How to add a system call in 2.6.x kernel

SungWon Chung · 11-26-2007, 12:32 AM

Hi. I have a problem about adding my own system call.

My system is fedora core 2.6.19.

I did all phases which should be done to add system call.

And I recompiled my kernel.

But I met a problem making an application program.

For using my system call, I wrote codes like this.

==========================================================================================
_syscall3(long, insert_iphash_entry, const uint32_t, category, const uint32_t, ip, const uint32_t, action);
_syscall3(long, modify_iphash_entry, const uint32_t, category, const uint32_t, ip, const uint32_t, action);
_syscall2(long, remove_iphash_entry, const uint32_t, category, const uint32_t, ip);
_syscall1(long, flush_iphash_bucket, const uint32_t, category);
==========================================================================================

Compiling this codes, error was occured like this.

iphash.c:37: error: expected declaration specifiers or '...' before 'insert_ipha sh_entry'
iphash.c:37: error: expected declaration specifiers or '...' before 'category'
iphash.c:37: error: expected declaration specifiers or '...' before 'ip'
iphash.c:37: error: expected declaration specifiers or '...' before 'action'
iphash.c:37: warning: data definition has no type or storage class
iphash.c:37: warning: type defaults to 'int' in declaration of '_syscall3'

I found the reason.

unistd.h which is in my system doesn't have a define like _syscall0, _syscall1 ...

What should I do?

Is this problem related with glibc version?

I don't know what I do.

Please help me.

osor · 11-26-2007, 04:53 PM

Quote:

Originally Posted by SungWon Chung

I did all phases which should be done to add system call.

And I recompiled my kernel.

What exactly were the steps you used?

Quote:

Originally Posted by SungWon Chung

For using my system call, I wrote codes like this.

==========================================================================================
_syscall3(long, insert_iphash_entry, const uint32_t, category, const uint32_t, ip, const uint32_t, action);
_syscall3(long, modify_iphash_entry, const uint32_t, category, const uint32_t, ip, const uint32_t, action);
_syscall2(long, remove_iphash_entry, const uint32_t, category, const uint32_t, ip);
_syscall1(long, flush_iphash_bucket, const uint32_t, category);
==========================================================================================

Well, that’s your problem… There are much better ways to use syscalls from C programs. There is a nice wrapper function called syscall() which comes with glibc (for more information, see “man 2 syscall”). For example (if you use glibc), try something like this for a header file:

Code:

#ifndef _MY_HEADER_H
#define _MY_HEADER_H

#include <sys/syscall.h>
#include <unistd.h>

#define __NR_insert_iphash_entry 300
#define __NR_modify_iphash_entry 301
#define __NR_remove_iphash_entry 302
#define __NR_flush_iphash_bucket 303

static inline int insert_iphash_entry(uint32_t category, uint32_t ip, uint32_t action)
{
	return syscall(__NR_insert_iphash_entry, category, ip, action);
}

static inline int modify_iphash_entry(uint32_t category, uint32_t ip, uint32_t action)
{
	return syscall(__NR_modify_iphash_entry, category, ip, action);
}

static inline int remove_iphash_entry(uint32_t category, uint32_t ip)
{
	return syscall(__NR_remove_iphash_entry, category, ip);
}

static inline int flush_iphash_bucket(uint32_t category)
{
	return syscall(__NR_flush_iphash_bucket, category);
}

#endif /* _MY_HEADER_H */

A few comments:

Obviously, the #defines are supposed to be the system call numbers themselves. If in implementing your system call, you already changed /usr/include/asm/unistd.h then you don’t need to replicate the #defines (in fact if you do, an error should be thrown). If you didn’t do this (or are distributing your program so that editing a system file would require too much permission), you’ll need to duplicate the macros. In fact, you might add some architecture-specific preprocessing since syscall numbers are arch-specific.
You don’t need to const-qualify the function arguments, since in C all arguments are passed by value.
Whereas in kernelspace, your system calls usually return long, in userspace they are interpreted as ints.
The wrapper function will automatically take care of errno for you. I.e., in case of a negative return value from the actual system call, the wrapper will return -1 and set errno accordingly.
Once you include a file with the above code, you may use the functions as if they were C functions instead of direct system calls.

SungWon Chung · 11-26-2007, 09:32 PM

Because of your help, I solved it completely.

I appreciate it.