Writing a malloc replacement with mmap

Most allocator tutorials stop at sbrk. I wanted to see what happens if you skip straight to mmap and build allocation, deallocation, defragmentation, and coalescing on top of it yourself.

Why mmap instead of sbrk

sbrk grows a single contiguous heap. mmap lets you request independent regions, which makes it easier to reason about large allocations and to return memory to the OS instead of just marking it free.

void *region = mmap(NULL, size, PROT_READ | PROT_WRITE,
                     MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);

The coalescing bug

The first version of block coalescing merged adjacent free blocks by walking a doubly linked list — but I wasn’t validating that “adjacent in the list” also meant “adjacent in memory.” Two blocks from different mmap regions that happened to sit next to each other in the free list got merged, corrupting the heap on the next allocation.

The fix was boring: store each block’s region boundaries and check physical adjacency before merging, not just list order.

What’s next

Custom heap sizing per allocation pattern is in, next up is instrumenting fragmentation over time so I can actually measure whether coalescing is paying for itself.