diff -r a11d3a5ca20b -r 1f9dd2360b17 src/hotspot/share/gc/cms/cmsCardTable.cpp --- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/src/hotspot/share/gc/cms/cmsCardTable.cpp Sat Mar 24 01:08:35 2018 +0100 @@ -0,0 +1,432 @@ +/* + * Copyright (c) 2007, 2018, 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/cms/cmsHeap.hpp" +#include "gc/shared/cardTableBarrierSet.hpp" +#include "gc/shared/cardTableRS.hpp" +#include "gc/shared/collectedHeap.hpp" +#include "gc/shared/space.inline.hpp" +#include "memory/allocation.inline.hpp" +#include "memory/virtualspace.hpp" +#include "oops/oop.inline.hpp" +#include "runtime/java.hpp" +#include "runtime/mutexLocker.hpp" +#include "runtime/orderAccess.inline.hpp" +#include "runtime/vmThread.hpp" + +void CardTableRS:: +non_clean_card_iterate_parallel_work(Space* sp, MemRegion mr, + OopsInGenClosure* cl, + CardTableRS* ct, + uint n_threads) { + assert(n_threads > 0, "expected n_threads > 0"); + assert(n_threads <= ParallelGCThreads, + "n_threads: %u > ParallelGCThreads: %u", n_threads, ParallelGCThreads); + + // 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); + + uint n_strides = n_threads * ParGCStridesPerThread; + SequentialSubTasksDone* pst = sp->par_seq_tasks(); + // Sets the condition for completion of the subtask (how many threads + // need to finish in order to be done). + pst->set_n_threads(n_threads); + pst->set_n_tasks(n_strides); + + uint 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 +CardTableRS:: +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"); + + // This function is used by the parallel card table iteration. + const bool parallel = true; + + DirtyCardToOopClosure* dcto_cl = sp->new_dcto_cl(cl, precision(), + cl->gen_boundary(), + parallel); + ClearNoncleanCardWrapper clear_cl(dcto_cl, ct, parallel); + + + // 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 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; + } +} + +void +CardTableRS:: +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 start_chunk_index = addr_to_chunk_index(chunk_mr.start()); + assert(start_chunk_index >= lowest_non_clean_base_chunk_index, "Bounds error."); + uintptr_t cur_chunk_index = start_chunk_index - lowest_non_clean_base_chunk_index; + + // 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 neighbor + // 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) { + assert(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; + } + } else { + // In this case we can help our neighbor by just asking them + // to stop at our first card (even though it may not be dirty). + 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; + } + + // 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(); + } 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(); + } 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)) { + 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"); + } 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"); + } + 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. However, we need to bound this checking to the largest + // entry in the LNC array: this is because the heap may expand + // after the LNC array has been created but before we reach this point, + // and the last block in our chunk may have been expanded to include + // the expansion delta (and possibly subsequently allocated from, so + // it wouldn't be sufficient to check whether that last block was + // or was not an object at this point). + uintptr_t last_chunk_index_to_check = addr_to_chunk_index(last_block + last_block_size - 1) + - lowest_non_clean_base_chunk_index; + const uintptr_t last_chunk_index = addr_to_chunk_index(used.last()) + - lowest_non_clean_base_chunk_index; + if (last_chunk_index_to_check > last_chunk_index) { + assert(last_block + last_block_size > used.end(), + "Inconsistency detected: last_block [" PTR_FORMAT "," PTR_FORMAT "]" + " does not exceed used.end() = " PTR_FORMAT "," + " yet last_chunk_index_to_check " INTPTR_FORMAT + " exceeds last_chunk_index " INTPTR_FORMAT, + p2i(last_block), p2i(last_block + last_block_size), + p2i(used.end()), + last_chunk_index_to_check, last_chunk_index); + assert(sp->used_region().end() > used.end(), + "Expansion did not happen: " + "[" PTR_FORMAT "," PTR_FORMAT ") -> [" PTR_FORMAT "," PTR_FORMAT ")", + p2i(sp->used_region().start()), p2i(sp->used_region().end()), + p2i(used.start()), p2i(used.end())); + 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) { + 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(); + } + 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 +} + +void +CardTableRS:: +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 = CMSHeap::heap()->total_collections(); + // Updated _last_LNC_resizing_collection[i] must not be visible before + // _lowest_non_clean and friends are visible. Therefore use acquire/release + // to guarantee this on non TSO architecures. + if (OrderAccess::load_acquire(&_last_LNC_resizing_collection[i]) != cur_collection) { + MutexLocker x(ParGCRareEvent_lock); + // This load_acquire is here for clarity only. The MutexLocker already fences. + if (OrderAccess::load_acquire(&_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, mtGC); + _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; + } + } + // Make sure this gets visible only after _lowest_non_clean* was initialized + OrderAccess::release_store(&_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]; +}