Memory

VirelaiOS builds its allocator from the memory map captured before UEFI Boot Services end. The kernel keeps an identity-mapped view for EL1; each process has a separate user address space with EL0 permissions granted only to its own mappings. The relevant implementations are alloc.zig, mmu.zig, process.zig, and exceptions.zig.

Address spaces and paging

The MMU uses 4 KiB pages and an identity map covering the low 4 GiB. Kernel mappings are EL1-only; each process root adds that process's user mappings.
Anonymous mmap reserves user virtual ranges. On the supported EL0 data faults, the kernel can allocate a zero-filled page on demand. Writes to copy-on-write mappings either make an unshared page writable or copy a shared page before resuming the process (M29).
The process registry bounds anonymous mappings to 16 regions and records up to 4,096 dynamically allocated pages per process. These are implementation limits, not promises that every process can use that much memory.
There is no swap or memory-balloon reclaim path. A failed page allocation does not imply that the kernel can evict a process's pages.

Fixed structures and bounds

Structure	Bound
Physical allocator bitmap	128 KiB BSS; tracks a 4 GiB span at 4 KiB per page
Page-table storage	512 pages (2 MiB) of fixed BSS
Scheduler task pool	16 task slots total
Anonymous mappings	16 regions per process
Dynamic page records	4,096 per process

The allocator pools whole pages from eligible captured-map regions and protects the kernel image, handoff data, stack, and captured-map buffer with exclusions. Reserved ranges and gaps remain unavailable. The live machine's usable-page total is derived from the EFI map; it is not a constant inferred from the VM's configured RAM size.

Syscall table

The runtime-built syscall table has 128 slots, of which 88 are registered (0–87). Anonymous mmap and munmap are slots 63 and 64; wall-clock setting and filesystem metadata are slots 78 and 79. See Userspace & syscalls for the complete interface.