duke@435: /*
stefank@2314:  * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
duke@435:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435:  *
duke@435:  * This code is free software; you can redistribute it and/or modify it
duke@435:  * under the terms of the GNU General Public License version 2 only, as
duke@435:  * published by the Free Software Foundation.
duke@435:  *
duke@435:  * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
duke@435:  * version 2 for more details (a copy is included in the LICENSE file that
duke@435:  * accompanied this code).
duke@435:  *
duke@435:  * You should have received a copy of the GNU General Public License version
duke@435:  * 2 along with this work; if not, write to the Free Software Foundation,
duke@435:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435:  *
trims@1907:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907:  * or visit www.oracle.com if you need additional information or have any
trims@1907:  * questions.
duke@435:  *
duke@435:  */
duke@435: 
stefank@2314: #include "precompiled.hpp"
stefank@2314: #include "gc_implementation/parNew/parGCAllocBuffer.hpp"
stefank@2314: #include "memory/sharedHeap.hpp"
stefank@2314: #include "oops/arrayOop.hpp"
stefank@2314: #include "oops/oop.inline.hpp"
duke@435: 
duke@435: ParGCAllocBuffer::ParGCAllocBuffer(size_t desired_plab_sz_) :
duke@435:   _word_sz(desired_plab_sz_), _bottom(NULL), _top(NULL),
duke@435:   _end(NULL), _hard_end(NULL),
duke@435:   _retained(false), _retained_filler(),
duke@435:   _allocated(0), _wasted(0)
duke@435: {
duke@435:   assert (min_size() > AlignmentReserve, "Inconsistency!");
coleenp@548:   // arrayOopDesc::header_size depends on command line initialization.
coleenp@548:   FillerHeaderSize = align_object_size(arrayOopDesc::header_size(T_INT));
coleenp@548:   AlignmentReserve = oopDesc::header_size() > MinObjAlignment ? FillerHeaderSize : 0;
duke@435: }
duke@435: 
coleenp@548: size_t ParGCAllocBuffer::FillerHeaderSize;
duke@435: 
duke@435: // If the minimum object size is greater than MinObjAlignment, we can
duke@435: // end up with a shard at the end of the buffer that's smaller than
duke@435: // the smallest object.  We can't allow that because the buffer must
duke@435: // look like it's full of objects when we retire it, so we make
duke@435: // sure we have enough space for a filler int array object.
coleenp@548: size_t ParGCAllocBuffer::AlignmentReserve;
duke@435: 
duke@435: void ParGCAllocBuffer::retire(bool end_of_gc, bool retain) {
duke@435:   assert(!retain || end_of_gc, "Can only retain at GC end.");
duke@435:   if (_retained) {
duke@435:     // If the buffer had been retained shorten the previous filler object.
duke@435:     assert(_retained_filler.end() <= _top, "INVARIANT");
jcoomes@916:     CollectedHeap::fill_with_object(_retained_filler);
duke@435:     // Wasted space book-keeping, otherwise (normally) done in invalidate()
duke@435:     _wasted += _retained_filler.word_size();
duke@435:     _retained = false;
duke@435:   }
duke@435:   assert(!end_of_gc || !_retained, "At this point, end_of_gc ==> !_retained.");
duke@435:   if (_top < _hard_end) {
jcoomes@916:     CollectedHeap::fill_with_object(_top, _hard_end);
duke@435:     if (!retain) {
duke@435:       invalidate();
duke@435:     } else {
duke@435:       // Is there wasted space we'd like to retain for the next GC?
duke@435:       if (pointer_delta(_end, _top) > FillerHeaderSize) {
duke@435:         _retained = true;
duke@435:         _retained_filler = MemRegion(_top, FillerHeaderSize);
duke@435:         _top = _top + FillerHeaderSize;
duke@435:       } else {
duke@435:         invalidate();
duke@435:       }
duke@435:     }
duke@435:   }
duke@435: }
duke@435: 
duke@435: void ParGCAllocBuffer::flush_stats(PLABStats* stats) {
duke@435:   assert(ResizePLAB, "Wasted work");
duke@435:   stats->add_allocated(_allocated);
duke@435:   stats->add_wasted(_wasted);
duke@435:   stats->add_unused(pointer_delta(_end, _top));
duke@435: }
duke@435: 
duke@435: // Compute desired plab size and latch result for later
duke@435: // use. This should be called once at the end of parallel
duke@435: // scavenge; it clears the sensor accumulators.
duke@435: void PLABStats::adjust_desired_plab_sz() {
duke@435:   assert(ResizePLAB, "Not set");
duke@435:   if (_allocated == 0) {
duke@435:     assert(_unused == 0, "Inconsistency in PLAB stats");
duke@435:     _allocated = 1;
duke@435:   }
duke@435:   double wasted_frac    = (double)_unused/(double)_allocated;
duke@435:   size_t target_refills = (size_t)((wasted_frac*TargetSurvivorRatio)/
duke@435:                                    TargetPLABWastePct);
duke@435:   if (target_refills == 0) {
duke@435:     target_refills = 1;
duke@435:   }
duke@435:   _used = _allocated - _wasted - _unused;
duke@435:   size_t plab_sz = _used/(target_refills*ParallelGCThreads);
duke@435:   if (PrintPLAB) gclog_or_tty->print(" (plab_sz = %d ", plab_sz);
duke@435:   // Take historical weighted average
duke@435:   _filter.sample(plab_sz);
duke@435:   // Clip from above and below, and align to object boundary
duke@435:   plab_sz = MAX2(min_size(), (size_t)_filter.average());
duke@435:   plab_sz = MIN2(max_size(), plab_sz);
duke@435:   plab_sz = align_object_size(plab_sz);
duke@435:   // Latch the result
duke@435:   if (PrintPLAB) gclog_or_tty->print(" desired_plab_sz = %d) ", plab_sz);
duke@435:   if (ResizePLAB) {
duke@435:     _desired_plab_sz = plab_sz;
duke@435:   }
duke@435:   // Now clear the accumulators for next round:
duke@435:   // note this needs to be fixed in the case where we
duke@435:   // are retaining across scavenges. FIX ME !!! XXX
duke@435:   _allocated = 0;
duke@435:   _wasted    = 0;
duke@435:   _unused    = 0;
duke@435: }
duke@435: 
duke@435: #ifndef PRODUCT
duke@435: void ParGCAllocBuffer::print() {
duke@435:   gclog_or_tty->print("parGCAllocBuffer: _bottom: %p  _top: %p  _end: %p  _hard_end: %p"
duke@435:              "_retained: %c _retained_filler: [%p,%p)\n",
duke@435:              _bottom, _top, _end, _hard_end,
duke@435:              "FT"[_retained], _retained_filler.start(), _retained_filler.end());
duke@435: }
duke@435: #endif // !PRODUCT
duke@435: 
duke@435: const size_t ParGCAllocBufferWithBOT::ChunkSizeInWords =
duke@435: MIN2(CardTableModRefBS::par_chunk_heapword_alignment(),
duke@435:      ((size_t)Generation::GenGrain)/HeapWordSize);
duke@435: const size_t ParGCAllocBufferWithBOT::ChunkSizeInBytes =
duke@435: MIN2(CardTableModRefBS::par_chunk_heapword_alignment() * HeapWordSize,
duke@435:      (size_t)Generation::GenGrain);
duke@435: 
duke@435: ParGCAllocBufferWithBOT::ParGCAllocBufferWithBOT(size_t word_sz,
duke@435:                                                  BlockOffsetSharedArray* bsa) :
duke@435:   ParGCAllocBuffer(word_sz),
duke@435:   _bsa(bsa),
duke@435:   _bt(bsa, MemRegion(_bottom, _hard_end)),
duke@435:   _true_end(_hard_end)
duke@435: {}
duke@435: 
duke@435: // The buffer comes with its own BOT, with a shared (obviously) underlying
duke@435: // BlockOffsetSharedArray. We manipulate this BOT in the normal way
duke@435: // as we would for any contiguous space. However, on accasion we
duke@435: // need to do some buffer surgery at the extremities before we
duke@435: // start using the body of the buffer for allocations. Such surgery
duke@435: // (as explained elsewhere) is to prevent allocation on a card that
duke@435: // is in the process of being walked concurrently by another GC thread.
duke@435: // When such surgery happens at a point that is far removed (to the
duke@435: // right of the current allocation point, top), we use the "contig"
duke@435: // parameter below to directly manipulate the shared array without
duke@435: // modifying the _next_threshold state in the BOT.
duke@435: void ParGCAllocBufferWithBOT::fill_region_with_block(MemRegion mr,
duke@435:                                                      bool contig) {
jcoomes@916:   CollectedHeap::fill_with_object(mr);
duke@435:   if (contig) {
duke@435:     _bt.alloc_block(mr.start(), mr.end());
duke@435:   } else {
duke@435:     _bt.BlockOffsetArray::alloc_block(mr.start(), mr.end());
duke@435:   }
duke@435: }
duke@435: 
duke@435: HeapWord* ParGCAllocBufferWithBOT::allocate_slow(size_t word_sz) {
duke@435:   HeapWord* res = NULL;
duke@435:   if (_true_end > _hard_end) {
duke@435:     assert((HeapWord*)align_size_down(intptr_t(_hard_end),
duke@435:                                       ChunkSizeInBytes) == _hard_end,
duke@435:            "or else _true_end should be equal to _hard_end");
duke@435:     assert(_retained, "or else _true_end should be equal to _hard_end");
duke@435:     assert(_retained_filler.end() <= _top, "INVARIANT");
jcoomes@916:     CollectedHeap::fill_with_object(_retained_filler);
duke@435:     if (_top < _hard_end) {
duke@435:       fill_region_with_block(MemRegion(_top, _hard_end), true);
duke@435:     }
duke@435:     HeapWord* next_hard_end = MIN2(_true_end, _hard_end + ChunkSizeInWords);
duke@435:     _retained_filler = MemRegion(_hard_end, FillerHeaderSize);
duke@435:     _bt.alloc_block(_retained_filler.start(), _retained_filler.word_size());
duke@435:     _top      = _retained_filler.end();
duke@435:     _hard_end = next_hard_end;
duke@435:     _end      = _hard_end - AlignmentReserve;
duke@435:     res       = ParGCAllocBuffer::allocate(word_sz);
duke@435:     if (res != NULL) {
duke@435:       _bt.alloc_block(res, word_sz);
duke@435:     }
duke@435:   }
duke@435:   return res;
duke@435: }
duke@435: 
duke@435: void
duke@435: ParGCAllocBufferWithBOT::undo_allocation(HeapWord* obj, size_t word_sz) {
duke@435:   ParGCAllocBuffer::undo_allocation(obj, word_sz);
duke@435:   // This may back us up beyond the previous threshold, so reset.
duke@435:   _bt.set_region(MemRegion(_top, _hard_end));
duke@435:   _bt.initialize_threshold();
duke@435: }
duke@435: 
duke@435: void ParGCAllocBufferWithBOT::retire(bool end_of_gc, bool retain) {
duke@435:   assert(!retain || end_of_gc, "Can only retain at GC end.");
duke@435:   if (_retained) {
duke@435:     // We're about to make the retained_filler into a block.
duke@435:     _bt.BlockOffsetArray::alloc_block(_retained_filler.start(),
duke@435:                                       _retained_filler.end());
duke@435:   }
duke@435:   // Reset _hard_end to _true_end (and update _end)
duke@435:   if (retain && _hard_end != NULL) {
duke@435:     assert(_hard_end <= _true_end, "Invariant.");
duke@435:     _hard_end = _true_end;
duke@435:     _end      = MAX2(_top, _hard_end - AlignmentReserve);
duke@435:     assert(_end <= _hard_end, "Invariant.");
duke@435:   }
duke@435:   _true_end = _hard_end;
duke@435:   HeapWord* pre_top = _top;
duke@435: 
duke@435:   ParGCAllocBuffer::retire(end_of_gc, retain);
duke@435:   // Now any old _retained_filler is cut back to size, the free part is
duke@435:   // filled with a filler object, and top is past the header of that
duke@435:   // object.
duke@435: 
duke@435:   if (retain && _top < _end) {
duke@435:     assert(end_of_gc && retain, "Or else retain should be false.");
duke@435:     // If the lab does not start on a card boundary, we don't want to
duke@435:     // allocate onto that card, since that might lead to concurrent
duke@435:     // allocation and card scanning, which we don't support.  So we fill
duke@435:     // the first card with a garbage object.
duke@435:     size_t first_card_index = _bsa->index_for(pre_top);
duke@435:     HeapWord* first_card_start = _bsa->address_for_index(first_card_index);
duke@435:     if (first_card_start < pre_top) {
duke@435:       HeapWord* second_card_start =
jmasa@736:         _bsa->inc_by_region_size(first_card_start);
duke@435: 
duke@435:       // Ensure enough room to fill with the smallest block
duke@435:       second_card_start = MAX2(second_card_start, pre_top + AlignmentReserve);
duke@435: 
duke@435:       // If the end is already in the first card, don't go beyond it!
duke@435:       // Or if the remainder is too small for a filler object, gobble it up.
duke@435:       if (_hard_end < second_card_start ||
duke@435:           pointer_delta(_hard_end, second_card_start) < AlignmentReserve) {
duke@435:         second_card_start = _hard_end;
duke@435:       }
duke@435:       if (pre_top < second_card_start) {
duke@435:         MemRegion first_card_suffix(pre_top, second_card_start);
duke@435:         fill_region_with_block(first_card_suffix, true);
duke@435:       }
duke@435:       pre_top = second_card_start;
duke@435:       _top = pre_top;
duke@435:       _end = MAX2(_top, _hard_end - AlignmentReserve);
duke@435:     }
duke@435: 
duke@435:     // If the lab does not end on a card boundary, we don't want to
duke@435:     // allocate onto that card, since that might lead to concurrent
duke@435:     // allocation and card scanning, which we don't support.  So we fill
duke@435:     // the last card with a garbage object.
duke@435:     size_t last_card_index = _bsa->index_for(_hard_end);
duke@435:     HeapWord* last_card_start = _bsa->address_for_index(last_card_index);
duke@435:     if (last_card_start < _hard_end) {
duke@435: 
duke@435:       // Ensure enough room to fill with the smallest block
duke@435:       last_card_start = MIN2(last_card_start, _hard_end - AlignmentReserve);
duke@435: 
duke@435:       // If the top is already in the last card, don't go back beyond it!
duke@435:       // Or if the remainder is too small for a filler object, gobble it up.
duke@435:       if (_top > last_card_start ||
duke@435:           pointer_delta(last_card_start, _top) < AlignmentReserve) {
duke@435:         last_card_start = _top;
duke@435:       }
duke@435:       if (last_card_start < _hard_end) {
duke@435:         MemRegion last_card_prefix(last_card_start, _hard_end);
duke@435:         fill_region_with_block(last_card_prefix, false);
duke@435:       }
duke@435:       _hard_end = last_card_start;
duke@435:       _end      = MAX2(_top, _hard_end - AlignmentReserve);
duke@435:       _true_end = _hard_end;
duke@435:       assert(_end <= _hard_end, "Invariant.");
duke@435:     }
duke@435: 
duke@435:     // At this point:
duke@435:     //   1) we had a filler object from the original top to hard_end.
duke@435:     //   2) We've filled in any partial cards at the front and back.
duke@435:     if (pre_top < _hard_end) {
duke@435:       // Now we can reset the _bt to do allocation in the given area.
duke@435:       MemRegion new_filler(pre_top, _hard_end);
duke@435:       fill_region_with_block(new_filler, false);
duke@435:       _top = pre_top + ParGCAllocBuffer::FillerHeaderSize;
duke@435:       // If there's no space left, don't retain.
duke@435:       if (_top >= _end) {
duke@435:         _retained = false;
duke@435:         invalidate();
duke@435:         return;
duke@435:       }
duke@435:       _retained_filler = MemRegion(pre_top, _top);
duke@435:       _bt.set_region(MemRegion(_top, _hard_end));
duke@435:       _bt.initialize_threshold();
duke@435:       assert(_bt.threshold() > _top, "initialize_threshold failed!");
duke@435: 
duke@435:       // There may be other reasons for queries into the middle of the
duke@435:       // filler object.  When such queries are done in parallel with
duke@435:       // allocation, bad things can happen, if the query involves object
duke@435:       // iteration.  So we ensure that such queries do not involve object
duke@435:       // iteration, by putting another filler object on the boundaries of
duke@435:       // such queries.  One such is the object spanning a parallel card
duke@435:       // chunk boundary.
duke@435: 
duke@435:       // "chunk_boundary" is the address of the first chunk boundary less
duke@435:       // than "hard_end".
duke@435:       HeapWord* chunk_boundary =
duke@435:         (HeapWord*)align_size_down(intptr_t(_hard_end-1), ChunkSizeInBytes);
duke@435:       assert(chunk_boundary < _hard_end, "Or else above did not work.");
duke@435:       assert(pointer_delta(_true_end, chunk_boundary) >= AlignmentReserve,
duke@435:              "Consequence of last card handling above.");
duke@435: 
duke@435:       if (_top <= chunk_boundary) {
duke@435:         assert(_true_end == _hard_end, "Invariant.");
duke@435:         while (_top <= chunk_boundary) {
duke@435:           assert(pointer_delta(_hard_end, chunk_boundary) >= AlignmentReserve,
duke@435:                  "Consequence of last card handling above.");
jcoomes@916:           _bt.BlockOffsetArray::alloc_block(chunk_boundary, _hard_end);
jcoomes@916:           CollectedHeap::fill_with_object(chunk_boundary, _hard_end);
jcoomes@916:           _hard_end = chunk_boundary;
duke@435:           chunk_boundary -= ChunkSizeInWords;
duke@435:         }
duke@435:         _end = _hard_end - AlignmentReserve;
duke@435:         assert(_top <= _end, "Invariant.");
duke@435:         // Now reset the initial filler chunk so it doesn't overlap with
duke@435:         // the one(s) inserted above.
duke@435:         MemRegion new_filler(pre_top, _hard_end);
duke@435:         fill_region_with_block(new_filler, false);
duke@435:       }
duke@435:     } else {
duke@435:       _retained = false;
duke@435:       invalidate();
duke@435:     }
duke@435:   } else {
duke@435:     assert(!end_of_gc ||
duke@435:            (!_retained && _true_end == _hard_end), "Checking.");
duke@435:   }
duke@435:   assert(_end <= _hard_end, "Invariant.");
duke@435:   assert(_top < _end || _top == _hard_end, "Invariant");
duke@435: }