3 Think OS A Brief Introduction to Operating Systems ::: Version 0.7.4 ::: 3 Think OS A Brief Introduction to Operating Systems ::: Version 0.7.4 Allen B. Downey Green Tea Press Needham, Massachusetts ::: Copyright © 2015 Allen B. Downey. Green Tea Press 9 Washburn Ave Needham MA 02492 ::: Permission is granted to copy, distribute, and/or modify this document under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License, which is available at . The LaTeX source for this book is available from . ::: Think OS: A Brief Introduction to Operating Systems Allen B. Downey Version 0.7.4 Copyright 2015 Allen B. Downey Permission is granted to copy, distribute, and/or modify this document under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Unported License, which is available at . ::: # Preface In many computer science programs, Operating Systems is an advanced topic. By the time students take it, they know how to program in C, and they have probably taken a class in Computer Architecture. Usually the goal of the class is to expose students to the design and implementation of operating systems, with the implied assumption that some of them will do research in this area, or write part of an OS. This book is intended for a different audience, and it has different goals. I developed it for a class at Olin College called Software Systems. Most students taking this class learned to program in Python, so one of the goals is to help them learn C. For that part of the class, I use Griffiths and Griffiths, *Head First C*, from O'Reilly Media. This book is meant to complement that one. Few of my students will ever write an operating system, but many of them will write low-level applications in C or work on embedded systems. My class includes material from operating systems, networks, databases, and embedded systems, but it emphasizes the topics programmers need to know. This book does not assume that you have studied Computer Architecture. As we go along, I will explain what we need. If this book is successful, it should give you a better understanding of what is happening when programs run, and what you can do to make them run better and faster. Chapter 1 explains some of the differences between compiled and interpreted languages, with some insight into how compilers work. Recommended reading: *Head First C* Chapter 1. Chapter 2 explains how the operating system uses processes to protect running programs from interfering with each other. Chapter 3 explains virtual memory and address translation. Recommended reading: *Head First C* Chapter 2. Chapter 4 is about file systems and data streams. Recommended reading: *Head First C* Chapter 3. Chapter 5 describes how numbers, letters, and other values are encoded, and presents the bitwise operators. Chapter 6 explains how to use dynamic memory management, and how it works. Recommended reading: *Head First C* Chapter 6. Chapter 7 is about caching and the memory hierarchy. Chapter 8 is about multitasking and scheduling. Chapter 9 is about POSIX threads and mutexes. Recommended reading: *Head First C* Chapter 12 and *Little Book of Semaphores* Chapters 1 and 2. Chapter 10 is about POSIX condition variables and the producer/consumer problem. Recommended reading: *Little Book of Semaphores* Chapters 3 and 4. Chapter 11 is about using POSIX semaphores and implementing semaphores in C. ## A note on this draft The current version of this book is an early draft. While I am working on the text, I have not yet included the figures. So there are a few places where, I'm sure, the explanation will be greatly improved when the figures are ready. ## Using the code Example code for this book is available from . Git is a version control system that allows you to keep track of the files that make up a project. A collection of files under Git's control is called a **repository**. GitHub is a hosting service that provides storage for Git repositories and a convenient web interface. The GitHub homepage for my repository provides several ways to work with the code: - You can create a copy of my repository on GitHub by pressing the Fork button. If you don't already have a GitHub account, you'll need to create one. After forking, you'll have your own repository on GitHub that you can use to keep track of code you write while working on this book. Then you can clone the repo, which means that you copy the files to your computer. - Or you could clone my repository. You don't need a GitHub account to do this, but you won't be able to write your changes back to GitHub. - If you don't want to use Git at all, you can download the files in a Zip file using the button in the lower-right corner of the GitHub page. ## Contributor List If you have a suggestion or correction, please send email to `downey@allendowney.com`. If I make a change based on your feedback, I will add you to the contributor list (unless you ask to be omitted). If you include at least part of the sentence the error appears in, that makes it easy for me to search. Page and section numbers are fine, too, but not quite as easy to work with. Thanks! - I am grateful to the students in Software Systems at Olin College, who tested an early draft of this book in Spring 2014. They corrected many errors and made many helpful suggestions. I appreciate their pioneering spirit! - James P Giannoules spotted a copy-and-paste error. - Andy Engle knows the difference between GB and GiB. - Aashish Karki noted some broken syntax. Other people who found typos and errors include Jim Tyson, Donald Robertson, Jeremy Vermast, Yuzhong Huang, Ian Hill. # Compilation ## Compiled and interpreted languages People often describe programming languages as either compiled or interpreted. "Compiled" means that programs are translated into machine language and then executed by hardware; "interpreted" means that programs are read and executed by a software interpreter. Usually C is considered a compiled language and Python is considered an interpreted language. But the distinction is not always clear-cut. First, many languages can be either compiled or interpreted. For example, there are C interpreters and Python compilers. Second, there are languages like Java that use a hybrid approach, compiling programs into an intermediate language and then running the translated program in an interpreter. Java uses an intermediate language called Java bytecode, which is similar to machine language, but it is executed by a software interpreter, the Java virtual machine (JVM). So being compiled or interpreted is not an intrinsic characteristic of a language; nevertheless, there are some general differences between compiled and interpreted languages. ## Static types Many interpreted languages support dynamic types, but compiled languages are usually limited to static types. In a statically-typed language, you can tell by looking at the program what type each variable refers to. In a dynamically-typed language, you don't always know the type of a variable until the program is running. In general, **static** refers to things that happen at compile time (while a program is being compiled), and **dynamic** refers to things that happen at run time (while a program is running). For example, in Python you can write a function like this: def add(x, y): return x + y Looking at this code, you can't tell what type `x` and `y` will refer to at run time. This function might be called several times, each time with values with different types. Any values that support the addition operator will work; any other types will cause an exception or **runtime error**. In C you would write the same function like this: int add(int x, int y) { return x + y; } The first line of the function includes **type declarations** for the parameters and the return value: `x` and `y` are declared to be integers, which means that we can check at compile time whether the addition operator is legal for this type (it is). The return value is also declared to be an integer. Because of these declarations, when this function is called elsewhere in the program, the compiler can check whether the arguments provided have the right type, and whether the return value is used correctly. These checks happen before the program starts executing, so errors can be found earlier. More importantly, errors can be found in parts of the program that have never run. Furthermore, these checks don't have to happen at run time, which is one of the reasons compiled languages generally run faster than interpreted languages. Declaring types at compile time also saves space. In dynamic languages, variable names are stored in memory while the program runs, and they are often accessible by the program. For example, in Python the built-in function `locals` returns a dictionary that contains variable names and their values. Here's an example in a Python interpreter: >>> x = 5 >>> print locals() {'x': 5, '__builtins__': , '__name__': '__main__', '__doc__': None, '__package__': None} This shows that the name of the variable is stored in memory while the program is running (along with some other values that are part of the default runtime environment). In compiled languages, variable names exist at compile-time but not at run time. The compiler chooses a location for each variable and records these locations as part of the compiled program.^1 The location of a variable is called its **address**. At run time, the value of each variable is stored at its address, but the names of the variables are not stored at all (unless they are added by the compiler for purposes of debugging). ## The compilation process As a programmer, you should have a mental model of what happens during compilation. If you understand the process, it will help you interpret error messages, debug your code, and avoid common pitfalls. The steps of compilation are: 1. Preprocessing: C is one of several languages that include **preprocessing directives** that take effect before the program is compiled. For example, the `#include` directive causes the source code from another file to be inserted at the location of the directive. 2. Parsing: During parsing, the compiler reads the source code and builds an internal representation of the program, called an **abstract syntax tree**. Errors detected during this step are generally syntax errors. 3. Static checking: The compiler checks whether variables and values have the right type, whether functions are called with the right number and type of arguments, etc. Errors detected during this step are sometimes called **static semantic** errors. 4. Code generation: The compiler reads the internal representation of the program and generates machine code or byte code. 5. Linking: If the program uses values and functions defined in a library, the compiler has to find the appropriate library and include the required code. 6. Optimization: At several points in the process, the compiler can transform the program to generate code that runs faster or uses less space. Most optimizations are simple changes that eliminate obvious waste, but some compilers perform sophisticated analyses and transformations. Normally when you run `gcc`, it runs all of these steps and generates an executable file. For example, here is a minimal C program: #include int main() { printf("Hello World\n"); } If you save this code in a file called `hello.c`, you can compile and run it like this: $ gcc hello.c $ ./a.out By default, `gcc` stores the executable code in a file called `a.out` (which originally stood for "assembler output"). The second line runs the executable. The prefix `./` tells the shell to look for it in the current directory. It is usually a good idea to use the `-o` flag to provide a better name for the executable: $ gcc hello.c -o hello $ ./hello ## Object code The `-c` flag tells `gcc` to compile the program and generate machine code, but not to link it or generate an executable: $ gcc hello.c -c The result is a file named `hello.o`, where the `o` stands for **object code**, which is the compiled program. Object code is not executable, but it can be linked into an executable. The UNIX command `nm` reads an object file and generates information about the names it defines and uses. For example: $ nm hello.o 0000000000000000 T main U puts This output indicates that `hello.o` defines the name `main` and uses a function named `puts`, which stands for "put string". In this example, `gcc` performs an optimization by replacing `printf`, which is a large and complicated function, with `puts`, which is relatively simple. You can control how much optimization `gcc` does with the `-O` flag. By default, it does very little optimization, which can make debugging easier. The option `-O1` turns on the most common and safe optimizations. Higher numbers turn on additional optimizations that require longer compilation time. In theory, optimization should not change the behavior of the program, other than to speed it up. But if your program has a subtle bug, you might find that optimization makes the bug appear or disappear. It is usually a good idea to turn off optimization while you are developing new code. Once the program is working and passing appropriate tests, you can turn on optimization and confirm that the tests still pass. ## Assembly code Similar to the `-c` flag, the `-S` flag tells `gcc` to compile the program and generate assembly code, which is basically a human-readable form of machine code. $ gcc hello.c -S The result is a file named `hello.s`, which might look something like this: .file "hello.c" .section .rodata .LC0: .string "Hello World" .text .globl main .type main, @function main: .LFB0: .cfi_startproc pushq %rbp .cfi_def_cfa_offset 16 .cfi_offset 6, -16 movq %rsp, %rbp .cfi_def_cfa_register 6 movl $.LC0, %edi call puts movl $0, %eax popq %rbp .cfi_def_cfa 7, 8 ret .cfi_endproc .LFE0: .size main, .-main .ident "GCC: (Ubuntu/Linaro 4.7.3-1ubuntu1) 4.7.3" .section .note.GNU-stack,"",@progbits `gcc` is usually configured to generate code for the machine you are running on, so for me it generates x86 assembly language, which runs on a wide variety of processors from Intel, AMD, and others. If you are running on a different architecture, you might see different code. ## Preprocessing Taking another step backward through the compilation process, you can use the `-E` flag to run the preprocessor only: $ gcc hello.c -E The result is the output from the preprocessor. In this example, it contains the included code from `stdio.h`, and all the files included from `stdio.h`, and all the files included from those files, and so on. On my machine, the total is more than 800 lines of code. Since almost every C program includes `stdio.h`, those 800 lines of code get compiled a lot. If, like many C programs, you also include `stdlib.h`, the result is more than 1800 lines of code. ## Understanding errors Now that we know the steps in the compilation process, it is easier to understand error messages. For example, if there is an error in a `#include` directive, you'll get a message from the preprocessor: hello.c:1:20: fatal error: stdioo.h: No such file or directory compilation terminated. If there's a syntax error, you get a message from the compiler: hello.c: In function 'main': hello.c:6:1: error: expected ';' before '}' token If you use a function that's not defined in any of the standard libraries, you get a message from the linker: /tmp/cc7iAUbN.o: In function `main': hello.c:(.text+0xf): undefined reference to `printff' collect2: error: ld returned 1 exit status `ld` is the name of the UNIX linker, so named because "loading" is another step in the compilation process that is closely related to linking. Once the program starts, C does very little runtime checking, so there are only a few runtime errors you are likely to see. If you divide by zero, or perform another illegal floating-point operation, you will get a "Floating point exception." And if you try to read or write an incorrect location in memory, you will get a "Segmentation fault." # Processes ## Abstraction and virtualization Before we talk about processes, I want to define a few words: - Abstraction: An abstraction is a simplified representation of something complicated. For example, if you drive a car, you understand that when you turn the wheel left, the car goes left, and vice versa. Of course, the steering wheel is connected to a sequence of mechanical and (often) hydraulic systems that turn the wheels, and the wheels interact with the road in ways that can be complex, but as a driver, you normally don't have to think about any of those details. You can get along very well with a simple mental model of steering. Your mental model is an abstraction. Similarly, when you use a web browser, you understand that when you click on a link, the browser displays the page the link refers to. The software and network communication that make that possible are complex, but as a user, you don't have to know the details. A large part of software engineering is designing abstractions like these that allow users and other programmers to use powerful and complicated systems without having to know about the details of their implementation. - Virtualization: An important kind of abstraction is virtualization, which is the process of creating a desirable illusion. For example, many public libraries participate in inter-library collaborations that allow them to borrow books from each other. When I request a book, sometimes the book is on the shelf at my local library, but other times it has to be transferred from another collection. Either way, I get a notification when it is available for pickup. I don't need to know where it came from, and I don't need to know which books my library has. As a whole, the system creates the illusion that my library has every book in the world. The collection physically located at my local library might be small, but the collection available to me virtually includes every book in the inter-library collaboration. As another example, most computers are only connected to one network, but that network is connected to others, and so on. What we call the Internet is a collection of networks and a set of protocols that forward packets from one network to the next. From the point of view of a user or programmer, the system behaves as if every computer on the Internet is connected to every other computer. The number of physical connections is small, but the number of virtual connections is very large. The word "virtual" is often used in the context of a virtual machine, which is software that creates the illusion of a dedicated computer running a particular operating system, when in reality the virtual machine might be running, along with many other virtual machines, on a computer running a different operating system. In the context of virtualization, we sometimes call what is really happening "physical", and what is virtually happening either "logical" or "abstract." ## Isolation One of the most important principles of engineering is isolation: when you are designing a system with multiple components, it is usually a good idea to isolate them from each other so that a change in one component doesn't have undesired effects on other components. One of the most important goals of an operating system is to isolate each running program from the others so that programmers don't have to think about every possible interaction. The software object that provides this isolation is a **process**. A process is a software object that represents a running program. I mean "software object" in the sense of object-oriented programming; in general, an object contains data and provides methods that operate on the data. A process is an object that contains the following data: - The text of the program, usually a sequence of machine language instructions. - Data associated with the program, including static data (allocated at compile time) and dynamic data (allocated at run time). - The state of any pending input/output operations. For example, if the process is waiting for data to be read from disk or for a packet to arrive on a network, the status of these operations is part of the process. - The hardware state of the program, which includes data stored in registers, status information, and the program counter, which indicates which instruction is currently executing. Usually one process runs one program, but it is also possible for a process to load and run a new program. It is also possible, and common, to run the same program in more than one process. In that case, the processes share the same program text but generally have different data and hardware states. Most operating systems provide a fundamental set of capabilities to isolate processes from each other: - Multitasking: Most operating systems have the ability to interrupt a running process at almost any time, save its hardware state, and then resume the process later. In general, programmers don't have to think about these interruptions. The program behaves as if it is running continuously on a dedicated processor, except that the time between instructions is unpredictable. - Virtual memory: Most operating systems create the illusion that each process has its own chunk of memory, isolated from all other processes. Again, programmers generally don't have to think about how virtual memory works; they can proceed as if every program has a dedicated chunk of memory. - Device abstraction: Processes running on the same computer share the disk drive, the network interface, the graphics card, and other hardware. If processes interacted with this hardware directly, without coordination, chaos would ensue. For example, network data intended for one process might be read by another. Or multiple processes might try to store data in the same location on a hard drive. It is up to the operating system to maintain order by providing appropriate abstractions. As a programmer, you don't need to know much about how these capabilities are implemented. But if you are curious, you will find a lot of interesting things going on under the metaphorical hood. And if you know what's going on, it can make you a better programmer. ## UNIX processes While I write this book, the process I am most aware of is my text editor, emacs. Every once in a while I switch to a terminal window, which is a window running a UNIX shell that provides a command-line interface. When I move the mouse, the window manager wakes up, sees that the mouse is over the terminal window, and wakes up the terminal. The terminal wakes up the shell. If I type `make` in the shell, it creates a new process to run Make, which creates another process to run LaTeX and then another process to display the results. If I need to look something up, I might switch to another desktop, which wakes up the window manager again. If I click on the icon for a web browser, the window manager creates a process to run the web browser. Some browsers, like Chrome, create a new process for each window and each tab. And those are just the processes I am aware of. At the same time there are many other processes running in the **background**. Many of them are performing operations related to the operating system. The UNIX command `ps` prints information about running processes. If you run it in a terminal, you might see something like this: PID TTY TIME CMD 2687 pts/1 00:00:00 bash 2801 pts/1 00:01:24 emacs 24762 pts/1 00:00:00 ps The first column is the unique numerical process ID. The second column is the terminal that created the process; "TTY" stands for teletypewriter, which was the original mechanical terminal. The third column is the total processor time used by the process, in hours, minutes, and seconds. The last column is the name of the running program. In this example, `bash` is the name of the shell that interprets the commands I type in the terminal, emacs is my text editor, and ps is the program generating this output. By default, `ps` lists only the processes associated with the current terminal. If you use the `-e` flag, you get every process (including processes belonging to other users, which is a security flaw, in my opinion). On my system there are currently 233 processes. Here are some of them: PID TTY TIME CMD 1 ? 00:00:17 init 2 ? 00:00:00 kthreadd 3 ? 00:00:02 ksoftirqd/0 4 ? 00:00:00 kworker/0:0 8 ? 00:00:00 migration/0 9 ? 00:00:00 rcu_bh 10 ? 00:00:16 rcu_sched 47 ? 00:00:00 cpuset 48 ? 00:00:00 khelper 49 ? 00:00:00 kdevtmpfs 50 ? 00:00:00 netns 51 ? 00:00:00 bdi-default 52 ? 00:00:00 kintegrityd 53 ? 00:00:00 kblockd 54 ? 00:00:00 ata_sff 55 ? 00:00:00 khubd 56 ? 00:00:00 md 57 ? 00:00:00 devfreq_wq `init` is the first process created when the operating system starts. It creates many of the other processes, and then sits idle until the processes it created are done. `kthreadd` is a process the operating system uses to create new **threads**. We'll talk more about threads later, but for now you can think of a thread as kind of a process. The `k` at the beginning stands for **kernel**, which is the part of the operating system responsible for core capabilities like creating threads. The extra `d` at the end stands for **daemon**, which is another name for processes like this that run in the background and provide operating system services. In this context, "daemon" is used in the sense of a helpful spirit, with no connotation of evil. Based on the name, you can infer that `ksoftirqd` is also a kernel daemon; specifically, it handles software interrupt requests, or "soft IRQ". `kworker` is a worker process created by the kernel to do some kind of processing for the kernel. There are often multiple processes running these kernel services. On my system at the moment, there are 8 `ksoftirqd` processes and 35 `kworker` processes. I won't go into more details about the other processes, but if you are interested you can search for more information about them. You should run `ps` on your system and compare your results to mine. # Virtual memory ## A bit of information theory A **bit** is a binary digit; it is also a unit of information. If you have one bit, you can specify one of two possibilities, usually written 0 and 1. If you have two bits, there are 4 possible combinations, 00, 01, 10, and 11. In general, if you have $b$ bits, you can indicate one of $2^b$ values. A **byte** is 8 bits, so it can hold one of 256 values. Going in the other direction, suppose you want to store a letter of the alphabet. There are 26 letters, so how many bits do you need? With 4 bits, you can specify one of 16 values, so that's not enough. With 5 bits, you can specify up to 32 values, so that's enough for all the letters, with a few values left over. In general, if you want to specify one of $N$ values, you should choose the smallest value of $b$ so that $2^b \ge N$. Taking the log base 2 of both sides yields $b \ge log_2 N$. Suppose I flip a coin and tell you the outcome. I have given you one bit of information. If I roll a six-sided die and tell you the outcome, I have given you $log_2 6$ bits of information. And in general, if the probability of the outcome is 1 in $N$, then the outcome contains $log_2 N$ bits of information. Equivalently, if the probability of the outcome is $p$, then the information content is $-log_2 p$. This quantity is called the **self-information** of the outcome. It measures how surprising the outcome is, which is why it is also called **surprisal**. If your horse has only one chance in 16 of winning, and he wins, you get 4 bits of information (along with the payout). But if the favorite wins 75% of the time, the news of the win contains only 0.42 bits. Intuitively, unexpected news carries a lot of information; conversely, if there is something you were already confident of, confirming it contributes only a small amount of information. For several topics in this book, we will need to be comfortable converting back and forth between the number of bits, $b$, and the number of values they can encode, $N = 2^b$. ## Memory and storage While a process is running, most of its data is held in **main memory**, which is usually some kind of random access memory (RAM). On most current computers, main memory is **volatile**, which means that when the computer shuts down, the contents of main memory are lost. A typical desktop computer has 2--8 GiB of memory. GiB stands for "gibibyte," which is $2^$ bytes. If the process reads and writes files, those files are usually stored on a hard disk drive (HDD) or solid state drive (SSD). These storage devices are **non-volatile**, so they are used for long-term storage. Currently a typical desktop computer has a HDD with a capacity of 500 GB to 2 TB. GB stands for "gigabyte," which is $10^9$ bytes. TB stands for "terabyte," which is $10^$ bytes. You might have noticed that I used the binary unit GiB for the size of main memory and the decimal units GB and TB for the size of the HDD. For historical and technical reasons, memory is measured in binary units, and disk drives are measured in decimal units. In this book I will be careful to distinguish binary and decimal units, but you should be aware that the word "gigabyte" and the abbreviation GB are often used ambiguously. In casual use, the term "memory" is sometimes used for HDDs and SSDs as well as RAM, but the properties of these devices are very different, so we will need to distinguish them. I will use **storage** to refer to HDDs and SSDs. ## Address spaces Each byte in main memory is specified by an integer **physical address**. The set of valid physical addresses is called the physical **address space**. It usually runs from 0 to $N-1$, where $N$ is the size of main memory. On a system with 1 GiB of physical memory, the highest valid address is $2^-1$, which is 1,073,741,823 in decimal, or 0x3fff ffff in hexadecimal (the prefix 0x indicates a hexadecimal number). However, most operating systems provide **virtual memory**, which means that programs never deal with physical addresses, and don't have to know how much physical memory is available. Instead, programs work with **virtual addresses**, which are numbered from 0 to $M-1$, where $M$ is the number of valid virtual addresses. The size of the virtual address space is determined by the operating system and the hardware it runs on. You have probably heard people talk about 32-bit and 64-bit systems. These terms indicate the size of the registers, which is usually also the size of a virtual address. On a 32-bit system, virtual addresses are 32 bits, which means that the virtual address space runs from 0 to 0xffff ffff. The size of this address space is $2^$ bytes, or 4 GiB. On a 64-bit system, the size of the virtual address space is $2^$ bytes, or $2^4 \cdot 1024^6$ bytes. That's 16 exbibytes, which is about a billion times bigger than current physical memories. It might seem strange that a virtual address space can be so much bigger than physical memory, but we will see soon how that works. When a program reads and writes values in memory, it generates virtual addresses. The hardware, with help from the operating system, translates to physical addresses before accessing main memory. This translation is done on a per-process basis, so even if two processes generate the same virtual address, they would map to different locations in physical memory. Thus, virtual memory is one important way the operating system isolates processes from each other. In general, a process cannot access data belonging to another process, because there is no virtual address it can generate that maps to physical memory allocated to another process. ## Memory segments The data of a running process is organized into five segments: - The **code segment** contains the program text; that is, the machine language instructions that make up the program. - The **static segment** contains immutable values, like string literals. For example, if your program contains the string `"Hello, World"`, those characters will be stored in the static segment. - The **global segment** contains global variables and local variables that are declared `static`. - The **heap segment** contains chunks of memory allocated at run time, most often by calling the C library function `malloc`. - The **stack segment** contains the call stack, which is a sequence of stack frames. Each time a function is called, a stack frame is allocated to contain the parameters and local variables of the function. When the function completes, its stack frame is removed from the stack. The arrangement of these segments is determined partly by the compiler and partly by the operating system. The details vary from one system to another, but in the most common arrangement: - The text segment is near the "bottom" of memory, that is, at addresses near 0. - The static segment is often just above the text segment, that is, at higher addresses. - The global segment is often just above the static segment. - The heap is often above the global segment. As it expands, it grows up toward larger addresses. - The stack is near the top of memory; that is, near the highest addresses in the virtual address space. As the stack expands, it grows down toward smaller addresses. To determine the layout of these segments on your system, try running this program, which is in `aspace.c` in the repository for this book (see Section 1.1(#code)). #include #include int global; int main () { int local = 5; void *p = malloc(128); char *s = "Hello, World"; printf ("Address of main is %p\n", main); printf ("Address of global is %p\n", &global); printf ("Address of local is %p\n", &local); printf ("p points to %p\n", p); printf ("s points to %p\n", s); } `main` is the name of a function; when it is used as a variable, it refers to the address of the first machine language instruction in `main`, which we expect to be in the text segment. `global` is a global variable, so we expect it to be in the global segment. `local` is a local variable, so we expect it to be on the stack. `s` refers to a "string literal\", which is a string that appears as part of the program (as opposed to a string that is read from a file, input by a user, etc.). We expect the location of the string to be in the static segment (as opposed to the pointer, `s`, which is a local variable). `p` contains an address returned by `malloc`, which allocates space in the heap. "malloc" stands for "memory allocate." The format sequence `%p` tells `printf` to format each address as a "pointer", so it displays the results in hexadecimal. When I run this program, the output looks like this (I added spaces to make it easier to read): Address of main is 0x 40057d Address of global is 0x 60104c Address of local is 0x7ffe6085443c p points to 0x 16c3010 s points to 0x 4006a4 As expected, the address of `main` is the lowest, followed by the location of the string literal. The location of `global` is next, then the address `p` points to. The address of `local` is much bigger. The largest address has 12 hexadecimal digits. Each hex digit corresponds to 4 bits, so it is a 48-bit address. That suggests that the usable part of the virtual address space is $2^$ bytes. As an exercise, run this program on your computer and compare your results to mine. Add a second call to `malloc` and check whether the heap on your system grows up (toward larger addresses). Add a function that prints the address of a local variable, and check whether the stack grows down. ## Static local variables Local variables on the stack are sometimes called **automatic**, because they are allocated automatically when a function is called, and freed automatically when the function returns. In C there is another kind of local variable, called **static**, which is allocated in the global segment. It is initialized when the program starts and keeps its value from one function call to the next. For example, the following function keeps track of how many times it has been called. int times_called() { static int counter = 0; counter++; return counter; } The keyword `static` indicates that `counter` is a static local variable. The initialization happens only once, when the program starts. If you add this function to `aspace.c` you can confirm that `counter` is allocated in the global segment along with global variables, not in the stack. ## Address translation How does a virtual address (VA) get translated to a physical address (PA)? The basic mechanism is simple, but a simple implementation would be too slow and take too much space. So actual implementations are a bit more complicated. Diagram of the address translation process. Most processors provide a memory management unit (MMU) that sits between the CPU and main memory. The MMU performs fast translation between VAs and PAs. 1. When a program reads or writes a variable, the CPU generates a VA. 2. The MMU splits the VA into two parts, called the page number and the offset. A "page" is a chunk of memory; the size of a page depends on the operating system and the hardware, but common sizes are 1--4 KiB. 3. The MMU looks up the page number in the translation lookaside buffer (TLB) and gets the corresponding physical page number. Then it combines the physical page number with the offset to produce a PA. 4. The PA is passed to main memory, which reads or writes the given location. The TLB contains cached copies of data from the page table (which is stored in kernel memory). The page table contains the mapping from virtual page numbers to physical page numbers. Since each process has its own page table, the TLB has to make sure it only uses entries from the page table of the process that's running. Figure 4.1(#addtrans) shows a diagram of this process. To see how it all works, suppose that the VA is 32 bits and the physical memory is 1 GiB, divided into 1 KiB pages. - Since 1 GiB is $2^$ bytes and 1 KiB is $2^$ bytes, there are $2^$ physical pages, sometimes called "frames." - The size of the virtual address space is $2^$ B and the size of a page is $2^$ B, so there are $2^$ virtual pages. - The size of the offset is determined by the page size. In this example the page size is $2^$ B, so it takes 10 bits to specify a byte on a page. - If a VA is 32 bits and the offset is 10 bits, the remaining 22 bits make up the virtual page number. - Since there are $2^$ physical pages, each physical page number is 20 bits. Adding in the 10 bit offset, the resulting PAs are 30 bits. So far this all seems feasible. But let's think about how big a page table might have to be. The simplest implementation of a page table is an array with one entry for each virtual page. Each entry would contain a physical page number, which is 20 bits in this example, plus some additional information about each frame. So we expect 3--4 bytes per entry. But with $2^$ virtual pages, the page table would require $2^$ bytes, or 16 MiB. And since we need a page table for each process, a system running 256 processes would need $2^$ bytes, or 4 GiB, just for page tables! And that's just with 32-bit virtual addresses. With 48- or 64-bit VAs, the numbers are ridiculous. Fortunately, we don't actually need that much space, because most processes don't use even a small fraction of their virtual address space. And if a process doesn't use a virtual page, we don't need an entry in the page table for it. Another way to say the same thing is that page tables are "sparse", which implies that the simple implementation, an array of page table entries, is a bad idea. Fortunately, there are several good implementations for sparse arrays. One option is a multilevel page table, which is what many operating systems, including Linux, use. Another option is an associative table, where each entry includes both the virtual page number and the physical page number. Searching an associative table can be slow in software, but in hardware we can search the entire table in parallel, so associative arrays are often used to represent the page table entries in the TLB. You can read more about these implementations at ; you might find the details interesting. But the fundamental idea is that page tables are sparse, so we have to choose a good implementation for sparse arrays. I mentioned earlier that the operating system can interrupt a running process, save its state, and then run another process. This mechanism is called a **context switch**. Since each process has its own page table, the operating system has to work with the MMU to make sure each process gets the right page table. In older machines, the page table information in the MMU had to be replaced during every context switch, which was expensive. In newer systems, each page table entry in the MMU includes the process ID, so page tables from multiple processes can be in the MMU at the same time. # Files and file systems When a process completes (or crashes), any data stored in main memory is lost. But data stored on a hard disk drive (HDD) or solid state drive (SSD) is "persistent;" that is, it survives after the process completes, even if the computer shuts down. Hard disk drives are complicated. Data is stored in blocks, which are laid out in sectors, which make up tracks, which are arranged in concentric circles on platters. Solid state drives are simpler in one sense, because blocks are numbered sequentially, but they raise a different complication: each block can be written a limited number of times before it becomes unreliable. As a programmer, you don't want to deal with these complications. What you want is an appropriate abstraction of persistent storage hardware. The most common abstraction is called a "file system." Abstractly: - A "file system" is a mapping from each file's name to its contents. If you think of the names as keys, and the contents as values, a file system is a kind of key-value database (see ). - A "file" is a sequence of bytes. File names are usually strings, and they are usually "hierarchical"; that is, the string specifies a path from a top-level directory (or folder), through a series of subdirectories, to a specific file. The primary difference between the abstraction and the underlying mechanism is that files are byte-based and persistent storage is block-based. The operating system translates byte-based file operations in the C library into block-based operations on storage devices. Typical block sizes are 1--8 KiB. For example, the following code opens a file and reads the first byte: FILE *fp = fopen("/home/downey/file.txt", "r"); char c = fgetc(fp); fclose(fp); When this code runs: 1. `fopen` uses the filename to find the top-level directory, called `/`, the subdirectory `home`, and the sub-subdirectory `downey`. 2. It finds the file named `file.txt` and "opens" it for reading, which means it creates a data structure that represents the file being read. Among other things, this data structure keeps track of how much of the file has been read, called the "file position". In DOS, this data structure is called a File Control Block, but I want to avoid that term because in UNIX it means something else. In UNIX, there seems to be no good name for it. It is an entry in the open file table, so I will call it an OpenFileTableEntry. 3. When we call `fgetc`, the operating system checks whether the next character of the file is already in memory. If so, it reads the next character, advances the file position, and returns the result. 4. If the next character is not in memory, the operating system issues an I/O request to get the next block. Disk drives are slow, so a process waiting for a block from disk is usually interrupted so another process can run until the data arrives. 5. When the I/O operation is complete, the new block of data is stored in memory, and the process resumes. It reads the first character and stores it as a local variable. 6. When the process closes the file, the operating system completes or cancels any pending operations, removes data stored in memory, and frees the OpenFileTableEntry. The process for writing a file is similar, but there are some additional steps. Here is an example that opens a file for writing and changes the first character. FILE *fp = fopen("/home/downey/file.txt", "w"); fputc('b', fp); fclose(fp); When this code runs: 1. Again, `fopen` uses the filename to find the file. If it does not already exist, it creates a new file and adds an entry in the parent directory, `/home/downey`. 2. The operating system creates an OpenFileTableEntry that indicates that the file is open for writing, and sets the file position to 0. 3. `fputc` attempts to write (or re-write) the first byte of the file. If the file already exists, the operating system has to load the first block into memory. Otherwise it allocates a new block in memory and requests a new block on disk. 4. After the block in memory is modified, it might not be copied back to the disk right away. In general, data written to a file is "buffered", which means it is stored in memory and only written to disk when there is at least one block to write. 5. When the file is closed, any buffered data is written to disk and the OpenFileTableEntry is freed. To summarize, the C library provides the abstraction of a file system that maps from file names to streams of bytes. This abstraction is built on top of storage devices that are actually organized in blocks. ## Disk performance I mentioned earlier that disk drives are slow. On current HDDs, the average time to read a block from disk to memory might be 5--25 ms (see ). SSDs are faster, taking 25 $\mu$s to read a 4 KiB block and 250 $\mu$s to write one (see ). To put these numbers in perspective, let's compare them to the clock cycle of the CPU. A processor with clock rate 2 GHz completes one clock cycle every 0.5 ns. The time to get a byte from memory to the CPU is typically around 100 ns. If the processor completes one instruction per clock cycle, it would complete 200 instructions while waiting for a byte from memory. In one microsecond, it would complete 2000 instructions, so while waiting 25 $\mu$s for a byte from an SSD, it would complete 50,000. In one millisecond, it would complete 2,000,000 instructions, so while waiting 20 ms for a byte from a HDD, it might complete 40 million. If there's nothing for the CPU to do while it waits, it would be idle. That's why the operating system generally switches to another process while it is waiting for data from disk. The gap in performance between main memory and persistent storage is one of the major challenges of computer system design. Operating systems and hardware provide several features intended to "fill in" this gap: - Block transfers: The time it takes to load a single byte from disk is 5--25 ms. By comparison, the additional time to load an 8 KiB block is negligible. So systems generally try to read large blocks each time they access the disk. - Prefetching: Sometimes the operating system can predict that a process will read a block and start loading it before it is requested. For example, if you open a file and read the first block, there is a good chance you will go on to read the second block. The operating system might start loading additional blocks before they are requested. - Buffering: As I mentioned, when you write a file, the operating system stores the data in memory and only writes it to disk later. If you modify the block several times while it is in memory, the system only has to write it to disk once. - Caching: If a process has used a block recently, it is likely to use it again soon. If the operating system keeps a copy of the block in memory, it can handle future requests at memory speed. Some of these features are also implemented in hardware. For example, some disk drives provide a cache that stores recently-used blocks, and many disk drives read more than one block at a time, even if only one is requested. These mechanisms generally improve the performance of programs, but they don't change the behavior. Usually programmers don't have to think about them, with two exceptions: (1) if the performance of a program is unexpectedly bad, you might have to know something about these mechanisms to diagnose the problem, and (2) when data is buffered, it can be harder to debug a program. For example, if a program prints a value and then crashes, the value might not appear, because it might be in a buffer. Similarly, if a program writes data to disk and then the computer loses power, the data might be lost if it is in a cache and not yet on disk. ## Disk metadata The blocks that make up a file might be arranged contiguously on disk, and file system performance is generally better if they are, but most operating systems don't require contiguous allocation. They are free to place a block anywhere on disk, and they use various data structures to keep track of them. In many UNIX file systems, that data structure is called an "inode," which stands for "index node". More generally, information about files, including the location of their blocks, is called "metadata". (The content of the file is data, so information about the file is data about data, hence "meta".) Since inodes reside on disk along with the rest of the data, they are designed to fit neatly into disk blocks. A UNIX inode contains information about a file, including the user ID of the file owner; permission flags indicating who is allowed to read, write, or execute it; and timestamps that indicate when it was last modified and accessed. In addition, it contains block numbers for the first 12 blocks that make up the file. If the block size is 8 KiB, the first 12 blocks make up 96 KiB. On most systems, that's big enough for a large majority of files, but it's definitely not big enough for all of them. That's why the inode also contains a pointer to an "indirection block", which contains nothing but pointers to other blocks. The number of pointers in an indirection block depends on the sizes of the blocks and the block numbers, but it is often 1024. With 1024 block numbers and 8 KiB blocks, an indirection block can address 8 MiB. That's big enough for all but the largest files, but still not big enough for all. That's why the inode also contains a pointer to a "double indirection block", which contains pointers to indirection blocks. With 1024 indirection blocks, we can address 8 GiB. And if that's not big enough, there is (finally) a triple indirection block, which contains pointers to double indirection blocks, yielding a maximum file size of 8 TiB. When UNIX inodes were designed, that seemed big enough to serve for a long time. But that was a long time ago. As an alternative to indirection blocks, some files systems, like FAT, use a File Allocation Table that contains one entry for each block, called a "cluster" in this context. A root directory contains a pointer to the first cluster in each file. The FAT entry for each cluster points to the next cluster in the file, similar to a linked list. For more details, see . ## Block allocation File systems have to keep track of which blocks belong to each file; they also have to keep track of which blocks are available for use. When a new file is created, the file system finds an available block and allocates it. When a file is deleted, the file system makes its blocks available for re-allocation. The goals of the block allocation system are: - Speed: Allocating and freeing blocks should be fast. - Minimal space overhead: The data structures used by the allocator should be small, leaving as much space as possible for data. - Minimal fragmentation: If some blocks are left unused, or some are only partially used, the unused space is called "fragmentation". - Maximum contiguity: Data that is likely to be used at the same time should be physically contiguous, if possible, to improve performance. It is hard to design a file system that achieves all of these goals, especially since file system performance depends on "workload characteristics" like file sizes, access patterns, etc. A file system that is well tuned for one workload might not perform as well for another. For this reason, most operating systems support several kinds of file systems, and file system design is an active area of research and development. In the last decade, Linux systems have migrated from ext2, which was a conventional UNIX file system, to ext3, a "journaling" file system intended to improve speed and contiguity, and more recently to ext4, which can handle larger files and file systems. Within the next few years, there might be another migration to the B-tree file system, Btrfs. ## Everything is a file? The file abstraction is really a "stream of bytes" abstraction, which turns out to be useful for many things, not just file systems. One example is the UNIX pipe, which is a simple form of inter-process communication. Processes can be set up so that output from one process is taken as input into another process. For the first process, the pipe behaves like a file open for writing, so it can use C library functions like `fputs` and `fprintf`. For the second process, the pipe behaves like a file open for reading, so it uses `fgets` and `fscanf`. Network communication also uses the stream of bytes abstraction. A UNIX socket is a data structure that represents a communication channel between processes on different computers (usually). Again, processes can read data from and write data to a socket using "file" handling functions. Reusing the file abstraction makes life easier for programmers, since they only have to learn one API (application program interface). It also makes programs more versatile, since a program intended to work with files can also work with data coming from pipes and other sources. # More bits and bytes ## Representing integers You probably know that computers represent numbers in base 2, also known as binary. For positive numbers, the binary representation is straightforward; for example, the representation for $5_$ is $b101$. For negative numbers, the most obvious representation uses a sign bit to indicate whether a number is positive or negative. But there is another representation, called "two's complement" that is much more common because it is easier to work with in hardware. To find the two's complement of a negative number, $-x$, find the binary representation of $x$, flip all the bits, and add 1. For example, to represent $-5_$, start with the representation of $5_$, which is $b0000 0101$ if we write the 8-bit version. Flipping all the bits and adding 1 yields $b1111 1011$. In two's complement, the leftmost bit acts like a sign bit; it is 0 for positive numbers and 1 for negative numbers. To convert from an 8-bit number to 16-bits, we have to add more 0's for a positive number and add 1's for a negative number. In effect, we have to copy the sign bit into the new bits. This process is called "sign extension". In C all integer types are signed (able to represent positive and negative numbers) unless you declare them `unsigned`. The difference, and the reason this declaration is important, is that operations on unsigned integers don't use sign extension. ## Bitwise operators People learning C are sometimes confused about the bitwise operators `&` and `|`. These operators treat integers as bit vectors and compute logical operations on corresponding bits. For example, `&` computes the AND operation, which yields 1 if both operands are 1, and 0 otherwise. Here is an example of `&` applied to two 4-bit numbers: 1100 & 1010 ---- 1000 In C, this means that the expression `12 & 10` has the value 8. Similarly, `|` computes the OR operation, which yields 1 if either operand is 1, and 0 otherwise. 1100 | 1010 ---- 1110 So the expression `12 | 10` has the value 14. Finally, `^` computes the XOR operation, which yields 1 if either operand is 1, but not both. 1100 ^ 1010 ---- 0110 So the expression `12 ^ 10` has the value 6. Most commonly, `&` is used to clear a set of bits from a bit vector, `|` is used to set bits, and `^` is used to flip, or "toggle" bits. Here are the details: **Clearing bits**: For any value $x$, $x \& 0$ is 0, and $x \& 1$ is $x$. So if you AND a vector with 3, it selects only the two rightmost bits, and sets the rest to 0. xxxx & 0011 ---- 00xx In this context, the value 3 is called a "mask" because it selects some bits and masks the rest. **Setting bits**: Similarly, for any $x$, $x | 0$ is x, and $x | 1$ is $1$. So if you OR a vector with 3, it sets the rightmost bits, and leaves the rest alone: xxxx | 0011 ---- xx11 **Toggling bits**: Finally, if you XOR a vector with 3, it flips the rightmost bits and leaves the rest alone. As an exercise, see if you can compute the two's complement of 12 using `^`. Hint: what's the two's complement representation of -1? C also provides shift operators, `>`, which shift bits left and right. Each left shift doubles a number, so `5 > 1` is 2 and `2 >> 1` is 1. ## Representing floating-point numbers Floating-point numbers are represented using the binary version of scientific notation. In decimal notation, large numbers are written as the product of a coefficient and 10 raised to an exponent. For example, the speed of light in m/s is approximately $2.998 \cdot 10^8$. Most computers use the IEEE standard for floating-point arithmetic. The C type `float` usually corresponds to the 32-bit IEEE standard; `double` usually corresponds to the 64-bit standard. In the 32-bit standard, the leftmost bit is the sign bit, $s$. The next 8 bits are the exponent, $q$, and the last 23 bits are the coefficient, $c$. The value of a floating-point number is $$(-1)^s c \cdot 2^q$$ Well, that's almost correct, but there's one more wrinkle. Floating-point numbers are usually normalized so that there is one digit before the point. For example, in base 10, we prefer $2.998 \cdot 10^8$ rather than $2998 \cdot 10^5$ or any other equivalent expression. In base 2, a normalized number always has the digit 1 before the binary point. Since the digit in this location is always 1, we can save space by leaving it out of the representation. For example, the integer representation of $13_$ is $b1101$. In floating point, that's $1.101 \cdot 2^3$, so the exponent is 3 and the part of the coefficient that would be stored is 101 (followed by 20 zeros). Well, that's almost correct, but there's one more wrinkle. The exponent is stored with a "bias". In the 32-bit standard, the bias is 127, so the exponent 3 would be stored as 130. To pack and unpack floating-point numbers in C, we can use a union and bitwise operations. Here's an example: union { float f; unsigned int u; } p; p.f = -13.0; unsigned int sign = (p.u >> 31) & 1; unsigned int exp = (p.u >> 23) & 0xff; unsigned int coef_mask = (1 << 23) - 1; unsigned int coef = p.u & coef_mask; printf("%d\n", sign); printf("%d\n", exp); printf("0x%x\n", coef); This code is in `float.c` in the repository for this book (see Section 1.1(#code)). The union allows us to store a floating-point value using `p.f` and then read it as an unsigned integer using `p.u`. To get the sign bit, we shift the bits to the right 31 places and then use a 1-bit mask to select only the rightmost bit. To get the exponent, we shift the bits 23 places, then select the rightmost 8 bits (the hexadecimal value `0xff` has eight 1's). To get the coefficient, we need to extract the 23 rightmost bits and ignore the rest. We do that by making a mask with 1s in the 23 rightmost places and 0s on the left. The easiest way to do that is by shifting 1 to the left by 23 places and then subtracting 1. The output of this program is: 1 130 0x500000 As expected, the sign bit for a negative number is 1. The exponent is 130, including the bias. And the coefficient, which I printed in hexadecimal, is $101$ followed by 20 zeros. As an exercise, try assembling or disassembling a `double`, which uses the 64-bit standard. See . ## Unions and memory errors There are two common uses of C unions. One, which we saw in the previous section, is to access the binary representation of data. Another is to store heterogeneous data. For example, you could use a union to represent a number that might be an integer, float, complex, or rational number. However, unions are error-prone. It is up to you, as the programmer, to keep track of what type of data is in the union; if you write a floating-point value and then interpret it as an integer, the result is usually nonsense. Actually, the same thing can happen if you read a location in memory incorrectly. One way that can happen is if you read past the end of an array. To see what happens, I'll start with a function that allocates an array on the stack and fills it with the numbers from 0 to 99. void f1() { int i; int array100; for (i=0; i<100; i++) { arrayi = i; } } Next I'll define a function that creates a smaller array and deliberately accesses elements before the beginning and after the end: void f2() { int x = 17; int array10; int y = 123; printf("%d\n", array-2); printf("%d\n", array-1); printf("%d\n", array10); printf("%d\n", array11); } If I call `f1` and then `f2`, I get these results: 17 123 98 99 The details here depend on the compiler, which arranges variables on the stack. From these results, we can infer that the compiler put `x` and `y` next to each other, "below" the array (at a lower address). And when we read past the array, it looks like we are getting values that were left on the stack by the previous function call. In this example, all of the variables are integers, so it is relatively easy to figure out what is going on. But in general when you read beyond the bounds of an array, the values you read might have any type. For example, if I change `f1` to make an array of floats, the results are: 17 123 1120141312 1120272384 The latter two values are what you get if you interpret a floating-point value as an integer. If you encountered this output while debugging, you would have a hard time figuring out what's going on. ## Representing strings Related issues sometimes come up with strings. First, remember that C strings are null-terminated. When you allocate space for a string, don't forget the extra byte at the end. Also, the letters *and numbers* in C strings are encoded in ASCII. The ASCII codes for the digits "0" through "9" are 48 through 57, *not* 0 through 9. The ASCII code 0 is the NUL character that marks the end of a string. And the ASCII codes 1 through 9 are special characters used in some communication protocols. ASCII code 7 is a bell; on some terminals, printing it makes a sound. The ASCII code for the letter "A" is 65; the code for "a" is 97. Here are those codes in binary: 65 = b0100 0001 97 = b0110 0001 A careful observer will notice that they differ by a single bit. And this pattern holds for the rest of the letters; the sixth bit (counting from the right) acts as a "case bit", 0 for upper-case letters and 1 for lower case letters. As an exercise, write a function that takes a string and converts from lower-case to upper-case by flipping the sixth bit. As a challenge, you can make a faster version by reading the string 32 or 64 bits at a time, rather than one character at a time. This optimization is made easier if the length of the string is a multiple of 4 or 8 bytes. If you read past the end of a string, you are likely to see strange characters. Conversely, if you write a string and then accidentally read it as an int or float, the results will be hard to interpret. For example, if you run: char array = "allen"; float *p = array; printf("%f\n", *p); You will find that the ASCII representation of the first 8 characters of my name, interpreted as a double-precision floating point number, is 69779713878800585457664. # Memory management C provides 4 functions for dynamic memory allocation: - `malloc`, which takes an integer size, in bytes, and returns a pointer to a newly-allocated chunk of memory with (at least) the given size. If it can't satisfy the request, it returns the special pointer value NULL. - `calloc`, which is the same as `malloc` except that it also clears the newly allocated chunk; that is, it sets all bytes in the chunk to 0. - `free`, which takes a pointer to a previously allocated chunk and deallocates it; that is, it makes the space available for future allocation. - `realloc`, which takes a pointer to a previously allocated chunk and a new size. It allocates a chunk of memory with the new size, copies data from the old chunk to the new, frees the old chunk, and returns a pointer to the new chunk. This API is notoriously error-prone and unforgiving. Memory management is one of the most challenging parts of designing large software systems, which is why most modern languages provide higher-level memory management features like garbage collection. ## Memory errors The C memory management API is a bit like Jasper Beardly, a minor character on the animated television program *The Simpsons*; in a few episodes, he appears as a strict substitute teacher who imposes corporal punishment --- a "paddlin"' --- for all infractions. Here are some of things a program can do that deserve a paddling: - If you access (read or write) any chunk that has not been allocated, that's a paddling. - If you free an allocated chunk and then access it, that's a paddling. - If you try to free a chunk that has not been allocated, that's a paddling. - If you free the same chunk more than once, that's a paddling. - If you call `realloc` with a chunk that was not allocated, or was allocated and then freed, that's a paddling. It might not sound difficult to follow these rules, but in a large program a chunk of memory might be allocated in one part of the program, used in several other parts, and freed in yet another part. So changes in one part of the program can require changes in many other parts. Also, there might be many aliases, or references to the same allocated chunk, in different parts of the program. The chunk should not be freed until all references to the chunk are no longer in use. Getting this right often requires careful analysis across all parts of the program, which is difficult and contrary to fundamental principles of good software engineering. Ideally, every function that allocates memory should include, as part of the documented interface, information about how that memory is supposed to be freed. Mature libraries often do this well, but in the real world, software engineering practice often falls short of this ideal. To make matters worse, memory errors can be difficult to find because the symptoms are unpredictable. For example: - If you read a value from an unallocated chunk, the system *might* detect the error, trigger a runtime error called a "segmentation fault", and stop the program. Or, the program might read unallocated memory without detecting the error; in that case, the value it gets is whatever happened to be stored at the accessed location, which is unpredictable, and might be different each time the program runs. - If you write a value to an unallocated chunk, and don't get a segmentation fault, things are even worse. After you write a value to an invalid location, a long time might pass before it is read and causes problems. At that point it will be very difficult to find the source of the problem. And things can be even worse than that! One of the most common problems with C-style memory management is that the data structures used to implement `malloc` and `free` (which we will see soon) are often stored along with the allocated chunks. So if you accidentally write past the end of a dynamically-allocated chunk, you are likely to mangle these data structures. The system usually won't detect the problem until later, when you call `malloc` or `free`, and those functions fail in some inscrutable way. One conclusion you should draw from this is that safe memory management requires design and discipline. If you write a library or module that allocates memory, you should also provide an interface to free it, and memory management should be part of the API design from the beginning. If you use a library that allocates memory, you should be disciplined in your use of the API. For example, if the library provides functions to allocate and deallocate storage, you should use those functions and not, for example, call `free` on a chunk you did not `malloc`. And you should avoid keeping multiple references to the same chunk in different parts of your program. Often there is a trade-off between safe memory management and performance. For example, the most common source of memory errors is writing beyond the bounds of an array. The obvious remedy for this problem is bounds checking; that is, every access to the array should check whether the index is out of bounds. High-level libraries that provide array-like structures usually perform bounds checking. But C arrays and most low-level libraries do not. ## Memory leaks There is one more memory error that may or may not deserve a paddling. If you allocate a chunk of memory and never free it, that's a "memory leak". For some programs, memory leaks are ok. For example, if your program allocates memory, performs computations on it, and then exits, it is probably not necessary to free the allocated memory. When the program exits, all of its memory is deallocated by the operating system. Freeing memory immediately before exiting might feel more responsible, but it is mostly a waste of time. But if a program runs for a long time and leaks memory, its total memory use will increase indefinitely. At that point, a few things might happen: - At some point, the system runs out of physical memory. On systems without virtual memory, the next call to `malloc` will fail, returning NULL. - On systems with virtual memory, the operating system can move another process's pages from memory to disk and then allocate more space to the leaking process. I explain this mechanism in Section 8.8(#paging). - There might be a limit on the amount of space a single process can allocate; beyond that, `malloc` returns NULL. - Eventually, a process might fill its virtual address space (or the usable part). After that, there are no more addresses to allocate, so `malloc` returns NULL. If `malloc` returns NULL, but you persist and access the chunk you think you allocated, you get a segmentation fault. For this reason, it is considered good style to check the result from `malloc` before using it. One option is to add a condition like this after every `malloc` call: void *p = malloc(size); if (p == NULL) { perror("malloc failed"); exit(-1); } `perror` is declared in `stdio.h`; it prints an error message and additional information about the last error that occurred. `exit`, which is declared in `stdlib.h`, causes the process to terminate. The argument is a status code that indicates how the process terminated. By convention, status code 0 indicates normal termination and -1 indicates an error condition. Sometimes other codes are used to indicate different error conditions. Error-checking code can be a nuisance, and it makes programs harder to read. You can mitigate these problems by wrapping library function calls and their error-checking code in your own functions. For example, here is a `malloc` wrapper that checks the return value. void *check_malloc(int size) { void *p = malloc (size); if (p == NULL) { perror("malloc failed"); exit(-1); } return p; } Because memory management is so difficult, most large programs, like web browsers, leak memory. To see which programs on your system are using the most memory, you can use the UNIX utilities `ps` and `top`. ## Implementation When a process starts, the system allocates space for the text segment and statically allocated data, space for the stack, and space for the heap, which contains dynamically allocated data. Not all programs allocate data dynamically, so the initial size of the heap might be small or zero. Initially the heap contains only one free chunk. When `malloc` is called, it checks whether it can find a free chunk that's big enough. If not, it has to request more memory from the system. The function that does that is `sbrk`, which sets the "program break", which you can think of as a pointer to the end of the heap. When `sbrk` is called, the OS allocates new pages of physical memory, updates the process's page table, and sets the program break. In theory, a program could call `sbrk` directly (without using `malloc`) and manage the heap itself. But `malloc` is easier to use and, for most memory-use patterns, it runs fast and uses memory efficiently. To implement the memory management API (that is, the functions `malloc`, `free`, `calloc`, and `realloc`), most Linux systems use `ptmalloc`, which is based on `dlmalloc`, written by Doug Lea. A short paper that describes key elements of the implementation is available at . For programmers, the most important elements to be aware of are: - The run time of `malloc` does not usually depend on the size of the chunk, but might depend on how many free chunks there are. `free` is usually fast, regardless of the number of free chunks. Because `calloc` clears every byte in the chunk, the run time depends on chunk size (as well as the number of free chunks). `realloc` is sometimes fast, if the new size is smaller than the current size, or if space is available to expand the existing chunk. If not, it has to copy data from the old chunk to the new; in that case, the run time depends on the size of the old chunk. - Boundary tags: When `malloc` allocates a chunk, it adds space at the beginning and end to store information about the chunk, including its size and the state (allocated or free). These bits of data are called "boundary tags". Using these tags, `malloc` can get from any chunk to the previous chunk and the next chunk in memory. In addition, free chunks are chained into a doubly-linked list; each free chunk contains pointers to the next and previous chunks in the "free list". The boundary tags and free list pointers make up `malloc`'s internal data structures. These data structures are interspersed with program data, so it is easy for a program error to damage them. - Space overhead: Boundary tags and free list pointers take up space. The minimum chunk size on most systems is 16 bytes. So for very small chunks, `malloc` is not space efficient. If your program requires large numbers of small structures, it might be more efficient to allocate them in arrays. - Fragmentation: If you allocate and free chunks with varied sizes, the heap will tend to become fragmented. That is, the free space might be broken into many small pieces. Fragmentation wastes space; it also slows the program down by making memory caches less effective. - Binning and caching: The free list is sorted by size into bins, so when `malloc` searches for a chunk with a particular size, it knows what bin to search in. If you free a chunk and then immediately allocate a chunk with the same size, `malloc` will usually be fast. # Caching ## How programs run In order to understand caching, you have to understand how computers execute programs. For a deep understanding of this topic, you should study computer architecture. My goal in this chapter is to provide a simple model of program execution. When a program starts, the code (or text) is usually on a hard disk or solid state drive. The operating system creates a new process to run the program, then the "loader" copies the text from storage into main memory and starts the program by calling `main`. While the program is running, most of its data is stored in main memory, but some of the data is in registers, which are small units of memory on the CPU. These registers include: - The program counter, or PC, which contains the address (in memory) of the next instruction in the program. - The instruction register, or IR, which contains the machine code instruction currently executing. - The stack pointer, or SP, which contains the address of the stack frame for the current function, which contains its parameters and local variables. - General-purpose registers that hold the data the program is currently working with. - A status register, or flag register, that contains information about the current computation. For example, the flag register usually contains a bit that is set if the result of the previous operation was zero. When a program is running, the CPU executes the following steps, called the "instruction cycle": - Fetch: The next instruction is fetched from memory and stored in the instruction register. - Decode: Part of the CPU, called the "control unit", decodes the instruction and sends signals to the other parts of the CPU. - Execute: Signals from the control unit cause the appropriate computation to occur. Most computers can execute a few hundred different instructions, called the "instruction set". But most instructions fall into a few general categories: - Load: Transfers a value from memory to a register. - Arithmetic/logic: Loads operands from registers, performs a mathematical operation, and stores the result in a register. - Store: Transfers a value from a register to memory. - Jump/branch: Changes the program counter, causing the flow of execution to jump to another location in the program. Branches are usually conditional, which means that they check a flag in the flag register and jump only if it is set. Some instructions sets, including the ubiquitous x86, provide instructions that combine a load and an arithmetic operation. During each instruction cycle, one instruction is read from the program text. In addition, about half of the instructions in a typical program load or store data. And therein lies one of the fundamental problems of computer architecture: the "memory bottleneck". In current computers, a typical core is capable of executing an instruction in less than 1 ns. But the time it takes to transfer data to and from memory is about 100 ns. If the CPU has to wait 100 ns to fetch the next instruction, and another 100 ns to load data, it would complete instructions 200 times slower than what's theoretically possible. For many computations, memory is the speed limiting factor, not the CPU. ## Cache performance The solution to this problem, or at least a partial solution, is caching. A "cache" is a small, fast memory that is physically close to the CPU, usually on the same chip. Actually, current computers typically have several levels of cache: the Level 1 cache, which is the smallest and fastest, might be 1--2 MiB with a access times near 1 ns; the Level 2 cache might have access times near 4 ns, and the Level 3 might take 16 ns. When the CPU loads a value from memory, it stores a copy in the cache. If the same value is loaded again, the CPU gets the cached copy and doesn't have to wait for memory. Eventually the cache gets full. Then, in order to bring something new in, we have to kick something out. So if the CPU loads a value and then loads it again much later, it might not be in cache any more. The performance of many programs is limited by the effectiveness of the cache. If the instructions and data needed by the CPU are usually in cache, the program can run close to the full speed of the CPU. If the CPU frequently needs data that are not in cache, the program is limited by the speed of memory. The cache "hit rate", $h$, is the fraction of memory accesses that find data in cache; the "miss rate", $m$, is the fraction of memory accesses that have to go to memory. If the time to process a cache hit is $T_h$ and the time for a cache miss is $T_m$, the average time for each memory access is $$h T_h + m T_m$$ Equivalently, we could define the "miss penalty" as the extra time to process a cache miss, $T_p = T_m - T_h$. Then the average access time is $$T_h + m T_p$$ When the miss rate is low, the average access time can be close to $T_h$. That is, the program can perform as if memory ran at cache speeds. ## Locality When a program reads a byte for the first time, the cache usually loads a "block" or "line" of data that includes the requested byte and some of its neighbors. If the program goes on to read one of the neighbors, it will already be in cache. As an example, suppose the block size is 64 B; you read a string with length 64, and the first byte of the string happens to fall at the beginning of a block. When you load the first byte, you incur a miss penalty, but after that the rest of the string will be in cache. After reading the whole string, the hit rate will be 63/64, about 98%. If the string spans two blocks, you would incur 2 miss penalties. But even then the hit rate would be 62/64, or almost 97%. If you then read the same string again, the hit rate would be 100%. On the other hand, if the program jumps around unpredictably, reading data from scattered locations in memory, and seldom accessing the same location twice, cache performance would be poor. The tendency of a program to use the same data more than once is called "temporal locality". The tendency to use data in nearby locations is called "spatial locality". Fortunately, many programs naturally display both kinds of locality: - Most programs contain blocks of code with no jumps or branches. Within these blocks, instructions run sequentially, so the access pattern has spatial locality. - In a loop, programs execute the same instructions many times, so the access pattern has temporal locality. - The result of one instruction is often used immediately as an operand of the next instruction, so the data access pattern has temporal locality. - When a program executes a function, its parameters and local variables are stored together on the stack; accessing these values has spatial locality. - One of the most common processing patterns is to read or write the elements of an array sequentially; this pattern also has spatial locality. The next section explores the relationship between a program's access pattern and cache performance. ## Measuring cache performance When I was a graduate student at U.C. Berkeley I was a teaching assistant for Computer Architecture with Brian Harvey. One of my favorite exercises involved a program that iterates through an array and measures the average time to read and write an element. By varying the size of the array, it is possible to infer the size of the cache, the block size, and some other attributes. My modified version of this program is in the `cache` directory of the repository for this book (see Section 1.1(#code)). The important part of the program is this loop: iters = 0; do { sec0 = get_seconds(); for (index = 0; index < limit; index += stride) arrayindex = arrayindex + 1; iters = iters + 1; sec = sec + (get_seconds() - sec0); } while (sec < 0.1); The inner `for` loop traverses the array. `limit` determines how much of the array it traverses; `stride` determines how many elements it skips over. For example, if `limit` is 16 and `stride` is 4, the loop would access elements 0, 4, 8, and 12. `sec` keeps track of the total CPU time used by the inner loop. The outer loop runs until `sec` exceeds 0.1 seconds, which is long enough that we can compute the average time with sufficient precision. `get_seconds` uses the system call `clock_gettime`, converts to seconds, and returns the result as a `double`: double get_seconds(){ struct timespec ts; clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts); return ts.tv_sec + ts.tv_nsec / 1e9; } Average miss penalty as a function of array size and stride. To isolate the time to access the elements of the array, the program runs a second loop that is almost identical except that the inner loop doesn't touch the array; it always increments the same variable: iters2 = 0; do { sec0 = get_seconds(); for (index = 0; index < limit; index += stride) temp = temp + index; iters2 = iters2 + 1; sec = sec - (get_seconds() - sec0); } while (iters2 < iters); The second loop runs the same number of iterations as the first. After each iteration, it *subtracts* the elapsed time from `sec`. When the loop completes, `sec` contains the total time for all array accesses, minus the total time it took to increment `temp`. This difference is the total miss penalty incurred by all accesses. Finally, we divide by the number of accesses to get the average miss penalty per access, in ns: sec * 1e9 / iters / limit * stride If you compile and run `cache.c` you should see output like this: Size: 4096 Stride: 8 read+write: 0.8633 ns Size: 4096 Stride: 16 read+write: 0.7023 ns Size: 4096 Stride: 32 read+write: 0.7105 ns Size: 4096 Stride: 64 read+write: 0.7058 ns If you have Python and `matplotlib` installed, you can use `graph_data.py` to graph the results. Figure 8.1(#cachedata) shows the results when I ran it on a Dell Optiplex 7010. Notice that the array size and stride are reported in bytes, not number of array elements. Take a minute to consider this graph, and see what you can infer about the cache. Here are some things to think about: - The program reads through the array many times, so it has plenty of temporal locality. If the entire array fits in cache, we expect the average miss penalty to be near 0. - When the stride is 4 bytes, we read every element of the array, so the program has plenty of spatial locality. If the block size is big enough to contain 64 elements, for example, the hit rate would be 63/64, even if the array does not fit in cache. - If the stride is equal to the block size (or greater), the spatial locality is effectively zero, because each time we read a block, we only access one element. In that case we expect to see the maximum miss penalty. In summary, we expect good cache performance if the array is smaller than the cache size *or* if the stride is smaller than the block size. Performance only degrades if the array is bigger than the cache *and* the stride is large. In Figure 8.1(#cachedata), cache performance is good, for all strides, as long as the array is less than $2^$ B. We can infer that the cache size is near 4 MiB; in fact, according to the specs, it is 3 MiB. When the stride is 8, 16, or 32 B, cache performance is good. At 64 B it starts to degrade, and for larger strides the average miss penalty is about 9 ns. We can infer that the block size near 128 B. Many processors use "multi-level caches" that include a small, fast cache and a bigger, slower cache. In this example, it looks like the miss penalty increases a little when the array size is bigger than $2^$ B, so it's possible that this processor also has a 16 KB cache with an access time less than 1 ns. ## Programming for cache performance Memory caching is implemented in hardware, so most of the time programmers don't need to know much about it. But if you know how caches work, you can write programs that use them more effectively. For example, if you are working with a large array, it might be faster to traverse the array once, performing several operations with each element, rather than traversing the array several times. If you are working with a 2-D array, it might be stored as an array of rows. If you traverse through the elements, it would be faster to go row-wise, with stride equal to the element size, rather than column-wise, with stride equal to the row length. Linked data structures don't always exhibit spatial locality, because the nodes aren't necessarily contiguous in memory. But if you allocate many nodes at the same time, they are usually co-located in the heap. Or, even better, if you allocate an array of nodes all at once, you know they will be contiguous. Recursive strategies like mergesort often have good cache behavior because they break big arrays into smaller pieces and then work with the pieces. Sometimes these algorithms can be tuned to take advantage of cache behavior. For applications where performance is critical, it is possible to design algorithms tailored to the size of the cache, the block size, and other hardware characterstics. Algorithms like that are called "cache-aware". The obvious drawback of cache-aware algorithms is that they are hardware-specific. ## The memory hierarchy At some point during this chapter, a question like the following might have occurred to you: "If caches are so much faster than main memory, why not make a really big cache and forget about memory?" Without going too far into computer architecture, there are two reasons: electronics and economics. Caches are fast because they are small and close to the CPU, which minimizes delays due to capacitance and signal propagation. If you make a cache big, it will be slower. Also, caches take up space on the processor chip, and bigger chips are more expensive. Main memory is usually dynamic random-access memory (DRAM), which uses only one transistor and one capacitor per bit, so it is possible to pack more memory into the same amount of space. But this way of implementing memory is slower than the way caches are implemented. Also main memory is usually packaged in a dual in-line memory module (DIMM) that includes 16 or more chips. Several small chips are cheaper than one big one. The trade-off between speed, size, and cost is the fundamental reason for caching. If there were one memory technology that was fast, big, and cheap, we wouldn't need anything else. The same principle applies to storage as well as memory. Solid state drives (SSD) are fast, but they are more expensive than hard drives (HDD), so they tend to be smaller. Tape drives are even slower than hard drives, but they can store large amounts of data relatively cheaply. The following table shows typical access times, sizes, and costs for each of these technologies. | | | | | |:---------|:-----------|:---------|:-------------| | Device | Access | Typical | Cost | | | time | size | | | Register | 0.5 ns | 256 B | ? | | Cache | 1 ns | 2 MiB | ? | | DRAM | 100 ns | 4 GiB | \$10 / GiB | | SSD | 10 $\mu$s | 100 GiB | \$1 / GiB | | HDD | 5 ms | 500 GiB | \$0.25 / GiB | | Tape | minutes | 1--2 TiB | \$0.02 / GiB | ::: The number and size of registers depends on details of the architecture. Current computers have about 32 general-purpose registers, each storing one "word". On a 32-bit computer, a word is 32 bits or 4 B. On a 64-bit computer, a word is 64 bits or 8 B. So the total size of the register file is 100--300 B. The cost of registers and caches is hard to quantify. They contribute to the cost of the chips they are on, but consumers don't see that cost directly. For the other numbers in the table, I looked at the specifications for typical hardware for sale from online computer hardware stores. By the time you read this, these numbers will be obsolete, but they give you an idea of what the performance and cost gaps looked like at one point in time. These technologies make up the "memory hierarchy" (note that this use of "memory" also includes storage). Each level of the hierarchy is bigger and slower than the one above it. And in some sense, each level acts as a cache for the one below it. You can think of main memory as a cache for programs and data that are stored permanently on SSDs and HHDs. And if you are working with very large datasets stored on tape, you could use hard drives to cache one subset of the data at a time. ## Caching policy The memory hierarchy suggests a framework for thinking about caching. At every level of the hierarchy, we have to address four fundamental questions of caching: - Who moves data up and down the hierarchy? At the top of the hierarchy, register allocation is usually done by the compiler. Hardware on the CPU handles the memory cache. Users implicitly move data from storage to memory when they execute programs and open files. But the operating system also moves data back and forth between memory and storage. At the bottom of the hierarchy, administrators move data explicitly between disk and tape. - What gets moved? In general, block sizes are small at the top of the hierarchy and bigger at the bottom. In a memory cache, a typical block size is 128 B. Pages in memory might be 4 KiB, but when the operating system reads a file from disk, it might read 10s or 100s of blocks at a time. - When does data get moved? In the most basic cache, data gets moved into cache when it is used for the first time. But many caches use some kind of "prefetching", meaning that data is loaded before it is explicitly requested. We have already seen one form of prefetching: loading an entire block when only part of it is requested. - Where in the cache does the data go? When the cache is full, we can't bring anything in without kicking something out. Ideally, we want to keep data that will be used again soon and replace data that won't. The answers to these questions make up the "cache policy". Near the top of the hierarchy, cache policies tend to be simple because they have to be fast and they are implemented in hardware. Near the bottom of the hierarchy, there is more time to make decisions, and well-designed policies can make a big difference. Most cache policies are based on the principle that history repeats itself; if we have information about the recent past, we can use it to predict the immediate future. For example, if a block of data has been used recently, we expect it to be used again soon. This principle suggests a replacement policy called "least recently used," or LRU, which removes from the cache a block of data that has not been used recently. For more on this topic, see . ## Paging In systems with virtual memory, the operating system can move pages back and forth between memory and storage. As I mentioned in Section 7.2(#leak), this mechanism is called "paging" or sometimes "swapping". Here's how the process works: 1. Suppose Process A calls `malloc` to allocate a chunk. If there is no free space in the heap with the requested size, `malloc` calls `sbrk` to ask the operating system for more memory. 2. If there is a free page in physical memory, the operating system adds it to the page table for Process A, creating a new range of valid virtual addresses. 3. If there are no free pages, the paging system chooses a "victim page" belonging to Process B. It copies the contents of the victim page from memory to disk, then it modifies the page table for Process B to indicate that this page is "swapped out". 4. Once the data from Process B is written, the page can be reallocated to Process A. To prevent Process A from reading Process B's data, the page should be cleared. 5. At this point the call to `sbrk` can return, giving `malloc` additional space in the heap. Then `malloc` allocates the requested chunk and returns. Process A can resume. 6. When Process A completes, or is interrupted, the scheduler might allow Process B to resume. When Process B accesses a page that has been swapped out, the memory management unit notices that the page is "invalid" and causes an interrupt. 7. When the operating system handles the interrupt, it sees that the page is swapped out, so it transfers the page back from disk to memory. 8. Once the page is swapped in, Process B can resume. When paging works well, it can greatly improve the utilization of physical memory, allowing more processes to run in less space. Here's why: - Most processes don't use all of their allocated memory. Many parts of the text segment are never executed, or execute once and never again. Those pages can be swapped out without causing any problems. - If a program leaks memory, it might leave allocated space behind and never access it again. By swapping those pages out, the operating system can effectively plug the leak. - On most systems, there are processes like daemons that sit idle most of the time and only occasionally "wake up" to respond to events. While they are idle, these processes can be swapped out. - A user might have many windows open, but only a few are active at a time. The inactive processes can be swapped out. - Also, there might be many processes running the same program. These processes can share the same text and static segments, avoiding the need to keep multiple copies in physical memory. If you add up the total memory allocated to all processes, it can greatly exceed the size of physical memory, and yet the system can still behave well. Up to a point. When a process accesses a page that's swapped out, it has to get the data back from disk, which can take several milliseconds. The delay is often noticeable. If you leave a window idle for a long time and then switch back to it, it might start slowly, and you might hear the disk drive working while pages are swapped in. Occasional delays like that might be acceptable, but if you have too many processes using too much space, they start to interfere with each other. When Process A runs, it evicts the pages Process B needs. Then when B runs, it evicts the pages A needs. When this happens, both processes slow to a crawl and the system can become unresponsive. This scenario is called "thrashing". In theory, operating systems could avoid thrashing by detecting an increase in paging and blocking or killing processes until the system is responsive again. But as far as I can tell, most systems don't do this, or don't do it well; it is often left to users to limit their use of physical memory or try to recover when thrashing occurs. # Multitasking In many current systems, the CPU contains multiple cores, which means it can run several processes at the same time. In addition, each core is capable of "multitasking", which means it can switch from one process to another quickly, creating the illusion that many processes are running at the same time. The part of the operating system that implements multitasking is the "kernel". In a nut or seed, the kernel is the innermost part, surrounded by a shell. In an operating system, the kernel is the lowest level of software, surrounded by several other layers, including an interface called a "shell." Computer scientists love extended metaphors. At its most basic, the kernel's job is to handle interrupts. An "interrupt" is an event that stops the normal instruction cycle and causes the flow of execution to jump to a special section of code called an "interrupt handler". A **hardware interrupt** is caused when a device sends a signal to the CPU. For example, a network interface might cause an interrupt when a packet of data arrives, or a disk drive might cause an interrupt when a data transfer is complete. Most systems also have timers that cause interrupts at regular intervals, or after an elapsed time. A **software interrupt** is caused by a running program. For example, if an instruction cannot complete for some reason, it might trigger an interrupt so the condition can be handled by the operating system. Some floating-point errors, like division by zero, are handled using interrupts. When a program needs to access a hardware device, it makes a **system call**, which is similar to a function call, except that instead of jumping to the beginning of the function, it executes a special instruction that triggers an interrupt, causing the flow of execution to jump to the kernel. The kernel reads the parameters of the system call, performs the requested operation, and then resumes the interrupted process. ## Hardware state Handling interrupts requires cooperation between hardware and software. When an interrupt occurs, there might be several instructions running on the CPU, data stored in registers, and other **hardware state**. Usually the hardware is responsible for bringing the CPU to a consistent state; for example, every instruction should either complete or behave as if it never started. No instruction should be left half complete. Also, the hardware is responsible for saving the program counter (PC), so the kernel knows where to resume. Then, usually, it is the responsibility of the interrupt handler to save the rest of the hardware state before it does anything that might modify it, and then restore the saved state before the interrupted process resumes. Here is an outline of this sequence of events: 1. When the interrupt occurs, the hardware saves the program counter in a special register and jumps to the appropriate interrupt handler. 2. The interrupt handler stores the program counter and the status register in memory, along with the contents of any data registers it plans to use. 3. The interrupt handler runs whatever code is needed to handle the interrupt. 4. Then it restores the contents of the saved registers. Finally, it restores the program counter of the interrupted process, which has the effect of jumping back to the interrupted instruction. If this mechanism works correctly, there is generally no way for the interrupted process to know there was an interrupt, unless it detects the change in time between instructions. ## Context switching Interrupt handlers can be fast because they don't have to save the entire hardware state; they only have to save registers they are planning to use. But when an interrupt occurs, the kernel does not always resume the interrupted process. It has the option of switching to another process. This mechanism is called a "context switch". In general, the kernel doesn't know which registers a process will use, so it has to save all of them. Also, when it switches to a new process, it might have to clear data stored in the memory management unit (see Section 4.6(#address_translation)). And after the context switch, it might take some time for the new process to load data into the cache. For these reasons, context switches are relatively slow, on the order of thousands of cycles, or a few microseconds. In a multi-tasking system, each process is allowed to run for a short period of time called a "time slice" or "quantum". During a context switch, the kernel sets a hardware timer that causes an interrupt at the end of the time slice. When the interrupt occurs, the kernel can switch to another process or allow the interrupted process to resume. The part of the operating system that makes this decision is the "scheduler". ## The process life cycle When a process is created, the operating system allocates a data structure that contains information about the process, called a "process control block" or PCB. Among other things, the PCB keeps track of the process state, which is one of: - Running, if the process is currently running on a core. - Ready, if the process could be running, but isn't, usually because there are more runnable processes than cores. - Blocked, if the process cannot run because it is waiting for a future event like network communication or a disk read. - Done, if the process has completed, but has exit status information that has not been read yet. Here are the events that cause a process to transition from one state to another: - A process is created when the running program executes a system call like `fork`. At the end of the system call, the new process is usually ready. Then the scheduler might resume the original process (the "parent") or start the new process (the "child"). - When a process is started or resumed by the scheduler, its state changes from ready to running. - When a process is interrupted and the scheduler chooses not to let it resume, its state changes from running to ready. - If a process executes a system call that cannot complete immediately, like a disk request, it becomes blocked and the scheduler usually chooses another process. - When an operation like a disk request completes, it causes an interrupt. The interrupt handler figures out which process was waiting for the request and switches its state from blocked to ready. Then the scheduler may or may not choose to resume the unblocked process. - When a process calls `exit`, the interrupt handler stores the exit code in the PCB and changes the process's state to done. ## Scheduling As we saw in Section 3.3(#unixps) there might be hundreds of processes on a computer, but usually most of them are blocked. Most of the time, there are only a few processes that are ready or running. When an interrupt occurs, the scheduler decides which process to start or resume. On a workstation or laptop, the primary goal of the scheduler is to minimize response time; that is, the computer should respond quickly to user actions. Response time is also important on a server, but in addition the scheduler might try to maximize throughput, which is the number of requests that complete per unit of time. Usually the scheduler doesn't have much information about what processes are doing, so its decisions are based on a few heuristics: - Processes might be limited by different resources. A process that does a lot of computation is probably CPU-bound, which means that its run time depends on how much CPU time it gets. A process that reads data from a network or disk might be I/O-bound, which means that it would run faster if data input and output went faster, but would not run faster with more CPU time. Finally, a process that interacts with the user is probably blocked, most of the time, waiting for user actions. The operating system can sometimes classify processes based on their past behavior, and schedule them accordingly. For example, when an interactive process is unblocked, it should probably run immediately, because a user is probably waiting for a reply. On the other hand, a CPU-bound process that has been running for a long time might be less time-sensitive. - If a process is likely to run for a short time and then make a blocking request, it should probably run immediately, for two reasons: (1) if the request takes some time to complete, we should start it as soon as possible, and (2) it is better for a long-running process to wait for a short one, rather than the other way around. As an analogy, suppose you are making an apple pie. The crust takes 5 minutes to prepare, but then it has to chill for half an hour. It takes 20 minutes to prepare the filling. If you prepare the crust first, you can prepare the filling while the crust is chilling, and you can finish the pie in 35 minutes. If you prepare the filling first, the process takes 55 minutes. Most schedulers use some form of priority-based scheduling, where each process has a priority that can be adjusted up or down over time. When the scheduler runs, it chooses the runnable process with the highest priority. Here are some of the factors that determine a process's priority: - A process usually starts with a relatively high priority so it starts running quickly. - If a process makes a request and blocks before its time slice is complete, it is more likely to be interactive or I/O-bound, so its priority should go up. - If a process runs for an entire time slice, it is more likely to be long-running and CPU-bound, so its priority should go down. - If a task blocks for a long time and then becomes ready, it should get a priority boost so it can respond to whatever it was waiting for. - If process A is blocked waiting for process B, for example if they are connected by a pipe, the priority of process B should go up. - The system call `nice` allows a process to decrease (but not increase) its own priority, allowing programmers to pass explicit information to the scheduler. For most systems running normal workloads, scheduling algorithms don't have a substantial effect on performance. Simple scheduling policies are usually good enough. ## Real-time scheduling However, for programs that interact with the real world, scheduling can be very important. For example, a program that reads data from sensors and controls motors might have to complete recurring tasks at some minimum frequency and react to external events with some maximum response time. These requirements are often expressed in terms of "tasks" that must be completed before "deadlines". Scheduling tasks to meet deadlines is called "real-time scheduling". For some applications, a general-purpose operating system like Linux can be modified to handle real-time scheduling. These modifications might include: - Providing richer APIs for controlling task priorities. - Modifying the scheduler to guarantee that the process with highest priority runs within a fixed amount of time. - Reorganizing interrupt handlers to guarantee a maximum completion time. - Modifying locks and other synchronization mechanisms (coming up in the next chapter) to allow a high-priority task to preempt a lower-priority task. - Choosing an implementation of dynamic memory allocation that guarantees a maximum completion time. For more demanding applications, especially in domains where real-time response is a matter of life and death, "real-time operating systems" provide specialized capabilities, often with much simpler designs than general purpose operating systems. # Threads When I mentioned threads in Section 3.3(#unixps), I said that a thread is a kind of process. Now I will provide a more careful explanation. When you create a process, the operating system creates a new address space, which includes the text segment, static segment, and heap; it also creates a new "thread of execution", which includes the program counter and other hardware state, and the call stack. The processes we have seen so far are "single-threaded", which means that only one thread of execution runs in each address space. In this chapter, you will learn about "multi-threaded" processes that have multiple threads running in the same address space. Within a single process, all threads share the same text segment, so they run the same code. But different threads often run different parts of the code. And they share the same static segment, so if one thread changes a global variable, other threads see the change. They also share the heap, so threads can share dynamically-allocated chunks. But each thread has its own stack, so threads can call functions without interfering with each other. Usually threads don't access each other's local variables (and sometimes they can't). The example code for this chapter is in the repository for this book, in a directory named `counter`. For information on downloading this code, see Section 1.1(#code). ## Creating threads The most popular threading standard used with C is POSIX Threads, or Pthreads for short. The POSIX standard defines a thread model and an interface for creating and controlling threads. Most versions of UNIX provide an implementation of Pthreads. Using Pthreads is like using most C libraries: - You include headers files at the beginning of your program. - You write code that calls functions defined by Pthreads. - When you compile the program, you link it with the Pthread library. For my examples, I include the following headers: #include #include #include #include The first two are standard; the third is for Pthreads and the fourth is for semaphores. To compile with the Pthread library in `gcc`, you can use the `-l` option on the command line: gcc -g -O2 -o array array.c -lpthread This compiles a source file named `array.c` with debugging info and optimization, links with the Pthread library, and generates an executable named `array`. ## Creating threads The Pthread function that creates threads is called `pthread_create`. The following function shows how to use it: pthread_t make_thread(void *(*entry)(void *), Shared *shared) { int n; pthread_t thread; n = pthread_create(&thread, NULL, entry, (void *)shared); if (n != 0) { perror("pthread_create failed"); exit(-1); } return thread; } `make_thread` is a wrapper I wrote to make `pthread_create` easier to use, and to provide error-checking. The return type from `pthread_create` is `pthread_t`, which you can think of as an id or "handle" for the new thread. If `pthread_create` succeeds, it returns 0 and `make_thread` returns the handle of the new thread. If an error occurs, `pthread_create` returns an error code and `make_thread` prints an error message and exits. The parameters of `make_thread` take some explaining. Starting with the second, `Shared` is a structure I defined to contain values shared between threads. The following `typedef` statement creates the new type: typedef struct { int counter; } Shared; In this case, the only shared variable is `counter`. `make_shared` allocates space for a `Shared` structure and initializes the contents: Shared *make_shared() { Shared *shared = check_malloc(sizeof (Shared)); shared->counter = 0; return shared; } Now that we have a shared data structure, let's get back to `make_thread`. The first parameter is a pointer to a function that takes a `void` pointer and returns a `void` pointer. If the syntax for declaring this type makes your eyes bleed, you are not alone. Anyway, the purpose of this parameter is to specify the function where the execution of the new thread will begin. By convention, this function is named `entry`: void *entry(void *arg) { Shared *shared = (Shared *) arg; child_code(shared); pthread_exit(NULL); } The parameter of `entry` has to be declared as a `void` pointer, but in this program we know that it is really a pointer to a `Shared` structure, so we can typecast it accordingly and then pass it along to `child_code`, which does the real work. As a simple example, `child_code` prints the value of the shared counter and increments it. void child_code(Shared *shared) { printf("counter = %d\n", shared->counter); shared->counter++; } When `child_code` returns, `entry` invokes `pthread_exit` which can be used to pass a value to the thread that joins with this thread. In this case, the child has nothing to say, so we pass `NULL`. Finally, here is the code that creates the child threads: int i; pthread_t childNUM_CHILDREN; Shared *shared = make_shared(1000000); for (i=0; imutex); printf("counter = %d\n", shared->counter); shared->counter++; mutex_unlock(shared->mutex); } Before any thread can access `counter`, it has to "lock" the mutex, which has the effect of barring all other threads. Suppose Thread A has locked the mutex and is in the middle of `child_code`. If Thread B arrives and executes `mutex_lock`, it blocks. When Thread A is done, it executes `mutex_unlock`, which allows Thread B to proceed. In effect, the threads line up to execute `child_code` one at a time, so they can't interfere with each other. When I run this code with 5 children, I get: counter = 0 counter = 1 counter = 2 counter = 3 counter = 4 And that satisfies the requirements. In order for this solution to work, I have to add the Mutex to the Shared struct: typedef struct { int counter; Mutex *mutex; } Shared; And initialize it in `make_shared` Shared *make_shared(int end) { Shared *shared = check_malloc(sizeof(Shared)); shared->counter = 0; shared->mutex = make_mutex(); //-- this line is new return shared; } The code in this section is in `counter_mutex.c`. The definition of `Mutex` is in `mutex.c`, which I explain in the next section. ## Mutex My definition of `Mutex` is a wrapper for a type called `pthread_mutex_t`, which is defined in the POSIX threads API. To create a POSIX mutex, you have to allocate space for a `pthread_mutex_t` type and then call `pthread_mutex_init`. One of the problems with this API is that `pthread_mutex_t` behaves like a structure, so if you pass it as an argument, it makes a copy, which makes the mutex behave incorrectly. To avoid that, you have to pass `pthread_mutex_t` by address. My code makes it easier to get that right. It defines a type, `Mutex`, which is just a more readable name for `pthread_mutex_t`: #include typedef pthread_mutex_t Mutex; Then it defines `make_mutex`, which allocates space and initializes the mutex: Mutex *make_mutex() { Mutex *mutex = check_malloc(sizeof(Mutex)); int n = pthread_mutex_init(mutex, NULL); if (n != 0) perror_exit("make_lock failed"); return mutex; } The return value is a pointer, which you can pass around as an argument without causing unwanted copying. The functions to lock and unlock the mutex are simple wrappers for POSIX functions: void mutex_lock(Mutex *mutex) { int n = pthread_mutex_lock(mutex); if (n != 0) perror_exit("lock failed"); } void mutex_unlock(Mutex *mutex) { int n = pthread_mutex_unlock(mutex); if (n != 0) perror_exit("unlock failed"); } This code is in `mutex.c` and the header file `mutex.h`. # Condition variables Many simple synchronization problems can be solved using mutexes as shown in the previous chapter. In this chapter I introduce a bigger challenge, the well-known "Producer-Consumer problem", and a new tool to solve it, the condition variable. ## The work queue In some multi-threaded programs, threads are organized to perform different tasks. Often they communicate with each other using a queue, where some threads, called "producers", put data into the queue and other threads, called "consumers", take data out. For example, in applications with a graphical user interface, there might be one thread that runs the GUI, responding to user events, and another thread that processes user requests. In that case, the GUI thread might put requests into a queue and the "back end" thread might take requests out and process them. To support this organization, we need a queue implementation that is "thread safe", which means that both threads (or more than two) can access the queue at the same time. And we need to handle the special cases when the queue is empty and, if the size of the queue is bounded, when the queue is full. I'll start with a simple queue that is not thread safe, then we'll see what goes wrong and fix it. The code for this example is in the repository for this book, in a folder called `queue`. The file `queue.c` contains a basic implementation of a circular buffer, which you can read about at . Here's the structure definition: typedef struct { int *array; int length; int next_in; int next_out; } Queue; `array` is the array that contains the elements of the queue. For this example the elements are ints, but more generally they would be structures that contain user events, items of work, etc. `length` is the length of the array. `next_in` is an index into the array that indices where the next element should be added; similarly, `next_out` is the index of the next element that should be removed. `make_queue` allocates space for this structure and initializes the fields: Queue *make_queue(int length) { Queue *queue = (Queue *) malloc(sizeof(Queue)); queue->length = length + 1; queue->array = (int *) malloc(length * sizeof(int)); queue->next_in = 0; queue->next_out = 0; return queue; } The initial value for `next_out` needs some explaining. Since the queue is initially empty, there is no next element to remove, so `next_out` is invalid. Setting `next_out == next_in` is a special case that indicates that the queue is empty, so we can write: int queue_empty(Queue *queue) { return (queue->next_in == queue->next_out); } Now we can add elements to the queue using `queue_push`: void queue_push(Queue *queue, int item) { if (queue_full(queue)) { perror_exit("queue is full"); } queue->arrayqueue->next_in = item; queue->next_in = queue_incr(queue, queue->next_in); } If the queue is full, `queue_push` prints an error message and exits. I will explain `queue_full` soon. If the queue is not full, `queue_push` inserts the new element and then increments `next_in` using `queue_incr`: int queue_incr(Queue *queue, int i) { return (i+1) % queue->length; } When the index, `i`, gets to the end of the array, it wraps around to 0. And that's where we run into a tricky part. If we keep adding elements to the queue, eventually `next_in` wraps around and catches up with `next_out`. But if `next_in == next_out`, we would incorrectly conclude that the queue was empty. To avoid that, we define another special case to indicate that the queue is full: int queue_full(Queue *queue) { return (queue_incr(queue, queue->next_in) == queue->next_out); } If incrementing `next_in` lands on `next_out`, that means we can't add another element without making the queue seem empty. So we stop one element before the "end" (keeping in mind that the end of the queue can be anywhere, not necessarily the end of the array). Now we can write `queue_pop`, which removes and returns the next element from the queue: int queue_pop(Queue *queue) { if (queue_empty(queue)) { perror_exit("queue is empty"); } int item = queue->arrayqueue->next_out; queue->next_out = queue_incr(queue, queue->next_out); return item; } If you try to pop from an empty queue, `queue_pop` prints an error message and exits. ## Producers and consumers Now let's make some threads to access this queue. Here's the producer code: void *producer_entry(void *arg) { Shared *shared = (Shared *) arg; for (int i=0; iqueue, i); } pthread_exit(NULL); } Here's the consumer code: void *consumer_entry(void *arg) { int item; Shared *shared = (Shared *) arg; for (int i=0; iqueue); printf("consuming item %d\n", item); } pthread_exit(NULL); } Here's the parent code that starts the threads and waits for them pthread_t childNUM_CHILDREN; Shared *shared = make_shared(); child0 = make_thread(producer_entry, shared); child1 = make_thread(consumer_entry, shared); for (int i=0; iqueue = make_queue(QUEUE_LENGTH); return shared; } The code we have so far is a good starting place, but it has several problems: - Access to the queue is not thread safe. Different threads could access `array`, `next_in`, and `next_out` at the same time and leave the queue in a broken, "inconsistent" state. - If the consumer is scheduled first, it finds the queue empty, print an error message, and exits. We would rather have the consumer block until the queue is not empty. Similarly, we would like the producer to block if the queue is full. In the next section, we solve the first problem with a `Mutex`. In the following section, we solve the second problem with condition variables. ## Mutual exclusion We can make the queue thread safe with a mutex. This version of the code is in `queue_mutex.c`. First we add a `Mutex` pointer to the queue structure: typedef struct { int *array; int length; int next_in; int next_out; Mutex *mutex; //-- this line is new } Queue; And initialize the `Mutex` in `make_queue`: Queue *make_queue(int length) { Queue *queue = (Queue *) malloc(sizeof(Queue)); queue->length = length; queue->array = (int *) malloc(length * sizeof(int)); queue->next_in = 0; queue->next_out = 0; queue->mutex = make_mutex(); //-- new return queue; } Next we add synchronization code to `queue_push`: void queue_push(Queue *queue, int item) { mutex_lock(queue->mutex); //-- new if (queue_full(queue)) { mutex_unlock(queue->mutex); //-- new perror_exit("queue is full"); } queue->arrayqueue->next_in = item; queue->next_in = queue_incr(queue, queue->next_in); mutex_unlock(queue->mutex); //-- new } Before checking whether the queue is full, we have to lock the `Mutex`. If the queue is full, we have to unlock the `Mutex` before exiting; otherwise the thread would leave it locked and no other threads could proceed. The synchronization code for `queue_pop` is similar: int queue_pop(Queue *queue) { mutex_lock(queue->mutex); if (queue_empty(queue)) { mutex_unlock(queue->mutex); perror_exit("queue is empty"); } int item = queue->arrayqueue->next_out; queue->next_out = queue_incr(queue, queue->next_out); mutex_unlock(queue->mutex); return item; } Note that the other `Queue` functions, `queue_full`, `queue_empty`, and `queue_incr` do not try to lock the mutex. Any thread that calls these functions is required to lock the mutex first; this requirement is part of the documented interface for these functions. With this additional code, the queue is thread safe; if you run it, you should not see any synchronization errors. But it is likely that the consumer will exit at some point because the queue is empty, or the producer will exit because the queue is full, or both. The next step is to add condition variables. ## Condition variables A condition variable is a data structure associated with a condition; it allows threads to block until the condition becomes true. For example, `thread_pop` might want check whether the queue is empty and, if so, wait for a condition like "queue not empty". Similarly, `thread_push` might want to check whether the queue is full and, if so, block until it is not full. I'll handle the first condition here, and you will have a chance to handle the second condition as an exercise. First we add a condition variable to the `Queue` structure: typedef struct { int *array; int length; int next_in; int next_out; Mutex *mutex; Cond *nonempty; //-- new } Queue; And initialize it in `make_queue`: Queue *make_queue(int length) { Queue *queue = (Queue *) malloc(sizeof(Queue)); queue->length = length; queue->array = (int *) malloc(length * sizeof(int)); queue->next_in = 0; queue->next_out = 0; queue->mutex = make_mutex(); queue->nonempty = make_cond(); //-- new return queue; } Now in `queue_pop`, if we find the queue empty, we don't exit; instead we use the condition variable to block: int queue_pop(Queue *queue) { mutex_lock(queue->mutex); while (queue_empty(queue)) { cond_wait(queue->nonempty, queue->mutex); //-- new } int item = queue->arrayqueue->next_out; queue->next_out = queue_incr(queue, queue->next_out); mutex_unlock(queue->mutex); cond_signal(queue->nonfull); //-- new return item; } `cond_wait` is complicated, so let's take it slow. The first argument is the condition variable; in this case, the condition we are waiting for is "queue not empty". The second argument is the mutex that protects the queue. When the thread that locked the mutex calls `cond_wait`, it unlocks the mutex and then blocks. This is important. If `cond_wait` did not unlock the mutex before blocking, no other thread would be able to access the queue, no more items could be added, and the queue would always be empty. So while the consumer is blocked on `nonempty`, the producer can run. Let's see what happens when the producer runs `queue_push`: void queue_push(Queue *queue, int item) { mutex_lock(queue->mutex); if (queue_full(queue)) { mutex_unlock(queue->mutex); perror_exit("queue is full"); } queue->arrayqueue->next_in = item; queue->next_in = queue_incr(queue, queue->next_in); mutex_unlock(queue->mutex); cond_signal(queue->nonempty); //-- new } Just as before, `queue_push` locks the `Mutex` and checks whether the queue is full. Assuming it is not, `queue_push` adds a new element to the queue and then unlocks the `Mutex`. But before returning, it does one more thing: it "signals" the condition variable `nonempty`. Signalling a condition variable usually indicates that the condition is true. If there are no threads waiting on the condition variable, the signal has no effect. If there are threads waiting on the condition variable, one of them gets unblocked and resumes execution of `cond_wait`. But before the awakened thread can return from `cond_wait`, it has to wait for and lock the `Mutex`, again. Now go back to `queue_pop` and see what happens when the thread returns from `cond_wait`. It loops back to the top of the while loop and checks the condition again. I'll explain why in just a second, but for now let's assume that the condition is true; that is, the queue is not empty. When the consumer thread exits the while loop, we know two things: (1) the condition is true, so there is at least one item in the queue, and (2) the `Mutex` is locked, so it is safe to access the queue. After removing an item, `queue_pop` unlocks the mutex and returns. In the next section I'll show you how my `Cond` code works, but first I want to answer two frequently-asked questions: - Why is `cond_wait` inside a while loop rather than an if statement; that is, why do we have to check the condition again after returning from `cond_wait`? The primary reason you have to re-check the condition is the possibility of an intercepted signal. Suppose Thread A is waiting on `nonempty`. Thread B adds an item to the queue and signals `nonempty`. Thread A wakes up an tries to lock the mutex, but before it gets the chance, Evil Thread C swoops in, locks the mutex, pops the item from the queue, and unlocks the mutex. Now the queue is empty again, but Thread A is not blocked any more. Thread A could lock the mutex and returns from `cond_wait`. If Thread A does not check the condition again, it would try to pop an element from an empty queue, and probably cause an error. - The other question that comes up when people learn about condition variables is "How does the condition variable know what condition it is associated with?" This question is understandable because there is no explicit connection between a `Cond` structure and the condition it relates to. The connection is implicit in the way it is used. Here's one way to think of it: the condition associated with a Cond is the thing that is false when you call `cond_wait` and true when you call `cond_signal`. Because threads have to check the condition when they return from `cond_wait`, it is not strictly necessary to call `cond_signal` only when the condition is true. If you have reason to think the condition *might* be true, you could call `cond_signal` as a suggestion that now is a good time to check. ## Condition variable implementation The Cond structure I used in the previous section is a wrapper for a type called `pthread_cond_t`, which is defined in the POSIX threads API. It is very similar to Mutex, which is a wrapper for `pthread_mutex_t`. Both wrappers are defined in `utils.c` and `utils.h`. Here's the typedef: typedef pthread_cond_t Cond; `make_cond` allocates space, initializes the condition variable, and returns a pointer: Cond *make_cond() { Cond *cond = check_malloc(sizeof(Cond)); int n = pthread_cond_init(cond, NULL); if (n != 0) perror_exit("make_cond failed"); return cond; } And here are the wrappers for `cond_wait` and `cond_signal`. void cond_wait(Cond *cond, Mutex *mutex) { int n = pthread_cond_wait(cond, mutex); if (n != 0) perror_exit("cond_wait failed"); } void cond_signal(Cond *cond) { int n = pthread_cond_signal(cond); if (n != 0) perror_exit("cond_signal failed"); } At this point there should be nothing too surprising there. # Semaphores in C Semaphores are a good way to learn about synchronization, but they are not as widely used, in practice, as mutexes and condition variables. Nevertheless, there are some synchronization problems that can be solved simply with semaphores, yielding solutions that are more demonstrably correct. This chapter presents a C API for working with semaphores and my code for making it easier to work with. And it presents a final challenge: can you write an implementation of a semaphore using mutexes and condition variables? The code for this chapter is in directory `semaphore` in the repository for this book (see Section 1.1(#code)). ## POSIX Semaphores A semaphore is a data structure used to help threads work together without interfering with each other. The POSIX standard specifies an interface for semaphores; it is not part of Pthreads, but most UNIXes that implement Pthreads also provide semaphores. POSIX semaphores have type `sem_t`. As usual, I put a wrapper around `sem_t` to make it easier to use. The interface is defined in `sem.h`: typedef sem_t Semaphore; Semaphore *make_semaphore(int value); void semaphore_wait(Semaphore *sem); void semaphore_signal(Semaphore *sem); `Semaphore` is a synonym for `sem_t`, but I find it more readable, and the capital letter reminds me to treat it like an object and pass it by pointer. The implementation of these functions is in `sem.c`: Semaphore *make_semaphore(int value) { Semaphore *sem = check_malloc(sizeof(Semaphore)); int n = sem_init(sem, 0, value); if (n != 0) perror_exit("sem_init failed"); return sem; } `make_semaphore` takes the initial value of the semaphore as a parameter. It allocates space for a Semaphore, initializes it, and returns a pointer to `Semaphore`. `sem_init` returns 0 if it succeeds and -1 if anything goes wrong. One nice thing about using wrapper functions is that you can encapsulate the error-checking code, which makes the code that uses these functions more readable. Here is the implementation of `semaphore_wait`: void semaphore_wait(Semaphore *sem) { int n = sem_wait(sem); if (n != 0) perror_exit("sem_wait failed"); } And here is `semaphore_signal`: void semaphore_signal(Semaphore *sem) { int n = sem_post(sem); if (n != 0) perror_exit("sem_post failed"); } I prefer to call this operation "signal" rather than "post", although both terms are common. Here's an example that shows how to use a semaphore as a mutex: Semaphore *mutex = make_semaphore(1); semaphore_wait(mutex); // protected code goes here semaphore_signal(mutex); When you use a semaphore as a mutex, you usually initialize it to 1 to indicate that the mutex is unlocked; that is, one thread can pass the semaphore without blocking. Here I am using the variable name `mutex` to indicate that the semaphore is being used as a mutex. But remember that the behavior of a semaphore is not the same as a Pthread mutex. ## Producers and consumers with semaphores Using these semaphore wrapper functions, we can write a solution to the Producer-Consumer problem from Section 11.2(#prodcon). The code in this section is in `queue_sem.c`. Here's the new definition of `Queue`, replacing the mutex and condition variables with semaphores: typedef struct { int *array; int length; int next_in; int next_out; Semaphore *mutex; //-- new Semaphore *items; //-- new Semaphore *spaces; //-- new } Queue; And here's the new version of `make_queue`: Queue *make_queue(int length) { Queue *queue = (Queue *) malloc(sizeof(Queue)); queue->length = length; queue->array = (int *) malloc(length * sizeof(int)); queue->next_in = 0; queue->next_out = 0; queue->mutex = make_semaphore(1); queue->items = make_semaphore(0); queue->spaces = make_semaphore(length-1); return queue; } `mutex` is used to guarantee exclusive access to the queue; the initial value is 1, so the mutex is initially unlocked. `items` is the number of items in the queue, which is also the number of consumer threads that can execute `queue_pop` without blocking. Initially there are no items in the queue. `spaces` is the number of empty spaces in the queue, which is the number of producer threads that can execute `queue_push` without blocking. Initially the number of spaces is the capacity of the queue, which is `length-1`, as explained in Section 11.1(#queue). Here is the new version of `queue_push`, which is run by producer threads: void queue_push(Queue *queue, int item) { semaphore_wait(queue->spaces); semaphore_wait(queue->mutex); queue->arrayqueue->next_in = item; queue->next_in = queue_incr(queue, queue->next_in); semaphore_signal(queue->mutex); semaphore_signal(queue->items); } Notice that `queue_push` doesn't have to call `queue_full` any more; instead, the semaphore keeps track of how many spaces are available and blocks producers if the queue is full. Here is the new version of `queue_pop`: int queue_pop(Queue *queue) { semaphore_wait(queue->items); semaphore_wait(queue->mutex); int item = queue->arrayqueue->next_out; queue->next_out = queue_incr(queue, queue->next_out); semaphore_signal(queue->mutex); semaphore_signal(queue->spaces); return item; } This solution is explained, using pseudo-code, in Chapter 4 of *The Little Book of Semaphores*. Using the code in the repository for this book, you should be able to compile and run this solution like this: $ make queue_sem $ ./queue_sem ## Make your own semaphores Any problem that can be solved with semaphores can also be solved with condition variables and mutexes. We can prove that's true by using condition variables and mutexes to implement a semaphore. Before you go on, you might want to try this as an exercise: write functions that implement the semaphore API in `sem.h` using using condition variables and mutexes. In the repository for this book, you'll find my solution in `mysem_soln.c` and `mysem_soln.h`. If you have trouble getting started, you can use the following structure definition, from my solution, as a hint: typedef struct { int value, wakeups; Mutex *mutex; Cond *cond; } Semaphore; `value` is the value of the semaphore. `wakeups` counts the number of pending signals; that is, the number of threads that have been woken but have not yet resumed execution. The reason for wakeups is to make sure that our semaphores have Property 3, described in `The Little Book of Semaphores`. `mutex` provides exclusive access to `value` and `wakeups`; `cond` is the condition variable threads wait on if they wait on the semaphore. Here is the initialization code for this structure: Semaphore *make_semaphore(int value) { Semaphore *semaphore = check_malloc(sizeof(Semaphore)); semaphore->value = value; semaphore->wakeups = 0; semaphore->mutex = make_mutex(); semaphore->cond = make_cond(); return semaphore; } ### Semaphore implementation Here is my implementation of semaphores using POSIX mutexes and condition variables: void semaphore_wait(Semaphore *semaphore) { mutex_lock(semaphore->mutex); semaphore->value--; if (semaphore->value < 0) { do { cond_wait(semaphore->cond, semaphore->mutex); } while (semaphore->wakeups < 1); semaphore->wakeups--; } mutex_unlock(semaphore->mutex); } When a thread waits on the semaphore, it has to lock the mutex before it decrements `value`. If the value of the semaphore becomes negative, the thread blocks until a "wakeup" is available. While it is blocked, the mutex is unlocked, so another thread can signal. Here is the code for `semaphore_signal`: void semaphore_signal(Semaphore *semaphore) { mutex_lock(semaphore->mutex); semaphore->value++; if (semaphore->value <= 0) { semaphore->wakeups++; cond_signal(semaphore->cond); } mutex_unlock(semaphore->mutex); } Again, a thread has to lock the mutex before it increments `value`. If the semaphore was negative, that means threads are waiting, so the signalling thread increments `wakeups` and signals the condition variable. At this point one of the waiting threads might wake up, but the mutex is still locked until the signalling thread unlocks it. At that point, one of the waiting threads returns from `cond_wait` and checks whether a wakeup is still available. If not, it loops and waits on the condition variable again. If so, it decrements `wakeups`, unlocks the mutex, and exits. One thing about this solution that might not be obvious is the use of a `do...while` loop. Can you figure out why it is not a more conventional `while` loop? What would go wrong? The problem is that with a `while` loop this implementation would not have Property 3. It would be possible for a thread to signal and then run around and catch its own signal. With the `do...while` loop, it is guaranteed^2 that when a thread signals, one of the waiting threads will get the signal, even if the signalling thread runs around and gets the mutex before one of the waiting threads resumes. ^1: This is a simplification; we will go into more detail later. ^2: Well, almost. It turns out that a well-timed spurious wakeup (see ) can violate this guarantee.