# Chapter 13 – I/O Systems
:::danger
:::spoiler
[TOC]
:::

# Overview
- 計算機的兩個主要工作
- I/O and Computation
- ==I/O devices==
- tape，HD，mouse，joystick搖桿，network card， screen，flash disks，etc
- ==I/O subsystem==
- 控制所有I/O devices的方法
- Two conflicting trends
- 統一（unified）及標準的HW/SW interfaces
- I/O devices有很多種類
- Device drivers
- a uniform device-access ==interface to the I/O subsystem==
- 在OS最下層要有個driver layer
- 相似於apps與OS之間的==system calls==
- Device 分類 -- 不同類型去設計不同的interface
- Storage devices：disks, tapes
- Transmission devices：network cards, modems
- Human-interface devices：keyboard, screen, mouse
- Specialized devices：joystick, touchpad
# I/O Hardware
## Port
- A ==connection point== between I/O devices and the host
- E.g.：USB ports
## Bus
- 一組==wire==和定義好的==protocol==規定messages如何透過這些wire做傳送
- E.g.：PCI bus
## Controller
- A collection of electronics可以操控port，bus，device
- controller可以有他自己的==processor和memory==, etc. (E.g.: SCSI controller)
## Typical PC Bus Structure
- 北橋晶片控制比較快的controller，bw、freq高
- 南橋晶片接傳統ide，比較慢的


# I/O Methods

## Port-mapped I/O
- 每個I/O port (device)被指定一個獨特的==port address==，一開始就燒死與軟體無關
- 每個I/O port由==四個registers (1~4Bytes)組成==
- Data-in register：read by the host to get input
- Data-out register：written by the host to send output
- Status register：read by the host to check I/O status
- Control register：written by the host to control the device
- Program與I/O port的互動透過==special I/O instructions==
- i/o instruction裡頭一定有port address

## I/O Methods Categorization
- Depending on ==how to address a device==
- ==Port-mapped I/O==
- 與memory用不一樣的address space
- 透過special I/O instruction (e.g. IN, OUT)做存取
- ==Memory-mapped I/O==
- 保留特別的memory space給device
- 透過標準的data-transfer instruction (e.g. MOV)做存取
- 寫到memory等於寫到某個device的buffer，要事先register好
- 優點：對大量的memory I/O是更有效率的(e.g. graphic card)
- 缺點：Vulnerable to accidental modification, error
- Depending on ==how to interact with a device==
- ==Poll (busy-waiting)==：processor週期性的檢查device的status register
- ==Interrupt==：device通知processor I/O的完成
- 若是資料量大，選擇用interrupt
- Depending on ==who to control the transfer==
- ==Programmed I/O==：transfer由CPU控制
- CPU會一直接收到interrupt，速度變慢
- ==Direct memory access== (DMA) I/O：由DMA controller控制(一個有特別目的的controller)
- 由I/O controller負責搬資料，不屬於任何一個I/O device
- 設計給大量資料傳輸
- 通常搭配使用memory-mapped I/O和interrupt I/O method
- 

## Interrupt
### Interrupt Vector Table
Intel Pentium Processor
### Interrupt Prioritization
- Maskable interrupt
- 參照IRQ
- 透過設定中斷遮罩暫存器（interrupt mask register; IMR）內的位元遮罩（bit-mask）值來==讓它可被系統忽略==
- 比當前較低優先度的interrupt不被認定直到pending interrupt完成
- Non-maskable interrupt (NMI)：highest-priority, never masked
- ==無法被系統忽略的==
- NMI通常使用於計時器（timers）
- Often used for power-down, memory error
### Interrupt-Driven I/O
## CPU and device Interrupt handshake
1. Device asserts ==interrupt request (IRQ)==
2. CPU在每個instruction cycle的開始檢查==interrupt request line==
3. ==Save== the status and address of ==interrupted process==
4. CPU認知道了interrupt，接著在interrupt vector table搜尋interrupt handler routines
5. CPU從==interrupt handler routine==抓下一個 instruction
6. 在完成執行interrupt handler routine後恢復interrupted process

## Note
- PnP（Plug and Play）技術可以協助解決 IRQ 線路衝突的問題
- PnP，簡化來說，就是讓周邊設備自己發出訊號去和所有其他支援 PnP 設計的周邊設備**協商，找到尚未被佔據的 IRQ 線路**
- USB 使得許多種不需要太快傳輸速度的周邊設備，例如：鍵盤、滑鼠、印表機、數據機、軟碟機、掃瞄機等等 (注意硬碟機不在此列，因為硬碟機需要越快越好的傳輸速度)， 可以串聯在一起，而且共用一個 IRQ 線路

# Kernel I/O Subsystem
## I/O Subsystem
- I/O Scheduling
- 藉由對I/O queue中的job排順序提昇系統效能
- e.g. disk I/O order scheduling
- Buffering
- 當要在I/O devices之間做transfer，將data存到memory（e.g. disk to network card）
- 調節I/O devices之間的速度
- 原因
- devices之間的==速度不一樣==
- Devices with ==different data-transfer sizes==
- Support copy semantics
- 要做多次的data copy
- Caching – fast memory that holds copies of data
- 永遠只是個copy
- Key to ==performance==，增加access的速度
- Spooling – holds **output** for a device
- 先load到spooling的空間去，再一次送到printer
- e.g. printing (cannot accept interleaved files)
- Error handling – when I/O error happens
- e.g. SCSI devices returns error information
- I/O protection
- Privileged instructions
## UNIX I/O Kernel Data Structure
- Linux treats all I/O devices like a ==file==
- 
## Device-status Table
## Blocking and Nonblocking I/O
- Blocking - process被中斷直到I/O完成
- Easy to use and understand
- Insufficient for some needs
- Use for ==synchronous== communication & I/O
- Nonblocking
- 透過==multi-threading==實作
- Returns quickly with count of bytes read or written
- Use for ==asynchronous== communication & I/O
- e.g. call interrupt告訴你他做完了，有write, write call為一組
## Life Cycle of An I/O Request

# Performance
## I/O在系統效能上是一個主要的因素
### 原因
- 讓CPU執行==device driver code==
- 每call一層driver，每一個subsystem都是不同的process，需要==context switches==
- 因為有一堆layer，controllers、buffer、 cache有一堆的==data copy==
- ==Interrupt handling== is a relatively expensive task
- 當I/O的時間少，Busy-waiting可以比interrupt-driven表現的更以效率
## Intercomputer Communications
- Network traffic could cause high context switch rate
- Interrupt generated during keyboard & network I/O
- Context switch occurs between prog. & kernel (drivers)
- 
## Improving Performance
- 減少==context switches==的次數
- 減少==data copying==
- 藉由large transfers，smart controllers，polling減少==interrupts==
- Use ==DMA==，CPU可以空出來
- Balance CPU，memory，bus，and I/O performance for highest throughput

# Application Interface
## A Kernel I/O Structure
### Device drivers
- a uniform device-access ==interface to the I/O subsystem==
- ==hide the differences== among **device controllers** from the **I/O sub-system of OS**
- 
## Characteristics of I/O Devices
## I/O Device Class
- Device class is fairly standard across different OS
- Block I/O
- Char-stream I/O
- Memory-mapped file access
- Network sockets
- Clock & timer interfaces
- ==Back-door== interfaces (e.g. ioctl() )，難以歸類到以上的類別
- call system calll：ioctl()完後，可以去操作任何的I/O device
- 使application可以任意access被device driver實作函式，而不用透過system call
## Block & Char Devices
- Block devices：disk drives
- system calls：read( ), write( ), seek( )，一整個memory buffer的概念
- Memory-mapped file can be layered on top
- 大的，用interrupt做I/O
- Char-stream devices：mouse, keyboard, serial ports
- system calls：get( ), put( )，以bytes為單位
- Libraries layered on top allow line editing
- 小的，用pulling做I/O
## Network Devices
- 有很多種形式（block、character...），對於個別有自己的interface
- System call: send(), recv(), select()
- select() returns which socket is waiting to send or receive, eliminates the need of busy waiting
- Many other approaches
- pipes, FIFOS, STREAMS, message queues