卡表（CardTable）在CMS中是最常见的概念之一，G1中不仅保留了这个概念，还引入了RSet。卡表到底是一个什么东西？GC最早引入卡表的目的是为了对内存的引用关系做标记，从而根据引用关系快速遍历活跃对象。举个简单的例子，有两个分区，假设分区大小都为1MB，分别为A和B。如果A中有一个对象objA，B中有一个对象objB，且objA.field=objB，那么这两个分区就有引用关系了，但是如果我们想找到分区A，要如何引用分区B？做法有两种：·遍历整个分区A，一个字一个字的移动（为什么以字为单位？原因是JVM中对象会对齐，所以不需要按字节移动），然后查看内存里面的值到底是不是指向B，这种方法效率太低，可以优化为一个对象一个对象地移动（这里涉及JVM如何识别对象，以及如何区分指针和立即数），但效率还是太低。
·借助额外的数据结构描述这种引用关系，例如使用类似位图（bitmap）的方法，记录A和B的内存块之间的引用关系，用一个位来描述一个字，假设在32位机器上（一个字为32位），需要32KB（32KB×32=1M）的空间来描述一个分区。那么我们就可以在这个对象ObjA所在分区A里面添加一个额外的指针，这个指针指向另外一个分区B的位图，如果我们可以把对象ObjA和指针关系进行映射，那么当访问ObjA的时候，顺便访问这个额外的指针，从这个指针指向的位图就能找到被ObjA引用的分区B对应的内存块。通常我们只需要判定位图里面对应的位是否有1，有的话则认为发生了引用。
class CardTable: public CHeapObj<mtGC> {friend class VMStructs;
public:typedef uint8_t CardValue;// All code generators assume that the size of a card table entry is one byte.// They need to be updated to reflect any change to this.// This code can typically be found by searching for the byte_map_base() method.STATIC_ASSERT(sizeof(CardValue) == 1);protected:// The declaration order of these const fields is important; see the// constructor before changing.const MemRegion _whole_heap; // the region covered by the card tableconst size_t _page_size; // page size used when mapping _byte_mapsize_t _byte_map_size; // in bytesCardValue* _byte_map; // the card marking arrayCardValue* _byte_map_base;// Some barrier sets create tables whose elements correspond to parts of// the heap; the CardTableBarrierSet is an example. Such barrier sets will// normally reserve space for such tables, and commit parts of the table// "covering" parts of the heap that are committed. At most one covered// region per generation is needed.static constexpr int max_covered_regions = 2;// The covered regions should be in address order.MemRegion _covered[max_covered_regions];// The last card is a guard card; never committed.MemRegion _guard_region;inline size_t compute_byte_map_size(size_t num_bytes);enum CardValues {clean_card = (CardValue)-1,dirty_card = 0,CT_MR_BS_last_reserved = 1};// a word's worth (row) of clean card valuesstatic const intptr_t clean_card_row = (intptr_t)(-1);// CardTable entry sizestatic uint _card_shift;static uint _card_size;static uint _card_size_in_words;size_t last_valid_index() const {return cards_required(_whole_heap.word_size()) - 1;}private:void initialize_covered_region(void* region0_start, void* region1_start);MemRegion committed_for(const MemRegion mr) const;
public:CardTable(MemRegion whole_heap);virtual ~CardTable() = default;void initialize(void* region0_start, void* region1_start);// *** Barrier set functions.// Initialization utilities; covered_words is the size of the covered region// in, um, words.inline size_t cards_required(size_t covered_words) const {assert(is_aligned(covered_words, _card_size_in_words), "precondition");return covered_words / _card_size_in_words;}// Dirty the bytes corresponding to "mr" (not all of which must be// covered.)void dirty_MemRegion(MemRegion mr);// Clear (to clean_card) the bytes entirely contained within "mr" (not// all of which must be covered.)void clear_MemRegion(MemRegion mr);// Return true if "p" is at the start of a card.bool is_card_aligned(HeapWord* p) {CardValue* pcard = byte_for(p);return (addr_for(pcard) == p);}// Mapping from address to card marking array entryCardValue* byte_for(const void* p) const {assert(_whole_heap.contains(p),"Attempt to access p = " PTR_FORMAT " out of bounds of "" card marking array's _whole_heap = [" PTR_FORMAT "," PTR_FORMAT ")",p2i(p), p2i(_whole_heap.start()), p2i(_whole_heap.end()));CardValue* result = &_byte_map_base[uintptr_t(p) >> _card_shift];assert(result >= _byte_map && result < _byte_map + _byte_map_size,"out of bounds accessor for card marking array");return result;}// The card table byte one after the card marking array// entry for argument address. Typically used for higher bounds// for loops iterating through the card table.CardValue* byte_after(const void* p) const {return byte_for(p) + 1;}void invalidate(MemRegion mr);// Provide read-only access to the card table array.const CardValue* byte_for_const(const void* p) const {return byte_for(p);}const CardValue* byte_after_const(const void* p) const {return byte_after(p);}// Mapping from card marking array entry to address of first wordHeapWord* addr_for(const CardValue* p) const {assert(p >= _byte_map && p < _byte_map + _byte_map_size,"out of bounds access to card marking array. p: " PTR_FORMAT" _byte_map: " PTR_FORMAT " _byte_map + _byte_map_size: " PTR_FORMAT,p2i(p), p2i(_byte_map), p2i(_byte_map + _byte_map_size));// As _byte_map_base may be "negative" (the card table has been allocated before// the heap in memory), do not use pointer_delta() to avoid the assertion failure.size_t delta = p - _byte_map_base;HeapWord* result = (HeapWord*) (delta << _card_shift);assert(_whole_heap.contains(result),"Returning result = " PTR_FORMAT " out of bounds of "" card marking array's _whole_heap = [" PTR_FORMAT "," PTR_FORMAT ")",p2i(result), p2i(_whole_heap.start()), p2i(_whole_heap.end()));return result;}// Mapping from address to card marking array index.size_t index_for(void* p) {assert(_whole_heap.contains(p),"Attempt to access p = " PTR_FORMAT " out of bounds of "" card marking array's _whole_heap = [" PTR_FORMAT "," PTR_FORMAT ")",p2i(p), p2i(_whole_heap.start()), p2i(_whole_heap.end()));return byte_for(p) - _byte_map;}CardValue* byte_for_index(const size_t card_index) const {return _byte_map + card_index;}// Resize one of the regions covered by the remembered set.void resize_covered_region(MemRegion new_region);// *** Card-table-RemSet-specific things.static uintx ct_max_alignment_constraint();static uint card_shift() {return _card_shift;}static uint card_size() {return _card_size;}static uint card_size_in_words() {return _card_size_in_words;}static constexpr CardValue clean_card_val() { return clean_card; }static constexpr CardValue dirty_card_val() { return dirty_card; }static intptr_t clean_card_row_val() { return clean_card_row; }// Initialize card sizestatic void initialize_card_size();// Card marking array base (adjusted for heap low boundary)// This would be the 0th element of _byte_map, if the heap started at 0x0.// But since the heap starts at some higher address, this points to somewhere// before the beginning of the actual _byte_map.CardValue* byte_map_base() const { return _byte_map_base; }virtual bool is_in_young(const void* p) const = 0;
};
class G1CardTable : public CardTable {friend class VMStructs;friend class G1CardTableChangedListener;G1CardTableChangedListener _listener;public:enum G1CardValues {g1_young_gen = CT_MR_BS_last_reserved << 1,// During evacuation we use the card table to consolidate the cards we need to// scan for roots onto the card table from the various sources. Further it is// used to record already completely scanned cards to avoid re-scanning them// when incrementally evacuating the old gen regions of a collection set.// This means that already scanned cards should be preserved.//// The merge at the start of each evacuation round simply sets cards to dirty// that are clean; scanned cards are set to 0x1.//// This means that the LSB determines what to do with the card during evacuation// given the following possible values://// 11111111 - clean, do not scan// 00000001 - already scanned, do not scan// 00000000 - dirty, needs to be scanned.//g1_card_already_scanned = 0x1};static const size_t WordAllClean = SIZE_MAX;static const size_t WordAllDirty = 0;STATIC_ASSERT(BitsPerByte == 8);static const size_t WordAlreadyScanned = (SIZE_MAX / 255) * g1_card_already_scanned;G1CardTable(MemRegion whole_heap): CardTable(whole_heap), _listener() {_listener.set_card_table(this);}static CardValue g1_young_card_val() { return g1_young_gen; }static CardValue g1_scanned_card_val() { return g1_card_already_scanned; }void verify_g1_young_region(MemRegion mr) PRODUCT_RETURN;void g1_mark_as_young(const MemRegion& mr);size_t index_for_cardvalue(CardValue const* p) const {return pointer_delta(p, _byte_map, sizeof(CardValue));}// Mark the given card as Dirty if it is Clean. Returns whether the card was// Clean before this operation. This result may be inaccurate as it does not// perform the dirtying atomically.inline bool mark_clean_as_dirty(CardValue* card);// Change Clean cards in a (large) area on the card table as Dirty, preserving// already scanned cards. Assumes that most cards in that area are Clean.inline void mark_range_dirty(size_t start_card_index, size_t num_cards);// Change the given range of dirty cards to "which". All of these cards must be Dirty.inline void change_dirty_cards_to(CardValue* start_card, CardValue* end_card, CardValue which);inline uint region_idx_for(CardValue* p);static size_t compute_size(size_t mem_region_size_in_words) {size_t number_of_slots = (mem_region_size_in_words / _card_size_in_words);return ReservedSpace::allocation_align_size_up(number_of_slots);}// Returns how many bytes of the heap a single byte of the Card Table corresponds to.static size_t heap_map_factor() { return _card_size; }void initialize(G1RegionToSpaceMapper* mapper);bool is_in_young(const void* p) const override;
};
以位为粒度的位图能准确描述每一个字的引用关系，但是一个位通常包含的信息太少，只能描述2个状态：引用还是未引用。实际应用中JVM在垃圾回收的时候需要更多的状态，如果增加至一个字节来描述状态，则位图需要256KB的空间，这个数字太大，开销占了25%。所以一个可能的做法位图不再描述一个字，而是一个区域，JVM选择512字节为单位，即用一个字节描述512字节的引用关系。选择一个区域除了空间利用率的问题之外，实际上还有现实的意义。我们知道Java对象实际上不是一个字能描述的（有一个参数可以控制对象最小对齐的大小，默认是8字节，实际上Java在JVM中还有一些附加信息，所以对齐后最小的Java对象是16字节），很多Java对象可能是几十个字节或者几百个字节，所以用一个字节描述一个区域是有意义的。但是我没有找到512的来源，为什么512效果最好？没有相应的数据来支持这个数字，而且这个值不可以配置，不能修改，但是有理由相信512字节的区域是为了节约内存额外开销。按照这个值，1MB的内存只需要2KB的额外空间就能描述引用关系。这又带来另一个问题，就是512字节里面的内存可能被引用多次，所以这是一个粗略的关系描述，那么在使用的时候需要遍历这512字节。
再举一个例子，假设有两个对象B、C都在这512字节的区域内。为了方便处理，记录对象引用关系的时候，都使用对象的起始位置，然后用这个地址和512对齐，因此B和C对象的卡表指针都指向这一个卡表的位置。那么对于引用处理也有可有两种处理方法：·处理的时候会以堆分区为处理单位，遍历整个堆分区，在遍历的时候，每次都会以对象大小为步长，结合卡表，如果该卡表中对应的位置被设置，则说明对象和其他分区的对象发生了引用。具体内容在后文中介绍Refine的时候还会详细介绍。·处理的时候借助于额外的数据结构，找到真正对象的位置，而不需要从头开始遍历。在后文的并发标记处理时就使用了这种方法，用于找到第一个对象的起始位置。在G1除了512字节粒度的卡表之外，还有bitMap，例如使用bitMap可以描述一个分区对另外一个分区的引用情况。在JVM中bitMap使用非常多，例如还可以描述内存的分配情况。
在G1除了512字节粒度的卡表之外，还有bitMap，例如使用bitMap可以描述一个分区对另外一个分区的引用情况。在JVM中bitMap使用非常多，例如还可以描述内存的分配情况。G1在混合收集算法中用到了并发标记。在并发标记的时候使用了bitMap来描述对象的分配情况。例如1MB的分区可以用16KB（16KB×ObjectAlignmentInBytes×8=1MB）来描述，即16KB额外的空间。其中ObjectAlignmentInBytes是8字节，指的是对象对齐，第二个8是指一个字节有8位。即每一个位可以描述64位。例如一个对象长度对齐之后为24字节，理论上它占用3个位来描述这个24字节已被使用了，实际上并不需要，在标记的时候只需要标记这3个位中的第一个位，再结合堆分区对象的大小信息就能准确找出。其最主要的目的是为了效率，标记一个位和标记3个位相比能节约不少时间，如果对象很大，则更划算。这些都是源码的实现细节，大家在阅读源码时需要细细斟酌。