OS Internals Interview Guide

Virtual Memory

• Each process sees contiguous virtual address space (typically 48-bit on x86_64)
• Page size: 4 KB (standard), 2 MB / 1 GB (huge pages)
• Virtual → Physical translation via page tables
• Demand paging: pages loaded on first access (page fault)

→ src/systems/os/virtual_memory.cpp

Page Tables

• Multi-level (x86_64: 4-level, PML4 → PDP → PD → PT)
• Each entry: physical frame number + flags (present, writable, user, dirty, accessed)
• Page walk: ~4 memory accesses without TLB

TLB (Translation Lookaside Buffer)

• Cache for recent virtual→physical translations
• L1 TLB: ~64 entries per core, ~1 cycle
• L2 TLB: ~1024-2048 entries, ~7 cycles
• TLB miss: full page walk (~20-100 cycles)
• Huge pages reduce TLB pressure (1 entry covers 2MB instead of 4KB)

Context Switching

Step	What Happens
1	Save CPU registers (general, FP, SIMD) to kernel stack
2	Save TLB state / flush (if different address space)
3	Switch page tables (CR3 on x86)
4	Restore target process registers
5	Return to user-space

Cost: ~1-10μs (depends on TLB flush, cache cold start)

Scheduling Algorithms

Algorithm	Type	Used In
CFS (Completely Fair Scheduler)	Weighted fair share, red-black tree	Linux default
RT (SCHED_FIFO / SCHED_RR)	Priority-based, preemptive	Real-time tasks
SCHED_DEADLINE	Earliest deadline first	Deadline-critical workloads
O(1) scheduler	Priority arrays, O(1) pick	Older Linux (pre-2.6.23)

→ src/systems/os/scheduler_sim.cpp

System Calls

• User-space → kernel transition via syscall instruction (x86_64)
• Cost: ~100-200ns (register save + mode switch + validation)
• vDSO: kernel-mapped into user-space for fast calls (clock_gettime, gettimeofday)
• Minimize syscalls on hot path (batch I/O, io_uring)

Interrupts

Type	Source	Example
Hardware (IRQ)	Devices	NIC packet arrival, timer tick
Software (trap)	CPU instruction	syscall, int 0x80
Exception	CPU error	Page fault, divide by zero, GP fault

Interrupt flow: save context → IDT lookup → handler → iret

DMA (Direct Memory Access)

• Device reads/writes main memory without CPU involvement
• CPU sets up DMA descriptor (source, dest, length) → device does transfer
• Interrupt on completion
• Essential for: disk I/O, network cards, GPU data transfer

Common OS Interview Questions

Question	Key Points
What happens on a page fault?	Check valid mapping → allocate frame → load from disk → update PT → restart instruction
Process vs thread?	Process = separate address space. Thread = shared address space, own stack/registers
What's a zombie process?	Terminated but parent hasn't called wait(). Entry remains in process table
Explain thrashing	Working set > physical memory → constant page faults → system crawls
Priority inversion	Low-priority thread holds lock needed by high-priority thread. Fix: priority inheritance
How does malloc work?	brk/mmap from kernel; user-space allocator manages free lists/arenas