Paging & Page Tables

Paging basics

Physical memory is divided into fixed frames; virtual memory into pages (typically 4 KiB, sometimes 2 MiB/1 GiB huge pages). A page table maps Virtual Page Number (VPN) → Physical Frame Number (PFN) plus permission bits (read, write, execute, user/supervisor).

Multi-level page tables (x86-64)

A flat table for 2⁴⁸ bytes would be impossibly large. Multi-level tables allocate inner levels only for used regions — sparse address spaces stay cheap. A page walk traverses 4 levels (PML4 → PDPT → PD → PT) on each TLB miss.

Page fault — not always a crash

Fault type	Kernel action
Page not present, on swap	Read from swap, resume
Page not present, never touched	Allocate frame (demand zero)
Copy-on-write after fork	Duplicate page on write
Invalid access	SIGSEGV to process

Demand paging loads code and data only when accessed. Copy-on-write makes fork cheap — parent and child share read-only pages until one writes.

Huge pages

2 MiB / 1 GiB pages reduce TLB pressure for large, contiguous workloads (databases, JVM heaps). Trade-off: internal fragmentation and harder compaction. hugetlbfs and transparent huge pages (THP) are the two Linux paths — THP can cause latency spikes when collapsing.

Senior-level signal

Memory-mapped databases and JVMs tune huge pages because TLB misses dominate at multi-GB working sets. Before enabling THP globally, measure — the compaction daemon has caused production tail-latency incidents.

Where this goes next

Swapping, Overcommit & OOM covers what happens when there aren't enough physical frames — and why free memory lying around doesn't mean you're safe from the OOM killer.