Memory management

PMM - the physical page store

kernel/mem/mem_pmm.{c,h}: a bitmap over 4 KiB page frames built from the multiboot2 memory map. Reserved regions (kernel image, framebuffer, the mbi itself) are marked used at init. O(1) alloc and free. memtest / pmmrace exercise it in-guest.

VMM - four-level paging

kernel/mem/mem_vmm.{c,h}: the kernel keeps boot.S's identity map of the low 4 GiB; VMM creates and edits additional address spaces for user processes. Protection flags map to read/write/nx/user bits. vmtest, vmmap, vmkernelpt, cr3test inspect it live.

Heap - kmalloc/kfree

kernel/mem/mem_heap.{c,h}: free-list, first-fit, coalescing; blocks are 16-byte aligned (the header is exactly 32 bytes - a historical alternating-parity bug made payload alignment depend on allocation parity, which then shifted ATA-DMA PRDT windows and broke FAT32 mounts; the header size is now static-asserted). 16 MiB fail-fast ceiling; heaptest + heapbounds are the tests, heap/frag the inspectors.

Growth by fault

Stack-growth and heap-growth #PF are handled in the exception path: a fault below a task's stack pointer within its stack region maps a new page; heap growth maps backing pages; anything else escalates to the crash path. exc proves the trap machinery each boot.

DMA constraints

Drivers that program physical addresses into device structures validate alignment and, for 32-bit DMA masks, that the address is below 4 GiB (xHCI refuses above-4GiB DMA at probe AND per-transfer; ATA-DMA masks the PRDT). This is why the heap's alignment guarantees are load-bearing.