void CardTableModRefBS:: process_stride(Space* sp, MemRegion used, jint stride, int n_strides, DirtyCardToOopClosure* dcto_cl, MemRegionClosure* cl, bool clear, jbyte** lowest_non_clean, uintptr_t lowest_non_clean_base_chunk_index, size_t lowest_non_clean_chunk_size) { // We don't have to go downwards here; it wouldn't help anyway, // because of parallelism. // 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) * CardsPerStrideChunk); } 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) * CardsPerStrideChunk); } while (chunk_card_start < end_card) { // We don't have to go downwards here; it wouldn't help anyway, // because of parallelism. (We take care with "min_done"; see below.) // Invariant: chunk_mr should be fully contained within the "used" region. jbyte* chunk_card_end = chunk_card_start + CardsPerStrideChunk; 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"); // 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); non_clean_card_iterate_work(chunk_mr, cl, clear); // Find the next chunk of the stride. chunk_card_start += CardsPerStrideChunk * n_strides; } }
void CardTableRS::clear_MemRegion(MemRegion mr) { jbyte* cur = byte_for(mr.start()); jbyte* last = byte_after(mr.last()); assert(addr_for(cur) == mr.start(), "region must be card aligned"); while (cur < last) { *cur = CardTableModRefBS::clean_card; cur++; } }
void CardTableRS::verify_space(Space* s, HeapWord* gen_boundary) { // We don't need to do young-gen spaces. if (s->end() <= gen_boundary) return; MemRegion used = s->used_region(); jbyte* cur_entry = byte_for(used.start()); jbyte* limit = byte_after(used.last()); while (cur_entry < limit) { if (*cur_entry == CardTableModRefBS::clean_card) { jbyte* first_dirty = cur_entry+1; while (first_dirty < limit && *first_dirty == CardTableModRefBS::clean_card) first_dirty++; // If the first object is a regular object, and it has a // young-to-old field, that would mark the previous card. HeapWord* boundary = addr_for(cur_entry); HeapWord* end = addr_for(first_dirty); HeapWord* boundary_block = s->block_start(boundary); HeapWord* begin = boundary; // Until proven otherwise. HeapWord* start_block = boundary_block; // Until proven otherwise. if (boundary_block < boundary) { if (s->block_is_obj(boundary_block)) { oop boundary_obj = oop(boundary_block); if (!boundary_obj->is_objArray() && !boundary_obj->is_typeArray()) { guarantee(cur_entry > byte_for(used.start()), "else boundary would be boundary_block"); if (*byte_for(boundary_block) != CardTableModRefBS::clean_card) { begin = boundary_block + s->block_size(boundary_block); start_block = begin; } } } } // Now traverse objects until end. HeapWord* cur = start_block; VerifyCleanCardClosure verify_blk(gen_boundary, begin, end); while (cur < end) { if (s->block_is_obj(cur)) { oop(cur)->oop_iterate(&verify_blk); } cur += s->block_size(cur); } cur_entry = first_dirty; } else { guarantee(*cur_entry != cur_youngergen_and_prev_nonclean_card, "Illegal CT value"); // If we're in the parallel case, the cur and prev values are // different, and we can't have left a prev in the table. guarantee(cur_youngergen_card_val() == youngergen_card || !is_prev_youngergen_card_val(*cur_entry), "Illegal CT value"); cur_entry++; } } }
void G1SATBCardTableModRefBS::g1_mark_as_young(const MemRegion& mr) { jbyte *const first = byte_for(mr.start()); jbyte *const last = byte_after(mr.last()); // Below we may use an explicit loop instead of memset() because on // certain platforms memset() can give concurrent readers phantom zeros. if (UseMemSetInBOT) { memset(first, g1_young_gen, last - first); } else { for (jbyte* i = first; i < last; i++) { *i = g1_young_gen; } } }
void CardTableRS::verify_aligned_region_empty(MemRegion mr) { if (!mr.is_empty()) { jbyte* cur_entry = byte_for(mr.start()); jbyte* limit = byte_after(mr.last()); // The region mr may not start on a card boundary so // the first card may reflect a write to the space // just prior to mr. if (!is_aligned(mr.start())) { cur_entry++; } for (;cur_entry < limit; cur_entry++) { guarantee(*cur_entry == CardTableModRefBS::clean_card, "Unexpected dirty card found"); } } }
void CardTableRS::verify_space(Space* s, HeapWord* gen_boundary) { // We don't need to do young-gen spaces. if (s->end() <= gen_boundary) return; MemRegion used = s->used_region(); jbyte* cur_entry = byte_for(used.start()); jbyte* limit = byte_after(used.last()); while (cur_entry < limit) { if (*cur_entry == CardTableModRefBS::clean_card) { jbyte* first_dirty = cur_entry+1; while (first_dirty < limit && *first_dirty == CardTableModRefBS::clean_card) { first_dirty++; } // If the first object is a regular object, and it has a // young-to-old field, that would mark the previous card. HeapWord* boundary = addr_for(cur_entry); HeapWord* end = (first_dirty >= limit) ? used.end() : addr_for(first_dirty); HeapWord* boundary_block = s->block_start(boundary); HeapWord* begin = boundary; // Until proven otherwise. HeapWord* start_block = boundary_block; // Until proven otherwise. if (boundary_block < boundary) { if (s->block_is_obj(boundary_block) && s->obj_is_alive(boundary_block)) { oop boundary_obj = oop(boundary_block); if (!boundary_obj->is_objArray() && !boundary_obj->is_typeArray()) { guarantee(cur_entry > byte_for(used.start()), "else boundary would be boundary_block"); if (*byte_for(boundary_block) != CardTableModRefBS::clean_card) { begin = boundary_block + s->block_size(boundary_block); start_block = begin; } } } } // Now traverse objects until end. if (begin < end) { MemRegion mr(begin, end); VerifyCleanCardClosure verify_blk(gen_boundary, begin, end); for (HeapWord* cur = start_block; cur < end; cur += s->block_size(cur)) { if (s->block_is_obj(cur) && s->obj_is_alive(cur)) { oop(cur)->oop_iterate(&verify_blk, mr); } } } cur_entry = first_dirty; } else { // We'd normally expect that cur_youngergen_and_prev_nonclean_card // is a transient value, that cannot be in the card table // except during GC, and thus assert that: // guarantee(*cur_entry != cur_youngergen_and_prev_nonclean_card, // "Illegal CT value"); // That however, need not hold, as will become clear in the // following... // We'd normally expect that if we are in the parallel case, // we can't have left a prev value (which would be different // from the current value) in the card table, and so we'd like to // assert that: // guarantee(cur_youngergen_card_val() == youngergen_card // || !is_prev_youngergen_card_val(*cur_entry), // "Illegal CT value"); // That, however, may not hold occasionally, because of // CMS or MSC in the old gen. To wit, consider the // following two simple illustrative scenarios: // (a) CMS: Consider the case where a large object L // spanning several cards is allocated in the old // gen, and has a young gen reference stored in it, dirtying // some interior cards. A young collection scans the card, // finds a young ref and installs a youngergenP_n value. // L then goes dead. Now a CMS collection starts, // finds L dead and sweeps it up. Assume that L is // abutting _unallocated_blk, so _unallocated_blk is // adjusted down to (below) L. Assume further that // no young collection intervenes during this CMS cycle. // The next young gen cycle will not get to look at this // youngergenP_n card since it lies in the unoccupied // part of the space. // Some young collections later the blocks on this // card can be re-allocated either due to direct allocation // or due to absorbing promotions. At this time, the // before-gc verification will fail the above assert. // (b) MSC: In this case, an object L with a young reference // is on a card that (therefore) holds a youngergen_n value. // Suppose also that L lies towards the end of the used // the used space before GC. An MSC collection // occurs that compacts to such an extent that this // card is no longer in the occupied part of the space. // Since current code in MSC does not always clear cards // in the unused part of old gen, this stale youngergen_n // value is left behind and can later be covered by // an object when promotion or direct allocation // re-allocates that part of the heap. // // Fortunately, the presence of such stale card values is // "only" a minor annoyance in that subsequent young collections // might needlessly scan such cards, but would still never corrupt // the heap as a result. However, it's likely not to be a significant // performance inhibitor in practice. For instance, // some recent measurements with unoccupied cards eagerly cleared // out to maintain this invariant, showed next to no // change in young collection times; of course one can construct // degenerate examples where the cost can be significant.) // Note, in particular, that if the "stale" card is modified // after re-allocation, it would be dirty, not "stale". Thus, // we can never have a younger ref in such a card and it is // safe not to scan that card in any collection. [As we see // below, we do some unnecessary scanning // in some cases in the current parallel scanning algorithm.] // // The main point below is that the parallel card scanning code // deals correctly with these stale card values. There are two main // cases to consider where we have a stale "younger gen" value and a // "derivative" case to consider, where we have a stale // "cur_younger_gen_and_prev_non_clean" value, as will become // apparent in the case analysis below. // o Case 1. If the stale value corresponds to a younger_gen_n // value other than the cur_younger_gen value then the code // treats this as being tantamount to a prev_younger_gen // card. This means that the card may be unnecessarily scanned. // There are two sub-cases to consider: // o Case 1a. Let us say that the card is in the occupied part // of the generation at the time the collection begins. In // that case the card will be either cleared when it is scanned // for young pointers, or will be set to cur_younger_gen as a // result of promotion. (We have elided the normal case where // the scanning thread and the promoting thread interleave // possibly resulting in a transient // cur_younger_gen_and_prev_non_clean value before settling // to cur_younger_gen. [End Case 1a.] // o Case 1b. Consider now the case when the card is in the unoccupied // part of the space which becomes occupied because of promotions // into it during the current young GC. In this case the card // will never be scanned for young references. The current // code will set the card value to either // cur_younger_gen_and_prev_non_clean or leave // it with its stale value -- because the promotions didn't // result in any younger refs on that card. Of these two // cases, the latter will be covered in Case 1a during // a subsequent scan. To deal with the former case, we need // to further consider how we deal with a stale value of // cur_younger_gen_and_prev_non_clean in our case analysis // below. This we do in Case 3 below. [End Case 1b] // [End Case 1] // o Case 2. If the stale value corresponds to cur_younger_gen being // a value not necessarily written by a current promotion, the // card will not be scanned by the younger refs scanning code. // (This is OK since as we argued above such cards cannot contain // any younger refs.) The result is that this value will be // treated as a prev_younger_gen value in a subsequent collection, // which is addressed in Case 1 above. [End Case 2] // o Case 3. We here consider the "derivative" case from Case 1b. above // because of which we may find a stale // cur_younger_gen_and_prev_non_clean card value in the table. // Once again, as in Case 1, we consider two subcases, depending // on whether the card lies in the occupied or unoccupied part // of the space at the start of the young collection. // o Case 3a. Let us say the card is in the occupied part of // the old gen at the start of the young collection. In that // case, the card will be scanned by the younger refs scanning // code which will set it to cur_younger_gen. In a subsequent // scan, the card will be considered again and get its final // correct value. [End Case 3a] // o Case 3b. Now consider the case where the card is in the // unoccupied part of the old gen, and is occupied as a result // of promotions during thus young gc. In that case, // the card will not be scanned for younger refs. The presence // of newly promoted objects on the card will then result in // its keeping the value cur_younger_gen_and_prev_non_clean // value, which we have dealt with in Case 3 here. [End Case 3b] // [End Case 3] // // (Please refer to the code in the helper class // ClearNonCleanCardWrapper and in CardTableModRefBS for details.) // // The informal arguments above can be tightened into a formal // correctness proof and it behooves us to write up such a proof, // or to use model checking to prove that there are no lingering // concerns. // // Clearly because of Case 3b one cannot bound the time for // which a card will retain what we have called a "stale" value. // However, one can obtain a Loose upper bound on the redundant // work as a result of such stale values. Note first that any // time a stale card lies in the occupied part of the space at // the start of the collection, it is scanned by younger refs // code and we can define a rank function on card values that // declines when this is so. Note also that when a card does not // lie in the occupied part of the space at the beginning of a // young collection, its rank can either decline or stay unchanged. // In this case, no extra work is done in terms of redundant // younger refs scanning of that card. // Then, the case analysis above reveals that, in the worst case, // any such stale card will be scanned unnecessarily at most twice. // // It is nonethelss advisable to try and get rid of some of this // redundant work in a subsequent (low priority) re-design of // the card-scanning code, if only to simplify the underlying // state machine analysis/proof. ysr 1/28/2002. XXX cur_entry++; } } }
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; } }