I don't know whether you found some exceptionally bad description or whether you expect too many details, I've been quite satisfied with the explanations I've seen. If the descriptions are brief and sound simplistic, that's because it really is a rather simple mechanism.
As you apparently already know, a generational garbage collector needs to be able to enumerate old objects that refer to young objects. It would be correct to scan all old objects, but that destroys the advantages of the generational approach, so you have to narrow it down. Regardless of how you do that, you need a write barrier - a piece of code executed whenever a member variable (of a reference type) is assigned/written to. If the new reference points to a young object and it's stored in an old object, the write barrier records that fact for the garbage collect. The difference lies in how it's recorded. There are exact schemes using so-called remembered sets, a collection of every old object that has (had at some point) a reference to a young object. As you can imagine, this takes quite a bit of space.
The card table is a trade-off: Instead of telling you which objects exactly contains young pointers (or at least did at some point), it groups objects into fixed-sized buckets and tracks which buckets contain objects with young pointers. This, of course, reduces space usage. For correctness, it doesn't really matter how you bucket the objects, as long as you're consistent about it. For efficiency, you just group them by their memory address (because you have that available for free), divided by some larger power of two (to make the division a cheap bitwise operation).
Also, instead of maintaining an explicit list of buckets, you reserve some space for each possible bucket up-front. Specifically, there is an array of N bits or bytes, where N is the number of buckets, so that the i
th value is 0 if the i
th bucket contains no young pointers, or 1 if it does contain young pointers. This is the card table proper. Typically this space is allocated and freed along with a large block of memory used as (part of) the heap. It may even be embedded in the start of the memory block, if it doesn't need to grow. Unless the entire address space is used as heap (which is very rare), the above formula gives numbers starting from start_of_memory_region >> K
instead of 0, so to get an index into the card table you have to subtract the start of the start address of the heap.
In summary, when the write barrier finds that the statement some_obj.field = other_obj;
stores a young pointer in an old object, it does this:
card_table[(&old_obj - start_of_heap) >> K] = 1;
Where &old_obj
is the address of the object that now has a young pointer (which is already in a register because it was just determined to refer to an old object).
During minor GC, the garbage collector looks at the card table to determine which heap regions to scan for young pointers:
for i from 0 to (heap_size >> K):
if card_table[i]:
scan heap[i << K .. (i + 1) << K] for young pointers