way to use a computer system is to spread a given computation over several processes.
The need for such communicating processes arises in parallel and distributed processing
contexts. Often it is possible to partition a computational task into segments which can be
distributed amongst several processes. Clearly, these processes would then form a set of
communicating processes which cooperate in advancing a solution. In a highly
distributed, multi-processor system, these processes may even be resident on different
machines. In such a case the communication is supported over a network. A
comprehensive coverage of distributed systems is beyond the scope of this book. There
are texts like Tanenbaum and Steen [8] which are exclusively devoted to this topic. All
the same, we shall study some of the basics to be able to do the following:
�� How to spawn (or create) a new process.
�� How to assign a task for execution to this newly spawned process.
�� A few mechanisms to enable communication amongst processes.
�� Synchronization amongst these processes.
In most cases an IPC package is used to establish inter-process communication.
Depending upon the nature of the chosen IPC, the package sets up a data structure in
kernel space. These data structures are often persistent. So once the purpose of IPC has
been fulfilled, this set-up needs to be deleted (a clean up operation). The usage pattern of
the IPC package in a system (like system V Unix) can be seen by using explicit
commands like ipcs. A user can also remove any unused kernel resources by using a
command like ipcrm.
For our discussions here, we shall assume Unix like environment. Hopefully, the
discussion here offers enough information to partially satisfy and raise the level of
curiosity about the distributed computing area.
7.1 Creating A New Process: The fork() System Call
One way to bring in a new process into an existing execution environment is to execute
fork() system call. Just to recap how system calls are handled, the reader may refer to
Figure 7.1. An application raises a system call using a library of call functions. A system
call in turn invokes its service (from the kernel) which may result in memory allocation,
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/2
device communication or a process creation. The system call fork() spawns a new process
which, in fact, is a copy of the parent process from where it was invoked!! The newly
spawned process inherits its parent's execution environment. In Table 7.1 we list some of
the attributes which the child process inherits from its parent.
Note that a child process is a process in its own right. It competes with the parent process
to get processor time for execution. In fact, this can be easily demonstrated (as we shall
later see). The questions one may raise are:
�� Can one identify when the processor is executing the parent and when it is
executing the child process?
�� What is the nature of communication between the child and parent processes?
The answer to the first question is yes. It is possible to identify when the parent or child is
in execution. The return value of fork() system call is used to determine this. Using the
return value, one can segment out the codes for execution in parent and child. We will
show that in an example later.
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/3
The most important communication from parent to child is the execution environment
which includes data and code segments. Also, when the child process terminates, the
parent process receives a signal. In fact, a signal of the termination of a child process, is
one feature very often exploited by programmers. For instance, one may choose to keep
parent process in wait mode till all of its own child processes have terminated. Signaling
is a very powerful inter-process communication mechanism (using signals) which we
shall learn in Section 7.3.5. The following program demonstrates how a child process
may be spawned.
The program: Demonstration of the use of fork() system call
main()
{ int i, j;
if ( fork() ) /* must be parent */
{ printf("\t\t In Parent \n");
printf("\t\t pid = %d and ppid = %d \n\n", getpid(), getppid());
for (i=0; i<100; i=i+5)
{ for (j=0; j<100000; j++);
printf("\t\t\t In Parent %d \n", i);
}
wait(0); /* wait for child to terminate */
printf("In Parent: Now the child has terminated \n");
}
else
{ printf("\t In child \n");
printf("\t pid = %d and ppid = %d \n\n", getpid(), getppid() );
for (i=0; i<100; i=i+10)
{ for (j=0; j<100000; j++);
printf("\t In child %d \n", i);
} } }
The reader should carefully examine the structure of the code above. In particular, note
how the return value of system call fork() is utilized. On perusing the code we note that,
the code is written to execute in different parts of the program code for the child and the
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/4
parent. The program makes use of true return value of fork() to print “In parent", i.e. if
the parent process is presently executing. The dummy loop not only slows down the
execution but also ensures that we obtain interleaved outputs with a manageable number
of lines on the viewing screen.
Response of this program:
[bhatt@iiitbsun IPC]$./a.out
In child
pid = 22484 and ppid = 22483
In child 0
In child 10
In child 20
In Parent
pid = 22483 and ppid = 22456
In Parent 0
In Parent 5
In Parent 10
In Parent 15
In child 30
.......
.......
In child 90
In Parent 20
In Parent 25
......
......
In Parent: Now the child has terminated;
Let us study the response. From the response, we can determine when the parent process
was executing and when the child process was executing. The final line shows the result
of the execution of line following wait command in parent. It executes after the child has
fallen through its code. Just as we used a wait command in the parent, we could have also
used an exit command explicitly in the child to exit its execution at any stage. The
command pair wait and exit are utilized to have inter-process communication. In
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/5
particular, these are used to synchronize activities in processes. This program
demonstrated how a process may be spawned. However, what one would wish to do is to
spawn a process and have it execute a planned task. Towards this objective, we shall next
populate the child code segment with a code for a specified task.
7.2 Assigning Task to a Newly Spawned Process
By now one thing should be obvious: if the child process is to execute some other code,
then we should first identify that executable (the one we wish to see executed). For our
example case, let us first generate such an executable. We compile a program entitled
get_int.c with the command line cc get_int.c -o int.o. So, when int.o executes, it reads in
an integer.
The program to get an integer :
#include
#include
int get_integer( n_p )
int *n_p;
{ int c;
int mul, sign;
int integer_part;
*n_p = 0;
mul = 10;
while( isspace( c = getchar() ) ); /* skipping white space */
if( !isdigit(c) && c != EOF && c != '+' && c != '-' )
{ /* ungetchar(c); */
printf("Found an invaild character in the integer description \n");
return 0;
}
if (c == '-') sign = -1.0;
if (c == '+') sign = 1.0;
if (c == '-' || c == '+' ) c = getchar();
for ( integer_part = 0; isdigit(c); c = getchar() )
{ integer_part = mul * integer_part + (c - '0');
};
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/6
*n_p = integer_part;
if ( sign == -1 ) *n_p = - *n_p;
if ( c == EOF ) return (*n_p);
}
main()
{ int no;
int get_integer();
printf("Input a number as signed or unsigned integer e.g. +5 or -6 or 23\n");
get_integer(&no);
printf("The no. that was input was %d \n", no);
}
Clearly, our second step is to have a process spawned and have it execute the program
int.o. Unix offers a way of directing the execution from a specified code segment by
using an exec command. In the program given below, we spawn a child process and
populate its code segment with the program int.o obtained earlier. We shall entitle this
program as int_wait.c.
Program int_wait.c
#include
main()
{
if (fork() == 0)
{ /* In child process execute the selected command */
execlp("./int.o", "./int.o", 0);
printf("command not found \n"); /* execlp failed */
fflush(stdout);
exit(1);
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/7
}
else
{ printf("Waiting for the child to finish \n");
wait(0);
printf("Waiting over as child has finished \n");
}
}
To see the programs in action follow the steps:
1. cc get_int.c -o int.o
2. cc int_wait.c
3. ./a.out
The main point to note here is that the forked child process gets populated by the code of
program int.o with the parent int_wait.c. Also, we should note the arguments
communicated in the exec command line.
Before we discuss some issues related to the new execution environment, a short
discussion on exec command is in order. The exec family of commands comes in several
flavors. We may choose an exec command to execute an identified executable defined
using a relative or absolute path name. The exec() command may use some other
arguments as well. Also, it may be executed with or without the inherited execution
environment.
Most Unix systems support exec commands with the description in Table 7.2. The
example above raises a few obvious questions. The first one is: Which are the properties
the child retains after it is populated by a different code segment? In Table 7.3 we note
that the process ID and user ID of the child process are carried over to the implanted
process. However, the data and code segments obtain new information. Though, usually,
a child process inherits open file descriptors from the parent, the implanted process may
have some restrictions based on file access controls.
With this example we now have a way to first spawn and then populate a child process
with the code of an arbitrary process. The implanted process still remains a child process
but has its code independent of the parent. A process may spawn any number of child
processes. However, much ingenuity lies in how we populate these processes and what
form of communication we establish amongst these to solve a problem.
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/8
7.3 Establishing Inter-process Communication
In this section we shall study a few inter-process communication mechanisms. Each of
these uses a different method to achieve communication amongst the processes. The first
mechanism we study employs pipes. Pipes, as used in commands like ls|more, direct the
output stream of one process to feed the input of another process. So for IPC, we need to
create a pipe and identify the direction of feeding the pipe. Another way to communicate
would be to use memory locations. We can have one process write into a memory
location and expect the other process to read from it. In this case the memory location is a
shared memory location. Finally, there is one more mechanism in which one may send a
message to another process. The receiving process may interpret the message. Usually,
the messages are used to communicate an event. We next study these mechanisms.
7.3.1 Pipes as a Mechanism for Inter-process Communication
Let us quickly recap the scheme which we used in section 7.2. The basic scheme has
three parts: spawn a process; populate it and use the wait command for synchronization.
Let us now examine what is involved in using pipes for establishing a communication
between two processes. As a first step we need to identify two executables that need to
communicate. As an example, consider a case where one process gets a character string
input and communicates it to the other process which reverses strings. Then we have two
processes which need to communicate. Next we define a pipe and connect it between the
processes to facilitate communication. One process gets input strings and writes into the
pipe. The other process, which reverses strings, gets its input (i.e. reads) from the pipe.
Figure 7.2 explains how the pipes are used. As shown in the upper part of the figure, a
pipe has an input end and an output end. One can write into a pipe from the input end and
read from the output end. A pipe descriptor, therefore, has an array that stores two
pointers. One pointer is for its input end and the other is for its output end. When a
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/9
process defines a pipe it gets both the addresses, as shown in the middle part of Figure
7.2. Let us suppose array pp is used to store the descriptors. pp[0] stores the write end
address and pp[1] stores the read end address. Suppose two processes, Process A and
Process B, need to communicate, then it is imperative that the process which writes
closes its read end of the pipe and the process which read closes its write end of the pipe.
Essentially, for a communication from Process A to process B the following should
happen. Process A should keep its write end open and close read end of the pipe.
Similarly, Process B should keep its read end open and close its write end. This is what is
shown in the lower part of Figure 7.2. Let us now describe how we may accomplish this.
1. First we have a parent process which declares a pipe in it.
2. Next we spawn two child processes. Both of these would get the pipe definition
which we have defined in the parent. The child processes, as well as the parent,
have both the write and read ends of the pipe open at this time.
3. Next, one child process, say Process A, closes its read end and the other child
process, Process B, closes its write end.
4. The parent process closes both write and read ends.
5. Next, Process A is populated with code to get a string and Process B is populated
to reverse a string.
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/10
With the above arrangement the output from Process A is piped as input to Process B.
The programs given below precisely achieve this.
In reading the programs, the following interpretations have to be borne in mind:
1. The pipe is defined by the declaration pipe(p_des).
2. The dup command replaces the standard I/O channels by pipe descriptors.
3. The execlp command is used to populate the child process with the desired code.
4. The close command closes the appropriate ends of the pipe.
5. The get_str and rev_str processes are pre-compiled to yield the required
executables.
The reader should be able to now assemble the programs correctly to see the operation of
the programs given below:
pipe.c
#include
#include
main()
{ int p_des[2];
pipe( p_des ); /* The pipe descriptor */
printf("Input a string \n");
if ( fork () == 0 )
{
dup2(p_des[1], 1);
close(p_des[0]); /* process-A closing read end of the pipe */
execlp("./get_str", "get_str", 0);
/*** exit(1); ***/
}
else
if ( fork () == 0 )
{ dup2(p_des[0], 0);
close(p_des[1]); /* process-B closing write end of the pipe */
execlp("./rev_str", "rev_str", 0);
/*** exit(1); ****/
}
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/11
else
{ close(p_des[1]); /* parent closing both the ends of pipe */
close(p_des[0]);
wait(0);
wait(0);
}
fflush(stdout);
}
get_str.c
#include
#include
void get_str(str)
char str[];
{ char c;
int ic;
c = getchar();
ic = 0;
while ( ic < 10 && ( c != EOF && c != '\n' && c != '\t' ))
{ str[ic] = c;
c = getchar();
ic++;
}
str[ic] = '\0';
return;
}
rev_str.c
void rev_str(str1, str2)
char str1[];
char str2[];
{ char c;
int ic;
int rc;
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/12
ic = 0;
c = str1[0];
while( ic < 10 && (c != EOF && c != '\0' && c != '\n') )
{ ic++;
c = str1[ic];
}
str2[ic] = '\0';
rc = ic - 1;
ic = 0;
while (rc-ic > -1)
{ str2[rc-ic] = str1[ic];
ic++;
}
return;
}
It is important to note the following about pipes as an IPC mechanism:
1. Unix pipes are buffers managed from within the kernel.
2. Note that as a channel of communication, a pipe operates in one direction only.
3. Some plumbing (closing of ends) is required to use a pipe.
4. Pipes are useful when both the processes are schedulable and are resident on the
same machine. So, pipes are not useful for processes across networks.
5. The read end of a pipe reads any way. It does not matter which process is
connected to the write end of the pipe. Therefore, this is a very insecure mode of
communication.
6. Pipes cannot support broadcast.
There is one other method of IPC using special files called “named pipes". We shall leave
out its details. Interested readers should explore the suggested reading list of books. In
particular, books by Stevenson, Chris Brown or Leach [28], [12], [24] are recommended.
7.3.2 Shared Files
One very commonly employed strategy for IPC is to share files. One process, identified
as a writer process, writes into a file. Another process, identified as a reader process,
reads from this file. The write process may continually make changes in a file and the
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/13
other may read these changes as these happen. Unlike other IPC methods, this method
does not require special system calls. It is, therefore, relatively easily portable. Of course,
for creating processes we shall use the standard system calls fork() and execlp(). Besides
these, there are no other system calls needed. However, we do need code for file creation,
access and operations on files. A word of caution is in order. If the reader is faster than
the writer, then this method shall have errors. Similarly, if a writer continues writing then
the file may grow to unbounded lengths. Both these situations result in errors. This
problem of a mismatch in the speed of reader and writer is called the reader writer
problem. We earlier learned to resolve similar problems using mutual exclusion of
resource sharing. In this case too we can program for mutually exclusive writes and
reads.
Shared file pointers: Another way to handle files would be to share file pointers instead
of files themselves. Sharing the file pointers with mutual exclusion could be easily done
using semaphores.
The shared file pointer method of IPC operates in two steps. In the first step, one process
positions a file pointer at a location in a file. In the second step, another process reads
from this file from the communicated location. Note that if the reader attempts to read a
file even before the writer has written something on a file, we shall have an error. So, in
our example we will ensure that the reader process sleeps for a while (so that the writer
has written some bytes). We shall use a semaphore simulation to achieve mutual
exclusion of access to the file pointer, and hence, to the file.
This method can be used when the two processes are related. This is because the shared
file pointer must be available to both. In our example, these two processes shall be a
parent and its child. Clearly, if a file has been opened before the child process is
spawned, then the file descriptors created by the parent are available to the child process
as well. Note that when a process tries to create a file which some other process has
already created, then an error is reported.
To understand the programs in the example, it is important to understand some
instructions for file operations. We shall use lseek() system command. It is used to access
a sequence of bytes from a certain offset in the file. The first byte in the file is considered
to have an offset of 0. It has the syntax long lseek(int fd, long offset, int arg) with the
following interpretation.
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/14
�� With arg = 0, the second argument is treated as an offset from the first byte in file.
�� With arg = 1, the current position of the file pointer is changed to sum of the
current file pointer and the value of the second argument.
�� With arg = 2, the current position of the file pointer is changed to the sum of the
size of file and value of the second argument. The value of the second argument
can be negative as long as the overall result of the sum is positive or zero.
The example here spans three programs, a main, a reader and a writer program. Let us
look at the code for the main program.
#include
#include
#define MAXBYTES 4096
void sem_simulation();
main(argc, argv)
int argc;
char *argv[];
{/* the program communicates from parent to child using a shared file pointer */
FILE *fp;
char message[MAXBYTES];
long i;
int mess_num, n_bytes, j, no_of_mess;
int sid, status;
if ( argc < 3 )
{ fputs("Bad argument count \n", stderr);
fputs("Usage: num_messages num_bytes \n", stderr);
exit(1);
}
no_of_mess = atoi(argv[1]);
n_bytes = atoi(argv[2]);
printf("no_of_mess : %6d and n_bytes : %6d \n", no_of_mess, n_bytes );
if(n_bytes > MAXBYTES)
{ fputs("Number of bytes exceeds maximum", stderr);
exit(1);
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/15
} /* open a file before creating a child process to share a file pointer*/
else if( ( fp = fopen("./temp_file", "w+" )) == NULL )
{ fputs("Cannot open temp_file for writing \n", stderr);
exit(1);
}
/* create processes and begin communication */
switch (fork ())
{ case -1: fputs("Error in fork ", stderr);
exit( 1 );
case 0: sleep(2);
if(execlp("./readfile", "./readfile", argv[1], argv[2], NULL) == -1)
fputs("Error in exec in child \n", stderr);
exit( 1 );
default: if(execlp("./writefile", "./writefile", argv[1], argv[2], NULL) == -1)
fputs("Error in exec in parent \n", stderr);
exit( 1 );
} /* end switch */
}
Now we describe the reader process.
#include
#include
#define MAXBYTES 4096
void sem_simulation()
{ if (creat( "creation", 0444) == -1)
{ fputs("Error in create \n", stderr);
system("rm creation");
}
else fputs(" No error in creat \n", stderr);
}
main (argc, argv)
int argc;
char *argv[];
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/16
{ FILE *fp;
long i;
char message[MAXBYTES];
int mess_num, n_bytes, j, no_of_mess;
int sid, status;
void sem_simulation();
no_of_mess = atoi(argv[1]);
n_bytes = atoi(argv[2]);
printf("in read_child \n");
/* read messages from the shared file */
for ( i=0; i < no_of_mess; i++ )
{ sem_simulation();
fseek(fp, i*n_bytes*1L, 0);
while((fgets(message, n_bytes+1, fp)) == NULL ) ;
fseek(fp, i*n_bytes*1L, 0);
sem_simulation();
} /* end of for loop */
exit(0);
}
Now let us describe the writer process.
#include
#include
#define MAXBYTES 4096
void sem_simulation()
{ if (creat( "creation", 0444) == -1)
{ fputs("Error in create \n", stderr);
system("rm creation");
}
else fputs(" No error in create \n", stderr);
}
main (argc, argv)
int argc;
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/17
char *argv[];
{ FILE *fp;
long i, j, status, message_num;
char message[MAXBYTES];
int n_bytes, no_of_mess;
void sem_simulation();
no_of_mess = atoi(argv[1]);
n_bytes = atoi(argv[2]);
printf("in parent with write option \n");
printf("no_of_mess : %6d n_bytes : %6d \n");
for ( i=0; i < no_of_mess; i++ )
{ /* Create a message with n_bytes */
message_num = i;
for ( j = message_num; j < n_bytes; j++ )
message[j] = 'd';
printf("%s \n", message);
/* Use semaphore to control synchronization, write to end of file */
sem_simulation();
fseek(fp, 0L, 2);
while( ( fputs(message, fp) ) == -1)
fputs("Cannot write message", stderr );
fseek(fp, 0L, 2);
sem_simulation();
}
wait(&status);
unlink("creation");
unlink("./temp_file");
fclose(fp);
}
The shared file pointer method is quite an elegant solution and is often a preferred
solution where files need to be shared. However, many parallel algorithms require that
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/18
“objects" be shared. The basic concept is to share memory. Our discussion shall now veer
to IPC using shared memory communication.
7.3.3 Shared Memory Communication
Ordinarily, processes use memory areas within the scope of virtual memory space.
However, memory management systems ensure that every process has a well-defined and
distinct data and code area. For shared memory communication, one process would write
into a certain commonly accessed area and another process would read subsequently from
that area. One other point which we can debate is: do the processes have to be related?
We have seen that a parent may share a data area or files with a child. Also, by using the
exec() function call we may be able to populate a process with another code segment or
data. Clearly, the shared memory method can allow access to a common data area even
amongst the processes that are not related. However, in that case an area like a process
stack may not be shareable. Also, it should be noted that it is important that the shared
data integrity may get compromised when an arbitrary sequence of reads and writes
occurs. To maintain data integrity, the access is planned carefully under a user program
control. That then is the key to shared memory protocol.
The shared memory model has the following steps of execution.
1. First we have to set up a shared memory mechanism in the kernel.
2. Next an identified \safe area" is attached to each of the processes.
3. Use this attached shared data space in a consistent manner.
4. When finished, detach the shared data space from all processes to which it was
attached.
5. Delete the information concerning the shared memory from the kernel.
Two important .h files in this context are: shm.h and ipc.h which are included in all the
process definitions. The first step is to set up shared memory mechanism in kernel. The
required data structure is obtained by using shmget() system call with the following
syntax.
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/19
int shmget( key_t key, int size, int flag );
The parameter key_t is usually a long int. It is declared internally as key_t key. key_t is
an alias defined in sys/types.h using a typedef structure. If this key is set to
IPC_PRIVATE, then it always creates a shared memory region. The second parameter,
size is the size of the sh-mem-region in bytes. The third parameter is a combination of
usual file access permissions of r/w/e for o/g/w with the interpretation of non-zero
constants as explained in Table 7.4.
A successful call results in the creation of a shared memory data structure with a defined
id. This data structure has the following information in it.
struct shmid_ds
{ struct ipc_perm shm_perm;
int shm_seg_segsz /* size of segments in bytes */
struct region *shm_reg; /* pointer to region struct */
char pad[4]; /* for swap compatibility */
ushort shm_lpid; /* pid of last shmop */
ushort shm_cpid; /* pid of creator */
ushort shm_nattch; /* used for shm_info */
ushort shm_cnattch; /* used for shm_info */
time_t shm_atime; /* last attach time */
time_t shm_dtime; /* last detach time */
time_t shm_ctime; /* last change time */
}
Once this is done we would have created a shared memory data space. The next step
requires that we attach it to processes that would share it. This can be done using the
system call shmat(). The system call shmat() has its syntax shown below.
char *shamt( int shmid, char *shmaddr, int shmflg );
The second argument should be set to zero as in (char *)0, if the kernel is to determine
the attachment. The system uses three possible flags which are: SHM_RND,
SHM_RDONLY and the combination SHM_RND | SHM_RDONLY. The
SHM_RDONLY flag indicates the shared region is read only. Otherwise, it is both for
read and write operations. The flag SHM_RND requires that the system enforces use of
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/20
the byte address of the shared memory region to coincide with a double word boundary
by rounding.
Now that we have a well-defined shared common area, reading and writing can be done
in this shared memory region. However, the user must write a code to ensure locking of
the shared region. For instance, we should be able to block a process attempting to write
while a reader process is reading. This can be done by using a synchronization method
such as semaphores. In most versions of Unix, semaphores are available to enforce
mutual exclusion. At some stage a process may have finished using the shared memory
region. In that case this region can be detached for that process. This is done by using the
shmdt() system call. This system call detaches that process from future access. This
information is kept within the kernel data-space. The system call shmdt() takes a single
argument, the address of the shared memory region. The return value from the system
call is rarely used except to check if an error has occurred (with -1 as the return value).
The last step is to clean up the kernel's data space using the system call shmctl(). The
system call shmctl() takes three parameters as input, a shared memory id, a set of flags,
and a buffer that allows copying between the user and the kernel data space.
A considerable amount of information is pointed to by the third parameter. A call to
shmctl() with the command parameter set to IPC_STAT gives the following information.
�� User's id
�� Creator's group id
�� Operation permissions
�� Key
�� segment size
�� Process id of creator *
�� Current number of attached segments in the memory.
�� Last time of attachment
�� User's group id
�� Creator's id
�� Last time of detachment
�� Last time of change
�� Current no. of segments attached
�� Process id of the last shared memory operation
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/21
Now let us examine the shmget() system call.
int shmget( key_t key, int region_size, int flags );
Here key is a user-defined integer, the size of the shared region to be attached is in bytes.
The flags usually turn on the bits in IPC_CREAT. Depending upon whether there is key
entry in the kernel's shared memory table, the shmget() call takes on one of the following
two actions. If there is an entry, then shmget() returns an integer indicating the position of
the entry. If there is no entry, then an entry is made in the kernel's shared memory table.
Also, note that the size of the shared memory is specified by the user. It, however, should
satisfy some system constraints which may be as follows.
struct shminfo
{ int shmmax, /* Maximum shared memory segment size 131072 for some */
shmmin, /* minimum shared memory segment size 1 for some */
shmni, /* No. of shared memory identifiers */
shmseg, /* Maximum attached segments per process */
shmall; /* Max. total shared memory system in pages */
};
The third parameter in shmget() corresponds to the flags which set access permissions as
shown below:
400 read by user ...... Typically in shm.h file as constant SHM_R
200 write by user .......Typically in shm.h file as constant SHM_W
040 read by group
020 write by group
004 read by others
002 read by others ......All these are octal constants.
For example, let us take a case where we have read/write permissions by the user's group
and no access by others. To be able to achieve this we use the following values.
SHM_R | SHM_W | 0040 | IPC_CREAT as a flag to a call to shmget().
Now consider the shmat() system call.
char *shmat( int shmid, char *address, int flags );
This system call returns a pointer to the shared memory region to be attached. It must be
preceded by a call to shmget(). The first argument is a shmid (returned by shmget()). It is
an integer. The second argument is an address. We can let the compiler decide where to
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/22
attach the shared memory data space by giving the second argument as (char *) 0. The
flags in arguments list are to communicate the permissions only as SHM_RND and
SHM_RDONLY. The shmdt() system call syntax is as follows:
int shmdt(char * addr );
This system call is used to detach. It must follow a call shmat() with the same base
address which is returned by shmat(). The last system call we need is shmctl(). It has the
following syntax.
int shmctl( int shmid, int command, struct shm_ds *buf_ptr );
The shmctl() call is used to change the ownership and permissions of the shared region.
The first argument is the one earlier returned by shmget() and is an integer. The command
argument has five possibilities:
• IPC_STAT : returns the status of the associated data structure for the shared
memory pointed by buffer pointer.
• IPC_RMID : used to remove the shared memory id.
• SHM_LOCK : used to lock
• SHM_UNLOCK : used to unlock
• IPC_SET : used to set permissions.
When a region is used as a shared memory data space it must be from a list of free data
space. Based on the above explanations, we can arrive at the code given below.
include
#include
#include
#include
#include
#include
#define MAXBYTES 4096 /* Maximum bytes per shared segment */
main(argc, argv)
int argc;
char *argv[];
{ /* Inter process communication using shared memory */
char message[MAXBYTES];
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/23
int i, message_num, j, no_of_mess, nbytes;
int key = getpid();
int semid;
int segid;
char *addr;
if (argc != 3) { printf("Usage : %s num_messages");
printf("num_of_bytes \n", argv[0]);
exit(1);
}
else
{ no_of_mess = atoi(argv[1]);
nbytes = atoi(argv[2]);
if (nbytes > MAXBYTES) nbytes = MAXBYTES;
if ( (semid=semget( (key_t)key, 1, 0666 | IPC_CREAT ))== -1)
{ printf("semget error \n");
exit(1);
}
/* Initialise the semaphore to 1 */
V(semid);
if ( (segid = shmget( (key_t) key, MAXBYTES, 0666 |
IPC_CREAT ) ) == -1 )
{ printf("shmget error \n");
exit(1);
}
/*if ( (addr = shmat(segid, (char * )0,0)) == (char *)-1) */
if ( (addr = shmat(segid, 0, 0)) == (char *) -1 )
{ printf("shmat error \n");
exit(1);
}
switch (fork())
{ case -1 : printf("Error in fork \n");
exit(1);
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/24
case 0 : /* Child process, receiving messages */
for (i=0; i < no_of_mess; i++)
if(receive(semid, message, sizeof(message)));
exit(0);
default : /* Parent process, sends messages */
for ( i=0; i < no_of_mess; i++)
{ for ( j=i; j < nbytes; j++)
message[j] = 'd';
if (!send(semid, message, sizeof(message)))
printf("Cannot send the message \n");
} /* end of for loop */
} /* end of switch */
} /* end of else part */
}
/* Semaphores */
#include
#include
#include
#include
#include
int sid;
cleanup(semid, segid, addr)
int semid, segid;
char *addr;
{ int status;
/* wait for the child process to die first */
/* removing semaphores */
wait(&status);
semctl(semid, 0, IPC_RMID, 0);
shmdt(addr);
shmctl(segid, 0, IPC_RMID, 0);
};
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/25
P(sid)
int sid;
{ /* Note the difference in this and previous structs */
struct sembuf *sb;
sb = (struct sembuf *) malloc(sizeof(struct sembuf *));
sb -> sem_num = 0;
sb -> sem_op = -1;
sb -> sem_flg = SEM_UNDO;
if( (semop(sid, sb, 1)) == -1) puts("semop error");
};
V(sid)
int sid;
{ struct sembuf *sb;
sb = (struct sembuf *) malloc(sizeof(struct sembuf *));
sb -> sem_num = 0;
sb -> sem_op = 1;
sb -> sem_flg = SEM_UNDO;
if( (semop(sid, sb, 1)) == -1) puts("semop error");
};
/* send message from addr to buf */
send(semid, addr, buf, nbytes)
int semid;
char *addr, *buf;
int nbytes;
{ P(semid);
memcpy(addr, buf, nbytes);
V(semid);
}
/* receive message from addr to buf */
receive(semid, addr, buf, nbytes)
int semid;
char *addr, *buf;
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/26
int nbytes;
{ P(semid);
memcpy(buf, addr, nbytes);
V(semid);
}
From the programs above, we notice that any process is capable of accessing the shared
memory area once the key is known to that process. This is one clear advantage over any
other method. Also, within the shared area the processes enjoy random access for the
stored information. This is a major reason why shared memory access is considered
efficient. In addition, shared memory can support many-to-many communication quite
easily. We shall next explore message-based IPC.
7.3.4 Message-Based IPC
Messages are a very general form of communication. Messages can be used to send and
receive formatted data streams between arbitrary processes. Messages may have types.
This helps in message interpretation. The type may specify appropriate permissions for
processes. Usually at the receiver end, messages are put in a queue. Messages may also
be formatted in their structure. This again is determined by the application process.
Messages are also the choice for many parallel computers such as Intel's hyper-cube. The
following four system calls achieve message transfers amongst processes.
�� msgget() returns (and possibly creates) message descriptor(s) to designate a
message queue for use in other systems calls.
�� msgctl() has options to set and return parameters associated with a message
descriptor. It also has an option to remove descriptors.
�� msgsnd() sends a message using a message queue.
�� msgrcv() receives a message using a message queue.
Let us now study some details of these system calls.
msgget() system call : The syntax of this call is as follows:
int msgget(key_t key, int flag);
The msgget() system call has one primary argument, the key, a second argument which is
a flag. It returns an integer called a qid which is the id of a queue. The returned qid is an
index to the kernel's message queue data-structure table. The call returns -1 if there is an
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/27
error. This call gets the resource, a message queue. The first argument key_t, is defined in
sys/types.h file as being a long. The second argument uses the following flags:
�� MSG_R : The process has read permission
�� MSG_W : The process has write permission
�� MSG_RWAIT : A reader is waiting to read a message from message queue
�� MSG_WWAIT : A writer is waiting to write a message to message queue
�� MSD_LOCKED : The msg queue is locked
�� MSG_LOCKWAIT : The msg queue is waiting for a lock
�� IPC_NOWAIT : Described earlier
�� IPC_EXCL : ....
In most cases these options can be used in bit-ored manner. It is important to have the
readers and writers of a message identify the relevant queue for message exchange. This
is done by associating and using the correct qid or key. The key can be kept relatively
private between processes by using a makekey() function (also used for data encryption).
For simple programs it is probably sufficient to use the process id of the creator process
(assuming that other processes wishing to access the queue know it). Usually, kernel uses
some algorithm to translate the key into qid. The access permissions for the IPC methods
are stored in IPC permissions structure which is a simple table. Entries in kernel's
message queue data structures are C structures. These resemble tables and have several
fields to describe permissions, size of queue, and other information. The message queue
data structure is as follows.
struct meqid_ds
{ struct ipc_perm meg_perm; /* permission structure */
struct msg *msg_first; /* pointer to first message */
struct msg *msg_last; /* ........... last ..........*/
ushort msg_cbytes; /* no. of bytes in queue */
ushort msg_qnum; /* no. of messages on queue */
ushort msg_qbytes; /* Max. no. of bytes on queue */
ushort msg_lspid; /* pid of last msgsnd */
ushort msg_lrpid; /* pid of the last msgrcv */
time_t msg_stime; /* last msgsnd time */
time_t msg_rtime; /* .....msgrcv................*/
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/28
time_t msg_ctime; /* last change time */
}
There is one message structure for each message that may be in the system.
struct msg
{ struct msg *msg_next; /* pointer to next message */
long msg_type; /* message type */
ushort msg_ts; /* message text size */
ushort msg_spot; /* internal address */
Note that several processes may send messages to the same message queue. The “type" of
message is used to determine which process amongst the processes is the originator of the
message received by some other process. This can be done by hard coding a particular
number for type or using process-id of the sender as the msg_type. The msgctl() function
call: This system call enables three basic actions. The most obvious one is to remove
message queue data structure from the kernel. The second action allows a user to
examine the contents of a message queue data structure by copying them into a buffer in
user's data area. The third action allows a user to set the contents of a message queue data
structure in the kernel by copying them from a buffer in the user's data area. The system
call has the following syntax.
int msgctl(int qid, int command, struct msqid_ds *ptr);
This system call is used to control the resource (a message queue). The first argument is
the qid which is assumed to exist before call to msgctl(). Otherwise the system is in error
state. Note that if msgget() and msgctl() are called by two different processes then there is
a potential for a \race" condition to occur. The second argument command is an integer
which must be one of the following constants (defined in the header file sys/msg.h).
�� IPC STAT: Places the contents of the kernel structure indexed by the first
argument, qid, into a data structure pointed to by the third argument, ptr. This
enables the user to examine and change the contents of a copy of the kernel's data
structure, as this is in user space.
�� IPC SET: Places the contents of the data structures in user space pointed to by the
third argument, ptr, into the kernel's data structure indexed by first argument qid,
thus enabling a user to change the contents of the kernel's data structure. The only
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/29
fields that a user can change are msg_perm.uid, msg_perm.gid, msg_perm.mode,
and msg_qbytes.
�� IPC RMID : Removes the kernel data structure entry indexed by qid.
The msgsnd() and msgrcv() system calls have the following syntax.
int msgsnd(int qid, struct msgbuf *msg_ptr, int message_size, int flag );
int msgrcv(int qid, struct msgbuf *msg_ptr, int message_size, int msgtype, int flag );
Both of these calls operate on a message queue by sending and receiving messages
respectively. The first three arguments are the same for both of these functions. The
syntax of the buffer structure is as follows.
struct msgbuf{ long mtype; char mtext[1]; }
This captures the message type and text. The flags specify the actions to be taken if the
queue is full, or if the total number of messages on all the message queues exceeds a
prescribed limit. With the flags the following actions take place. If IPC_NOWAIT is set,
no message is sent and the calling process returns without any error action. If
IPC_NOWAIT is set to 0, then the calling process suspends until any of the following
two events occur.
1. A message is removed from this or from other queue.
2. The queue is removed by another process. If the message data structure indexed
by qid is removed when the flag argument is 0, an error occurs (msgsnd() returns
-1).
The fourth arg to msgrcv() is a message type. It is a long integer. The type argument is
used as follows.
o If the value is 0, the first message on the queue is received.
o If the value is positive, the queue is scanned till the first message of this
type is received. The pointer is then set to the first message of the queue.
o If the value is -ve, the message queue is scanned to find the first message
with a type whose value is less than, or equal to, this argument.
The flags in the msgrcv() are treated the same way as for msgsnd().
A successful execution of either msgsnd(), or msgrcv() always updates the appropriate
entries in msgid_ds data structure. With the above explanation, let us examine the
message passing program which follows.
#include
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/30
#include
#include
main(argc, argv)
int argc;
char *argv[];
{ int status, pid, pid1;
if (( pid=fork())==0) execlp("./messender", "messender", argv[1], argv[2], 0);
if (( pid1=fork())==0) execlp("./mesrec", "mesrec", argv[1], 0);
wait(&status); /* wait for some child to terminate */
wait(&status); /* wait for some child to terminate */
}
Next we give the message sender program.
#include
#include
#include
main(argc, argv)
int argc;
char *argv[];
/* This is the sender. It sends messages using IPC system V messages queues.*/
/* It takes two arguments : */
/* No. of messages and no. of bytes */
/* key_t MSGKEY = 100; */
/* struct msgformat {long mtype; int mpid; char mtext[256]} msg; */
{
key_t MSGKEY = 100;
struct msgformat { long mtype;
int mpid;
char mtext[256];
} msg;
int i ;
int msgid;
int loop, bytes;
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/31
extern cleanup();
loop = atoi(argv[1]);
bytes = atoi(argv[2]);
printf("In the sender child \n");
for ( i = 0; i < bytes; i++ ) msg.mtext[i] = 'm';
printf("the number of 'm' s is : %6d \n", i);
msgid = msgget(MSGKEY, 0660 | IPC_CREAT);
msg.mtype = 1;
msg.mpid = getpid();
/* Send number of messages specified by user argument */
for (i=0; i
/* Cleaning up; maximum number queues 32 */
for (i=0; i<32; i++) signal(i, cleanup);
}
cleanup()
{ int msgid;
msgctl(msgid, IPC_RMID, 0);
exit(0);
}
|Now we give the receiver program listing.
#include
#include
#include
main(argc, argv)
int argc;
char *argv[];
/* The receiver of the two processes communicating message using */
/* IPC system V messages queues. */
/* It takes two arguments: No. of messages and no. of bytes */
/* key_t MSGKEY = 100; */
/* struct msgformat {long mtype; int mpid; char mtext[256]} msg; */
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/32
{
key_t MSGKEY = 100;
struct msgformat { long mtype;
int mpid;
char mtext[256];
} msg;
int i, pid, *pint;
int msgid;
int loop, bytes;
msgid = msgget(MSGKEY, 0777);
loop = atoi(argv[1]);
bytes = atoi(argv[2]);
for ( i = 0; i <= bytes; i++ )
{ printf("receiving a message \n");
msgrcv(msgid, &msg, 256, 2, 0);
} }
If there are multiple writer processes and a single reader process, then the code shall be
somewhat along the following lines.
if ( mesg_type == 1) { search mesg_queue for type 1; process msg_type type 1 }
.
.
if ( mesg_type == n) { search mesg_queue for type n; process msg_type type n }
The number and size of messages available is limited by some constant in the IPC
package.
In fact this can be set in the system V IPC package when it is installed. Typically the
constants and structure are as follows.
MSGPOOL 8
MSGMNB 2048 /* Max. no. of bytes on queue */
MSGMNI 50 /* No. of msg. queue identifiers */
MSGTQL 50 /* No. of system message headers */
MSGMAP 100 /* No. of entries in msg map */
MSGMAX ( MSGPOOL *1024 ) /* Maximum message size */
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/33
MSGSSZ 8 /* Message segment size */
MSGSEG (( MSGPOOL *1024 ) / MSGSSZ ) /* No. of msg. segments */
Finally, we may note that the message queue information structure is as follows.
struct msginfo{int msgmap, msgmax, msgmnb, msgmni, msgssz, msgtql; ushort msgseg}
From the programs above, it should be obvious that the message-based IPC can also be
used for merging multiple data streams (multiplexing). As messages carry senders' id it
should also be possible to do de-multiplexing. The message type may also capture
priorities. Prioritizing messages can be very useful in some application contexts. Also,
note that the communicating parties need not be active at the same time. In our program
descriptions we used signals. Note that signals too, are messages! Signals are important
and so we shall discuss these in the next subsection.
7.3.5 Signals as IPC
Within the suite of IPC mechanisms, signals stand out for one very good reason. A signal,
as a mechanism, is one clean way to communicate asynchronous events. In fact, we use
signals more often than any other means of IPC. Every time we abort a program using ^c,
a signal is generated to break. Similarly, if an unexpected value for a pointer is generated,
we have core dump and a segmentation fault recognized. When we change a window
size, a signal is generated. Note that in all these examples, an event happens within the
process or the process receives it as an input. In general, a process may send a signal to
another process. In all these situations the process receiving a signal needs to respond.
We shall first enumerate typical sources of signal, and later examine the possible forms
of responses that are generated. Below we list the sources for signal during a process
execution:
1. From the terminal: Consider a process which has been launched from a terminal
and is running. Now if we input the interrupt character, ^c, from the keyboard
then we have a signal SIGINT initiated. Suppose, we have disconnect of the
terminal line (this may happen when we may close the window for instance), then
there is a signal SIGHUP to capture the hanging up of the line.
2. From window manager: This may be any of the mouse activity that may happen
in the selected window. In case of change of size of the window the signal is
SIGWINCH.
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/34
3. From other subsystems: This may be from memory or other subsystems. For
instance, if a memory reference is out of the process's data or code space, then
there shall be a signal SIGSEGV.
4. From kernel: The typical usage of time in processes can be used to set an alarm.
The alarm signal is SIGALARM.
5. From the processes: It is not unusual to kill a child process. In fact, sometimes we
may kill a job which may have entered an infinite loop. There may be other
reasons to abort a process. The typical kill signal is SIGKILL. One of the uses is
when a terminal hangs, the best thing to do is to log in from another terminal and
kill the hanging process. One may also look upon the last case as a shell initiated
signal. Note that a shell is it self a process.
Above we have noted various sources from where signals may be generated. Usually this
helps to define the signal type. A process may expect certain types of signals and make a
provision for handling these by defining a set of signal handlers. The signal handlers can
offer a set of responses which may even include ignoring certain signals! So next, we
shall study the different kind of signal responses which processes may generate.
In Figure 7.3 we see a program statement signal(SIGXXX, sighandler) to define how this
process should respond to a signal. In this statement SIGXXX identifies the signal and
sighandler identifies a signal service routine. In general, a process may respond to a given
signal in one of the following ways.
1. Ignore it: A process may choose to ignore some kinds of signal. Since processes
may receive signals from any source, it is quite possible that a process would
authenticate the process before honoring the signal. In some cases then a process
may simply ignore the signal and offer no response at all.
Operating Systems/Inter-Process Communication Lecture Notes
PCP Bhatt/IISc, Bangalore M7/V1/June 04/35
2. Respond to it: This is quite often the case in the distributed computing scenarios
where processes communicate to further computations in steps. These signals may
require some response. The response is encoded in the signal handler. For
instance, a debugger and the process being debugged would require signal
communication quite often. Another usage might be to advise a clean-up
operation. For instance, we need to clean-up following the shared memory mode
of IPC. Users of Java would recognize that response for exception handling falls
in the same category.
3. Reconfigure: This is required whenever system services are dynamically
reconfigured. This happens often in fault-tolerant systems or networked systems.
The following is a good example of dynamic configuration. Suppose we have
several application servers (like WebSphere) provisioning services. A dispatcher
system allocates the servers. During operations, some server may fail. This entails
redeployment by the dispatcher. The failure needs to be recognized and
dispatching reconfigured for future.
4. Turn on/off options: During debugging as well as profiling (as discussed in
chapter on “Other Tools") we may turn some options \On" or \Off" and this may
require some signals to be generated.
5. Timer information: In real-time systems, we may have several timers to keep a tab
on periodic events. The ideas is to periodically generate required signals to set up
services, set alarms or offer other time-based services.
In this chapter we examined the ways to establish communication amongst processes. Its
a brief exposure. To comprehend the distributed computing field, it is important to look
up the suggested reading list. Interested readers should explore PVM (parallel virtual
machine) [30] and MPI (message passing interface) [31] as distributed computing
environments.
No comments:
Post a Comment