaptiveSizePolicy, "sanity check");
assert(requested_eden_size > 0 && requested_survivor_size > 0,
"just checking");
CollectedHeap* heap = Universe::heap();
assert(heap->kind() == CollectedHeap::GenCollectedHeap, "Sanity");
// We require eden and to space to be empty
if ((!eden()->is_empty()) || (!to()->is_empty())) {
return;
}
size_t cur_eden_size = eden()->capacity();
if (PrintAdaptiveSizePolicy && Verbose) {
gclog_or_tty->print_cr("ASParNew::resize_spaces(requested_eden_size: "
SIZE_FORMAT
", requested_survivor_size: " SIZE_FORMAT ")",
requested_eden_size, requested_survivor_size);
gclog_or_tty->print_cr(" eden: [" PTR_FORMAT ".." PTR_FORMAT ") "
SIZE_FORMAT,
eden()->bottom(),
eden()->end(),
pointer_delta(eden()->end(),
eden()->bottom(),
sizeof(char)));
gclog_or_tty->print_cr(" from: [" PTR_FORMAT ".." PTR_FORMAT ") "
SIZE_FORMAT,
from()->bottom(),
from()->end(),
pointer_delta(from()->end(),
from()->bottom(),
sizeof(char)));
gclog_or_tty->print_cr(" to: [" PTR_FORMAT ".." PTR_FORMAT ") "
SIZE_FORMAT,
to()->bottom(),
to()->end(),
pointer_delta( to()->end(),
to()->bottom(),
sizeof(char)));
}
// There's nothing to do if the new sizes are the same as the current
if (requested_survivor_size == to()->capacity() &&
requested_survivor_size == from()->capacity() &&
requested_eden_size == eden()->capacity()) {
if (PrintAdaptiveSizePolicy && Verbose) {
gclog_or_tty->print_cr(" capacities are the right sizes, returning");
}
return;
}
char* eden_start = (char*)eden()->bottom();
char* eden_end = (char*)eden()->end();
char* from_start = (char*)from()->bottom();
char* from_end = (char*)from()->end();
char* to_start = (char*)to()->bottom();
char* to_end = (char*)to()->end();
const size_t alignment = os::vm_page_size();
const bool maintain_minimum =
(requested_eden_size + 2 * requested_survivor_size) <= min_gen_size();
// Check whether from space is below to space
if (from_start < to_start) {
// Eden, from, to
if (PrintAdaptiveSizePolicy && Verbose) {
gclog_or_tty->print_cr(" Eden, from, to:");
}
// Set eden
// "requested_eden_size" is a goal for the size of eden
// and may not be attainable. "eden_size" below is
// calculated based on the location of from-space and
// the goal for the size of eden. from-space is
// fixed in place because it contains live data.
// The calculation is done this way to avoid 32bit
// overflow (i.e., eden_start + requested_eden_size
// may too large for representation in 32bits).
size_t eden_size;
if (maintain_minimum) {
// Only make eden larger than the requested size if
// the minimum size of the generation has to be maintained.
// This could be done in general but policy at a higher
// level is determining a requested size for eden and that
// should be honored unless there is a fundamental reason.
eden_size = pointer_delta(from_start,
eden_start,
sizeof(char));
} else {
eden_size = MIN2(requested_eden_size,
pointer_delta(from_start, eden_start, sizeof(char)));
}
eden_size = align_size_down(eden_size, alignment);
eden_end = eden_start + eden_size;
assert(eden_end >= eden_start, "addition overflowed");
// To may resize into from space as long as it is clear of live data.
// From space must remain page aligned, though, so we need to do some
// extra calculations.
// First calculate an optimal to-space
to_end = (char*)virtual_space()->high();
to_start = (char*)pointer_delta(to_end, (char*)requested_survivor_size,
sizeof(char));
// Does the optimal to-space overlap from-space?
if (to_start < (char*)from()->end()) {
// Calculate the minimum offset possible for from_end
size_t from_size = pointer_delta(from()->top(), from_start, sizeof(char));
// Should we be in this method if from_space is empty? Why not the set_space method? FIX ME!
if (from_size == 0) {
from_size = alignment;
} else {
from_size = align_size_up(from_size, alignment);
}
from_end = from_start + from_size;
assert(from_end > from_start, "addition overflow or from_size problem");
guarantee(from_end <= (char*)from()->end(), "from_end moved to the right");
// Now update to_start with the new from_end
to_start = MAX2(from_end, to_start);
} else {
// If shrinking, move to-space down to abut the end of from-space
// so that shrinking will move to-space down. If not shrinking
// to-space is moving up to allow for growth on the next expansion.
if (requested_eden_size <= cur_eden_size) {
to_start = from_end;
if (to_start + requested_survivor_size > to_start) {
to_end = to_start + requested_survivor_size;
}
}
// else leave to_end pointing to the high end of the virtual space.
}
guarantee(to_start != to_end, "to space is zero sized");
if (PrintAdaptiveSizePolicy && Verbose) {
gclog_or_tty->print_cr(" [eden_start .. eden_end): "
"[" PTR_FORMAT " .. " PTR_FORMAT ") " SIZE_FORMAT,
eden_start,
eden_end,
pointer_delta(eden_end, eden_start, sizeof(char)));
gclog_or_tty->print_cr(" [from_start .. from_end): "
"[" PTR_FORMAT " .. " PTR_FORMAT ") " SIZE_FORMAT,
from_start,
from_end,
pointer_delta(from_end, from_start, sizeof(char)));
gclog_or_tty->print_cr(" [ to_start .. to_end): "
"[" PTR_FORMAT " .. " PTR_FORMAT ") " SIZE_FORMAT,
to_start,
to_end,
pointer_delta( to_end, to_start, sizeof(char)));
}
} else {
// Eden, to, from
if (PrintAdaptiveSizePolicy && Verbose) {
gclog_or_tty->print_cr(" Eden, to, from:");
}
// Calculate the to-space boundaries based on
// the start of from-space.
to_end = from_start;
to_start = (char*)pointer_delta(from_start,
(char*)requested_survivor_size,
sizeof(char));
// Calculate the ideal eden boundaries.
// eden_end is already at the bottom of the generation
assert(eden_start == virtual_space()->low(),
"Eden is not starting at the low end of the virtual space");
if (eden_start + requested_eden_size >= eden_start) {
eden_end = eden_start + requested_eden_size;
} else {
eden_end = to_start;
}
// Does eden intrude into to-space? to-space
// gets priority but eden is not allowed to shrink
// to 0.
if (eden_end > to_start) {
eden_end = to_start;
}
// Don't let eden shrink down to 0 or less.
eden_end = MAX2(eden_end, eden_start + alignment);
assert(eden_start + alignment >= eden_start, "Overflow");
size_t eden_size;
if (maintain_minimum) {
// Use all the space available.
eden_end = MAX2(eden_end, to_start);
eden_size = pointer_delta(eden_end, eden_start, sizeof(char));
eden_size = MIN2(eden_size, cur_eden_size);
} else {
eden_size = pointer_delta(eden_end, eden_start, sizeof(char));
}
eden_size = align_size_down(eden_size, alignment);
assert(maintain_minimum || eden_size <= requested_eden_size,
"Eden size is too large");
assert(eden_size >= alignment, "Eden size is too small");
eden_end = eden_start + eden_size;
// Move to-space down to eden.
if (requested_eden_size < cur_eden_size) {
to_start = eden_end;
if (to_start + requested_survivor_size > to_start) {
to_end = MIN2(from_start, to_start + requested_survivor_size);
} else {
to_end = from_start;
}
}
// eden_end may have moved so again make sure
// the to-space and eden don't overlap.
to_start = MAX2(eden_end, to_start);
// from-space
size_t from_used = from()->used();
if (requested_survivor_size > from_used) {
if (from_start + requested_survivor_size >= from_start) {
from_end = from_start + requested_survivor_size;
}
if (from_end > virtual_space()->high()) {
from_end = virtual_space()->high();
}
}
assert(to_start >= eden_end, "to-space should be above eden");
if (PrintAdaptiveSizePolicy && Verbose) {
gclog_or_tty->print_cr(" [eden_start .. eden_end): "
"[" PTR_FORMAT " .. " PTR_FORMAT ") " SIZE_FORMAT,
eden_start,
eden_end,
pointer_delta(eden_end, eden_start, sizeof(char)));
gclog_or_tty->print_cr(" [ to_start .. to_end): "
"[" PTR_FORMAT " .. " PTR_FORMAT ") " SIZE_FORMAT,
to_start,
to_end,
pointer_delta( to_end, to_start, sizeof(char)));
gclog_or_tty->print_cr(" [from_start .. from_end): "
"[" PTR_FORMAT " .. " PTR_FORMAT ") " SIZE_FORMAT,
from_start,
from_end,
pointer_delta(from_end, from_start, sizeof(char)));
}
}
guarantee((HeapWord*)from_start <= from()->bottom(),
"from start moved to the right");
guarantee((HeapWord*)from_end >= from()->top(),
"from end moved into live data");
assert(is_object_aligned((intptr_t)eden_start), "checking alignment");
assert(is_object_aligned((intptr_t)from_start), "checking alignment");
assert(is_object_aligned((intptr_t)to_start), "checking alignment");
MemRegion edenMR((HeapWord*)eden_start, (HeapWord*)eden_end);
MemRegion toMR ((HeapWord*)to_start, (HeapWord*)to_end);
MemRegion fromMR((HeapWord*)from_start, (HeapWord*)from_end);
// Let's make sure the call to initialize doesn't reset "top"!
HeapWord* old_from_top = from()->top();
// For PrintAdaptiveSizePolicy block below
size_t old_from = from()->capacity();
size_t old_to = to()->capacity();
// If not clearing the spaces, do some checking to verify that
// the spaces are already mangled.
// Must check mangling before the spaces are reshaped. Otherwise,
// the bottom or end of one space may have moved into another
// a failure of the check may not correctly indicate which space
// is not properly mangled.
if (ZapUnusedHeapArea) {
HeapWord* limit = (HeapWord*) virtual_space()->high();
eden()->check_mangled_unused_area(limit);
from()->check_mangled_unused_area(limit);
to()->check_mangled_unused_area(limit);
}
// The call to initialize NULL's the next compaction space
eden()->initialize(edenMR,
SpaceDecorator::Clear,
SpaceDecorator::DontMangle);
eden()->set_next_compaction_space(from());
to()->initialize(toMR ,
SpaceDecorator::Clear,
SpaceDecorator::DontMangle);
from()->initialize(fromMR,
SpaceDecorator::DontClear,
SpaceDecorator::DontMangle);
assert(from()->top() == old_from_top, "from top changed!");
if (PrintAdaptiveSizePolicy) {
GenCollectedHeap* gch = GenCollectedHeap::heap();
assert(gch->kind() == CollectedHeap::GenCollectedHeap, "Sanity");
gclog_or_tty->print("AdaptiveSizePolicy::survivor space sizes: "
"collection: %d "
"(" SIZE_FORMAT ", " SIZE_FORMAT ") -> "
"(" SIZE_FORMAT ", " SIZE_FORMAT ") ",
gch->total_collections(),
old_from, old_to,
from()->capacity(),
to()->capacity());
gclog_or_tty->cr();
}
}
void ASParNewGeneration::compute_new_size() {
GenCollectedHeap* gch = GenCollectedHeap::heap();
assert(gch->kind() == CollectedHeap::GenCollectedHeap,
"not a CMS generational heap");
CMSAdaptiveSizePolicy* size_policy =
(CMSAdaptiveSizePolicy*)gch->gen_policy()->size_policy();
assert(size_policy->is_gc_cms_adaptive_size_policy(),
"Wrong type of size policy");
size_t survived = from()->used();
if (!survivor_overflow()) {
// Keep running averages on how much survived
size_policy->avg_survived()->sample(survived);
} else {
size_t promoted =
(size_t) next_gen()->gc_stats()->avg_promoted()->last_sample();
assert(promoted < gch->capacity(), "Conversion problem?");
size_t survived_guess = survived + promoted;
size_policy->avg_survived()->sample(survived_guess);
}
size_t survivor_limit = max_survivor_size();
_tenuring_threshold =
size_policy->compute_survivor_space_size_and_threshold(
_survivor_overflow,
_tenuring_threshold,
survivor_limit);
size_policy->avg_young_live()->sample(used());
size_policy->avg_eden_live()->sample(eden()->used());
size_policy->compute_young_generation_free_space(eden()->capacity(),
max_gen_size());
resize(size_policy->calculated_eden_size_in_bytes(),
size_policy->calculated_survivor_size_in_bytes());
if (UsePerfData) {
CMSGCAdaptivePolicyCounters* counters =
(CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters();
assert(counters->kind() ==
GCPolicyCounters::CMSGCAdaptivePolicyCountersKind,
"Wrong kind of counters");
counters->update_tenuring_threshold(_tenuring_threshold);
counters->update_survivor_overflowed(_survivor_overflow);
counters->update_young_capacity(capacity());
}
}
#ifndef PRODUCT
// Changes from PSYoungGen version
// value of "alignment"
void ASParNewGeneration::space_invariants() {
const size_t alignment = os::vm_page_size();
// Currently, our eden size cannot shrink to zero
guarantee(eden()->capacity() >= alignment, "eden too small");
guarantee(from()->capacity() >= alignment, "from too small");
guarantee(to()->capacity() >= alignment, "to too small");
// Relationship of spaces to each other
char* eden_start = (char*)eden()->bottom();
char* eden_end = (char*)eden()->end();
char* from_start = (char*)from()->bottom();
char* from_end = (char*)from()->end();
char* to_start = (char*)to()->bottom();
char* to_end = (char*)to()->end();
guarantee(eden_start >= virtual_space()->low(), "eden bottom");
guarantee(eden_start < eden_end, "eden space consistency");
guarantee(from_start < from_end, "from space consistency");
guarantee(to_start < to_end, "to space consistency");
// Check whether from space is below to space
if (from_start < to_start) {
// Eden, from, to
guarantee(eden_end <= from_start, "eden/from boundary");
guarantee(from_end <= to_start, "from/to boundary");
guarantee(to_end <= virtual_space()->high(), "to end");
} else {
// Eden, to, from
guarantee(eden_end <= to_start, "eden/to boundary");
guarantee(to_end <= from_start, "to/from boundary");
guarantee(from_end <= virtual_space()->high(), "from end");
}
// More checks that the virtual space is consistent with the spaces
assert(virtual_space()->committed_size() >=
(eden()->capacity() +
to()->capacity() +
from()->capacity()), "Committed size is inconsistent");
assert(virtual_space()->committed_size() <= virtual_space()->reserved_size(),
"Space invariant");
char* eden_top = (char*)eden()->top();
char* from_top = (char*)from()->top();
char* to_top = (char*)to()->top();
assert(eden_top <= virtual_space()->high(), "eden top");
assert(from_top <= virtual_space()->high(), "from top");
assert(to_top <= virtual_space()->high(), "to top");
}
#endif
parNewGeneration.cpp
/*
* Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp"
#include "gc_implementation/parNew/parGCAllocBuffer.hpp"
#include "gc_implementation/parNew/parNewGeneration.hpp"
#include "gc_implementation/parNew/parOopClosures.inline.hpp"
#include "gc_implementation/shared/adaptiveSizePolicy.hpp"
#include "gc_implementation/shared/ageTable.hpp"
#include "gc_implementation/shared/spaceDecorator.hpp"
#include "memory/defNewGeneration.inline.hpp"
#include "memory/genCollectedHeap.hpp"
#include "memory/genOopClosures.inline.hpp"
#include "memory/generation.hpp"
#include "memory/generation.inline.hpp"
#include "memory/referencePolicy.hpp"
#include "memory/resourceArea.hpp"
#include "memory/sharedHeap.hpp"
#include "memory/space.hpp"
#include "oops/objArrayOop.hpp"
#include "oops/oop.inline.hpp"
#include "oops/oop.pcgc.inline.hpp"
#include "runtime/handles.hpp"
#include "runtime/handles.inline.hpp"
#include "runtime/java.hpp"
#include "runtime/thread.hpp"
#include "utilities/copy.hpp"
#include "utilities/globalDefinitions.hpp"
#include "utilities/workgroup.hpp"
#ifdef _MSC_VER
#pragma warning( push )
#pragma warning( disable:4355 ) // 'this' : used in base member initializer list
#endif
ParScanThreadState::ParScanThreadState(Space* to_space_,
ParNewGeneration* gen_,
Generation* old_gen_,
int thread_num_,
ObjToScanQueueSet* work_queue_set_,
Stack<oop>* overflow_stacks_,
size_t desired_plab_sz_,
ParallelTaskTerminator& term_) :
_to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
_work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
_overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
_ageTable(false), // false ==> not the global age table, no perf data.
_to_space_alloc_buffer(desired_plab_sz_),
_to_space_closure(gen_, this), _old_gen_closure(gen_, this),
_to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
_older_gen_closure(gen_, this),
_evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
&_to_space_root_closure, gen_, &_old_gen_root_closure,
work_queue_set_, &term_),
_is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
_keep_alive_closure(&_scan_weak_ref_closure),
_promotion_failure_size(0),
_strong_roots_time(0.0), _term_time(0.0)
{
#if TASKQUEUE_STATS
_term_attempts = 0;
_overflow_refills = 0;
_overflow_refill_objs = 0;
#endif // TASKQUEUE_STATS
_survivor_chunk_array =
(ChunkArray*) old_gen()->get_data_recorder(thread_num());
_hash_seed = 17; // Might want to take time-based random value.
_start = os::elapsedTime();
_old_gen_closure.set_generation(old_gen_);
_old_gen_root_closure.set_generation(old_gen_);
}
#ifdef _MSC_VER
#pragma warning( pop )
#endif
void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
size_t plab_word_size) {
ChunkArray* sca = survivor_chunk_array();
if (sca != NULL) {
// A non-null SCA implies that we want the PLAB data recorded.
sca->record_sample(plab_start, plab_word_size);
}
}
bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
return new_obj->is_objArray() &&
arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
new_obj != old_obj;
}
void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
assert(old->is_objArray(), "must be obj array");
assert(old->is_forwarded(), "must be forwarded");
assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
assert(!old_gen()->is_in(old), "must be in young generation.");
objArrayOop obj = objArrayOop(old->forwardee());
// Process ParGCArrayScanChunk elements now
// and push the remainder back onto queue
int start = arrayOop(old)->length();
int end = obj->length();
int remainder = end - start;
assert(start <= end, "just checking");
if (remainder > 2 * ParGCArrayScanChunk) {
// Test above combines last partial chunk with a full chunk
end = start + ParGCArrayScanChunk;
arrayOop(old)->set_length(end);
// Push remainder.
bool ok = work_queue()->push(old);
assert(ok, "just popped, push must be okay");
} else {
// Restore length so that it can be used if there
// is a promotion failure and forwarding pointers
// must be removed.
arrayOop(old)->set_length(end);
}
// process our set of indices (include header in first chunk)
// should make sure end is even (aligned to HeapWord in case of compressed oops)
if ((HeapWord *)obj < young_old_boundary()) {
// object is in to_space
obj->oop_iterate_range(&_to_space_closure, start, end);
} else {
// object is in old generation
obj->oop_iterate_range(&_old_gen_closure, start, end);
}
}
void ParScanThreadState::trim_queues(int max_size) {
ObjToScanQueue* queue = work_queue();
do {
while (queue->size() > (juint)max_size) {
oop obj_to_scan;
if (queue->pop_local(obj_to_scan)) {
if ((HeapWord *)obj_to_scan < young_old_boundary()) {
if (obj_to_scan->is_objArray() &&
obj_to_scan->is_forwarded() &&
obj_to_scan->forwardee() != obj_to_scan) {
scan_partial_array_and_push_remainder(obj_to_scan);
} else {
// object is in to_space
obj_to_scan->oop_iterate(&_to_space_closure);
}
} else {
// object is in old generation
obj_to_scan->oop_iterate(&_old_gen_closure);
}
}
}
// For the case of compressed oops, we have a private, non-shared
// overflow stack, so we eagerly drain it so as to more evenly
// distribute load early. Note: this may be good to do in
// general rather than delay for the final stealing phase.
// If applicable, we'll transfer a set of objects over to our
// work queue, allowing them to be stolen and draining our
// private overflow stack.
} while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
}
bool ParScanThreadState::take_from_overflow_stack() {
assert(ParGCUseLocalOverflow, "Else should not call");
assert(young_gen()->overflow_list() == NULL, "Error");
ObjToScanQueue* queue = work_queue();
Stack<oop>* const of_stack = overflow_stack();
const size_t num_overflow_elems = of_stack->size();
const size_t space_available = queue->max_elems() - queue->size();
const size_t num_take_elems = MIN3(space_available / 4,
ParGCDesiredObjsFromOverflowList,
num_overflow_elems);
// Transfer the most recent num_take_elems from the overflow
// stack to our work queue.
for (size_t i = 0; i != num_take_elems; i++) {
oop cur = of_stack->pop();
oop obj_to_push = cur->forwardee();
assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
if (should_be_partially_scanned(obj_to_push, cur)) {
assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
obj_to_push = cur;
}
bool ok = queue->push(obj_to_push);
assert(ok, "Should have succeeded");
}
assert(young_gen()->overflow_list() == NULL, "Error");
return num_take_elems > 0; // was something transferred?
}
void ParScanThreadState::push_on_overflow_stack(oop p) {
assert(ParGCUseLocalOverflow, "Else should not call");
overflow_stack()->push(p);
assert(young_gen()->overflow_list() == NULL, "Error");
}
HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
// Otherwise, if the object is small enough, try to reallocate the
// buffer.
HeapWord* obj = NULL;
if (!_to_space_full) {
ParGCAllocBuffer* const plab = to_space_alloc_buffer();
Space* const sp = to_space();
if (word_sz * 100 <
ParallelGCBufferWastePct * plab->word_sz()) {
// Is small enough; abandon this buffer and start a new one.
plab->retire(false, false);
size_t buf_size = plab->word_sz();
HeapWord* buf_space = sp->par_allocate(buf_size);
if (buf_space == NULL) {
const size_t min_bytes =
ParGCAllocBuffer::min_size() << LogHeapWordSize;
size_t free_bytes = sp->free();
while(buf_space == NULL && free_bytes >= min_bytes) {
buf_size = free_bytes >> LogHeapWordSize;
assert(buf_size == (size_t)align_object_size(buf_size),
"Invariant");
buf_space = sp->par_allocate(buf_size);
free_bytes = sp->free();
}
}
if (buf_space != NULL) {
plab->set_word_size(buf_size);
plab->set_buf(buf_space);
record_survivor_plab(buf_space, buf_size);
obj = plab->allocate(word_sz);
// Note that we cannot compare buf_size < word_sz below
// because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
assert(obj != NULL || plab->words_remaining() < word_sz,
"Else should have been able to allocate");
// It's conceivable that we may be able to use the
// buffer we just grabbed for subsequent small requests
// even if not for this one.
} else {
// We're used up.
_to_space_full = true;
}
} else {
// Too large; allocate the object individually.
obj = sp->par_allocate(word_sz);
}
}
return obj;
}
void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
size_t word_sz) {
// Is the alloc in the current alloc buffer?
if (to_space_alloc_buffer()->contains(obj)) {
assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
"Should contain whole object.");
to_space_alloc_buffer()->undo_allocation(obj, word_sz);
} else {
CollectedHeap::fill_with_object(obj, word_sz);
}
}
void ParScanThreadState::print_and_clear_promotion_failure_size() {
if (_promotion_failure_size != 0) {
if (PrintPromotionFailure) {
gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
_thread_num, _promotion_failure_size);
}
_promotion_failure_size = 0;
}
}
class ParScanThreadStateSet: private ResourceArray {
public:
// Initializes states for the specified number of threads;
ParScanThreadStateSet(int num_threads,
Space& to_space,
ParNewGeneration& gen,
Generation& old_gen,
ObjToScanQueueSet& queue_set,
Stack<oop>* overflow_stacks_,
size_t desired_plab_sz,
ParallelTaskTerminator& term);
~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
inline ParScanThreadState& thread_state(int i);
void reset(bool promotion_failed);
void flush();
#if TASKQUEUE_STATS
static void
print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
void print_termination_stats(outputStream* const st = gclog_or_tty);
static void
print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
void reset_stats();
#endif // TASKQUEUE_STATS
private:
ParallelTaskTerminator& _term;
ParNewGeneration& _gen;
Generation& _next_gen;
};
ParScanThreadStateSet::ParScanThreadStateSet(
int num_threads, Space& to_space, ParNewGeneration& gen,
Generation& old_gen, ObjToScanQueueSet& queue_set,
Stack<oop>* overflow_stacks,
size_t desired_plab_sz, ParallelTaskTerminator& term)
: ResourceArray(sizeof(ParScanThreadState), num_threads),
_gen(gen), _next_gen(old_gen), _term(term)
{
assert(num_threads > 0, "sanity check!");
assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
"overflow_stack allocation mismatch");
// Initialize states.
for (int i = 0; i < num_threads; ++i) {
new ((ParScanThreadState*)_data + i)
ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
overflow_stacks, desired_plab_sz, term);
}
}
inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
{
assert(i >= 0 && i < length(), "sanity check!");
return ((ParScanThreadState*)_data)[i];
}
void ParScanThreadStateSet::reset(bool promotion_failed)
{
_term.reset_for_reuse();
if (promotion_failed) {
for (int i = 0; i < length(); ++i) {
thread_state(i).print_and_clear_promotion_failure_size();
}
}
}
#if TASKQUEUE_STATS
void
ParScanThreadState::reset_stats()
{
taskqueue_stats().reset();
_term_attempts = 0;
_overflow_refills = 0;
_overflow_refill_objs = 0;
}
void ParScanThreadStateSet::reset_stats()
{
for (int i = 0; i < length(); ++i) {
thread_state(i).reset_stats();
}
}
void
ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
{
st->print_raw_cr("GC Termination Stats");
st->print_raw_cr(" elapsed --strong roots-- "
"-------termination-------");
st->print_raw_cr("thr ms ms % "
" ms % attempts");
st->print_raw_cr("--- --------- --------- ------ "
"--------- ------ --------");
}
void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
{
print_termination_stats_hdr(st);
for (int i = 0; i < length(); ++i) {
const ParScanThreadState & pss = thread_state(i);
const double elapsed_ms = pss.elapsed_time() * 1000.0;
const double s_roots_ms = pss.strong_roots_time() * 1000.0;
const double term_ms = pss.term_time() * 1000.0;
st->print_cr("%3d %9.2f %9.2f %6.2f "
"%9.2f %6.2f " SIZE_FORMAT_W(8),
i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
}
}
// Print stats related to work queue activity.
void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
{
st->print_raw_cr("GC Task Stats");
st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
}
void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
{
print_taskqueue_stats_hdr(st);
TaskQueueStats totals;
for (int i = 0; i < length(); ++i) {
const ParScanThreadState & pss = thread_state(i);
const TaskQueueStats & stats = pss.taskqueue_stats();
st->print("%3d ", i); stats.print(st); st->cr();
totals += stats;
if (pss.overflow_refills() > 0) {
st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
SIZE_FORMAT_W(10) " overflow objects",
pss.overflow_refills(), pss.overflow_refill_objs());
}
}
st->print("tot "); totals.print(st); st->cr();
DEBUG_ONLY(totals.verify());
}
#endif // TASKQUEUE_STATS
void ParScanThreadStateSet::flush()
{
// Work in this loop should be kept as lightweight as
// possible since this might otherwise become a bottleneck
// to scaling. Should we add heavy-weight work into this
// loop, consider parallelizing the loop into the worker threads.
for (int i = 0; i < length(); ++i) {
ParScanThreadState& par_scan_state = thread_state(i);
// Flush stats related to To-space PLAB activity and
// retire the last buffer.
par_scan_state.to_space_alloc_buffer()->
flush_stats_and_retire(_gen.plab_stats(),
false /* !retain */);
// Every thread has its own age table. We need to merge
// them all into one.
ageTable *local_table = par_scan_state.age_table();
_gen.age_table()->merge(local_table);
// Inform old gen that we're done.
_next_gen.par_promote_alloc_done(i);
_next_gen.par_oop_since_save_marks_iterate_done(i);
}
if (UseConcMarkSweepGC && ParallelGCThreads > 0) {
// We need to call this even when ResizeOldPLAB is disabled
// so as to avoid breaking some asserts. While we may be able
// to avoid this by reorganizing the code a bit, I am loathe
// to do that unless we find cases where ergo leads to bad
// performance.
CFLS_LAB::compute_desired_plab_size();
}
}
ParScanClosure::ParScanClosure(ParNewGeneration* g,
ParScanThreadState* par_scan_state) :
OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g)
{
assert(_g->level() == 0, "Optimized for youngest generation");
_boundary = _g->reserved().end();
}
void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
ParScanThreadState* par_scan_state)
: ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
{}
void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
#ifdef WIN32
#pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
#endif
ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
ParScanThreadState* par_scan_state_,
ParScanWithoutBarrierClosure* to_space_closure_,
ParScanWithBarrierClosure* old_gen_closure_,
ParRootScanWithoutBarrierClosure* to_space_root_closure_,
ParNewGeneration* par_gen_,
ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
ObjToScanQueueSet* task_queues_,
ParallelTaskTerminator* terminator_) :
_par_scan_state(par_scan_state_),
_to_space_closure(to_space_closure_),
_old_gen_closure(old_gen_closure_),
_to_space_root_closure(to_space_root_closure_),
_old_gen_root_closure(old_gen_root_closure_),
_par_gen(par_gen_),
_task_queues(task_queues_),
_terminator(terminator_)
{}
void ParEvacuateFollowersClosure::do_void() {
ObjToScanQueue* work_q = par_scan_state()->work_queue();
while (true) {
// Scan to-space and old-gen objs until we run out of both.
oop obj_to_scan;
par_scan_state()->trim_queues(0);
// We have no local work, attempt to steal from other threads.
// attempt to steal work from promoted.
if (task_queues()->steal(par_scan_state()->thread_num(),
par_scan_state()->hash_seed(),
obj_to_scan)) {
bool res = work_q->push(obj_to_scan);
assert(res, "Empty queue should have room for a push.");
// if successful, goto Start.
continue;
// try global overflow list.
} else if (par_gen()->take_from_overflow_list(par_scan_state())) {
continue;
}
// Otherwise, offer termination.
par_scan_state()->start_term_time();
if (terminator()->offer_termination()) break;
par_scan_state()->end_term_time();
}
assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
"Broken overflow list?");
// Finish the last termination pause.
par_scan_state()->end_term_time();
}
ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen,
HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
AbstractGangTask("ParNewGeneration collection"),
_gen(gen), _next_gen(next_gen),
_young_old_boundary(young_old_boundary),
_state_set(state_set)
{}
void ParNewGenTask::work(int i) {
GenCollectedHeap* gch = GenCollectedHeap::heap();
// Since this is being done in a separate thread, need new resource
// and handle marks.
ResourceMark rm;
HandleMark hm;
// We would need multiple old-gen queues otherwise.
assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
Generation* old_gen = gch->next_gen(_gen);
ParScanThreadState& par_scan_state = _state_set->thread_state(i);
par_scan_state.set_young_old_boundary(_young_old_boundary);
par_scan_state.start_strong_roots();
gch->gen_process_strong_roots(_gen->level(),
true, // Process younger gens, if any,
// as strong roots.
false, // no scope; this is parallel code
false, // not collecting perm generation.
SharedHeap::SO_AllClasses,
&par_scan_state.to_space_root_closure(),
true, // walk *all* scavengable nmethods
&par_scan_state.older_gen_closure());
par_scan_state.end_strong_roots();
// "evacuate followers".
par_scan_state.evacuate_followers_closure().do_void();
}
#ifdef _MSC_VER
#pragma warning( push )
#pragma warning( disable:4355 ) // 'this' : used in base member initializer list
#endif
ParNewGeneration::
ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
: DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
_overflow_list(NULL),
_is_alive_closure(this),
_plab_stats(YoungPLABSize, PLABWeight)
{
NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
NOT_PRODUCT(_num_par_pushes = 0;)
_task_queues = new ObjToScanQueueSet(ParallelGCThreads);
guarantee(_task_queues != NULL, "task_queues allocation failure.");
for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
ObjToScanQueue *q = new ObjToScanQueue();
guarantee(q != NULL, "work_queue Allocation failure.");
_task_queues->register_queue(i1, q);
}
for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
_task_queues->queue(i2)->initialize();
_overflow_stacks = NULL;
if (ParGCUseLocalOverflow) {
_overflow_stacks = NEW_C_HEAP_ARRAY(Stack<oop>, ParallelGCThreads);
for (size_t i = 0; i < ParallelGCThreads; ++i) {
new (_overflow_stacks + i) Stack<oop>();
}
}
if (UsePerfData) {
EXCEPTION_MARK;
ResourceMark rm;
const char* cname =
PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
ParallelGCThreads, CHECK);
}
}
#ifdef _MSC_VER
#pragma warning( pop )
#endif
// ParNewGeneration::
ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
template <class T>
void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
#ifdef ASSERT
{
assert(!oopDesc::is_null(*p), "expected non-null ref");
oop obj = oopDesc::load_decode_heap_oop_not_null(p);
// We never expect to see a null reference being processed
// as a weak reference.
assert(obj->is_oop(), "expected an oop while scanning weak refs");
}
#endif // ASSERT
_par_cl->do_oop_nv(p);
if (Universe::heap()->is_in_reserved(p)) {
oop obj = oopDesc::load_decode_heap_oop_not_null(p);
_rs->write_ref_field_gc_par(p, obj);
}
}
void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
// ParNewGeneration::
KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
DefNewGeneration::KeepAliveClosure(cl) {}
template <class T>
void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
#ifdef ASSERT
{
assert(!oopDesc::is_null(*p), "expected non-null ref");
oop obj = oopDesc::load_decode_heap_oop_not_null(p);
// We never expect to see a null reference being processed
// as a weak reference.
assert(obj->is_oop(), "expected an oop while scanning weak refs");
}
#endif // ASSERT
_cl->do_oop_nv(p);
if (Universe::heap()->is_in_reserved(p)) {
oop obj = oopDesc::load_decode_heap_oop_not_null(p);
_rs->write_ref_field_gc_par(p, obj);
}
}
void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
T heap_oop = oopDesc::load_heap_oop(p);
if (!oopDesc::is_null(heap_oop)) {
oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
if ((HeapWord*)obj < _boundary) {
assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
oop new_obj = obj->is_forwarded()
? obj->forwardee()
: _g->DefNewGeneration::copy_to_survivor_space(obj);
oopDesc::encode_store_heap_oop_not_null(p, new_obj);
}
if (_gc_barrier) {
// If p points to a younger generation, mark the card.
if ((HeapWord*)obj < _gen_boundary) {
_rs->write_ref_field_gc_par(p, obj);
}
}
}
}
void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
class ParNewRefProcTaskProxy: public AbstractGangTask {
typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
public:
ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
Generation& next_gen,
HeapWord* young_old_boundary,
ParScanThreadStateSet& state_set);
private:
virtual void work(int i);
private:
ParNewGeneration& _gen;
ProcessTask& _task;
Generation& _next_gen;
HeapWord* _young_old_boundary;
ParScanThreadStateSet& _state_set;
};
ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(
ProcessTask& task, ParNewGeneration& gen,
Generation& next_gen,
HeapWord* young_old_boundary,
ParScanThreadStateSet& state_set)
: AbstractGangTask("ParNewGeneration parallel reference processing"),
_gen(gen),
_task(task),
_next_gen(next_gen),
_young_old_boundary(young_old_boundary),
_state_set(state_set)
{
}
void ParNewRefProcTaskProxy::work(int i)
{
ResourceMark rm;
HandleMark hm;
ParScanThreadState& par_scan_state = _state_set.thread_state(i);
par_scan_state.set_young_old_boundary(_young_old_boundary);
_task.work(i, par_scan_state.is_alive_closure(),
par_scan_state.keep_alive_closure(),
par_scan_state.evacuate_followers_closure());
}
class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
EnqueueTask& _task;
public:
ParNewRefEnqueueTaskProxy(EnqueueTask& task)
: AbstractGangTask("ParNewGeneration parallel reference enqueue"),
_task(task)
{ }
virtual void work(int i)
{
_task.work(i);
}
};
void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
{
GenCollectedHeap* gch = GenCollectedHeap::heap();
assert(gch->kind() == CollectedHeap::GenCollectedHeap,
"not a generational heap");
WorkGang* workers = gch->workers();
assert(workers != NULL, "Need parallel worker threads.");
ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
_generation.reserved().end(), _state_set);
workers->run_task(&rp_task);
_state_set.reset(_generation.promotion_failed());
}
void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
{
GenCollectedHeap* gch = GenCollectedHeap::heap();
WorkGang* workers = gch->workers();
assert(workers != NULL, "Need parallel worker threads.");
ParNewRefEnqueueTaskProxy enq_task(task);
workers->run_task(&enq_task);
}
void ParNewRefProcTaskExecutor::set_single_threaded_mode()
{
_state_set.flush();
GenCollectedHeap* gch = GenCollectedHeap::heap();
gch->set_par_threads(0); // 0 ==> non-parallel.
gch->save_marks();
}
ScanClosureWithParBarrier::
ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
ScanClosure(g, gc_barrier) {}
EvacuateFollowersClosureGeneral::
EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
OopsInGenClosure* cur,
OopsInGenClosure* older) :
_gch(gch), _level(level),
_scan_cur_or_nonheap(cur), _scan_older(older)
{}
void EvacuateFollowersClosureGeneral::do_void() {
do {
// Beware: this call will lead to closure applications via virtual
// calls.
_gch->oop_since_save_marks_iterate(_level,
_scan_cur_or_nonheap,
_scan_older);
} while (!_gch->no_allocs_since_save_marks(_level));
}
bool ParNewGeneration::_avoid_promotion_undo = false;
void ParNewGeneration::adjust_desired_tenuring_threshold() {
// Set the desired survivor size to half the real survivor space
_tenuring_threshold =
age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize);
}
// A Generation that does parallel young-gen collection.
void ParNewGeneration::collect(bool full,
bool clear_all_soft_refs,
size_t size,
bool is_tlab) {
assert(full || size > 0, "otherwise we don't want to collect");
GenCollectedHeap* gch = GenCollectedHeap::heap();
assert(gch->kind() == CollectedHeap::GenCollectedHeap,
"not a CMS generational heap");
AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
WorkGang* workers = gch->workers();
_next_gen = gch->next_gen(this);
assert(_next_gen != NULL,
"This must be the youngest gen, and not the only gen");
assert(gch->n_gens() == 2,
"Par collection currently only works with single older gen.");
// Do we have to avoid promotion_undo?
if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) {
set_avoid_promotion_undo(true);
}
// If the next generation is too full to accomodate worst-case promotion
// from this generation, pass on collection; let the next generation
// do it.
if (!collection_attempt_is_safe()) {
gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
return;
}
assert(to()->is_empty(), "Else not collection_attempt_is_safe");
init_assuming_no_promotion_failure();
if (UseAdaptiveSizePolicy) {
set_survivor_overflow(false);
size_policy->minor_collection_begin();
}
TraceTime t1("GC", PrintGC && !PrintGCDetails, true, gclog_or_tty);
// Capture heap used before collection (for printing).
size_t gch_prev_used = gch->used();
SpecializationStats::clear();
age_table()->clear();
to()->clear(SpaceDecorator::Mangle);
gch->save_marks();
assert(workers != NULL, "Need parallel worker threads.");
ParallelTaskTerminator _term(workers->total_workers(), task_queues());
ParScanThreadStateSet thread_state_set(workers->total_workers(),
*to(), *this, *_next_gen, *task_queues(),
_overflow_stacks, desired_plab_sz(), _term);
ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set);
int n_workers = workers->total_workers();
gch->set_par_threads(n_workers);
gch->rem_set()->prepare_for_younger_refs_iterate(true);
// It turns out that even when we're using 1 thread, doing the work in a
// separate thread causes wide variance in run times. We can't help this
// in the multi-threaded case, but we special-case n=1 here to get
// repeatable measurements of the 1-thread overhead of the parallel code.
if (n_workers > 1) {
GenCollectedHeap::StrongRootsScope srs(gch);
workers->run_task(&tsk);
} else {
GenCollectedHeap::StrongRootsScope srs(gch);
tsk.work(0);
}
thread_state_set.reset(promotion_failed());
// Process (weak) reference objects found during scavenge.
ReferenceProcessor* rp = ref_processor();
IsAliveClosure is_alive(this);
ScanWeakRefClosure scan_weak_ref(this);
KeepAliveClosure keep_alive(&scan_weak_ref);
ScanClosure scan_without_gc_barrier(this, false);
ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
&scan_without_gc_barrier, &scan_with_gc_barrier);
rp->setup_policy(clear_all_soft_refs);
if (rp->processing_is_mt()) {
ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
rp->process_discovered_references(&is_alive, &keep_alive,
&evacuate_followers, &task_executor);
} else {
thread_state_set.flush();
gch->set_par_threads(0); // 0 ==> non-parallel.
gch->save_marks();
rp->process_discovered_references(&is_alive, &keep_alive,
&evacuate_followers, NULL);
}
if (!promotion_failed()) {
// Swap the survivor spaces.
eden()->clear(SpaceDecorator::Mangle);
from()->clear(SpaceDecorator::Mangle);
if (ZapUnusedHeapArea) {
// This is now done here because of the piece-meal mangling which
// can check for valid mangling at intermediate points in the
// collection(s). When a minor collection fails to collect
// sufficient space resizing of the young generation can occur
// an redistribute the spaces in the young generation. Mangle
// here so that unzapped regions don't get distributed to
// other spaces.
to()->mangle_unused_area();
}
swap_spaces();
// A successful scavenge should restart the GC time limit count which is
// for full GC's.
size_policy->reset_gc_overhead_limit_count();
assert(to()->is_empty(), "to space should be empty now");
} else {
assert(_promo_failure_scan_stack.is_empty(), "post condition");
_promo_failure_scan_stack.clear(true); // Clear cached segments.
remove_forwarding_pointers();
if (PrintGCDetails) {
gclog_or_tty->print(" (promotion failed)");
}
// All the spaces are in play for mark-sweep.
swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
from()->set_next_compaction_space(to());
gch->set_incremental_collection_failed();
// Inform the next generation that a promotion failure occurred.
_next_gen->promotion_failure_occurred();
// Reset the PromotionFailureALot counters.
NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
}
// set new iteration safe limit for the survivor spaces
from()->set_concurrent_iteration_safe_limit(from()->top());
to()->set_concurrent_iteration_safe_limit(to()->top());
adjust_desired_tenuring_threshold();
if (ResizePLAB) {
plab_stats()->adjust_desired_plab_sz();
}
if (PrintGC && !PrintGCDetails) {
gch->print_heap_change(gch_prev_used);
}
if (PrintGCDetails && ParallelGCVerbose) {
TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
}
if (UseAdaptiveSizePolicy) {
size_policy->minor_collection_end(gch->gc_cause());
size_policy->avg_survived()->sample(from()->used());
}
update_time_of_last_gc(os::javaTimeMillis());
SpecializationStats::print();
rp->set_enqueuing_is_done(true);
if (rp->processing_is_mt()) {
ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
rp->enqueue_discovered_references(&task_executor);
} else {
rp->enqueue_discovered_references(NULL);
}
rp->verify_no_references_recorded();
}
static int sum;
void ParNewGeneration::waste_some_time() {
for (int i = 0; i < 100; i++) {
sum += i;
}
}
static const oop ClaimedForwardPtr = oop(0x4);
// Because of concurrency, there are times where an object for which
// "is_forwarded()" is true contains an "interim" forwarding pointer
// value. Such a value will soon be overwritten with a real value.
// This method requires "obj" to have a forwarding pointer, and waits, if
// necessary for a real one to be inserted, and returns it.
oop ParNewGeneration::real_forwardee(oop obj) {
oop forward_ptr = obj->forwardee();
if (forward_ptr != ClaimedForwardPtr) {
return forward_ptr;
} else {
return real_forwardee_slow(obj);
}
}
oop ParNewGeneration::real_forwardee_slow(oop obj) {
// Spin-read if it is claimed but not yet written by another thread.
oop forward_ptr = obj->forwardee();
while (forward_ptr == ClaimedForwardPtr) {
waste_some_time();
assert(obj->is_forwarded(), "precondition");
forward_ptr = obj->forwardee();
}
return forward_ptr;
}
#ifdef ASSERT
bool ParNewGeneration::is_legal_forward_ptr(oop p) {
return
(_avoid_promotion_undo && p == ClaimedForwardPtr)
|| Universe::heap()->is_in_reserved(p);
}
#endif
void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
if (m->must_be_preserved_for_promotion_failure(obj)) {
// We should really have separate per-worker stacks, rather
// than use locking of a common pair of stacks.
MutexLocker ml(ParGCRareEvent_lock);
preserve_mark(obj, m);
}
}
// Multiple GC threads may try to promote an object. If the object
// is successfully promoted, a forwarding pointer will be installed in
// the object in the young generation. This method claims the right
// to install the forwarding pointer before it copies the object,
// thus avoiding the need to undo the copy as in
// copy_to_survivor_space_avoiding_with_undo.
oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
// In the sequential version, this assert also says that the object is
// not forwarded. That might not be the case here. It is the case that
// the caller observed it to be not forwarded at some time in the past.
assert(is_in_reserved(old), "shouldn't be scavenging this oop");
// The sequential code read "old->age()" below. That doesn't work here,
// since the age is in the mark word, and that might be overwritten with
// a forwarding pointer by a parallel thread. So we must save the mark
// word in a local and then analyze it.
oopDesc dummyOld;
dummyOld.set_mark(m);
assert(!dummyOld.is_forwarded(),
"should not be called with forwarding pointer mark word.");
oop new_obj = NULL;
oop forward_ptr;
// Try allocating obj in to-space (unless too old)
if (dummyOld.age() < tenuring_threshold()) {
new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
if (new_obj == NULL) {
set_survivor_overflow(true);
}
}
if (new_obj == NULL) {
// Either to-space is full or we decided to promote
// try allocating obj tenured
// Attempt to install a null forwarding pointer (atomically),
// to claim the right to install the real forwarding pointer.
forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
if (forward_ptr != NULL) {
// someone else beat us to it.
return real_forwardee(old);
}
new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
old, m, sz);
if (new_obj == NULL) {
// promotion failed, forward to self
_promotion_failed = true;
new_obj = old;
preserve_mark_if_necessary(old, m);
// Log the size of the maiden promotion failure
par_scan_state->log_promotion_failure(sz);
}
old->forward_to(new_obj);
forward_ptr = NULL;
} else {
// Is in to-space; do copying ourselves.
Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
forward_ptr = old->forward_to_atomic(new_obj);
// Restore the mark word copied above.
new_obj->set_mark(m);
// Increment age if obj still in new generation
new_obj->incr_age();
par_scan_state->age_table()->add(new_obj, sz);
}
assert(new_obj != NULL, "just checking");
if (forward_ptr == NULL) {
oop obj_to_push = new_obj;
if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
// Length field used as index of next element to be scanned.
// Real length can be obtained from real_forwardee()
arrayOop(old)->set_length(0);
obj_to_push = old;
assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
"push forwarded object");
}
// Push it on one of the queues of to-be-scanned objects.
bool simulate_overflow = false;
NOT_PRODUCT(
if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
// simulate a stack overflow
simulate_overflow = true;
}
)
if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
// Add stats for overflow pushes.
if (Verbose && PrintGCDetails) {
gclog_or_tty->print("queue overflow!\n");
}
push_on_overflow_list(old, par_scan_state);
TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
}
return new_obj;
}
// Oops. Someone beat us to it. Undo the allocation. Where did we
// allocate it?
if (is_in_reserved(new_obj)) {
// Must be in to_space.
assert(to()->is_in_reserved(new_obj), "Checking");
if (forward_ptr == ClaimedForwardPtr) {
// Wait to get the real forwarding pointer value.
forward_ptr = real_forwardee(old);
}
par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
}
return forward_ptr;
}
// Multiple GC threads may try to promote the same object. If two
// or more GC threads copy the object, only one wins the race to install
// the forwarding pointer. The other threads have to undo their copy.
oop ParNewGeneration::copy_to_survivor_space_with_undo(
ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
// In the sequential version, this assert also says that the object is
// not forwarded. That might not be the case here. It is the case that
// the caller observed it to be not forwarded at some time in the past.
assert(is_in_reserved(old), "shouldn't be scavenging this oop");
// The sequential code read "old->age()" below. That doesn't work here,
// since the age is in the mark word, and that might be overwritten with
// a forwarding pointer by a parallel thread. So we must save the mark
// word here, install it in a local oopDesc, and then analyze it.
oopDesc dummyOld;
dummyOld.set_mark(m);
assert(!dummyOld.is_forwarded(),
"should not be called with forwarding pointer mark word.");
bool failed_to_promote = false;
oop new_obj = NULL;
oop forward_ptr;
// Try allocating obj in to-space (unless too old)
if (dummyOld.age() < tenuring_threshold()) {
new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
if (new_obj == NULL) {
set_survivor_overflow(true);
}
}
if (new_obj == NULL) {
// Either to-space is full or we decided to promote
// try allocating obj tenured
new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
old, m, sz);
if (new_obj == NULL) {
// promotion failed, forward to self
forward_ptr = old->forward_to_atomic(old);
new_obj = old;
if (forward_ptr != NULL) {
return forward_ptr; // someone else succeeded
}
_promotion_failed = true;
failed_to_promote = true;
preserve_mark_if_necessary(old, m);
// Log the size of the maiden promotion failure
par_scan_state->log_promotion_failure(sz);
}
} else {
// Is in to-space; do copying ourselves.
Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
// Restore the mark word copied above.
new_obj->set_mark(m);
// Increment age if new_obj still in new generation
new_obj->incr_age();
par_scan_state->age_table()->add(new_obj, sz);
}
assert(new_obj != NULL, "just checking");
// Now attempt to install the forwarding pointer (atomically).
// We have to copy the mark word before overwriting with forwarding
// ptr, so we can restore it below in the copy.
if (!failed_to_promote) {
forward_ptr = old->forward_to_atomic(new_obj);
}
if (forward_ptr == NULL) {
oop obj_to_push = new_obj;
if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
// Length field used as index of next element to be scanned.
// Real length can be obtained from real_forwardee()
arrayOop(old)->set_length(0);
obj_to_push = old;
assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
"push forwarded object");
}
// Push it on one of the queues of to-be-scanned objects.
bool simulate_overflow = false;
NOT_PRODUCT(
if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
// simulate a stack overflow
simulate_overflow = true;
}
)
if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
// Add stats for overflow pushes.
push_on_overflow_list(old, par_scan_state);
TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
}
return new_obj;
}
// Oops. Someone beat us to it. Undo the allocation. Where did we
// allocate it?
if (is_in_reserved(new_obj)) {
// Must be in to_space.
assert(to()->is_in_reserved(new_obj), "Checking");
par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
} else {
assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
_next_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
(HeapWord*)new_obj, sz);
}
return forward_ptr;
}
#ifndef PRODUCT
// It's OK to call this multi-threaded; the worst thing
// that can happen is that we'll get a bunch of closely
// spaced simulated oveflows, but that's OK, in fact
// probably good as it would exercise the overflow code
// under contention.
bool ParNewGeneration::should_simulate_overflow() {
if (_overflow_counter-- <= 0) { // just being defensive
_overflow_counter = ParGCWorkQueueOverflowInterval;
return true;
} else {
return false;
}
}
#endif
// In case we are using compressed oops, we need to be careful.
// If the object being pushed is an object array, then its length
// field keeps track of the "grey boundary" at which the next
// incremental scan will be done (see ParGCArrayScanChunk).
// When using compressed oops, this length field is kept in the
// lower 32 bits of the erstwhile klass word and cannot be used
// for the overflow chaining pointer (OCP below). As such the OCP
// would itself need to be compressed into the top 32-bits in this
// case. Unfortunately, see below, in the event that we have a
// promotion failure, the node to be pushed on the list can be
// outside of the Java heap, so the heap-based pointer compression
// would not work (we would have potential aliasing between C-heap
// and Java-heap pointers). For this reason, when using compressed
// oops, we simply use a worker-thread-local, non-shared overflow
// list in the form of a growable array, with a slightly different
// overflow stack draining strategy. If/when we start using fat
// stacks here, we can go back to using (fat) pointer chains
// (although some performance comparisons would be useful since
// single global lists have their own performance disadvantages
// as we were made painfully aware not long ago, see 6786503).
#define BUSY (oop(0x1aff1aff))
void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
assert(is_in_reserved(from_space_obj), "Should be from this generation");
if (ParGCUseLocalOverflow) {
// In the case of compressed oops, we use a private, not-shared
// overflow stack.
par_scan_state->push_on_overflow_stack(from_space_obj);
} else {
assert(!UseCompressedOops, "Error");
// if the object has been forwarded to itself, then we cannot
// use the klass pointer for the linked list. Instead we have
// to allocate an oopDesc in the C-Heap and use that for the linked list.
// XXX This is horribly inefficient when a promotion failure occurs
// and should be fixed. XXX FIX ME !!!
#ifndef PRODUCT
Atomic::inc_ptr(&_num_par_pushes);
assert(_num_par_pushes > 0, "Tautology");
#endif
if (from_space_obj->forwardee() == from_space_obj) {
oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1);
listhead->forward_to(from_space_obj);
from_space_obj = listhead;
}
oop observed_overflow_list = _overflow_list;
oop cur_overflow_list;
do {
cur_overflow_list = observed_overflow_list;
if (cur_overflow_list != BUSY) {
from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
} else {
from_space_obj->set_klass_to_list_ptr(NULL);
}
observed_overflow_list =
(oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
} while (cur_overflow_list != observed_overflow_list);
}
}
bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
bool res;
if (ParGCUseLocalOverflow) {
res = par_scan_state->take_from_overflow_stack();
} else {
assert(!UseCompressedOops, "Error");
res = take_from_overflow_list_work(par_scan_state);
}
return res;
}
// *NOTE*: The overflow list manipulation code here and
// in CMSCollector:: are very similar in shape,
// except that in the CMS case we thread the objects
// directly into the list via their mark word, and do
// not need to deal with special cases below related
// to chunking of object arrays and promotion failure
// handling.
// CR 6797058 has been filed to attempt consolidation of
// the common code.
// Because of the common code, if you make any changes in
// the code below, please check the CMS version to see if
// similar changes might be needed.
// See CMSCollector::par_take_from_overflow_list() for
// more extensive documentation comments.
bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
ObjToScanQueue* work_q = par_scan_state->work_queue();
// How many to take?
size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
(size_t)ParGCDesiredObjsFromOverflowList);
assert(!UseCompressedOops, "Error");
assert(par_scan_state->overflow_stack() == NULL, "Error");
if (_overflow_list == NULL) return false;
// Otherwise, there was something there; try claiming the list.
oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
// Trim off a prefix of at most objsFromOverflow items
Thread* tid = Thread::current();
size_t spin_count = (size_t)ParallelGCThreads;
size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
// someone grabbed it before we did ...
// ... we spin for a short while...
os::sleep(tid, sleep_time_millis, false);
if (_overflow_list == NULL) {
// nothing left to take
return false;
} else if (_overflow_list != BUSY) {
// try and grab the prefix
prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
}
}
if (prefix == NULL || prefix == BUSY) {
// Nothing to take or waited long enough
if (prefix == NULL) {
// Write back the NULL in case we overwrote it with BUSY above
// and it is still the same value.
(void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
}
return false;
}
assert(prefix != NULL && prefix != BUSY, "Error");
size_t i = 1;
oop cur = prefix;
while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
i++; cur = oop(cur->klass());
}
// Reattach remaining (suffix) to overflow list
if (cur->klass_or_null() == NULL) {
// Write back the NULL in lieu of the BUSY we wrote
// above and it is still the same value.
if (_overflow_list == BUSY) {
(void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
}
} else {
assert(cur->klass_or_null() != BUSY, "Error");
oop suffix = oop(cur->klass()); // suffix will be put back on global list
cur->set_klass_to_list_ptr(NULL); // break off suffix
// It's possible that the list is still in the empty(busy) state
// we left it in a short while ago; in that case we may be
// able to place back the suffix.
oop observed_overflow_list = _overflow_list;
oop cur_overflow_list = observed_overflow_list;
bool attached = false;
while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
observed_overflow_list =
(oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
if (cur_overflow_list == observed_overflow_list) {
attached = true;
break;
} else cur_overflow_list = observed_overflow_list;
}
if (!attached) {
// Too bad, someone else got in in between; we'll need to do a splice.
// Find the last item of suffix list
oop last = suffix;
while (last->klass_or_null() != NULL) {
last = oop(last->klass());
}
// Atomically prepend suffix to current overflow list
observed_overflow_list = _overflow_list;
do {
cur_overflow_list = observed_overflow_list;
if (cur_overflow_list != BUSY) {
// Do the splice ...
last->set_klass_to_list_ptr(cur_overflow_list);
} else { // cur_overflow_list == BUSY
last->set_klass_to_list_ptr(NULL);
}
observed_overflow_list =
(oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
} while (cur_overflow_list != observed_overflow_list);
}
}
// Push objects on prefix list onto this thread's work queue
assert(prefix != NULL && prefix != BUSY, "program logic");
cur = prefix;
ssize_t n = 0;
while (cur != NULL) {
oop obj_to_push = cur->forwardee();
oop next = oop(cur->klass_or_null());
cur->set_klass(obj_to_push->klass());
// This may be an array object that is self-forwarded. In that case, the list pointer
// space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
if (!is_in_reserved(cur)) {
// This can become a scaling bottleneck when there is work queue overflow coincident
// with promotion failure.
oopDesc* f = cur;
FREE_C_HEAP_ARRAY(oopDesc, f);
} else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
obj_to_push = cur;
}
bool ok = work_q->push(obj_to_push);
assert(ok, "Should have succeeded");
cur = next;
n++;
}
TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
#ifndef PRODUCT
assert(_num_par_pushes >= n, "Too many pops?");
Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
#endif
return true;
}
#undef BUSY
void ParNewGeneration::ref_processor_init()
{
if (_ref_processor == NULL) {
// Allocate and initialize a reference processor
_ref_processor =
new ReferenceProcessor(_reserved, // span
ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
(int) ParallelGCThreads, // mt processing degree
refs_discovery_is_mt(), // mt discovery
(int) ParallelGCThreads, // mt discovery degree
refs_discovery_is_atomic(), // atomic_discovery
NULL, // is_alive_non_header
false); // write barrier for next field updates
}
}
const char* ParNewGeneration::name() const {
return "par new generation";
}
bool ParNewGeneration::in_use() {
return UseParNewGC && ParallelGCThreads > 0;
}
parCardTableModRefBS.cpp
Copyright (c) 2007, 2011, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "memory/allocation.inline.hpp"
#include "memory/cardTableModRefBS.hpp"
#include "memory/cardTableRS.hpp"
#include "memory/sharedHeap.hpp"
#include "memory/space.inline.hpp"
#include "memory/universe.hpp"
#include "oops/oop.inline.hpp"
#include "runtime/java.hpp"
#include "runtime/mutexLocker.hpp"
#include "runtime/virtualspace.hpp"
void CardTableModRefBS::non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr,
OopsInGenClosure* cl,
CardTableRS* ct,
int n_threads) {
assert(n_threads > 0, "Error: expected n_threads > 0");
assert((n_threads == 1 && ParallelGCThreads == 0) ||
n_threads <= (int)ParallelGCThreads,
"# worker threads != # requested!");
// Make sure the LNC array is valid for the space.
jbyte** lowest_non_clean;
uintptr_t lowest_non_clean_base_chunk_index;
size_t lowest_non_clean_chunk_size;
get_LNC_array_for_space(sp, lowest_non_clean,
lowest_non_clean_base_chunk_index,
lowest_non_clean_chunk_size);
int n_strides = n_threads * ParGCStridesPerThread;
SequentialSubTasksDone* pst = sp->par_seq_tasks();
pst->set_n_threads(n_threads);
pst->set_n_tasks(n_strides);
int stride = 0;
while (!pst->is_task_claimed(/* reference */ stride)) {
process_stride(sp, mr, stride, n_strides, cl, ct,
lowest_non_clean,
lowest_non_clean_base_chunk_index,
lowest_non_clean_chunk_size);
}
if (pst->all_tasks_completed()) {
// Clear lowest_non_clean array for next time.
intptr_t first_chunk_index = addr_to_chunk_index(mr.start());
uintptr_t last_chunk_index = addr_to_chunk_index(mr.last());
for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) {
intptr_t ind = ch - lowest_non_clean_base_chunk_index;
assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size,
"Bounds error");
lowest_non_clean[ind] = NULL;
}
}
}
void
CardTableModRefBS::
process_stride(Space* sp,
MemRegion used,
jint stride, int n_strides,
OopsInGenClosure* cl,
CardTableRS* ct,
jbyte** lowest_non_clean,
uintptr_t lowest_non_clean_base_chunk_index,
size_t lowest_non_clean_chunk_size) {
// We go from higher to lower addresses here; it wouldn't help that much
// because of the strided parallelism pattern used here.
// Find the first card address of the first chunk in the stride that is
// at least "bottom" of the used region.
jbyte* start_card = byte_for(used.start());
jbyte* end_card = byte_after(used.last());
uintptr_t start_chunk = addr_to_chunk_index(used.start());
uintptr_t start_chunk_stride_num = start_chunk % n_strides;
jbyte* chunk_card_start;
if ((uintptr_t)stride >= start_chunk_stride_num) {
chunk_card_start = (jbyte*)(start_card +
(stride - start_chunk_stride_num) *
ParGCCardsPerStrideChunk);
} else {
// Go ahead to the next chunk group boundary, then to the requested stride.
chunk_card_start = (jbyte*)(start_card +
(n_strides - start_chunk_stride_num + stride) *
ParGCCardsPerStrideChunk);
}
while (chunk_card_start < end_card) {
// Even though we go from lower to higher addresses below, the
// strided parallelism can interleave the actual processing of the
// dirty pages in various ways. For a specific chunk within this
// stride, we take care to avoid double scanning or missing a card
// by suitably initializing the "min_done" field in process_chunk_boundaries()
// below, together with the dirty region extension accomplished in
// DirtyCardToOopClosure::do_MemRegion().
jbyte* chunk_card_end = chunk_card_start + ParGCCardsPerStrideChunk;
// Invariant: chunk_mr should be fully contained within the "used" region.
MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start),
chunk_card_end >= end_card ?
used.end() : addr_for(chunk_card_end));
assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)");
assert(used.contains(chunk_mr), "chunk_mr should be subset of used");
DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(),
cl->gen_boundary());
ClearNoncleanCardWrapper clear_cl(dcto_cl, ct);
// Process the chunk.
process_chunk_boundaries(sp,
dcto_cl,
chunk_mr,
used,
lowest_non_clean,
lowest_non_clean_base_chunk_index,
lowest_non_clean_chunk_size);
// We want the LNC array updates above in process_chunk_boundaries
// to be visible before any of the card table value changes as a
// result of the dirty card iteration below.
OrderAccess::storestore();
// We do not call the non_clean_card_iterate_serial() version because
// we want to clear the cards: clear_cl here does the work of finding
// contiguous dirty ranges of cards to process and clear.
clear_cl.do_MemRegion(chunk_mr);
// Find the next chunk of the stride.
chunk_card_start += ParGCCardsPerStrideChunk * n_strides;
}
}
// If you want a talkative process_chunk_boundaries,
// then #define NOISY(x) x
#ifdef NOISY
#error "Encountered a global preprocessor flag, NOISY, which might clash with local definition to follow"
#else
#define NOISY(x)
#endif
void
CardTableModRefBS::
process_chunk_boundaries(Space* sp,
DirtyCardToOopClosure* dcto_cl,
MemRegion chunk_mr,
MemRegion used,
jbyte** lowest_non_clean,
uintptr_t lowest_non_clean_base_chunk_index,
size_t lowest_non_clean_chunk_size)
{
// We must worry about non-array objects that cross chunk boundaries,
// because such objects are both precisely and imprecisely marked:
// .. if the head of such an object is dirty, the entire object
// needs to be scanned, under the interpretation that this
// was an imprecise mark
// .. if the head of such an object is not dirty, we can assume
// precise marking and it's efficient to scan just the dirty
// cards.
// In either case, each scanned reference must be scanned precisely
// once so as to avoid cloning of a young referent. For efficiency,
// our closures depend on this property and do not protect against
// double scans.
uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start());
cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index;
NOISY(tty->print_cr("===========================================================================");)
NOISY(tty->print_cr(" process_chunk_boundary: Called with [" PTR_FORMAT "," PTR_FORMAT ")",
chunk_mr.start(), chunk_mr.end());)
// First, set "our" lowest_non_clean entry, which would be
// used by the thread scanning an adjoining left chunk with
// a non-array object straddling the mutual boundary.
// Find the object that spans our boundary, if one exists.
// first_block is the block possibly straddling our left boundary.
HeapWord* first_block = sp->block_start(chunk_mr.start());
assert((chunk_mr.start() != used.start()) || (first_block == chunk_mr.start()),
"First chunk should always have a co-initial block");
// Does the block straddle the chunk's left boundary, and is it
// a non-array object?
if (first_block < chunk_mr.start() // first block straddles left bdry
&& sp->block_is_obj(first_block) // first block is an object
&& !(oop(first_block)->is_objArray() // first block is not an array (arrays are precisely dirtied)
|| oop(first_block)->is_typeArray())) {
// Find our least non-clean card, so that a left neighbour
// does not scan an object straddling the mutual boundary
// too far to the right, and attempt to scan a portion of
// that object twice.
jbyte* first_dirty_card = NULL;
jbyte* last_card_of_first_obj =
byte_for(first_block + sp->block_size(first_block) - 1);
jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last());
jbyte* last_card_to_check =
(jbyte*) MIN2((intptr_t) last_card_of_cur_chunk,
(intptr_t) last_card_of_first_obj);
// Note that this does not need to go beyond our last card
// if our first object completely straddles this chunk.
for (jbyte* cur = first_card_of_cur_chunk;
cur <= last_card_to_check; cur++) {
jbyte val = *cur;
if (card_will_be_scanned(val)) {
first_dirty_card = cur; break;
} else {
assert(!card_may_have_been_dirty(val), "Error");
}
}
if (first_dirty_card != NULL) {
NOISY(tty->print_cr(" LNC: Found a dirty card at " PTR_FORMAT " in current chunk",
first_dirty_card);)
assert(0 <= cur_chunk_index && cur_chunk_index < lowest_non_clean_chunk_size,
"Bounds error.");
assert(lowest_non_clean[cur_chunk_index] == NULL,
"Write exactly once : value should be stable hereafter for this round");
lowest_non_clean[cur_chunk_index] = first_dirty_card;
} NOISY(else {
tty->print_cr(" LNC: Found no dirty card in current chunk; leaving LNC entry NULL");
// In the future, we could have this thread look for a non-NULL value to copy from its
// right neighbour (up to the end of the first object).
if (last_card_of_cur_chunk < last_card_of_first_obj) {
tty->print_cr(" LNC: BEWARE!!! first obj straddles past right end of chunk:\n"
" might be efficient to get value from right neighbour?");
}
})
} else {
// In this case we can help our neighbour by just asking them
// to stop at our first card (even though it may not be dirty).
NOISY(tty->print_cr(" LNC: first block is not a non-array object; setting LNC to first card of current chunk");)
assert(lowest_non_clean[cur_chunk_index] == NULL, "Write once : value should be stable hereafter");
jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start());
lowest_non_clean[cur_chunk_index] = first_card_of_cur_chunk;
}
NOISY(tty->print_cr(" process_chunk_boundary: lowest_non_clean[" INTPTR_FORMAT "] = " PTR_FORMAT
" which corresponds to the heap address " PTR_FORMAT,
cur_chunk_index, lowest_non_clean[cur_chunk_index],
(lowest_non_clean[cur_chunk_index] != NULL)
? addr_for(lowest_non_clean[cur_chunk_index])
: NULL);)
NOISY(tty->print_cr("---------------------------------------------------------------------------");)
// Next, set our own max_to_do, which will strictly/exclusively bound
// the highest address that we will scan past the right end of our chunk.
HeapWord* max_to_do = NULL;
if (chunk_mr.end() < used.end()) {
// This is not the last chunk in the used region.
// What is our last block? We check the first block of
// the next (right) chunk rather than strictly check our last block
// because it's potentially more efficient to do so.
HeapWord* const last_block = sp->block_start(chunk_mr.end());
assert(last_block <= chunk_mr.end(), "In case this property changes.");
if ((last_block == chunk_mr.end()) // our last block does not straddle boundary
|| !sp->block_is_obj(last_block) // last_block isn't an object
|| oop(last_block)->is_objArray() // last_block is an array (precisely marked)
|| oop(last_block)->is_typeArray()) {
max_to_do = chunk_mr.end();
NOISY(tty->print_cr(" process_chunk_boundary: Last block on this card is not a non-array object;\n"
" max_to_do left at " PTR_FORMAT, max_to_do);)
} else {
assert(last_block < chunk_mr.end(), "Tautology");
// It is a non-array object that straddles the right boundary of this chunk.
// last_obj_card is the card corresponding to the start of the last object
// in the chunk. Note that the last object may not start in
// the chunk.
jbyte* const last_obj_card = byte_for(last_block);
const jbyte val = *last_obj_card;
if (!card_will_be_scanned(val)) {
assert(!card_may_have_been_dirty(val), "Error");
// The card containing the head is not dirty. Any marks on
// subsequent cards still in this chunk must have been made
// precisely; we can cap processing at the end of our chunk.
max_to_do = chunk_mr.end();
NOISY(tty->print_cr(" process_chunk_boundary: Head of last object on this card is not dirty;\n"
" max_to_do left at " PTR_FORMAT,
max_to_do);)
} else {
// The last object must be considered dirty, and extends onto the
// following chunk. Look for a dirty card in that chunk that will
// bound our processing.
jbyte* limit_card = NULL;
const size_t last_block_size = sp->block_size(last_block);
jbyte* const last_card_of_last_obj =
byte_for(last_block + last_block_size - 1);
jbyte* const first_card_of_next_chunk = byte_for(chunk_mr.end());
// This search potentially goes a long distance looking
// for the next card that will be scanned, terminating
// at the end of the last_block, if no earlier dirty card
// is found.
assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == ParGCCardsPerStrideChunk,
"last card of next chunk may be wrong");
for (jbyte* cur = first_card_of_next_chunk;
cur <= last_card_of_last_obj; cur++) {
const jbyte val = *cur;
if (card_will_be_scanned(val)) {
NOISY(tty->print_cr(" Found a non-clean card " PTR_FORMAT " with value 0x%x",
cur, (int)val);)
limit_card = cur; break;
} else {
assert(!card_may_have_been_dirty(val), "Error: card can't be skipped");
}
}
if (limit_card != NULL) {
max_to_do = addr_for(limit_card);
assert(limit_card != NULL && max_to_do != NULL, "Error");
NOISY(tty->print_cr(" process_chunk_boundary: Found a dirty card at " PTR_FORMAT
" max_to_do set at " PTR_FORMAT " which is before end of last block in chunk: "
PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
limit_card, max_to_do, last_block, last_block_size, (last_block+last_block_size));)
} else {
// The following is a pessimistic value, because it's possible
// that a dirty card on a subsequent chunk has been cleared by
// the time we get to look at it; we'll correct for that further below,
// using the LNC array which records the least non-clean card
// before cards were cleared in a particular chunk.
limit_card = last_card_of_last_obj;
max_to_do = last_block + last_block_size;
assert(limit_card != NULL && max_to_do != NULL, "Error");
NOISY(tty->print_cr(" process_chunk_boundary: Found no dirty card before end of last block in chunk\n"
" Setting limit_card to " PTR_FORMAT
" and max_to_do " PTR_FORMAT " + " PTR_FORMAT " = " PTR_FORMAT,
limit_card, last_block, last_block_size, max_to_do);)
}
assert(0 < cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size,
"Bounds error.");
// It is possible that a dirty card for the last object may have been
// cleared before we had a chance to examine it. In that case, the value
// will have been logged in the LNC for that chunk.
// We need to examine as many chunks to the right as this object
// covers.
const uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1)
- lowest_non_clean_base_chunk_index;
DEBUG_ONLY(const uintptr_t last_chunk_index = addr_to_chunk_index(used.last())
- lowest_non_clean_base_chunk_index;)
assert(last_chunk_index_to_check <= last_chunk_index,
err_msg("Out of bounds: last_chunk_index_to_check " INTPTR_FORMAT
" exceeds last_chunk_index " INTPTR_FORMAT,
last_chunk_index_to_check, last_chunk_index));
for (uintptr_t lnc_index = cur_chunk_index + 1;
lnc_index <= last_chunk_index_to_check;
lnc_index++) {
jbyte* lnc_card = lowest_non_clean[lnc_index];
if (lnc_card != NULL) {
// we can stop at the first non-NULL entry we find
if (lnc_card <= limit_card) {
NOISY(tty->print_cr(" process_chunk_boundary: LNC card " PTR_FORMAT " is lower than limit_card " PTR_FORMAT,
" max_to_do will be lowered to " PTR_FORMAT " from " PTR_FORMAT,
lnc_card, limit_card, addr_for(lnc_card), max_to_do);)
limit_card = lnc_card;
max_to_do = addr_for(limit_card);
assert(limit_card != NULL && max_to_do != NULL, "Error");
}
// In any case, we break now
break;
} // else continue to look for a non-NULL entry if any
}
assert(limit_card != NULL && max_to_do != NULL, "Error");
}
assert(max_to_do != NULL, "OOPS 1 !");
}
assert(max_to_do != NULL, "OOPS 2!");
} else {
max_to_do = used.end();
NOISY(tty->print_cr(" process_chunk_boundary: Last chunk of this space;\n"
" max_to_do left at " PTR_FORMAT,
max_to_do);)
}
assert(max_to_do != NULL, "OOPS 3!");
// Now we can set the closure we're using so it doesn't to beyond
// max_to_do.
dcto_cl->set_min_done(max_to_do);
#ifndef PRODUCT
dcto_cl->set_last_bottom(max_to_do);
#endif
NOISY(tty->print_cr("===========================================================================\n");)
}
#undef NOISY
void
CardTableModRefBS::
get_LNC_array_for_space(Space* sp,
jbyte**& lowest_non_clean,
uintptr_t& lowest_non_clean_base_chunk_index,
size_t& lowest_non_clean_chunk_size) {
int i = find_covering_region_containing(sp->bottom());
MemRegion covered = _covered[i];
size_t n_chunks = chunks_to_cover(covered);
// Only the first thread to obtain the lock will resize the
// LNC array for the covered region. Any later expansion can't affect
// the used_at_save_marks region.
// (I observed a bug in which the first thread to execute this would
// resize, and then it would cause "expand_and_allocate" that would
// increase the number of chunks in the covered region. Then a second
// thread would come and execute this, see that the size didn't match,
// and free and allocate again. So the first thread would be using a
// freed "_lowest_non_clean" array.)
// Do a dirty read here. If we pass the conditional then take the rare
// event lock and do the read again in case some other thread had already
// succeeded and done the resize.
int cur_collection = Universe::heap()->total_collections();
if (_last_LNC_resizing_collection[i] != cur_collection) {
MutexLocker x(ParGCRareEvent_lock);
if (_last_LNC_resizing_collection[i] != cur_collection) {
if (_lowest_non_clean[i] == NULL ||
n_chunks != _lowest_non_clean_chunk_size[i]) {
// Should we delete the old?
if (_lowest_non_clean[i] != NULL) {
assert(n_chunks != _lowest_non_clean_chunk_size[i],
"logical consequence");
FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]);
_lowest_non_clean[i] = NULL;
}
// Now allocate a new one if necessary.
if (_lowest_non_clean[i] == NULL) {
_lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks);
_lowest_non_clean_chunk_size[i] = n_chunks;
_lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start());
for (int j = 0; j < (int)n_chunks; j++)
_lowest_non_clean[i][j] = NULL;
}
}
_last_LNC_resizing_collection[i] = cur_collection;
}
}
// In any case, now do the initialization.
lowest_non_clean = _lowest_non_clean[i];
lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i];
lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i];
}
parGCAllocBuffer.cpp
Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "gc_implementation/parNew/parGCAllocBuffer.hpp"
#include "memory/sharedHeap.hpp"
#include "oops/arrayOop.hpp"
#include "oops/oop.inline.hpp"
ParGCAllocBuffer::ParGCAllocBuffer(size_t desired_plab_sz_) :
_word_sz(desired_plab_sz_), _bottom(NULL), _top(NULL),
_end(NULL), _hard_end(NULL),
_retained(false), _retained_filler(),
_allocated(0), _wasted(0)
{
assert (min_size() > AlignmentReserve, "Inconsistency!");
// arrayOopDesc::header_size depends on command line initialization.
FillerHeaderSize = align_object_size(arrayOopDesc::header_size(T_INT));
AlignmentReserve = oopDesc::header_size() > MinObjAlignment ? FillerHeaderSize : 0;
}
size_t ParGCAllocBuffer::FillerHeaderSize;
// If the minimum object size is greater than MinObjAlignment, we can
// end up with a shard at the end of the buffer that's smaller than
// the smallest object. We can't allow that because the buffer must
// look like it's full of objects when we retire it, so we make
// sure we have enough space for a filler int array object.
size_t ParGCAllocBuffer::AlignmentReserve;
void ParGCAllocBuffer::retire(bool end_of_gc, bool retain) {
assert(!retain || end_of_gc, "Can only retain at GC end.");
if (_retained) {
// If the buffer had been retained shorten the previous filler object.
assert(_retained_filler.end() <= _top, "INVARIANT");
CollectedHeap::fill_with_object(_retained_filler);
// Wasted space book-keeping, otherwise (normally) done in invalidate()
_wasted += _retained_filler.word_size();
_retained = false;
}
assert(!end_of_gc || !_retained, "At this point, end_of_gc ==> !_retained.");
if (_top < _hard_end) {
CollectedHeap::fill_with_object(_top, _hard_end);
if (!retain) {
invalidate();
} else {
// Is there wasted space we'd like to retain for the next GC?
if (pointer_delta(_end, _top) > FillerHeaderSize) {
_retained = true;
_retained_filler = MemRegion(_top, FillerHeaderSize);
_top = _top + FillerHeaderSize;
} else {
invalidate();
}
}
}
}
void ParGCAllocBuffer::flush_stats(PLABStats* stats) {
assert(ResizePLAB, "Wasted work");
stats->add_allocated(_allocated);
stats->add_wasted(_wasted);
stats->add_unused(pointer_delta(_end, _top));
}
// Compute desired plab size and latch result for later
// use. This should be called once at the end of parallel
// scavenge; it clears the sensor accumulators.
void PLABStats::adjust_desired_plab_sz() {
assert(ResizePLAB, "Not set");
if (_allocated == 0) {
assert(_unused == 0, "Inconsistency in PLAB stats");
_allocated = 1;
}
double wasted_frac = (double)_unused/(double)_allocated;
size_t target_refills = (size_t)((wasted_frac*TargetSurvivorRatio)/
TargetPLABWastePct);
if (target_refills == 0) {
target_refills = 1;
}
_used = _allocated - _wasted - _unused;
size_t plab_sz = _used/(target_refills*ParallelGCThreads);
if (PrintPLAB) gclog_or_tty->print(" (plab_sz = %d ", plab_sz);
// Take historical weighted average
_filter.sample(plab_sz);
// Clip from above and below, and align to object boundary
plab_sz = MAX2(min_size(), (size_t)_filter.average());
plab_sz = MIN2(max_size(), plab_sz);
plab_sz = align_object_size(plab_sz);
// Latch the result
if (PrintPLAB) gclog_or_tty->print(" desired_plab_sz = %d) ", plab_sz);
if (ResizePLAB) {
_desired_plab_sz = plab_sz;
}
// Now clear the accumulators for next round:
// note this needs to be fixed in the case where we
// are retaining across scavenges. FIX ME !!! XXX
_allocated = 0;
_wasted = 0;
_unused = 0;
}
#ifndef PRODUCT
void ParGCAllocBuffer::print() {
gclog_or_tty->print("parGCAllocBuffer: _bottom: %p _top: %p _end: %p _hard_end: %p"
"_retained: %c _retained_filler: [%p,%p)\n",
_bottom, _top, _end, _hard_end,
"FT"[_retained], _retained_filler.start(), _retained_filler.end());
}
#endif // !PRODUCT
const size_t ParGCAllocBufferWithBOT::ChunkSizeInWords =
MIN2(CardTableModRefBS::par_chunk_heapword_alignment(),
((size_t)Generation::GenGrain)/HeapWordSize);
const size_t ParGCAllocBufferWithBOT::ChunkSizeInBytes =
MIN2(CardTableModRefBS::par_chunk_heapword_alignment() * HeapWordSize,
(size_t)Generation::GenGrain);
ParGCAllocBufferWithBOT::ParGCAllocBufferWithBOT(size_t word_sz,
BlockOffsetSharedArray* bsa) :
ParGCAllocBuffer(word_sz),
_bsa(bsa),
_bt(bsa, MemRegion(_bottom, _hard_end)),
_true_end(_hard_end)
{}
// The buffer comes with its own BOT, with a shared (obviously) underlying
// BlockOffsetSharedArray. We manipulate this BOT in the normal way
// as we would for any contiguous space. However, on accasion we
// need to do some buffer surgery at the extremities before we
// start using the body of the buffer for allocations. Such surgery
// (as explained elsewhere) is to prevent allocation on a card that
// is in the process of being walked concurrently by another GC thread.
// When such surgery happens at a point that is far removed (to the
// right of the current allocation point, top), we use the "contig"
// parameter below to directly manipulate the shared array without
// modifying the _next_threshold state in the BOT.
void ParGCAllocBufferWithBOT::fill_region_with_block(MemRegion mr,
bool contig) {
CollectedHeap::fill_with_object(mr);
if (contig) {
_bt.alloc_block(mr.start(), mr.end());
} else {
_bt.BlockOffsetArray::alloc_block(mr.start(), mr.end());
}
}
HeapWord* ParGCAllocBufferWithBOT::allocate_slow(size_t word_sz) {
HeapWord* res = NULL;
if (_true_end > _hard_end) {
assert((HeapWord*)align_size_down(intptr_t(_hard_end),
ChunkSizeInBytes) == _hard_end,
"or else _true_end should be equal to _hard_end");
assert(_retained, "or else _true_end should be equal to _hard_end");
assert(_retained_filler.end() <= _top, "INVARIANT");
CollectedHeap::fill_with_object(_retained_filler);
if (_top < _hard_end) {
fill_region_with_block(MemRegion(_top, _hard_end), true);
}
HeapWord* next_hard_end = MIN2(_true_end, _hard_end + ChunkSizeInWords);
_retained_filler = MemRegion(_hard_end, FillerHeaderSize);
_bt.alloc_block(_retained_filler.start(), _retained_filler.word_size());
_top = _retained_filler.end();
_hard_end = next_hard_end;
_end = _hard_end - AlignmentReserve;
res = ParGCAllocBuffer::allocate(word_sz);
if (res != NULL) {
_bt.alloc_block(res, word_sz);
}
}
return res;
}
void
ParGCAllocBufferWithBOT::undo_allocation(HeapWord* obj, size_t word_sz) {
ParGCAllocBuffer::undo_allocation(obj, word_sz);
// This may back us up beyond the previous threshold, so reset.
_bt.set_region(MemRegion(_top, _hard_end));
_bt.initialize_threshold();
}
void ParGCAllocBufferWithBOT::retire(bool end_of_gc, bool retain) {
assert(!retain || end_of_gc, "Can only retain at GC end.");
if (_retained) {
// We're about to make the retained_filler into a block.
_bt.BlockOffsetArray::alloc_block(_retained_filler.start(),
_retained_filler.end());
}
// Reset _hard_end to _true_end (and update _end)
if (retain && _hard_end != NULL) {
assert(_hard_end <= _true_end, "Invariant.");
_hard_end = _true_end;
_end = MAX2(_top, _hard_end - AlignmentReserve);
assert(_end <= _hard_end, "Invariant.");
}
_true_end = _hard_end;
HeapWord* pre_top = _top;
ParGCAllocBuffer::retire(end_of_gc, retain);
// Now any old _retained_filler is cut back to size, the free part is
// filled with a filler object, and top is past the header of that
// object.
if (retain && _top < _end) {
assert(end_of_gc && retain, "Or else retain should be false.");
// If the lab does not start on a card boundary, we don't want to
// allocate onto that card, since that might lead to concurrent
// allocation and card scanning, which we don't support. So we fill
// the first card with a garbage object.
size_t first_card_index = _bsa->index_for(pre_top);
HeapWord* first_card_start = _bsa->address_for_index(first_card_index);
if (first_card_start < pre_top) {
HeapWord* second_card_start =
_bsa->inc_by_region_size(first_card_start);
// Ensure enough room to fill with the smallest block
second_card_start = MAX2(second_card_start, pre_top + AlignmentReserve);
// If the end is already in the first card, don't go beyond it!
// Or if the remainder is too small for a filler object, gobble it up.
if (_hard_end < second_card_start ||
pointer_delta(_hard_end, second_card_start) < AlignmentReserve) {
second_card_start = _hard_end;
}
if (pre_top < second_card_start) {
MemRegion first_card_suffix(pre_top, second_card_start);
fill_region_with_block(first_card_suffix, true);
}
pre_top = second_card_start;
_top = pre_top;
_end = MAX2(_top, _hard_end - AlignmentReserve);
}
// If the lab does not end on a card boundary, we don't want to
// allocate onto that card, since that might lead to concurrent
// allocation and card scanning, which we don't support. So we fill
// the last card with a garbage object.
size_t last_card_index = _bsa->index_for(_hard_end);
HeapWord* last_card_start = _bsa->address_for_index(last_card_index);
if (last_card_start < _hard_end) {
// Ensure enough room to fill with the smallest block
last_card_start = MIN2(last_card_start, _hard_end - AlignmentReserve);
// If the top is already in the last card, don't go back beyond it!
// Or if the remainder is too small for a filler object, gobble it up.
if (_top > last_card_start ||
pointer_delta(last_card_start, _top) < AlignmentReserve) {
last_card_start = _top;
}
if (last_card_start < _hard_end) {
MemRegion last_card_prefix(last_card_start, _hard_end);
fill_region_with_block(last_card_prefix, false);
}
_hard_end = last_card_start;
_end = MAX2(_top, _hard_end - AlignmentReserve);
_true_end = _hard_end;
assert(_end <= _hard_end, "Invariant.");
}
// At this point:
// 1) we had a filler object from the original top to hard_end.
// 2) We've filled in any partial cards at the front and back.
if (pre_top < _hard_end) {
// Now we can reset the _bt to do allocation in the given area.
MemRegion new_filler(pre_top, _hard_end);
fill_region_with_block(new_filler, false);
_top = pre_top + ParGCAllocBuffer::FillerHeaderSize;
// If there's no space left, don't retain.
if (_top >= _end) {
_retained = false;
invalidate();
return;
}
_retained_filler = MemRegion(pre_top, _top);
_bt.set_region(MemRegion(_top, _hard_end));
_bt.initialize_threshold();
assert(_bt.threshold() > _top, "initialize_threshold failed!");
// There may be other reasons for queries into the middle of the
// filler object. When such queries are done in parallel with
// allocation, bad things can happen, if the query involves object
// iteration. So we ensure that such queries do not involve object
// iteration, by putting another filler object on the boundaries of
// such queries. One such is the object spanning a parallel card
// chunk boundary.
// "chunk_boundary" is the address of the first chunk boundary less
// than "hard_end".
HeapWord* chunk_boundary =
(HeapWord*)align_size_down(intptr_t(_hard_end-1), ChunkSizeInBytes);
assert(chunk_boundary < _hard_end, "Or else above did not work.");
assert(pointer_delta(_true_end, chunk_boundary) >= AlignmentReserve,
"Consequence of last card handling above.");
if (_top <= chunk_boundary) {
assert(_true_end == _hard_end, "Invariant.");
while (_top <= chunk_boundary) {
assert(pointer_delta(_hard_end, chunk_boundary) >= AlignmentReserve,
"Consequence of last card handling above.");
_bt.BlockOffsetArray::alloc_block(chunk_boundary, _hard_end);
CollectedHeap::fill_with_object(chunk_boundary, _hard_end);
_hard_end = chunk_boundary;
chunk_boundary -= ChunkSizeInWords;
}
_end = _hard_end - AlignmentReserve;
assert(_top <= _end, "Invariant.");
// Now reset the initial filler chunk so it doesn't overlap with
// the one(s) inserted above.
MemRegion new_filler(pre_top, _hard_end);
fill_region_with_block(new_filler, false);
}
} else {
_retained = false;
invalidate();
}
} else {
assert(!end_of_gc ||
(!_retained && _true_end == _hard_end), "Checking.");
}
assert(_end <= _hard_end, "Invariant.");
assert(_top < _end || _top == _hard_end, "Invariant");
}