Day48 - Epoll + Timerfd + Idle Timeout

Overview

In this lab, we extend the epoll-based TCP server by integrating timerfd to implement idle timeout handling for multiple clients.

The server now supports:

• Multi-client connection (epoll)
• Non-blocking sockets
• TX queue with partial send handling
• Periodic timer event using timerfd
• Idle timeout detection and automatic client cleanup

---

timerfd Basics

Key APIs

• timerfd_create()
• timerfd_settime()
• read()

Important Concept

read(timerfd) does NOT return time.

It returns:

number of expirations since last read

---

Example

timer interval = 1 sec

If read every second:
    expirations = 1

If read after 5 seconds:
    expirations = 5

---

Why timerfd with epoll?

Compared to other approaches:

Method	Problem
sleep()	blocks thread
signal	complex, race conditions
thread timer	synchronization overhead

timerfd advantages:

• file descriptor based
• works naturally with epoll
• no signal handling
• no extra thread

---

Idle Timeout Design

Architecture

timerfd (1 sec tick)
    ↓
epoll event loop
    ↓
scan all clients
    ↓
check idle time
    ↓
destroy timed-out clients

---

Key Design Choice

We use:

Single global timerfd
+ per-client timestamp

Instead of:

Per-client timerfd

---

Trade-offs

Approach	Pros	Cons
Global timer + scan	simple, scalable	O(N) scan
Per-client timerfd	precise timeout	complex, many fd

---

Client Data Structure

struct client_ctx {
    int fd;
    uint64_t last_active_ms;

    struct tx_queue txq;

    struct client_ctx *next;
};

---

Server Data Structure

struct server_ctx {
    int listen_fd;
    int epoll_fd;
    int timer_fd;

    uint64_t timer_ms;

    struct client_ctx *client_list;
};

---

Lifecycle Management

Accept

accept()
    → create client_ctx
    → add epoll
    → set last_active_ms
    → add to client_list

---

Activity Update

EPOLLIN (recv > 0)
    → update last_active_ms

---

Timeout Detection

if ((now - client->last_active_ms) >= TIMEOUT_MS)
    client_destroy(...)

---

Destroy Client

client_destroy():
    → remove from list
    → epoll_ctl(DEL)
    → close(fd)
    → clear TX queue
    → free(client)

---

epoll vs Client List

Important concept:

epoll = event notification
client_list = state tracking

epoll only tells:

"which fd has event"

It does NOT track all active clients.

---

Time Source

We use:

clock_gettime(CLOCK_MONOTONIC)

Reasons:

• not affected by system time changes
• safe for timeout calculation

---

Timeout Precision

Timeout is not exact.

Example:

timeout = 10 sec
timer interval = 1 sec

Actual timeout ≈ 10 ~ 11 sec

---

Summary

This lab introduces:

• fd-based timer (timerfd)
• integrating timer into epoll loop
• scalable idle timeout design
• clean client lifecycle management

This is a fundamental building block for real-world servers.