Skip to content
Permalink
Browse files

8253237: [REDO] Improve large object handling during evacuation

8253238: [REDO] Improve object array chunking test in G1's copy_to_survivor_space
8253236: [REDO] Investigate fast-path for scanning only objects with references during gc

Generate multiple partial array tasks for large objArrays.

Reviewed-by: sjohanss, ayang, tschatzl
  • Loading branch information
Kim Barrett
Kim Barrett committed Sep 22, 2020
1 parent d1f9b8a commit 0e98fc1ccd3a4053ccb178045aad97ee45fdc1ca
@@ -31,10 +31,12 @@
#include "gc/g1/g1RootClosures.hpp"
#include "gc/g1/g1StringDedup.hpp"
#include "gc/g1/g1Trace.hpp"
#include "gc/shared/partialArrayTaskStepper.inline.hpp"
#include "gc/shared/taskqueue.inline.hpp"
#include "memory/allocation.inline.hpp"
#include "oops/access.inline.hpp"
#include "oops/oop.inline.hpp"
#include "runtime/atomic.hpp"
#include "runtime/prefetch.inline.hpp"
#include "utilities/globalDefinitions.hpp"
#include "utilities/macros.hpp"
@@ -50,6 +52,7 @@
G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
G1RedirtyCardsQueueSet* rdcqs,
uint worker_id,
uint n_workers,
size_t young_cset_length,
size_t optional_cset_length)
: _g1h(g1h),
@@ -70,6 +73,8 @@ G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
_surviving_young_words(NULL),
_surviving_words_length(young_cset_length + 1),
_old_gen_is_full(false),
_partial_objarray_chunk_size(ParGCArrayScanChunk),
_partial_array_stepper(n_workers),
_num_optional_regions(optional_cset_length),
_numa(g1h->numa()),
_obj_alloc_stat(NULL)
@@ -212,48 +217,60 @@ void G1ParScanThreadState::do_partial_array(PartialArrayScanTask task) {

assert(_g1h->is_in_reserved(from_obj), "must be in heap.");
assert(from_obj->is_objArray(), "must be obj array");
objArrayOop from_obj_array = objArrayOop(from_obj);
// The from-space object contains the real length.
int length = from_obj_array->length();

assert(from_obj->is_forwarded(), "must be forwarded");
oop to_obj = from_obj->forwardee();

oop to_obj = from_obj->forwardee();
assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
objArrayOop to_obj_array = objArrayOop(to_obj);
// We keep track of the next start index in the length field of the
// to-space object.
int next_index = to_obj_array->length();
assert(0 <= next_index && next_index < length,
"invariant, next index: %d, length: %d", next_index, length);

int start = next_index;
int end = length;
int remainder = end - start;
// We'll try not to push a range that's smaller than ParGCArrayScanChunk.
if (remainder > 2 * ParGCArrayScanChunk) {
end = start + ParGCArrayScanChunk;
to_obj_array->set_length(end);
// Push the remainder before we process the range in case another
// worker has run out of things to do and can steal it.
assert(to_obj->is_objArray(), "must be obj array");
objArrayOop to_array = objArrayOop(to_obj);

PartialArrayTaskStepper::Step step
= _partial_array_stepper.next(objArrayOop(from_obj),
to_array,
_partial_objarray_chunk_size);
for (uint i = 0; i < step._ncreate; ++i) {
push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
} else {
assert(length == end, "sanity");
// We'll process the final range for this object. Restore the length
// so that the heap remains parsable in case of evacuation failure.
to_obj_array->set_length(end);
}

HeapRegion* hr = _g1h->heap_region_containing(to_obj);
HeapRegion* hr = _g1h->heap_region_containing(to_array);
G1ScanInYoungSetter x(&_scanner, hr->is_young());
// Process indexes [start,end). It will also process the header
// along with the first chunk (i.e., the chunk with start == 0).
// Note that at this point the length field of to_obj_array is not
// correct given that we are using it to keep track of the next
// start index. oop_iterate_range() (thankfully!) ignores the length
// field and only relies on the start / end parameters. It does
// however return the size of the object which will be incorrect. So
// we have to ignore it even if we wanted to use it.
to_obj_array->oop_iterate_range(&_scanner, start, end);
// Process claimed task. The length of to_array is not correct, but
// fortunately the iteration ignores the length field and just relies
// on start/end.
to_array->oop_iterate_range(&_scanner,
step._index,
step._index + _partial_objarray_chunk_size);
}

MAYBE_INLINE_EVACUATION
void G1ParScanThreadState::start_partial_objarray(G1HeapRegionAttr dest_attr,
oop from_obj,
oop to_obj) {
assert(from_obj->is_objArray(), "precondition");
assert(from_obj->is_forwarded(), "precondition");
assert(from_obj->forwardee() == to_obj, "precondition");
assert(from_obj != to_obj, "should not be scanning self-forwarded objects");
assert(to_obj->is_objArray(), "precondition");

objArrayOop to_array = objArrayOop(to_obj);

PartialArrayTaskStepper::Step step
= _partial_array_stepper.start(objArrayOop(from_obj),
to_array,
_partial_objarray_chunk_size);

// Push any needed partial scan tasks. Pushed before processing the
// intitial chunk to allow other workers to steal while we're processing.
for (uint i = 0; i < step._ncreate; ++i) {
push_on_queue(ScannerTask(PartialArrayScanTask(from_obj)));
}

G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
// Process the initial chunk. No need to process the type in the
// klass, as it will already be handled by processing the built-in
// module. The length of to_array is not correct, but fortunately
// the iteration ignores that length field and relies on start/end.
to_array->oop_iterate_range(&_scanner, 0, step._index);
}

MAYBE_INLINE_EVACUATION
@@ -409,7 +426,10 @@ oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const regio
assert(region_attr.is_in_cset(),
"Unexpected region attr type: %s", region_attr.get_type_str());

const size_t word_sz = old->size();
// Get the klass once. We'll need it again later, and this avoids
// re-decoding when it's compressed.
Klass* klass = old->klass();
const size_t word_sz = old->size_given_klass(klass);

uint age = 0;
G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
@@ -476,6 +496,20 @@ oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const regio
obj->set_mark_raw(old_mark);
}

// Most objects are not arrays, so do one array check rather than
// checking for each array category for each object.
if (klass->is_array_klass()) {
if (klass->is_objArray_klass()) {
start_partial_objarray(dest_attr, old, obj);
} else {
// Nothing needs to be done for typeArrays. Body doesn't contain
// any oops to scan, and the type in the klass will already be handled
// by processing the built-in module.
assert(klass->is_typeArray_klass(), "invariant");
}
return obj;
}

if (G1StringDedup::is_enabled()) {
const bool is_from_young = region_attr.is_young();
const bool is_to_young = dest_attr.is_young();
@@ -489,17 +523,10 @@ oop G1ParScanThreadState::do_copy_to_survivor_space(G1HeapRegionAttr const regio
obj);
}

if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
// We keep track of the next start index in the length field of
// the to-space object. The actual length can be found in the
// length field of the from-space object.
arrayOop(obj)->set_length(0);
do_partial_array(PartialArrayScanTask(old));
} else {
G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
obj->oop_iterate_backwards(&_scanner);
}
G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
obj->oop_iterate_backwards(&_scanner);
return obj;

} else {
_plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
return forward_ptr;
@@ -518,7 +545,9 @@ G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id)
assert(worker_id < _n_workers, "out of bounds access");
if (_states[worker_id] == NULL) {
_states[worker_id] =
new G1ParScanThreadState(_g1h, _rdcqs, worker_id, _young_cset_length, _optional_cset_length);
new G1ParScanThreadState(_g1h, _rdcqs,
worker_id, _n_workers,
_young_cset_length, _optional_cset_length);
}
return _states[worker_id];
}
@@ -32,6 +32,7 @@
#include "gc/g1/g1RemSet.hpp"
#include "gc/g1/heapRegionRemSet.hpp"
#include "gc/shared/ageTable.hpp"
#include "gc/shared/partialArrayTaskStepper.hpp"
#include "gc/shared/taskqueue.hpp"
#include "memory/allocation.hpp"
#include "oops/oop.hpp"
@@ -79,6 +80,9 @@ class G1ParScanThreadState : public CHeapObj<mtGC> {
// Indicates whether in the last generation (old) there is no more space
// available for allocation.
bool _old_gen_is_full;
// Size (in elements) of a partial objArray task chunk.
int _partial_objarray_chunk_size;
PartialArrayTaskStepper _partial_array_stepper;

G1RedirtyCardsQueue& redirty_cards_queue() { return _rdcq; }
G1CardTable* ct() { return _ct; }
@@ -105,6 +109,7 @@ class G1ParScanThreadState : public CHeapObj<mtGC> {
G1ParScanThreadState(G1CollectedHeap* g1h,
G1RedirtyCardsQueueSet* rdcqs,
uint worker_id,
uint n_workers,
size_t young_cset_length,
size_t optional_cset_length);
virtual ~G1ParScanThreadState();
@@ -157,6 +162,7 @@ class G1ParScanThreadState : public CHeapObj<mtGC> {

private:
void do_partial_array(PartialArrayScanTask task);
void start_partial_objarray(G1HeapRegionAttr dest_dir, oop from, oop to);

HeapWord* allocate_copy_slow(G1HeapRegionAttr* dest_attr,
oop old,
@@ -249,9 +255,6 @@ class G1ParScanThreadStateSet : public StackObj {
G1ParScanThreadState* state_for_worker(uint worker_id);

const size_t* surviving_young_words() const;

private:
G1ParScanThreadState* new_par_scan_state(uint worker_id, size_t young_cset_length);
};

#endif // SHARE_GC_G1_G1PARSCANTHREADSTATE_HPP
@@ -0,0 +1,53 @@
/*
* Copyright (c) 2020, 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/shared/partialArrayTaskStepper.hpp"
#include "oops/arrayOop.hpp"
#include "utilities/globalDefinitions.hpp"

static uint compute_task_limit(uint n_workers) {
// Don't need more than n_workers tasks at a time. But allowing up to
// that maximizes available parallelism.
return n_workers;
}

static uint compute_task_fanout(uint task_limit) {
assert(task_limit > 0, "precondition");
// There is a tradeoff between providing parallelism more quickly and
// number of enqueued tasks. A constant fanout may be too slow when
// parallelism (and so task_limit) is large. A constant fraction might
// be overly eager. Using log2 attempts to balance between those.
uint result = log2_uint(task_limit);
// result must be > 0. result should be > 1 if task_limit > 1, to
// provide some potentially parallel tasks. But don't just +1 to
// avoid otherwise increasing rate of task generation.
if (result < 2) ++result;
return result;
}

PartialArrayTaskStepper::PartialArrayTaskStepper(uint n_workers) :
_task_limit(compute_task_limit(n_workers)),
_task_fanout(compute_task_fanout(_task_limit))
{}
@@ -0,0 +1,82 @@
/*
* Copyright (c) 2020, 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.
*
*/

#ifndef SHARE_GC_SHARED_PARTIALARRAYTASKSTEPPER_HPP
#define SHARE_GC_SHARED_PARTIALARRAYTASKSTEPPER_HPP

#include "oops/arrayOop.hpp"
#include "utilities/globalDefinitions.hpp"

// Helper for handling PartialArrayTasks.
//
// When an array is large, we want to split it up into chunks that can be
// processed in parallel. Each task (implicitly) represents such a chunk.
// We can enqueue multiple tasks at the same time. We want to enqueue
// enough tasks to benefit from the available parallelism, while not so many
// as to substantially expand the task queues.
//
// A task directly refers to the from-space array. The from-space array's
// forwarding pointer refers to the associated to-space array, and its
// length is the actual length. The to-space array's length field is used to
// indicate processing progress. It is the starting index of the next chunk
// to process, or equals the actual length when there are no more chunks to
// be processed.
class PartialArrayTaskStepper {
public:
PartialArrayTaskStepper(uint n_workers);

struct Step {
int _index; // Array index for the step.
uint _ncreate; // Number of new tasks to create.
};

// Set to's length to the end of the initial chunk, which is the start of
// the first partial task if the array is large enough to need splitting.
// Returns a Step with _index being that index and _ncreate being the
// initial number of partial tasks to enqueue.
inline Step start(arrayOop from, arrayOop to, int chunk_size) const;

// Increment to's length by chunk_size to claim the next chunk. Returns a
// Step with _index being the starting index of the claimed chunk and
// _ncreate being the number of additional partial tasks to enqueue.
// precondition: chunk_size must be the same as used to start the task sequence.
inline Step next(arrayOop from, arrayOop to, int chunk_size) const;

class TestSupport; // For unit tests

private:
// Limit on the number of partial array tasks to create for a given array.
uint _task_limit;
// Maximum number of new tasks to create when processing an existing task.
uint _task_fanout;

// Split start/next into public part dealing with oops and private
// impl dealing with lengths and pointers to lengths, for unit testing.
// length is the actual length obtained from the from-space object.
// to_length_addr is the address of the to-space object's length value.
inline Step start_impl(int length, int* to_length_addr, int chunk_size) const;
inline Step next_impl(int length, int* to_length_addr, int chunk_size) const;
};

#endif // SHARE_GC_SHARED_PARTIALARRAYTASKSTEPPER_HPP

1 comment on commit 0e98fc1

@bridgekeeper

This comment has been minimized.

Copy link

@bridgekeeper bridgekeeper bot commented on 0e98fc1 Sep 22, 2020

Please sign in to comment.
You can’t perform that action at this time.