Interprocess Communication in Unix Paulo Marques Department of Informatics Engineering University of Coimbra [email protected] 2006/2007 An Introduction

InterprocessCommunication in Unix

Paulo MarquesDepartment of Informatics EngineeringUniversity of [email protected]

2006

/200

7

An Introduction toConcurrent Programming

2

Contents

Concurrent Programming Concepts Process Creation and Management Signals Pipes, Named Pipes and Unix Sockets IO Multiplexing Message Queues Shared Memory Semaphores POSIX Threads

Creation and Management Synchronization

3

Standards

There are many UNIX flavors around... Linux, HP-UX, Solaris, AIX, BSD, IRIX, etc. Different capabilities, different APIs for advanced

functionalities

Vendors tried to find a common ground: POSIX = Portable Operating System Interface

Standards? UNIX System V (Release 4) BSD 4.3 (1988) POSIX.1 (1994) POSIX Spec. 1170 (2001) IEEE Std. 1003.1-2001 (or POSIX)

4

POSIX Extensions

There is the 2001 base standard and there’s its extensions Implementations that comply with the base standard

define _POSIX_VERSION to 200112L in <unistd.h>

5

POSIX Extensions (2)

And many, many more...

6

Man pages typically indicate compliance



2006

/200

7

Process Creation and Management

8

Process Model

Process creation in Unix is based on spawning child processes which inherit all the characteristics of their fathers Variables, program counter,

open files, etc. Spawning a process is done

using the fork() system call

After forking, each process will be executing having different variables and different state. The Program Counter will be

pointing to the next instruction

Changing a variable in the child program doesn’t affect its father (and vice-versa)

a = f();fork();b = g();

State

a = f();fork();b = g();

State

a = f();fork();b = g();

State

Child Process

Original Process

Original Process

9

Process Management

Each process has an unique identifier (PID). Each process has a father, which is also identified (PPID).

pid_t getpid(void);Returns the PID of the current process.

pid_t getppid(void);Returns the PID of the parent process.

pid_t fork(void);Creates a new process which inherits all its father’s state. It returns 0 on the original process and the child’s PID in the spawned process.

pid_t wait(int* status);Waits until a child process exits. The status of the child is set in status. (status is the return value of the process)

pid_t waitpid(pid_t who, int* status, int options);Same as wait() but allows to wait for a particular child. In options, by using WNOHANG in options, allows for checking if a child has already exited without blocking. 0 in who means “wait for any child”.

10

Using processes for doing different things

The key idea is to create asymmetry between processes

#include <stdio.h>#include <unistd.h>#include <sys/wait.h>#include <sys/types.h>

int main(){ pid_t id;

id = fork(); if (id == 0) { printf("[%d] I'm the son!\n", getpid()); printf("[%d] My parent is: %d\n", getpid(), getppid()); } else { printf("[%d] I'm the father!\n", getpid()); wait(NULL); } return 0;}

11

And the result is...

But why did we do wait(NULL)?

...printf("[%d] I'm the father!\n", getpid());wait(NULL);...

12

Process Termination in UNIX

A process is only truly eliminated by the operating system when it’s father calls wait()/waitpid() on it. This allows the parent check things like the exit code of its

son’s

Zombie Process: One that has died and it’s parent has not acknowledged its death (by calling wait()) Be careful with this if your are designing servers. They are

eating up resources!!

Orphan Process: One whose original parent has died. In that case, it’s parent becomes init (process 1).

13

Let’s generate some Zombies

#include <stdio.h>#include <stdlib.h>#include <unistd.h>#include <sys/wait.h>#include <sys/types.h>

void worker() { printf("[%d] Hi, I'm a worker process! Going to die...\n", getpid());}

int main() { for (int i=0; i<10; i++) { if (fork() == 0) { worker(); exit(0); } } printf("[%d] Big father is sleeping!\n", getpid()); sleep(10);

return 0;}

14

Zombies (2)

15

Let’s see adoption by init

#include (...)

void worker(){ sleep(10); printf("[%d] Let's see who my dady is: %d\n", getpid(), getppid());}

int main(){ for (int i=0; i<10; i++) { if (fork() == 0) {

worker();exit(0);

} }

printf("[%d] Big dady is going away!\n", getpid()); return 0;}

16


17

How to structure code

(...)

if ((id = fork()) == 0){ // Huge amount of code // ...}else{ // Huge amount of code // ...}

(...)

Fairly common...Don’t do it!

18

How to structure code

void client(int id){ // Client code // ...}

if ((id = fork()) == 0){ client(id); exit(0);}else if (id == -1){ error();}

// Original process code// ...

Server

Client Client Client

Note: You still have to consider how to take care of zombies

19

How a process becomes another executable

Somehow the OS must be able to execute code starting from an executable file e.g. how does the shell (bash) becomes ‘ls’?

int execl(const char *path, const char *arg, ...);int execlp(const char *file, const char *arg, ...);int execle(const char *path, const char *arg, ..., char *const envp[]);int execv(const char *path, char *const argv[]);int execvp(const char *file, char *const argv[]);

“exec family” of functions Allow to substitute the current process executable image by

another one. The substitution is complete! The functions that have a ‘p’ make use of the environment

PATH; The functions that have a ‘v’ make use of a pointer to an array on parameters; The functions that have an ‘l’ have the parameters passed separated by commas

Make sure that the first parameter is the name of the program!

20

Example

Simple program that lists the files in the current directory

Note: A successful exec() never returns The code, the stack, the heap, it’s all replaced by the new

executable

Original Code “ls code”

“ls code”

exec()

21

The corresponding code...

#include <stdio.h>#include <stdlib.h>#include <unistd.h>#include <sys/types.h>

int main(){ if (execlp("ls", "ls", "-a", NULL) == -1) perror("Error executing ls: "); else printf("This cannot happen!\n");

return 0;}

char* ls_param[] = { "ls", "-a", NULL };

if (execvp(ls_param[0], ls_param) == -1) perror("Error executing ls: ");

Using an arraycan be more flexible...



2006

/200

7

Asynchronous Events:Signals

23

Signals

A signal represents an asynchronous event which an application must (should? can?) process The programmer can register a routine to handle such

events Examples:

The user hits Ctrl+C SIGINT The system requests the application to terminate

SIGTERM The program tried to write to a closed channel

SIGPIPEProcess

normal flowof execution

int sigint_handler() { // process signal}

SIGINT

24

Signals (2)

Signals can be in one of four states: Blocked:

Upon arrival, they are stored in a queue until the process unblocks them. Then, they are delivered.

Ignored:Upon arrival, they are discarded. It is as if they had never existed.

Being Handled:They are redirected to a signal handler which is called.

None of the above:Non-handled, non-blocked or ignored signals. Upon arrival, they cause program termination.

Some signals cannot be ignored or handled (e.g. SIGKILL)

When a process starts, signals are on their “default behavior”. Some are ignored, most are in the “non-handled, non-blocked

nor ignored state”. If a signal occurs, the process will die.

25

Basic Signal Routines

typedef void (*sighandler_t)(int); sighandler_t signal(int signum, sighandler_t handler);

Redirects a certain signal (signum) to a handler routine.

int kill(pid_t pid, int sig); Sends a signal to a certain process identified by a PID. (Note: if pid is 0, sends to all processes in the current process group.)

int pause();Blocks the process until a signal is received.

prototype of the handler routine

int sigaction(); int sigprocmask(); int sigpending(); int sigsuspend();

Note: These are the recommended POSIX routines. We are not going to cover them here. The problem with signal() is that in certain cases its behavior is undefined across systems.

26

Handling a signal

void sigint(int signum) { char option[2]; printf("\n ^C pressed. Do you want to abort? ");

scanf("%1s", option); if (option[0] == 'y') { printf("Ok, bye bye!\n"); exit(0); }}

int main(){ // Redirects SIGINT to sigint() signal(SIGINT, sigint);

// Do some work! while (1) { printf("Doing some work...\n"); sleep(1); } return 0;}

27

Handling a signal (2)

28

Special constants in signal()

signal(SIGINT, SIG_IGN)Ignores SIGINT

signal(SIGINT, SIG_DFL)Restores SIGINT to its

“default” handling

29

The problem with signals

They make programming extremely hard It’s completely asynchronous: you never know when you

are going to get a signal This means that you have to protect all calls!

After calling a standard function, it may return -1 indicating an error errno==EINTR means that a certain routine was

interrupted and has to be tried again. Other routines return other things. It you are using signals, you have to protect them against

all that!

30

The problem with signals (2)

For instance, simply to try to read a “struct person” from disk...

struct person p;...int n, total = 0;while (total < sizeof(p)) { n = read(fd, (char*)p + total, sizeof(p)-total); if (n == -1) { if (errno == EAGAIN) continue; else { // True error! } } total+= n;}

And you have to do something like this for all calls being done in your program!

31

Sending a signal

It’s just a question of calling kill() with the PID of the target process...

void master(pid_t pid_son){ printf("Master sleeping for a while...\n"); sleep(3); printf("Master says: Hello son!\n"); kill(pid_son, SIGUSR1);}

int main() { pid_t son;

// Creates a worker process if ((son=fork()) == 0) { worker(); exit(0); } // The master master(son); wait(NULL); return 0;}

32

The code of the child process...

void dady_call(int signum) { printf("Dady has just called in!\n");}

void worker() { // Redirect "user signal 1" to a handler routine signal(SIGUSR1, dady_call);

// Do some work printf("Child process, life is good...\n"); for (int i=0; i<10; i++) { printf("Child doing some work\n"); sleep(1); } printf("Child saying bye bye!\n");}

33


34

Danger!!!

What do you think it will happen if you receive a signal inside a signal handler?? In most systems, upon entering a signal handling routine, all

signals of that type become blocked (i.e. they are queued). [Well, for “normal” signals, a finite set of them are queued (typically 1); for “real time signals”, all are...]

The other signals are still processed asynchronously if they arrive.

This behavior is not consistent across systems. In fact, in some systems, that signal type resets to its default behavior. This means that if, meanwhile, the program receives a signal of the same type it may die! On that type of system, the first thing that you must do is to once again set the signal handler.

Well... doesn’t really solve the problem, it just makes it less likely.

The new POSIX routines address this – use them. Also, most system nowadays don’t reset the signal handler.

void dady_call(int signum) { signal(SIGUSR1, dady_call); printf("Dady has just called in!\n");}

35

Beware

Signal numbers vary across operating systems and architectures. Don’t rely on them, use symbolic constants

Linux, i386

“man 7 signal”

36

Some Signals



2006

/200

7

Pipes, Named Pipes andUNIX Sockets

38

Stream mode of communication

Pipes, Named Pipes and Stream Unix SOCKETS allow processes to communicate using “streams of data” A “pipe” is a connection between two processes. You can

send things through the pipe, you can try to receive things from the pipe.

A pipe acts like a synchronous finite buffer. If a process tries to write to a pipe that is full, it blocks If a process tries to read from a pipe that is empty, it blocks

write() read()

pipe

39

Pipes

Provides for communication amount processes that are hierarchically related (i.e. father-child) Pipes must be created prior to creating child processes

Whenever a pipe is created, using pipe(), two file descriptors are opened: one for reading (fd[0]), one for writing (fd[1]) Unused file descriptors should be closed!

Pipes are unidirectional (well... normally)

fd[1] fd[0]int fd[2];

pipe(fd);readingwriting

40

Example

typedef struct { int a; int b;} numbers;

// File descriptors for the pipe channelint channel[2];(...)

int main() { // Create a pipe pipe(channel);

// Create the processes if (fork() == 0) { worker(); exit(0); } master(); wait(NULL); return 0;}

41

Example (cont.)

void worker() { numbers n;

close(channel[1]);

while (1) { read(channel[0], &n, sizeof(numbers)); printf("[WORKER] Received (%d,%d) from master to add. Result=%d\n", n.a, n.b, n.a+n.b); }}

void master(){ numbers n; close(channel[0]); while (1) { n.a = rand() % 100; n.b = rand() % 100; printf("[MASTER] Sending (%d,%d) for WORKER to add\n", n.a, n.b); write(channel[1], &n, sizeof(numbers)); sleep(2); }}

42

The result is...

43

Be careful!

A pipe is a finite buffer. If you try to write too much too quickly into it, the process will block until some space clears up.

Atomicity is something to be dealt with If you try to write less that PIPE_BUF bytes into a pipe, you are

guarantied that it will be written atomically It you try to write more, you have no guaranties! If several processes

are writing at the same time, the writes can be interleaved Also, when a process tries to read from a pipe, you are not guarantied

that it will be able to read everything Meaning...

You must synchronize your writes when you’re writing a lot of data! You must ensure that you read complete messages!

struct person p;

int n, total = 0;while (total < sizeof(p)) { n = read(fd[0], (char*)p + total, sizeof(p)-total); total+= n;}

44

Controlling File Descriptors

Each process has a file descriptor table. By default, entries 0, 1 and 2 are: stdin, stdout, stderr.

Each time a file is open, an entry is added to this table. Each time a file is closed, the corresponding entry becomes available.

The process descriptor table, in fact, contains only references to the OS global file descriptor table.

<free>

<free>

0

1

2

3

4

5

6

stdin

stdout

stderr

f2

f3 File Descriptor Table after:open(“f1”)open(“f2”)open(“f3”)close(“f1”)

45

Controlling File Descriptors (2)

Two routines are useful for controlling file descriptors: int dup(int fd)

Duplicates file descriptor “fd” on the first available position of the file descriptor table.

int dup2(int fd, int newfd)Duplicates file descriptor “fd” on the “newfd” position, closing it if necessary.

Note that after a file descriptor is duplicated, the original and the duplicate can be used interchangeably. They share the file pointers, the buffers, locks, etc. Careful: Closing one file descriptor doesn’t close all other

that have been duplicated!

46

Implementing a pipe between two processes

Implementing a pipe between two processes is quite easy. It’s only necessary to associate the standard output of one process with the standard input of another.

Simple example: “ls | sort”.

Note: closing one file descriptor doesn’t close all other that have been duplicated!

47

Resulting in...

48

Named Pipes (also known as FIFOs)

Similar to pipes but allow communication between unrelated processes. Each pipe has a name (string). The pipe is written persistently in the file system. For creating a named pipe, use the “mkfifo” command or

call mkfifo(const char* filename, mode_t mode);

Typically, like pipes, they are half-duplex Means that the must be open read-only or write-only They are opened like files, but they are not files You cannot fseek() a named pipe; write() always appends

to the pipe, read() always returns data from the beginning of the pipe.

After data is read from the named pipe, it’s no longer there. It’s not a file, it’s an object in the unix kernel!

49

Unrelated client/server program (np_server.c)

#define PIPE_NAME "np_client_server"(...)

int main(){ // Creates the named pipe if it doesn't exist yet if ((mkfifo(PIPE_NAME, O_CREAT|O_EXCL|0600)<0) && (errno!= EEXIST)) { perror("Cannot create pipe: "); exit(0); }

// Opens the pipe for reading int fd; if ((fd = open(PIPE_NAME, O_RDONLY)) < 0) { perror("Cannot open pipe for reading: "); exit(0); }

// Do some work numbers n; while (1) { read(fd, &n, sizeof(numbers)); printf("[SERVER] Received (%d,%d), adding it: %d\n", n.a, n.b, n.a+n.b); } return 0;}

50

Unrelated client/server program (np_client.c)

#define PIPE_NAME "np_client_server"(...)

int main(){ // Opens the pipe for writing int fd; if ((fd = open(PIPE_NAME, O_WRONLY)) < 0) { perror("Cannot open pipe for writing: "); exit(0); }

// Do some work while (1) { numbers n; n.a = rand() % 100; n.b = rand() % 100; printf("[CLIENT] Sending (%d,%d) for adding\n", n.a, n.b); write(fd, &n, sizeof(numbers)); sleep(2); }

return 0;}

51

Executing them...

52

Some interesting issues...

If you get a SIGPIPE signal, this means that you are trying to read/write from a closed pipe

A named pipe is a connection between two processes. A process blocks until the other party open the pipe... Being it for reading or writing. It’s possible to bypass this behavior (open it non-blocking –

O_NONBLOCK), but be very, very careful: if not properly programmed, it can lead to busy waiting. If a named pipe is open non-blocking, EOF is indicated when read() returns 0.

When designing a client/server multiple client application, this means that either the pipe is re-opened after each client disconnects, or the pipe is open read-write.

If opened “read-write”, the server will not block until the other party connects (since, he itself is also another party!)

53

Unix Domain Sockets [Very Briefly!]

Similar to Internet sockets but on the local machine Both connection-oriented and datagram Datagram sockets are reliable (not lost nor delivered out-

of-order)

It’s a little bit of a mix between sockets and pipes It’s possible to use a “network-oriented socket interface”

to use them It’s easy to create a pair of unnamed, connected domain

sockets between two processes

They are much lighter than TCP/IP Sockets No protocol processing, no network headers, no

checksums, no sequence numbers... just plain memcpy-ing.

54

Simple (non-concurrent) Echo Server

#define SOCK_NAME "/tmp/my_echo"#define OK_MESSAGE "Ok, got it!"#define BUF_SIZE 4096#define BACKLOG 5

// Our server socket descriptorint sock;

void create_socket() { // Creates the unix socket and binds it assert( (sock = socket(AF_UNIX, SOCK_STREAM, 0)) != -1 ); struct sockaddr_un local; local.sun_family = AF_UNIX; strcpy(local.sun_path, SOCK_NAME); unlink(local.sun_path); int len = strlen(local.sun_path) + sizeof(local.sun_family); assert( bind(sock, (struct sockaddr *) &local, len) != -1 ); assert( listen(sock, BACKLOG) != -1);}

int main() { create_socket(); process_clients(); return 0;}

55

Simple (non-concurrent) Echo Server (2)

void process_clients(){ struct sockaddr_un remote; int connection; socklen_t addr_len; while (1) { printf("Waiting for a connection...\n"); addr_len = sizeof(remote); assert( (connection = accept(sock, (struct sockaddr *) &remote, &addr_len)) != -1); printf("Connected.\n");

// Receive and send a message (highly simplified) char buf[BUF_SIZE]; int n = read(connection, buf, BUF_SIZE); buf[n] = '\0'; printf("Read from socket: %s\n", buf); write(connection, OK_MESSAGE, strlen(OK_MESSAGE)+1); close(connection); }}

56

Simple (non-concurrent) Echo Client

(...)

#define SOCK_NAME "/tmp/my_echo"#define BUF_SIZE 4096

// Our socket descriptorint sock;

void connect_socket() { struct sockaddr_un remote; socklen_t addr_len; assert( (sock = socket(AF_UNIX, SOCK_STREAM, 0)) != -1); printf("Trying to connect...\n");

remote.sun_family = AF_UNIX; strcpy(remote.sun_path, SOCK_NAME); addr_len = strlen(remote.sun_path) + sizeof(remote.sun_family); if (connect(sock, (struct sockaddr *) &remote, addr_len) == -1) { perror("Could not connect: "); exit(1); }

printf("Connected.\n");}

57

Simple (non-concurrent) Echo Client (2)

void send_message(const char* msg){ write(sock, msg, strlen(msg)+1); char buf[BUF_SIZE]; int n = read(sock, buf, BUF_SIZE); buf[n] = '\0'; printf("Read from server: %s\n", buf); close(sock);}

int main(int argc, char* argv[]){ char* msg = "Hello server!"; if (argc == 2) msg = argv[1];

connect_socket(); send_message(msg);}

58

The result is:

59

Checking active Unix sockets

“netstat -a”



2006

/200

7

I/O Multiplexing

61

Interesting Problem

A printer daemon is connected to a physical printer There are 3 named-pipes which allow automatic

formatted printing

PrinterDaemon

/printer/a4_double_sided

/printer/a4_single_sided

/printer/a3_single_sided

Pipes are blocking, so this doesn’t work!!!

62

I/O Multiplexing

I/O Multiplexing: The ability to examine several file descriptors at the same time select() and pselect()

int select(int n, fd_set* readfd, fd_set* writefd, fd_set* exceptfd, struct timeval* timeout)

Blocks until activity is detected or a timeout occurs.

Greatest fd plus oneFor reading activityFor writing activityFor out-of-band activity

The fd_set variables are input/output. Upon return, they indicateif there was activity in a certain descriptor or not.

63

select()

Careful: n is the number of the highest filedescriptor added of one. It’s not the number of file descriptors

fd_setA bit set representing file descriptors

FD_ZERO(fd_set* set)Cleans up the file descriptor set

FD_SET(int fd, fd_set* set)Sets a bit in the file descriptor set

FD_CLEAR(int fd, fd_set* set)Clears a bit in the file descriptor set

FD_ISSET(int fd, fd_set* set)Tests if a file descriptor is set

64

Example (printerd.c)

(...)

#define BUF_SIZE 4096#define NUM_PRINTERS 3

const char* PRINTER_NAME[] = { "printer1", "printer2", "printer3" };

// The printer file descriptorsint printer[NUM_PRINTERS];

void create_printers() { for (int i=0; i<NUM_PRINTERS; i++) { unlink(PRINTER_NAME[i]); mkfifo(PRINTER_NAME[i], O_CREAT|O_EXCL|0666); printer[i] = open(PRINTER_NAME[i], O_RDONLY|O_NONBLOCK); assert(printer[i] >= 0); }}

int main(int argc, char* argv[]) { create_printers(); accept_requests();}

65

Example (printerd.c) (2)

void accept_requests() { while (1) { fd_set read_set; FD_ZERO(&read_set); for (int i=0; i<NUM_PRINTERS; i++) FD_SET(printer[i], &read_set);

if ( select(printer[NUM_PRINTERS-1]+1, &read_set, NULL, NULL, NULL) > 0 ) { for (int i=0; i<NUM_PRINTERS; i++) { if (FD_ISSET(printer[i], &read_set)) { printf("[<%s> PRINTING]: ", PRINTER_NAME[i]);

char buf[BUF_SIZE]; int n = 0; do { n = read(printer[i], buf, BUF_SIZE); if (n > 0) { buf[n] = '\0'; printf("%s", buf); } } while (n > 0);

close(printer[i]); printer[i] = open(PRINTER_NAME[i], O_RDONLY|O_NONBLOCK); } } } } }

66

Resulting in...



2006

/200

7

Message Queues

68

Types of communication

Streams represent “a flow” of bytes. There are no fixed data boundaries. The sender requests the transmission of N bytes The data starts flowing, the receiver starts getting it The receiver may get several chucks of less then N bytes

Messages represent a complete fixed structure of data It’s like sending a letter. Either you get if fully or you

don’t. You don’t get “half a letter”.

169

69

Message Queues

Another IPC mechanism Based on messages, not no data streams

Completely asynchronous A process can start executing, write some messages to a

message queue and die; another process can latter on come alive an receive them.

Sharp contrast with all the mechanisms that we’ve seen so far, which require both the sender and the receiver to be present at the same time

Message queues are maintained by the operating system. They are not destroyed if a process dies!

msgsnd() msgrcv()

70

Message Queues – System V

int msgget(key_t key, int flags)Obtains an identifier to an existing message queue or creates a new one. “key” can be IPC_PRIVATE (which creates a new unique

identifier), or an existing identifier. ftok() can be used to generate a number based on a filename.

“flags”, normal mode flags. When ORed with IPC_CREAT creates a new one.

int msgctl(int mqid, int cmd, struct msqid_ds* buff)Provides a variety of control operations on the message queue. “mqid” is the value returned by msgget() “cmd” is the command (most usually: IPC_RMID to remove it) “buff” a structure used in some control operations

71

Message Queues – System V (2)

int msgsnd(int mqid, const void* message, size_t length, int flags)Sends a message to a certain key. “mqid” is the value returned by msgget() “message” it’s a pointer to the message to send “length” represents the length of the payload of the message

(not total) “flags”: 0 or IPC_NOWAIT (non-blocking)

int msgrcv(int mqid, void* message, size_t length, int type, int flags)Retrieves a message from a message queue. “mqid” is the value returned by msgget() “message” it’s a pointer to the message to receive “length” represents the maximum payload we are willing to

receive “type” represent the type of message to receive (0 FIFO) “flags”: 0 or IPC_NOWAIT

72

Messages a Message Payload

In System V a message can be anything. But, it must always have a “long” integer in the beginning This long is called a message type identifier

typedef struct{ long mtype; int first; int second;} numbers_message;

Message type (must be >0)!

Payload

73

mq_pong.c (1)

typedef struct { long mtype; int first, second;} numbers_msg;

// Message queue idint id;

void cleanup(int signum) { msgctl(id, IPC_RMID, NULL); exit(0);}

void main(int argc, char* argv[]) { assert( (id = msgget(IPC_PRIVATE, IPC_CREAT|0700)) != 0 ); signal(SIGINT, cleanup); if (fork() == 0) ping(); else pong();}

74

mq_pong.c (2)

void ping(){ numbers_msg msg; msg.first = rand() % 100; msg.second = rand() % 100; while (1) { msg.mtype = 1; printf("[A] Sending (%d,%d)\n", msg.first, msg.second); msgsnd(id, &msg, sizeof(msg)-sizeof(long), 0); msgrcv(id, &msg, sizeof(msg)-sizeof(long), 2, 0); printf("[A] Received (%d,%d)\n", msg.first, msg.second); ++msg.first; ++msg.second; sleep(3); }}

75

mq_pong.c (3)

void pong(){ numbers_msg msg;

while (1) { msgrcv(id, &msg, sizeof(msg)-sizeof(long), 1, 0); printf("[B] Received (%d,%d)\n", msg.first, msg.second);

msg.mtype = 2; ++msg.first; ++msg.second;

printf("[B] Sending (%d,%d)\n", msg.first, msg.second); msgsnd(id, &msg, sizeof(msg)-sizeof(long), 0); }}

76

Resulting in...

77

IPC Resources

Remember, IPC resources are not automatically cleaned

This can lead so serious resource leaks “ipcs” allows you to see the current System V IPCs in use “ipcrm” allows you to manually delete resources

78

POSIX Message Queues

Very similar to System V except: Message types represent priorities A read from a POSIX message queue always returns the

oldest message of the largest type (priority) POSIX message queues allow a signal to be raised when a

message is put on an empty queue or the initiation of a thread

Messages queues are represented by names on the file system (like named pipes)

mq_open() mq_close() mq_unlink() mq_send() mq_receive()



2006

/200

7

Shared Memory and Semaphores

80

Why shared memory?

Up until now... System calls are slow! Copying thought the kernel is slow!

P1 P2

buffer

write() read()

User Space

Kernel Space

81

Why shared memory?

Shared Memory (Almost) No kernel involvement! Fast! Very Fast!

P1 P2

sharedbuffer

User Space

Kernel Space

Dangerous, very dangerous!

82

How does it work

Each process has an address space Each address space corresponds to a page table. There

are as many page tables as there are processes Shared memory corresponds to putting the same

“real memory pages” in the page tables of two different processes

0 0

4Gb 4Gb

Address SpaceProcess A

Address SpaceProcess B

1000 1000

Physical Memory0

256Mb5000

Page Translation

(slightly simplified)

83

Shared Memory – System V

int shmget(key_t key, int size, int flags)Obtains an identifier to an existing shared memory or creates a new one. “key” can be IPC_PRIVATE (which creates a new unique

identifier), or an existing identifier. ftok() can be used to generate a number based on a filename.

“size” its the shared memory size in bytes “flags”, normal mode flags. When ORed with IPC_CREAT creates

a new one.

int shmctl(int shmid, int cmd, struct shmid_ds* buff)Provides a variety of control operations on the shared memory. “shmid” is the value returned by shmget() “cmd” is the command (most usually: IPC_RMID to remove it) “buff” a structure used in some control operations

84

Shared Memory – System V (2)

int shmat(int shmid, const void* where, int flags)Maps a certain shared memory region into the current process address space. “shmid” represents the shared memory identifier “shmid”

returned by shmget() “where” represents an unused address space location where to

map the shared memory (normally, use NULL) “flags” represent different ways of doing the mapping (typically

0)

int shmdt(const void* where)Unmaps a certain shared memory region from the current address space. “where” represents an unused address space location where to

map the shared memory (normally, use NULL)

85

How does attaching work

Address Space A Address Space B

“Real Memory”

int id = shmget(1234, 2, IPC_CREAT|0777)

char* p = shmat(id, NULL, 0);

p[0] = 19;

int id = shmget(1234, 2, 0777)

char* p = shmat(id, NULL, 0);

p[1] = 59;

p

p(id.. 324)

19

19

19

59

5959

86

What’s wrong with this routine?

P1 P2

print_work(a, 12); print_work(b, 65);

87

Synchronization – Semaphores

A semaphore is a synchronization object Controlled access to a counter (a value) Two operations are supported: wait() and post()

wait() If the semaphore is positive, decrement it and continue If not, block the calling process (thread)

post() Increment the semaphore value If there was any process (thread) blocked due to the semaphore, unblock

one of them.

5

P1

value

P6 P3 NULLblockedprocesslist

“A semaphore”

88

Corrected version

Mutual Exclusion: Only one process can be in here!

You always have to synchronize, even if you areonly reading or writing one byte!

89

Synchronization – Semaphores

System V Semaphores Works with semaphore arrays semget(), semctl(), semop() A little bit hard to use by themselves

Use a library to encapsulate them!

POSIX Semaphores Quite easy to use sem_init(), sem_close(), sem_post(), sem_wait() Also work with threads! But... in Linux, they only work with threads (kernel 2.4)

Semaphores are very useful for other things besides mutual exclusion: they can be used to count things!

90

semaphore.h

My Semaphore Library

91

Example – Producer/Consumer

A producer puts elements on a finite buffer. If the buffer is full, it blocks until there’s space.

The consumer retrieves elements. If the buffer is empty, it blocks until something comes along.

We will need three semaphores One to count the empty slots One to count the full slots One to provide for mutual exclusion to the shared buffer

Producer Consumer

92

Example – Producer/Consumer

Producer Consumer

empty full

read_pos write_pos

put_element(e) { sem_wait(empty); sem_wait(mutex); buf[write_pos] = e; write_pos = (write_pos+1) % N; sem_post(mutex); sem_post(full);}

mutex

get_element() { sem_wait(full); sem_wait(mutex); e = buf[read_pos]; read_pos = (read_pos+1) % N; sem_post(mutex); sem_post(empty); return e;}

93

The result of executing it!

94

And the main part of its code...

void producer() { for (int i=TOTAL_VALUES; i>0; i--) { printf("[PRODUCER] Writing %d\n", i); put_element(i); }}

void consumer() { for (int i=0; i<TOTAL_VALUES; i++) { int e = get_element(); printf("[CONSUMER] Retrieved %d\n", e); sleep(1); } terminate();}

void main(int argc, char* argv[]) { init();

if (fork() == 0) { producer(); exit(0); } else { consumer(); exit(0); }}

95

put_element() and get_element()

void put_element(int e) { sem_wait(sem, EMPTY); sem_wait(sem, MUTEX);

buf[write_pos] = e; write_pos = (write_pos+1) % N;

sem_post(sem, MUTEX); sem_post(sem, FULL);}

int get_element() { sem_wait(sem, FULL); sem_wait(sem, MUTEX);

int e = buf[read_pos]; read_pos = (read_pos+1) % N;

sem_post(sem, MUTEX); sem_post(sem, EMPTY);

return e;}

96

init() and terminate()

int sem, shmid;int write_pos, read_pos;int* buf;

void init() { sem = sem_get(3, 0); sem_setvalue(sem, EMPTY, N); // N is the number of slots sem_setvalue(sem, FULL, 0); sem_setvalue(sem, MUTEX, 1);

write_pos = read_pos = 0;

shmid = shmget(IPC_PRIVATE, N*sizeof(int), IPC_CREAT|0700); buf = (int*) shmat(shmid, NULL, 0);}

void terminate() { sem_close(sem); shmctl(shmid, IPC_RMID, NULL);}

97

Remember

Always cleanup...

98

Implementation of semlib (semlib.c)

99

Implementation of semlib (semlib.c) (2)

100

Implementation of semlib (semlib.c) (3)



2006

/200

7

POSIX Threads

102

Threads

Flows of execution inside of a program They share all the address space of a process Each thread has its own stack and local variables

(even so, they can access other’s threads variables)Process

Thread 1 Thread 2 Thread 3

103

Why use threads?

They are very light weight compared to processes Light context switches Fast to create and terminate Fast to synchronize

Much easier to program than shared memory! Everything is already shared Be careful to synchronize accesses!

104

POSIX Threads – Thread Management

105

Simple thread creation example (simple_thread.c)

106

Running the example

gcc –lpthread –D_REENTRANT –Wall fich.c –o fich

107

Compiling with threads

Linux

-D_REENTRANT is quite important in LinuxThreads (Kernel 2.4) It instructs the compiler to use special re-entrant routine

functions If you don’t... it ONLY appears to work, until you get in

trouble!

gcc –lpthread –D_REENTRANT –Wall fich.c –o fich

Beware: Many routines are not re-entrant, they cannot be directly used with threads since they use common storage in an unsynchronized way (e.g. stktok())!

In some cases, there are re-entrant versions (e.g. strtok_r()). Check the manual! Don’t trust common sense.

108

Example of a non-reentrant routine

What happens if this is called from two different threads at the same time??

109

Things to beware of

Doesn’t work, “i” is on the stack and constantly changing

Doesn’t work, (1) after main() dies, its variables disappear – race condition with the starting threads; (2) main() dies everything dies!

110

If you need to terminate the main() thread...

This is OK!

Note: the other threads continue to execute.

111

Synchronization

Mutexes Provide mutual exclusion zones between threads In fact, these are just fast binary semaphores

POSIX Semaphores Used to signal events across threads Used to count objects in an synchronized way

Condition Variables Allow a thread to block or to notify others on any

condition Semaphores are a kind of condition variable:

the implicit condition is the semaphore being greater than 0

112

Mutexes

113

POSIX Semaphores

114

Producer/Consumer Revisited

Producer Consumer

Producer

Producer

115

prod_cons_threads.c

116

prod_cons_threads.c (2)

117

Which results in...

118

Synchronization – Condition Variables

Condition variables allow the programmer to suspend or notify a thread on any condition!

119

Synchronization – Condition Variables (2)

Important rule: A condition variable always has an associated mutex. Always check the condition variable in mutual exclusion.

The mutex must be locked.

How does it work?Makes thread Acheck its conditionagain

120

Synchronization – Condition Variables (3)

The thread tests a condition in mutual exclusion. If the condition is false, pthread_cond_wait() atomically releases de mutex AND waits until someone signals that the condition should be tested again.

When the condition is signaled AND the mutex is available, pthread_cond_wait() atomically reacquires the mutex AND releases the thread.

pthread_cond_signal() indicates that exactly one blocked thread should test the condition again. Note that this is note a semaphore. If there is no thread blocked, the “signal” is lost.

If all threads should re-check the condition, use pthread_cond_broadcast(). Since a mutex is involved, each one will test it one at a time, in mutual exclusion.

121

Example

Suppose a buffer that can hold a maximum of N elements. When it is full, it should immediately be emptied. While the buffer is being emptied, no thread can put things into it.

Producer

CleanerProducer

Producer

122

bounded_buffer.c

123

bounded_buffer.c (2)

124

With 3 producers and one cleaner...

125

Condition Variables – Rules (!)

The condition must always be tested with a while loop, never an if! Being unlocked out of a condition variable only means that the condition must be re-checked, not that it is true

The condition must always be checked and signaled inside a locked mutex.

While condition() may betrue while Thread B is executing,something may happen betweenthe time that the condition is signaledand Thread A is unblock (e.g. anotherthread may change the condition)

126

Synchronization – Some basic rules...

Never Interlock waits! Locks should always be taken in the same order Always release locks in the reverse order they have

been taken

sem_wait(A)sem_wait(B)

// Critical Section

sem_post(B)sem_post(A)

sem_wait(B)sem_wait(A)

// Critical Section

sem_post(A)sem_post(B)

Deadlock!

One way to assure that you always take locks in the same order is to create a lock hierarchy. I.e. associate a number to each lock using a table and always lock in increasing order using that table as reference (index).

127

Synchronization – Some basic rules... (2)

Sometimes it is not possible to know each order to take when locking (or using semaphores) Example: you are using

two resources owned by the operating system. They are controlled by locks. You cannot be sure another application is not using exactly the same resources and locking in reverse order.

In that case, use pthread_mutex_trylock() or sem_trywait() and back off if you are unsuccessful. Allow the system to make

progress and not deadlock

128


Mutexes are for implementing mutual exclusion, not for signaling across threads! One the thread that has locked a mutex can unlock it. Not doing

so will probably result in a core dump! To signal across threads use semaphores!

lock(&m)

unlock(&m)

lock(&m)

unlock(&m)

CORRECT!

lock(&m)

unlock(&m)

INCORRECT!(Use a semaphore for this!)

A B A B

(in mutual exclusion)

(in mutual exclusion)

(blocked)

129


Don’t mess with thread priorities unless you really need to. Although apparently “correct” and “logical” in design, it’s quite

easy to make a system deadlock, live-lock or some threads to starve.

130

The Mars PathFinder Problem

Priority Inversion Problem (in Mars Path Finder): Low priority thread locks a

semaphore. High priority thread starts to

execute and tries to lock the same semaphore. It’s suspended since it cannot violate the other thread lock.

Medium priority threads comes to execute and preempts the low priority one. Since it doesn’t need the semaphore, it continues to execute.

Meanwhile the high priority one is starving. After a while, a watchdog timer detects that the high priority one is not executing and resets the machine.

A

B

C

starts toexecuteand getsthe lock

starts toexecute

tries to get the lock and is suspended

continuesto execute

is preempted asmedium prioritygets to execute

watchdog timer resets the machine as the high priorityone doesn’t get toexecute

131

To Learn More

Advanced Programming in the UNIX Environment, 2nd Edition by W. Richard Stevens, Stephen A. Rago

Addison-Wesley, June 2005

UNIX Network Programming, Volume 2: Interprocess Communications, 2nd Ed. by W. Richard Stevens

Prentice Hall, August 1998

Unix Systems Programming: Communication, Concurrency and Threads, 2nd Edition by Kay Robbins, Steve Robbins

Prentice Hall, June 2003

132

IMPORTANT NOTICE

YOU ARE FREE TO USE THIS MATERIAL FOR YOUR PERSONAL LERNING OR REFERENCE, DISTRIBUTE IT AMONG COLLEGUES OR EVEN USE IT FOR TEACHING CLASSES. YOU MAY EVEN MODIFY IT, INCLUDING MORE INFORMATION OR CORRECTING STANDING ERRORS.

THIS RIGHT IS GIVEN TO YOU AS LONG AS YOU KEEP THIS NOTICE AND GIVE PROPER CREDIT TO THE AUTHOR. YOU CANNOT REMOVE THE REFERENCES TO THE AUTHOR OR TO THE INFORMATICS ENGINEERING DEPARTMENT OF THE UNIVERSITY OF COIMBRA.

(c) 2005 – Paulo Marques, [email protected]

Documents

Interprocess Communication in Unix Paulo Marques Department of Informatics Engineering University of Coimbra [email protected] 2006/2007 An Introduction