duke@435: /* duke@435: * Copyright (c) 2007 Sun Microsystems, Inc. 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: * duke@435: * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, duke@435: * CA 95054 USA or visit www.sun.com if you need additional information or duke@435: * have any questions. duke@435: * duke@435: */ duke@435: duke@435: # include "incls/_precompiled.incl" duke@435: # include "incls/_parCardTableModRefBS.cpp.incl" duke@435: duke@435: void CardTableModRefBS::par_non_clean_card_iterate_work(Space* sp, MemRegion mr, duke@435: DirtyCardToOopClosure* dcto_cl, duke@435: MemRegionClosure* cl, duke@435: bool clear, duke@435: int n_threads) { duke@435: if (n_threads > 0) { duke@435: assert(n_threads == (int)ParallelGCThreads, "# worker threads != # requested!"); duke@435: duke@435: // Make sure the LNC array is valid for the space. duke@435: jbyte** lowest_non_clean; duke@435: uintptr_t lowest_non_clean_base_chunk_index; duke@435: size_t lowest_non_clean_chunk_size; duke@435: get_LNC_array_for_space(sp, lowest_non_clean, duke@435: lowest_non_clean_base_chunk_index, duke@435: lowest_non_clean_chunk_size); duke@435: duke@435: int n_strides = n_threads * StridesPerThread; duke@435: SequentialSubTasksDone* pst = sp->par_seq_tasks(); duke@435: pst->set_par_threads(n_threads); duke@435: pst->set_n_tasks(n_strides); duke@435: duke@435: int stride = 0; duke@435: while (!pst->is_task_claimed(/* reference */ stride)) { duke@435: process_stride(sp, mr, stride, n_strides, dcto_cl, cl, clear, duke@435: lowest_non_clean, duke@435: lowest_non_clean_base_chunk_index, duke@435: lowest_non_clean_chunk_size); duke@435: } duke@435: if (pst->all_tasks_completed()) { duke@435: // Clear lowest_non_clean array for next time. duke@435: intptr_t first_chunk_index = addr_to_chunk_index(mr.start()); duke@435: uintptr_t last_chunk_index = addr_to_chunk_index(mr.last()); duke@435: for (uintptr_t ch = first_chunk_index; ch <= last_chunk_index; ch++) { duke@435: intptr_t ind = ch - lowest_non_clean_base_chunk_index; duke@435: assert(0 <= ind && ind < (intptr_t)lowest_non_clean_chunk_size, duke@435: "Bounds error"); duke@435: lowest_non_clean[ind] = NULL; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: void duke@435: CardTableModRefBS:: duke@435: process_stride(Space* sp, duke@435: MemRegion used, duke@435: jint stride, int n_strides, duke@435: DirtyCardToOopClosure* dcto_cl, duke@435: MemRegionClosure* cl, duke@435: bool clear, duke@435: jbyte** lowest_non_clean, duke@435: uintptr_t lowest_non_clean_base_chunk_index, duke@435: size_t lowest_non_clean_chunk_size) { duke@435: // We don't have to go downwards here; it wouldn't help anyway, duke@435: // because of parallelism. duke@435: duke@435: // Find the first card address of the first chunk in the stride that is duke@435: // at least "bottom" of the used region. duke@435: jbyte* start_card = byte_for(used.start()); duke@435: jbyte* end_card = byte_after(used.last()); duke@435: uintptr_t start_chunk = addr_to_chunk_index(used.start()); duke@435: uintptr_t start_chunk_stride_num = start_chunk % n_strides; duke@435: jbyte* chunk_card_start; duke@435: duke@435: if ((uintptr_t)stride >= start_chunk_stride_num) { duke@435: chunk_card_start = (jbyte*)(start_card + duke@435: (stride - start_chunk_stride_num) * duke@435: CardsPerStrideChunk); duke@435: } else { duke@435: // Go ahead to the next chunk group boundary, then to the requested stride. duke@435: chunk_card_start = (jbyte*)(start_card + duke@435: (n_strides - start_chunk_stride_num + stride) * duke@435: CardsPerStrideChunk); duke@435: } duke@435: duke@435: while (chunk_card_start < end_card) { duke@435: // We don't have to go downwards here; it wouldn't help anyway, duke@435: // because of parallelism. (We take care with "min_done"; see below.) duke@435: // Invariant: chunk_mr should be fully contained within the "used" region. duke@435: jbyte* chunk_card_end = chunk_card_start + CardsPerStrideChunk; duke@435: MemRegion chunk_mr = MemRegion(addr_for(chunk_card_start), duke@435: chunk_card_end >= end_card ? duke@435: used.end() : addr_for(chunk_card_end)); duke@435: assert(chunk_mr.word_size() > 0, "[chunk_card_start > used_end)"); duke@435: assert(used.contains(chunk_mr), "chunk_mr should be subset of used"); duke@435: duke@435: // Process the chunk. duke@435: process_chunk_boundaries(sp, duke@435: dcto_cl, duke@435: chunk_mr, duke@435: used, duke@435: lowest_non_clean, duke@435: lowest_non_clean_base_chunk_index, duke@435: lowest_non_clean_chunk_size); duke@435: duke@435: non_clean_card_iterate_work(chunk_mr, cl, clear); duke@435: duke@435: // Find the next chunk of the stride. duke@435: chunk_card_start += CardsPerStrideChunk * n_strides; duke@435: } duke@435: } duke@435: duke@435: void duke@435: CardTableModRefBS:: duke@435: process_chunk_boundaries(Space* sp, duke@435: DirtyCardToOopClosure* dcto_cl, duke@435: MemRegion chunk_mr, duke@435: MemRegion used, duke@435: jbyte** lowest_non_clean, duke@435: uintptr_t lowest_non_clean_base_chunk_index, duke@435: size_t lowest_non_clean_chunk_size) duke@435: { duke@435: // We must worry about the chunk boundaries. duke@435: duke@435: // First, set our max_to_do: duke@435: HeapWord* max_to_do = NULL; duke@435: uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start()); duke@435: cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index; duke@435: duke@435: if (chunk_mr.end() < used.end()) { duke@435: // This is not the last chunk in the used region. What is the last duke@435: // object? duke@435: HeapWord* last_block = sp->block_start(chunk_mr.end()); duke@435: assert(last_block <= chunk_mr.end(), "In case this property changes."); duke@435: if (last_block == chunk_mr.end() duke@435: || !sp->block_is_obj(last_block)) { duke@435: max_to_do = chunk_mr.end(); duke@435: duke@435: } else { duke@435: // It is an object and starts before the end of the current chunk. duke@435: // last_obj_card is the card corresponding to the start of the last object duke@435: // in the chunk. Note that the last object may not start in duke@435: // the chunk. duke@435: jbyte* last_obj_card = byte_for(last_block); duke@435: if (!card_may_have_been_dirty(*last_obj_card)) { duke@435: // The card containing the head is not dirty. Any marks in duke@435: // subsequent cards still in this chunk must have been made duke@435: // precisely; we can cap processing at the end. duke@435: max_to_do = chunk_mr.end(); duke@435: } else { duke@435: // The last object must be considered dirty, and extends onto the duke@435: // following chunk. Look for a dirty card in that chunk that will duke@435: // bound our processing. duke@435: jbyte* limit_card = NULL; duke@435: size_t last_block_size = sp->block_size(last_block); duke@435: jbyte* last_card_of_last_obj = duke@435: byte_for(last_block + last_block_size - 1); duke@435: jbyte* first_card_of_next_chunk = byte_for(chunk_mr.end()); duke@435: // This search potentially goes a long distance looking duke@435: // for the next card that will be scanned. For example, duke@435: // an object that is an array of primitives will not duke@435: // have any cards covering regions interior to the array duke@435: // that will need to be scanned. The scan can be terminated duke@435: // at the last card of the next chunk. That would leave duke@435: // limit_card as NULL and would result in "max_to_do" duke@435: // being set with the LNC value or with the end duke@435: // of the last block. duke@435: jbyte* last_card_of_next_chunk = first_card_of_next_chunk + duke@435: CardsPerStrideChunk; duke@435: assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) duke@435: == CardsPerStrideChunk, "last card of next chunk may be wrong"); duke@435: jbyte* last_card_to_check = (jbyte*) MIN2(last_card_of_last_obj, duke@435: last_card_of_next_chunk); duke@435: for (jbyte* cur = first_card_of_next_chunk; duke@435: cur <= last_card_to_check; cur++) { duke@435: if (card_will_be_scanned(*cur)) { duke@435: limit_card = cur; break; duke@435: } duke@435: } duke@435: assert(0 <= cur_chunk_index+1 && duke@435: cur_chunk_index+1 < lowest_non_clean_chunk_size, duke@435: "Bounds error."); duke@435: // LNC for the next chunk duke@435: jbyte* lnc_card = lowest_non_clean[cur_chunk_index+1]; duke@435: if (limit_card == NULL) { duke@435: limit_card = lnc_card; duke@435: } duke@435: if (limit_card != NULL) { duke@435: if (lnc_card != NULL) { duke@435: limit_card = (jbyte*)MIN2((intptr_t)limit_card, duke@435: (intptr_t)lnc_card); duke@435: } duke@435: max_to_do = addr_for(limit_card); duke@435: } else { duke@435: max_to_do = last_block + last_block_size; duke@435: } duke@435: } duke@435: } duke@435: assert(max_to_do != NULL, "OOPS!"); duke@435: } else { duke@435: max_to_do = used.end(); duke@435: } duke@435: // Now we can set the closure we're using so it doesn't to beyond duke@435: // max_to_do. duke@435: dcto_cl->set_min_done(max_to_do); duke@435: #ifndef PRODUCT duke@435: dcto_cl->set_last_bottom(max_to_do); duke@435: #endif duke@435: duke@435: // Now we set *our" lowest_non_clean entry. duke@435: // Find the object that spans our boundary, if one exists. duke@435: // Nothing to do on the first chunk. duke@435: if (chunk_mr.start() > used.start()) { duke@435: // first_block is the block possibly spanning the chunk start duke@435: HeapWord* first_block = sp->block_start(chunk_mr.start()); duke@435: // Does the block span the start of the chunk and is it duke@435: // an object? duke@435: if (first_block < chunk_mr.start() && duke@435: sp->block_is_obj(first_block)) { duke@435: jbyte* first_dirty_card = NULL; duke@435: jbyte* last_card_of_first_obj = duke@435: byte_for(first_block + sp->block_size(first_block) - 1); duke@435: jbyte* first_card_of_cur_chunk = byte_for(chunk_mr.start()); duke@435: jbyte* last_card_of_cur_chunk = byte_for(chunk_mr.last()); duke@435: jbyte* last_card_to_check = duke@435: (jbyte*) MIN2((intptr_t) last_card_of_cur_chunk, duke@435: (intptr_t) last_card_of_first_obj); duke@435: for (jbyte* cur = first_card_of_cur_chunk; duke@435: cur <= last_card_to_check; cur++) { duke@435: if (card_will_be_scanned(*cur)) { duke@435: first_dirty_card = cur; break; duke@435: } duke@435: } duke@435: if (first_dirty_card != NULL) { duke@435: assert(0 <= cur_chunk_index && duke@435: cur_chunk_index < lowest_non_clean_chunk_size, duke@435: "Bounds error."); duke@435: lowest_non_clean[cur_chunk_index] = first_dirty_card; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: void duke@435: CardTableModRefBS:: duke@435: get_LNC_array_for_space(Space* sp, duke@435: jbyte**& lowest_non_clean, duke@435: uintptr_t& lowest_non_clean_base_chunk_index, duke@435: size_t& lowest_non_clean_chunk_size) { duke@435: duke@435: int i = find_covering_region_containing(sp->bottom()); duke@435: MemRegion covered = _covered[i]; duke@435: size_t n_chunks = chunks_to_cover(covered); duke@435: duke@435: // Only the first thread to obtain the lock will resize the duke@435: // LNC array for the covered region. Any later expansion can't affect duke@435: // the used_at_save_marks region. duke@435: // (I observed a bug in which the first thread to execute this would duke@435: // resize, and then it would cause "expand_and_allocates" that would duke@435: // Increase the number of chunks in the covered region. Then a second duke@435: // thread would come and execute this, see that the size didn't match, duke@435: // and free and allocate again. So the first thread would be using a duke@435: // freed "_lowest_non_clean" array.) duke@435: duke@435: // Do a dirty read here. If we pass the conditional then take the rare duke@435: // event lock and do the read again in case some other thread had already duke@435: // succeeded and done the resize. duke@435: int cur_collection = Universe::heap()->total_collections(); duke@435: if (_last_LNC_resizing_collection[i] != cur_collection) { duke@435: MutexLocker x(ParGCRareEvent_lock); duke@435: if (_last_LNC_resizing_collection[i] != cur_collection) { duke@435: if (_lowest_non_clean[i] == NULL || duke@435: n_chunks != _lowest_non_clean_chunk_size[i]) { duke@435: duke@435: // Should we delete the old? duke@435: if (_lowest_non_clean[i] != NULL) { duke@435: assert(n_chunks != _lowest_non_clean_chunk_size[i], duke@435: "logical consequence"); duke@435: FREE_C_HEAP_ARRAY(CardPtr, _lowest_non_clean[i]); duke@435: _lowest_non_clean[i] = NULL; duke@435: } duke@435: // Now allocate a new one if necessary. duke@435: if (_lowest_non_clean[i] == NULL) { duke@435: _lowest_non_clean[i] = NEW_C_HEAP_ARRAY(CardPtr, n_chunks); duke@435: _lowest_non_clean_chunk_size[i] = n_chunks; duke@435: _lowest_non_clean_base_chunk_index[i] = addr_to_chunk_index(covered.start()); duke@435: for (int j = 0; j < (int)n_chunks; j++) duke@435: _lowest_non_clean[i][j] = NULL; duke@435: } duke@435: } duke@435: _last_LNC_resizing_collection[i] = cur_collection; duke@435: } duke@435: } duke@435: // In any case, now do the initialization. duke@435: lowest_non_clean = _lowest_non_clean[i]; duke@435: lowest_non_clean_base_chunk_index = _lowest_non_clean_base_chunk_index[i]; duke@435: lowest_non_clean_chunk_size = _lowest_non_clean_chunk_size[i]; duke@435: }