Turns out Windows is more conservative both in initial allocation when creating new heap and also in growing the heap.
Comments:
- Without patch 1 the related test in test_block_layout fails on x64 for me ("unexpected padding") after the last patch. I don't think the last patches really changes much in this regard, looks like the test just depends on the uninitialized data; - Without patch 2 test_HeapCreate fails (line 601 where it checks the last block pointer). I think it is also only triggered by my patch, I think that last block detection only works correctly now when commit size is smaller than the subheap size (which happens to be not the case in the test after the last patch); - The last patch extends a bit a todo in test_block_layout(). The actual reason for (extended) todo is that we are probably not handling pending free blocks quite like on Windows. I watched the pointers on Windows in the same test and they are always withing the same subheap allocation which suggests that Windows might be reusing pending free block once it can't find normal free block and before extending the heap. So the last patch changes subheap sizes and that triggers different subheaps for the allocated pointers in Wine. So it looks like not an actual issue with my patch. Reusing pending free pointers is apparently possible to implement but it seems to me that it doesn't worth the trouble until anything depends on that (since those pending free blocks are only there when heap debug flags are enabled which is rarely the case outside of debugging).
-- v2: ntdll: Better match Windows subheap sizes. kernel32/tests: Add tests for subheap sizes. ntdll: Fix last block detection in heap_walk_blocks(). ntdll: Fix tail padding in mark_block_tail().