Lecture 04: Files, Memory, and Processes

Principles of Computer Systems

Winter 2020

Stanford University

Computer Science Department

Lecturer: Chris Gregg and

Nick Troccoli

PDF of this presentation

You've seen how a file system (e.g., the System 6 file system, s6fs) layers on top of a block device to present an abstraction of files and directories
Layering: decomposing systems into components with well-defined responsibilities, specifying repcise APIs between them (above and below)
- s6fs works on top of anything that provides a block interface: hard disk, solid state disk, RAM disk, loopback disk, etc.
- Many different file systems can sit on top of a block device: s6fs, ext2fs, ext4fs, btfs, ntfs, etc., etc.
Abstraction: defining an API of an underlying resource that is simultaneously simple to use, allows great flexibility in implementation, and can perform well
- Userspace programs operate on files and directories
- File system has great flexibility in how it represents files and directories on a disk
- Not always perfect: block interface for flash and FTLs
Names and name resolution: files are resources, directory entries (file names) are the way we name and refer to those resources

Recap of Lectures 1-3

This diagram shows a vertical list of connections from userspace processes down to harder. Specifically: userspace processes<--system calls-->operating system<--vnode interface-->file system (ext4fs, s6fs, btfs)<--block interface-->device driver (SATA, iSCSI)<--hardare interface-->hardware (PCI-E, DMA, RDMA, etc.)

How files on disk are presented to a program as file descriptors
How programs open, control and manipulate files
- How does the command below work?
- $cat people.txt | uniq | sort > list.txt
File descriptors vs. open files (many-to-one mapping)
vnode abstraction of a file within the kernel
Memory mapped files and the buffer cache
The concept of a process and what it represents
- Address space: virtualization of memory
- Seamless thread(s) of execution: virtualization of CPU
Creating and managing processes
Concurrency: challenges when you have multiple processes running and how you manage them

Today's Lecture in 3 Parts

Linux maintains a data structure for each active process. These data structures are called process control blocks, and they are stored in the process table
- We'll explain exactly what a process is later in lecture
Process control blocks store many things (the user who launched it, what time it was launched, CPU state, etc.). Among the many items it stores is the file descriptor table
A file descriptor (used by your program) is a small integer that's an index into this table
- Descriptors 0, 1, and 2 are standard input, standard output, and standard error, but there are
  no predefined meanings for descriptors 3 and up. When you run a program from the terminal, descriptors 0, 1, and 2 are most often bound to the terminal

File Descriptor Table and File Descriptors

This diagram shows the "Process Control Blocks", handled by the OS. Specifically, it shows the "descriptor table for process ID 1000", "descriptor table for process ID 1001", and "descriptor table for process ID 1002". Each sub-diagram has an array from 0 to 9.

A file descriptor is the identifier needed to interact with a resource (most often a file) via system calls (e.g., read, write, and close)
A name has semantic meaning, an address denotes a location; an identifier has no meaning
- /etc/passwd vs.34.196.104.129 vs. file descriptor 5
Many system calls allocate file descriptors
- read: open a file
- pipe: create two unidirectional byte streams (one read, one write) between processes
- accept: accept a TCP connection request, returns descriptor to new socket
When allocating a new file descriptor, kernel chooses the smallest available number
- These semantics are important! If you close stdout (1) then open a file, it will be assigned to file descriptor 1 so act as stdout (this is how $ cat in.txt > out.txt works)

Creating and Using File Descriptors

E.g., a file table entry (for a regular file) keeps track of a current position in the file
- If you read 1000 bytes, the next read will be from 1000 bytes after the preceding one
- If you write 380 bytes, the next write will start 380 bytes after the preceding one
If you want multiple processes to write to the same log file and have the results be intelligible, then you have all of them share a single file table tnry: their calls to write will be serialized and occur in some linear order

File Descriptor vs. File Table Entries

A entry in the file descriptor table is just a pointer to a file table entry
Multiple entries in a table can point to the same file table entry
Entries in different file descriptor tables (different processes!) can point to the same file table entry

Almost the same as previous, with the addition of an expansion of one of the elements in process 1001. The expansion has an entry that reads: "mode: r, cursor: 0, refcount: 1, vnode: (blank)"

File Descriptors vs. File Table Entries Example

$ ./main 1> log.txt 2> log.txt

$ ./main 1> log.txt 2>&1

Opens log.txt twice (two file table entries)

Opens log.txt once, two descriptors for same file table entry

// file: testfd.c
#include <stdio.h>
#include <unistd.h>
#include <string.h>

int main(int argc, char **argv)
{
    const char* error = "One plus one is\ntwo.\n";
    const char* msg   = "One plus two is\n";

    write(2, error, strlen(error));
    write(1, msg, strlen(msg));
    return 0;
}

File Descriptors vs. File Table Entries Example

pos: 0

log.txt

fd table

file table

vnode

pos: 0

log.txt

fd table

file table

vnode

cgregg@myth60:$ ./testfd 1> log.txt 2> log.txt
cgregg@myth60:$ cat log.txt
One plus two is
two.
cgregg@myth60:$

cgregg@myth60:$ ./testfd 1> log.txt 2>&1
cgregg@myth60:$ cat log.txt
One plus one is
two.
One plus two is
cgregg@myth60:$

Each process maintains its own descriptor table, but there is one, system-wide open ﬁle table. This allows for ﬁle resources to be shared between processes, as we've seen
As drawn above, descriptors 0, 1, and 2 in each of the three PCBs alias the same three open files. That's why each of the referred table entries have refcounts of 3 instead of 1.
This shouldn't surprise you. If your bash shell calls make, which itself calls g++, each of them inserts text into the same terminal window: those three files could be stdin, stdout, and stderr for a terminal

File Table Details

Almost the same as the previous diagram, except that there are many descriptor table entries. The idea is that file descriptior 0 from each descriptor table points to the same open file table entry, corresponding to STDIN. Likewise for descriptor 1 (all point to the same open file table entry, for STDOUT, which has a reference count of 3, because there are three files that have STDOUT open), and for descriptor 2 (STDERR). Each descriptor table also has some entries (above 2) that point to open files that are not shared between other processes.

vnodes

The vnode is the kernel's abstraction of an actual file: it includes information on what kind of file it is, how many file table entries reference it, and function pointers for performing operations.
A vnode's interface is file-system independent, but its implementation is file-system specific; any file system (or file abstraction) can put state it needs to in the vnode (e.g., inode number)
The term vnode comes from BSD UNIX; in Linux source it's called a generic inode (CONFUSING!)

Each open ﬁle entry has a pointer to a vnode, which is a structure housing static information about a ﬁle or ﬁle-like resource.

Almost the same as the last image, except that one of the vnode entries is further expanded to show the details in the "vnode table". In particular: "type: regfile, refcount: 1, fnptrs ***, inode: 0644; 8:23pm; cgregg"

File Decriptors -> File Table -> vnode Table

There is one system-wide vnode table for the same reason there is one system-wide open ﬁle table. Independent ﬁle sessions reading from the same ﬁle don't need independent copies of the vnode. They can all alias the same one.

Almost the same as the last diagram, except now many of the open file table vnode entries are expanded for the vnode table.

UNIX File Abstractions Summary

Userspace programs interact through files through file descriptors, small integers which are indexes into a per-process file descriptor table
Many file descriptors can point to the same file table entry
- They share seek pointers (writes and reads are serialized)
- Multiple programs share stdout/stderr in a terminal
Many file table entries can point to the same file (vnode)
- They concurrently access the file with different seek pointers
- You run two instances of a python script in parallel: each invocation of Python opens the file separately, with a different file table entry
Exactly how vnodes are implemented is filesystem/resource dependent
- A terminal (tty) vnode is different than an ext4fs one
Reference counting throughout
- Free a file table entry when the last file descriptor closes it
- Free a vnode when the last file table entry is freed
- Free a file when its reference count is 0 and there is no vnode
Key principles: abstraction, layers, naming

Same diagram as earlier: This diagram shows a vertical list of connections from userspace processes down to harder. Specifically: userspace processes<--system calls-->operating system<--vnode interface-->file system (ext4fs, s6fs, btfs)<--block interface-->device driver (SATA, iSCSI)<--hardare interface-->hardware (PCI-E, DMA, RDMA, etc.)

Questions about files in UNIX?

Memory mapped files

read(2) is not the only way to for a program to access a file
Read requires making a copy: program provides a buffer to read into
- What if many programs want read-only access to the file at the same time?
Example: libc.so
- Almost program wants to read libc.so for some of the functions it provides
- Imagine if every program had to read() all of its libraries into local memory
- Let's use pmap to look at how much memory libraries take up
Solution: memory mapped files
- Ask the operating system: "please map this file into memory for me"

Process Address Spaces

Recall that each process operates as if it owns all of main memory.
The diagram on the right presents a 64-bit process's general memory playground that stretches from address 0 up through and including 2⁶⁴ - 1.
CS107 and CS107-like intro-to-architecture courses present the diagram on the right, and discuss how various portions of the address space are cordoned oﬀ to manage traditional function call and return, dynamically allocated memory, access global data, and machine code storage and execution.
No process actually uses all 2⁶⁴ bytes of its address space. In fact, the vast majority of processes use a miniscule fraction of what they otherwise think they own.
The OS virtualizes memory: each process thinks it as the complete system memory (but obviously it doesn't)

This diagram shows main memory layout as a large, vertical rectangle, going from address 0x0 to address 0xFFFFFFFFFFFFFFF

Memory Regions in a Process

Most of a process's memory isn't used: valid regions are defined by segments, blocks of memory for a particular use
- Quick quiz: what's a SEGV (segmentation violation)?
Some segments you know quite well are the stack, heap, BSS, data, rodata, and code (where your executable is)
- Quick quiz: differences between bss, data, and rodata?
There are also segments for shared libraries
- We just pmapped a process on myth
Code is usually not read in through read: instead, it's memory mapped
A memory mapped file acts like the whole file is read into a segment of memory, but it a single copy can be shared across many processes

This image is another vertical memory diagram, with the segment details, as follows, from top to bottom (starting in the middle): "user stack, shared libraries, heap, bss, data, rodata, code"

Memory Mapped Files

void *mmap(void *addr,
size_t len,
int prot,
int flags,
int fd,
off_t offset);
"The mmap() system call causes the pages starting at addr and continuing for at most len bytes to be mapped from the object described by fd, starting at byte offset offset. If offset or len is not a multiple of the pagesize, the mapped region may extend past the specified range. Any extension beyond the end of the mapped object will be zero-filled."
- A page (typically 4kB) is an operating system's unit of memory management, defined by hardware
You can also mmap() anonymous memory, memory that has no backing file: pages in an anonymous region are zero (until written)
- This is how the heap, stack, data, and bss are set up

0xFFFFFFFFFFFFFFFF

0x0

libc.so

bash

heap

stack

libdl.so

data

Memory Mapped Files and the Buffer Cache

The operating system maintains a buffer cache: a pool of (page-sized) pieces of files that are in memory
- Caching: keeping pieces of used data in faster storage to improve performance
- Buffer cache: keeping parts of files in use in memory so you don't have to hit disk
Calls to read() and write() operate on the buffer cache
Blocks are read from disk and put into the buffer cache as needed
Dirty pages in the buffer cache are written to disk when needed
- sync() system call flushes buffers associated with file
Two memory maps of the same file can point to the same buffer cache entry
There's a bit more to this, but you'll have to wait for CS140 (we could spend 4 lectures on just the basics)

libc.so

Process A

Process B

Buffer Cache

Memory Mapped Files Summary

A program can map a file into its memory with mmap()
- Virtualization: every process thinks it has its own copy, but in reality there's a single one in memory (exception: MAP_PRIVATE)
Memory mapped files unify the idea of the process address space with its file descriptors
Used parts of files are kept in memory in the buffer cache
- Caching: don't force every process to read every file in entirety when it loads

libc.so

Process A

Process B

Buffer Cache

Questions about mmap()?

Introduction to Multiprocessing

In the CS curriculum so far, your programs have operated in a single process, meaning, basically, that one program was running your code. The operating system gives your program the illusion that it was the only thing running, and that was that.

Now, we are going to move into the realm of multiprocessing, where you control more than one process at a time with your programs. You will ask the OS, “do these things concurrently”, and it will.

Process

What is a process?
- When you start a program, it runs in a single process. It has a process id (an integer) that the OS assigns. A program can get its own process id with the getpid system call:

The operating system schedules runnable processes to run on CPU cores.
- Some processes are not runnable: waiting on sleep() , a read(), etc.
There may be tens, hundreds, or thousands of processes "running" at once, but on a single-core system, only one can run at a time, and the OS decides which.
On multicore machines (like most modern computers), multiple programs can literally run at the same time, one on each core.

// file: getpidEx.c
#include<stdio.h>
#include<stdlib.h>
#include <unistd.h> // getpid

int main(int argc, char **argv)
{
    pid_t pid = getpid();
    printf("My process id: %d\n",pid);
    return 0;
}

cgregg@myth57$ ./getpidEx
My process id: 7526

fork()

The fork() system call creates a new process

If a program wants to launch a second process, it uses fork()
fork() does exactly this:
- It creates a new process that starts on the following instruction
  after the original, parent, process. The parent process also
  continues on the following instruction, as well.
- The fork call returns a pid_t (an integer) to both processes. Neither is the actual pid of the process that receives it:
  - The parent process gets a return value that is the pid of the child process.
  - The child process gets a return value of 0, indicating that it is the child.
    - The child process does, indeed, have its own pid, but it would need to call getpid itself to retrieve it.
- Everything is duplicated in the child process
  - File descriptors (increasing reference counts on file table entries)
  - Mapped memory regions (the address space)
    - But anonymous regions like stack, heap, etc. are copied (why?)

Fork Return Value

The reason that the parent and its child get different return values from fork is twofold:
- It is useful for the parent to know its child's pid. There is no other way for a parent to easily get its children's process ids (if a child wants to get its parent's pid, it can call getppid)
- It differentiates them. It is almost a certainty that the parent and child will have different objectives after the fork, and it is useful for a process to know whether it is the parent or the child.
- Here's a simple program that knows how to spawn new processes. It uses the system calls fork, getpid, and getppid. The full program can be viewed right here.

int main(int argc, char *argv[]) {
  printf("Greetings from process %d! (parent %d)\n", getpid(), getppid());
  pid_t pid = fork();
  assert(pid >= 0);
  printf("Bye-bye from process %d! (parent %d)\n", getpid(), getppid());
  return 0;
}

Fork in Action!

Here is the output of two consecutive runs of the program:

myth60$ ./basic-fork 
Greetings from process 29686! (parent 29351)
Bye-bye from process 29686! (parent 29351)
Bye-bye from process 29687! (parent 29686)

myth60$ ./basic-fork 
Greetings from process 29688! (parent 29351)
Bye-bye from process 29688! (parent 29351)
Bye-bye from process 29689! (parent 29688

There are a couple of things to note about this program:
- The original process has a parent, which is the shell -- that is the program that you run in the terminal.
- The ordering of the parent and child output is nondeterministic. Sometimes the parent prints first, and sometimes the child prints first - why?

int main(int argc, char *argv[]) {
  printf("Greetings from process %d! (parent %d)\n", getpid(), getppid());
  pid_t pid = fork();
  assert(pid >= 0);
  printf("Bye-bye from process %d! (parent %d)\n", getpid(), getppid());
  return 0;
}

Debugging Multiprocess Programs

You might be asking yourself, How do I debug two processes at once? This is a very good question! gdb has built-in support for debugging multiple processes, as follows:
- set detach-on-fork off
  - This tells gdb to capture any fork'd processes, though it pauses them upon the fork.
- info inferiors
  - This lists the processes that gdb has captured.
- inferior X
  - Switch to a different process to debug it.
- detach inferior X
  - Tell gdb to stop watching the process, and continue it
- You can see an entire debugging session on the basic-fork program right here.

A Tree of Forks

Second example: A tree of fork calls
- While you rarely have reason to use fork this way, it's instructive to trace through a short program where spawned processes themselves call fork. The full program can be viewed right here.

static const char const *kTrail = "abcd";
int main(int argc, char *argv[]) {
  size_t trailLength = strlen(kTrail);
  for (size_t i = 0; i < trailLength; i++) {
    printf("%c\n", kTrail[i]);
    pid_t pid = fork();
    assert(pid >= 0);
  }
  return 0;
}

A Tree of Forks, Continued

Second example: A tree of fork calls
- Two samples runs on the right
- Reasonably obvious: A single a is printed by the soon-to-be-great-grandparent process.
- Less obvious: The first child and the parent each return from fork and continue running in mirror processes, each with their own copy of the global "abcd" string, and each advancing to the i++ line within a loop that promotes a 0 to 1. It's hopefully clear now that two b's will be printed.
- Key questions to answer:
  - Why aren't the two b's always consecutive?
  - How many c's get printed?
  - How many d's get printed?
  - Why is there a shell prompt in the middle of the output of the second run on the right?

myth60$ ./fork-puzzle 
a
b
c
b
d
c
d
c
c
d
d
d
d
d
d
myth60$

myth60$ ./fork-puzzle
a
b
b
c
d
c
d
c
d
d
c
d
myth60$ d 
d
d

Where and When Do We Fork?

Fork is used pervasively in applications. A few examples:
- Running a program in a shell: the shell forks a new process to run the program
- Servers: most network servers run many copies of the server in different processes (why?)
Fork is used pervasively in systems. A few examples:
- When your kernel boots, it starts the system.d program, which forks off all of the services and systems for your computer
  - Let's take a look with pstree
- Your window manager spawns processes when you start programs
- Network servers spawn processes when they receive connections
  - E.g., when you ssh into myth, sshd spawns a process to run your shell in (after setting up file descriptors for your terminals over ssh)
Processes are the first step in understanding concurrency, another key principle in computer
systems; we'll look at other forms of concurrency later in the quarter

Managing Your Progeny

You've spawned a child and let it loose in the world: now what?

Managing Your Progeny

waitpid can be used to temporarily block a process until a child process exits.
- The first argument specifies whom to wait on, which for the moment is just the id of a child process that needs to complete before waitpid can return.
- The second argument supplies the address of an integer where termination information can be placed (or we can pass in NULL if we don't care for the information).
- The third argument is a collection of bitwise-or'ed flags we'll study later. For the time being, we'll just go with 0 as the required parameter value, which means that waitpid should only return when a process in the supplied wait set exits.
- The return value is the pid of the child that exited, or -1 if waitpid was called and there were no child processes in the supplied wait set.

pid_t waitpid(pid_t pid, int *status, int options);

Example of waitpid()

Third example: Synchronizing between parent and child using waitpid
- Consider the following program, which is more representative of how fork really gets used in practice (full program, with error checking, is right here):

int main(int argc, char *argv[]) {
  printf("Before.\n");
  pid_t pid = fork();
  printf("After.\n");
  if (pid == 0) {
    printf("I am the child, and the parent will wait up for me.\n");
    return 110; // contrived exit status
  } else {
    int status;
    waitpid(pid, &status, 0)
    if (WIFEXITED(status)) {
      printf("Child exited with status %d.\n", WEXITSTATUS(status));
    } else {
      printf("Child terminated abnormally.\n");
    }
    return 0;
  }
 }

Work with neighbor: what do you expect the output to be?

Synchronizing Between Parent and Child Using waitpid()

The output is the same every single time the above program executes.
The implementation directs the child process one way, the parent another.
The parent process correctly waits for the child to complete using waitpid.
The parent lifts child exit information out of the waitpid call, and uses the WIFEXITED
macro to examine some high-order bits of its argument to confirm the process exited normally, and it uses the WEXITSTATUS macro to extract the lower eight bits of its argument to produce the child return value (which we can see is, and should be, 110).
The waitpid call also donates child process-oriented resources back to the system.

myth60$ ./separate 
Before.
After.
After.
I am the child, and the parent will wait up for me.
Child exited with status 110.
myth60$

How Deep is the Copy?

This next example is more of a brain teaser, but it illustrates just how deep a clone the process created by fork really is (full program, with more error checking, is right here).
The code emulates a coin flip to seduce exactly one of the two processes to sleep for a second, which is more than enough time for the child process to finish.
The parent waits for the child to exit before it allows itself to exit. It's akin to the parent not being able to fall asleep until he/she knows the child has, and it's emblematic of the types of synchronization directives we'll be seeing a lot of this quarter.
The final printf gets executed twice. The child is always the first to execute it, because
the parent is blocked in its waitpid call until the child executes everything.

int main(int argc, char *argv[]) {
  printf("I'm unique and just get printed once.\n");
  bool parent = fork() != 0;
  if ((random() % 2 == 0) == parent) sleep(1); // force exactly one of the two to sleep
  if (parent) waitpid(pid, NULL, 0); // parent shouldn't exit until child has finished
  printf("I get printed twice (this one is being printed from the %s).\n", 
         parent  ? "parent" : "child");
  return 0;
}

Multiple Children

Spawning and synchronizing with multiple child processes
- A parent can call fork multiple times, provided it reaps the child processes (via waitpid) once they exit. If we want to reap processes as they exit without concern for the order they were spawned, then this does the trick (full program checking right here):

int main(int argc, char *argv[]) {
  for (size_t i = 0; i < 8; i++) {
    if (fork() == 0) exit(110 + i);
  } 
  while (true) {
    int status;
    pid_t pid = waitpid(-1, &status, 0);
    if (pid == -1) { assert(errno == ECHILD); break; }
    if (WIFEXITED(status)) {
      printf("Child %d exited: status %d\n", pid, WEXITSTATUS(status));
    } else {
      printf("Child %d exited abnormally.\n", pid);
    }
  }
  return 0;
}

Calling waitpid() on Multiple Children

Pass -1 as the first argument to waitpid. man waitpid:
- "The value of pid can be:
  - < -1 meaning wait for any child process whose process group ID is equal to the absolute value of pid.
  - -1 meaning wait for any child process.
  - 0 meaning wait for any child process whose process group ID is equal to that of the calling process.
  - > 0 meaning wait for the child whose process ID is equal to the value of pid.
- Eventually, all children exit and waitpid correctly returns -1 to signal there are no more processes under the parent's jurisdiction.
- When waitpid returns -1, it sets a global variable called errno to the constant ECHILD to signal waitpid returned -1 because all child processes have terminated. That's the "error" we want.

What's the difference?

What's different about these two different invocations at the shell?
- ```
$ vim main.c
```
- ```
$ vim main.c &
```
What does & make the shell do differently?

Questions about fork()?

Lecture Review

The UNIX file system provides a naming and name resolution system for user data, through files and directories
The UNIX file system is designed to use abstraction and layering so that it's easy to use new file systems and use existing file systems on new devices
- Processes use the abstraction of a file descriptor, which refers to an open file in the file entry table
- A file entry refers to a vnode, which describes the actual file itself
- There can be many file descriptors for the same single file table entry, and many file table entries for the same vnode
- This layering has important semantics which give you a lot of power in how you manipulate files
Processes can directly map files into their memory with mmap(), which allows the OS to use caching
- In-use parts of files are kept in memory by a system called the buffer cache
- The buffer cache allows many processes to share a single copy of data (e.g., library code)
- Processes virtualize memory and the CPU
Processes create new processes with fork(), the processes run concurrently
- The parent and child are identical except for the return value
  - Share file descriptors, memory regions, etc. (some memory, like the heap, is copied)
- Parent can wait for children to exit (and learn about their exit) with waitpid()
- This concurrency lets your program use more than one core: much of the quarter will look at the complications