// Card marking void inline_write_ref_field_gc(void* field, oop new_val) { jbyte* byte = byte_for(field); *byte = youngergen_card; }
// // Note! ending cards are exclusive! HeapWord* slice_start = addr_for(worker_start_card); HeapWord* slice_end = MIN2((HeapWord*) sp_top, addr_for(worker_end_card)); // If there are not objects starting within the chunk, skip it. if (!start_array->object_starts_in_range(slice_start, slice_end)) { continue; } // Update our beginning addr HeapWord* first_object = start_array->object_start(slice_start); debug_only(oop* first_object_within_slice = (oop*) first_object;) if (first_object < slice_start) { last_scanned = (oop*)(first_object + oop(first_object)->size()); debug_only(first_object_within_slice = last_scanned;) worker_start_card = byte_for(last_scanned); } // Update the ending addr if (slice_end < (HeapWord*)sp_top) { // The subtraction is important! An object may start precisely at slice_end. HeapWord* last_object = start_array->object_start(slice_end - 1); slice_end = last_object + oop(last_object)->size(); // worker_end_card is exclusive, so bump it one past the end of last_object's // covered span. worker_end_card = byte_for(slice_end) + 1; if (worker_end_card > end_card) worker_end_card = end_card; }
void set_card_newgen(void* addr) { jbyte* p = byte_for(addr); *p = verify_card; }
// We get passed the space_top value to prevent us from traversing into // the old_gen promotion labs, which cannot be safely parsed. void CardTableExtension::scavenge_contents(ObjectStartArray* start_array, MutableSpace* sp, HeapWord* space_top, PSPromotionManager* pm) { assert(start_array != NULL && sp != NULL && pm != NULL, "Sanity"); assert(start_array->covered_region().contains(sp->used_region()), "ObjectStartArray does not cover space"); if (sp->not_empty()) { oop* sp_top = (oop*)space_top; oop* prev_top = NULL; jbyte* current_card = byte_for(sp->bottom()); jbyte* end_card = byte_for(sp_top - 1); // sp_top is exclusive // scan card marking array while (current_card <= end_card) { jbyte value = *current_card; // skip clean cards if (card_is_clean(value)) { current_card++; } else { // we found a non-clean card jbyte* first_nonclean_card = current_card++; oop* bottom = (oop*)addr_for(first_nonclean_card); // find object starting on card oop* bottom_obj = (oop*)start_array->object_start((HeapWord*)bottom); // bottom_obj = (oop*)start_array->object_start((HeapWord*)bottom); assert(bottom_obj <= bottom, "just checking"); // make sure we don't scan oops we already looked at if (bottom < prev_top) bottom = prev_top; // figure out when to stop scanning jbyte* first_clean_card; oop* top; bool restart_scanning; do { restart_scanning = false; // find a clean card while (current_card <= end_card) { value = *current_card; if (card_is_clean(value)) break; current_card++; } // check if we reached the end, if so we are done if (current_card >= end_card) { first_clean_card = end_card + 1; current_card++; top = sp_top; } else { // we have a clean card, find object starting on that card first_clean_card = current_card++; top = (oop*)addr_for(first_clean_card); oop* top_obj = (oop*)start_array->object_start((HeapWord*)top); // top_obj = (oop*)start_array->object_start((HeapWord*)top); assert(top_obj <= top, "just checking"); if (oop(top_obj)->is_objArray() || oop(top_obj)->is_typeArray()) { // an arrayOop is starting on the clean card - since we do exact store // checks for objArrays we are done } else { // otherwise, it is possible that the object starting on the clean card // spans the entire card, and that the store happened on a later card. // figure out where the object ends top = top_obj + oop(top_obj)->size(); jbyte* top_card = CardTableModRefBS::byte_for(top - 1); // top is exclusive if (top_card > first_clean_card) { // object ends a different card current_card = top_card + 1; if (card_is_clean(*top_card)) { // the ending card is clean, we are done first_clean_card = top_card; } else { // the ending card is not clean, continue scanning at start of do-while restart_scanning = true; } } else { // object ends on the clean card, we are done. assert(first_clean_card == top_card, "just checking"); } } } } while (restart_scanning); // we know which cards to scan, now clear them while (first_nonclean_card < first_clean_card) { *first_nonclean_card++ = clean_card; } // scan oops in objects do { oop(bottom_obj)->push_contents(pm); bottom_obj += oop(bottom_obj)->size(); assert(bottom_obj <= sp_top, "just checking"); } while (bottom_obj < top); pm->drain_stacks_cond_depth(); // remember top oop* scanned prev_top = top; } } } }
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; } }
void CardTableModRefBS:: 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 the chunk boundaries. // First, set our max_to_do: HeapWord* max_to_do = NULL; uintptr_t cur_chunk_index = addr_to_chunk_index(chunk_mr.start()); cur_chunk_index = cur_chunk_index - lowest_non_clean_base_chunk_index; if (chunk_mr.end() < used.end()) { // This is not the last chunk in the used region. What is the last // object? HeapWord* 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() || !sp->block_is_obj(last_block)) { max_to_do = chunk_mr.end(); } else { // It is an object and starts before the end of the current 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* last_obj_card = byte_for(last_block); if (!card_may_have_been_dirty(*last_obj_card)) { // The card containing the head is not dirty. Any marks in // subsequent cards still in this chunk must have been made // precisely; we can cap processing at the end. 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; size_t last_block_size = sp->block_size(last_block); jbyte* last_card_of_last_obj = byte_for(last_block + last_block_size - 1); jbyte* 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. For example, // an object that is an array of primitives will not // have any cards covering regions interior to the array // that will need to be scanned. The scan can be terminated // at the last card of the next chunk. That would leave // limit_card as NULL and would result in "max_to_do" // being set with the LNC value or with the end // of the last block. jbyte* last_card_of_next_chunk = first_card_of_next_chunk + CardsPerStrideChunk; assert(byte_for(chunk_mr.end()) - byte_for(chunk_mr.start()) == CardsPerStrideChunk, "last card of next chunk may be wrong"); jbyte* last_card_to_check = (jbyte*) MIN2(last_card_of_last_obj, last_card_of_next_chunk); for (jbyte* cur = first_card_of_next_chunk; cur <= last_card_to_check; cur++) { if (card_will_be_scanned(*cur)) { limit_card = cur; break; } } assert(0 <= cur_chunk_index+1 && cur_chunk_index+1 < lowest_non_clean_chunk_size, "Bounds error."); // LNC for the next chunk jbyte* lnc_card = lowest_non_clean[cur_chunk_index+1]; if (limit_card == NULL) { limit_card = lnc_card; } if (limit_card != NULL) { if (lnc_card != NULL) { limit_card = (jbyte*)MIN2((intptr_t)limit_card, (intptr_t)lnc_card); } max_to_do = addr_for(limit_card); } else { max_to_do = last_block + last_block_size; } } } assert(max_to_do != NULL, "OOPS!"); } else { max_to_do = used.end(); } // 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 // Now we set *our" lowest_non_clean entry. // Find the object that spans our boundary, if one exists. // Nothing to do on the first chunk. if (chunk_mr.start() > used.start()) { // first_block is the block possibly spanning the chunk start HeapWord* first_block = sp->block_start(chunk_mr.start()); // Does the block span the start of the chunk and is it // an object? if (first_block < chunk_mr.start() && sp->block_is_obj(first_block)) { 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); for (jbyte* cur = first_card_of_cur_chunk; cur <= last_card_to_check; cur++) { if (card_will_be_scanned(*cur)) { first_dirty_card = cur; break; } } if (first_dirty_card != NULL) { assert(0 <= cur_chunk_index && cur_chunk_index < lowest_non_clean_chunk_size, "Bounds error."); lowest_non_clean[cur_chunk_index] = first_dirty_card; } } } }