Playing with ptrace, Part I

Using ptrace allows you to set up system call interception and modification at the user level.

Have you ever wondered how system calls can be intercepted? Have you ever tried fooling the kernel by changing system call arguments? Have you ever wondered how debuggers stop a running process and let you take control of the process?

If you are thinking of using complex kernel programming to accomplish tasks, think again. Linux provides an elegant mechanism to achieve all of these things: the ptrace (Process Trace) system call. ptrace provides a mechanism by which a parent process may observe and control the execution of another process. It can examine and change its core image and registers and is used primarily to implement breakpoint debugging and system call tracing.

In this article, we learn how to intercept a system call and change its arguments. In Part II of the article we will study advanced techniques—setting breakpoints and injecting code into a running program. We will peek into the child process' registers and data segment and modify the contents. We will also describe a way to inject code so the process can be stopped and execute arbitrary instructions.


Operating systems offer services through a standard mechanism called system calls. They provide a standard API for accessing the underlying hardware and low-level services, such as the filesystems. When a process wants to invoke a system call, it puts the arguments to system calls in registers and calls soft interrupt 0x80. This soft interrupt is like a gate to the kernel mode, and the kernel will execute the system call after examining the arguments.

On the i386 architecture (all the code in this article is i386-specific), the system call number is put in the register %eax. The arguments to this system call are put into registers %ebx, %ecx, %edx, %esi and %edi, in that order. For example, the call:

write(2, "Hello", 5)

roughly would translate into

movl   $4, %eax
movl   $2, %ebx
movl   $hello,%ecx
movl   $5, %edx
int    $0x80
where $hello points to a literal string “Hello”.

So where does ptrace come into picture? Before executing the system call, the kernel checks whether the process is being traced. If it is, the kernel stops the process and gives control to the tracking process so it can examine and modify the traced process' registers.

Let's clarify this explanation with an example of how the process works:

#include <sys/ptrace.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <unistd.h>
#include <linux/user.h>   /* For constants
                                   ORIG_EAX etc */
int main()
{   pid_t child;
    long orig_eax;
    child = fork();
    if(child == 0) {
        ptrace(PTRACE_TRACEME, 0, NULL, NULL);
        execl("/bin/ls", "ls", NULL);
    else {
        orig_eax = ptrace(PTRACE_PEEKUSER,
                          child, 4 * ORIG_EAX,
        printf("The child made a "
               "system call %ld\n", orig_eax);
        ptrace(PTRACE_CONT, child, NULL, NULL);
    return 0;

When run, this program prints:

The child made a system call 11
along with the output of ls. System call number 11 is execve, and it's the first system call executed by the child. For reference, system call numbers can be found in /usr/include/asm/unistd.h.

As you can see in the example, a process forks a child and the child executes the process we want to trace. Before running exec, the child calls ptrace with the first argument, equal to PTRACE_TRACEME. This tells the kernel that the process is being traced, and when the child executes the execve system call, it hands over control to its parent. The parent waits for notification from the kernel with a wait() call. Then the parent can check the arguments of the system call or do other things, such as looking into the registers.

When the system call occurs, the kernel saves the original contents of the eax register, which contains the system call number. We can read this value from child's USER segment by calling ptrace with the first argument PTRACE_PEEKUSER, shown as above.

After we are done examining the system call, the child can continue with a call to ptrace with the first argument PTRACE_CONT, which lets the system call continue.

ptrace Parameters

ptrace is called with four arguments:

long ptrace(enum __ptrace_request request,
            pid_t pid,
            void *addr,
            void *data);

The first argument determines the behaviour of ptrace and how other arguments are used. The value of request should be one of PTRACE_TRACEME, PTRACE_PEEKTEXT, PTRACE_PEEKDATA, PTRACE_PEEKUSER, PTRACE_POKETEXT, PTRACE_POKEDATA, PTRACE_POKEUSER, PTRACE_GETREGS, PTRACE_GETFPREGS, PTRACE_SETREGS, PTRACE_SETFPREGS, PTRACE_CONT, PTRACE_SYSCALL, PTRACE_SINGLESTEP, PTRACE_DETACH. The significance of each of these requests will be explained in the rest of the article.



Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.

Ptrace for multi-thread application

Solon Chen's picture

After struggled a long time, I got a true way to make my ptrace worked correct with multi-thread application. Here're my sample codes, hope it can help others whom have the same confusion.

char trapCode[] = {0, 0, 0, 0};
int status;

ptrace(PTRACE_ATTACH, childProcess, NULL, NULL); //childProcess is the main thread

printf("\nchild %d created\n", childProcess);

long ptraceOption = PTRACE_O_TRACECLONE;
ptrace(PTRACE_SETOPTIONS, childProcess, NULL, ptraceOption);

struct user_regs_struct regs;

for(unsigned int i = 0; i < m_breakPoints.size(); i++)
BreakPoint_Info breakPointInfo = m_breakPoints[i];

unsigned int index = breakPointInfo.m_checkPointIndex;
if(m_bytesBackup.find(m_checkPoints[index].m_offset) != m_bytesBackup.end())

unsigned long readAddr = m_checkPoints[index].m_offset;
One_Byte_With_Result *oneByte = new One_Byte_With_Result;
getData(childProcess, readAddr, trapCode, 4);
oneByte->m_char = trapCode[0];
trapCode[0] = 0xcc;
putData(childProcess, readAddr, trapCode, 4);

m_bytesBackup.insert(std::make_pair(m_checkPoints[index].m_offset, oneByte));

std::set allThreads;
std::set::iterator allThreadsIter;

int rec = ptrace(PTRACE_CONT, childProcess, NULL, NULL);

pid_t child_waited = waitpid(-1, &status, __WALL);

if(child_waited == -1)

if(allThreads.find(child_waited) == allThreads.end())
printf("\nreceived unknown child %d\t", child_waited);

if(WIFSTOPPED(status) && WSTOPSIG(status) == SIGTRAP)
pid_t new_child;
if(((status >> 16) & 0xffff) == PTRACE_EVENT_CLONE)
if(ptrace(PTRACE_GETEVENTMSG, child_waited, 0, &new_child) != -1)
ptrace(PTRACE_CONT, new_child, 0, 0);

printf("\nchild %d created\t", new_child);
ptrace(PTRACE_CONT, child_waited, 0, 0);

printf("\nchild %d exited with status %d\t", child_waited, WEXITSTATUS(status));

if(allThreads.size() == 0)
else if(WIFSIGNALED(status))
printf("\nchild %d killed by signal %d\t", child_waited, WTERMSIG(status));

if(allThreads.size() == 0)
else if(WIFSTOPPED(status))
int stopCode = WSTOPSIG(status);
if(stopCode == SIGTRAP)
ptrace(PTRACE_GETREGS, child_waited, NULL, &regs);
unsigned long currentEip = regs.eip;
//printf("%d\t%08x\n", child_waited, currentEip);

Address_Bytes_Map::iterator iter = m_bytesBackup.find(currentEip - 1);
if(iter != m_bytesBackup.end())
iter->second->m_result = true;

regs.eip = regs.eip - 1;
getData(child_waited, regs.eip, trapCode, 4);
trapCode[0] = iter->second->m_char;
putData(child_waited, regs.eip, trapCode, 4);
rec = ptrace(PTRACE_SETREGS, child_waited, NULL, &regs);

rec = ptrace(PTRACE_CONT, child_waited, 1, NULL);


Breakpoint at line number

eager's picture

Karan Verma asked how to put a breakpoint at a particular line number. This is exactly what a debugger does.

There are many nuances and details, but I'll simplify this to the bare minimum:

1) Find where the program is loaded in memory.
2) Read the executable and locate the debug data which corresponds to the source file containing the line at which you want to place the breakpoint.
3) Interpret the line number table in the debug data to locate the address which corresponds to the desired line.
4) Adjust the address from (3) to make it correspond with (1) if necessary.
5) Copy and save the original instruction at the breakpoint address.
6) Write the breakpoint instruction at the breakpoint address

Naturally, this is usually done when the child process is stopped. After inserting the breakpoint, the child is allowed to run.

When the breakpoint is executed, the child process will stop and the parent will receive a signal. The parent process needs to replace the breakpoint instruction the original instruction before it can allow the child process to continue.

GDB and LLDB are open source debuggers which give examples of how this is done. Reading GDB is not for the faint of heart -- there's a lot of complexity in handling many different object file formats and many different target architectures.

Michael Eager, Consultant, Embedded Systems, Compilers, Debuggers

Walking the call stack

eager's picture

Sandeep mentioned wanting to use ptrace() to walk the call stack.

Ptrace() provides access to the memory and registers of a child process. It doesn't tell you how that memory is organized.

Information on how the stack is organized is usually contained in the ABI (Application Binary Interface) for each processor. DWARF debugging information (see Call Frame Information (CFI) describes where each call has saved registers.

If you want to write a routine which walks the call stack, I suggest that you start with one which will walk the stack in the current process and later convert it to accessing a child process. The first step is to find the start of the current stack frame, then find the previous frame.

Michael Eager, Consultant, Embedded Systems, Compilers, Debuggers

Answer: Where is ORIG_EAX?

eager's picture

ORIG_EAX is in . Replace with this path. (Oh, and add #include to avoid compiler complaints about printf definition conflicts.)

Michael Eager, Consultant, Embedded Systems, Compilers, Debuggers

Answer: Where is ORIG_EAX?

eager's picture

ORIG_EAX is in . Replace with this path. (Oh, and add #include to avoid compiler complaints about printf definition conflicts.)

Michael Eager, Consultant, Embedded Systems, Compilers, Debuggers

ptrace multi thread application

Solon Chen's picture

This article helps me a lot. I'm trying to create a line coverage tool using ptrace.

One problem is ptrace only resolve single thread, and I don't know how to deal with multi thread application.

I set option to catch clone event, it can help me to find all lwp's pid.
I also try to make all thread continue, but it seems only child thread will go on until sleep. The parent thread has no affect on continue command, /proc//status shows it in "tracing stop".

Could you tell me how to make all thread continue? (I have restore the breakpoint in runtime memory)


Pradeep, this example sucks!

Anonymous's picture

This example is useless. The author try to impress with complicated function calls which are absolutely useless.

Sounds to me like your

Anonymous's picture

Sounds to me like your frustrated with your incompetence

Compilation Errors for some constants

qtp's picture

i am getting Compilation Errors for few constants, like "ORIG_EAX".

it is probably in User.h. but when i search for user.h, i dont fid it declared there.

can i define these variables to some value from my code? which values should i assign and what is the significance of those?


Excellent article

Anonymous's picture

Thanks Pradeep, great intro to ptrace().

I've used strace() for debugging for a couple of years and never knew this was what it used; I'm hoping to create a simulator using ptrace() soon for automatic integration testing of an embedded project and your article is a great start to see some real code :-)


C++ stack trace on linux

Anonymous's picture


i am working on an application which can dump call stack of C++ for me.
i am trying to use ptrace() for the same but not getting proper direction or the steps i must follow for the same.

could someone guide me on the same?


Error in putdata()

Aliphany's picture

Please be aware that putdata() contains a serious mistake. If len is not a multiple of four, then putdata() should read the final long value, replace one, two, or three bytes, and then write the value. This mistake causes the second example in Part II to seg fault.


ChrisO's picture

Great article, thanks!

One question though, why do you multiply the addresses of registers by 4 when reading them with PTRACE_PEEKDATA?

Indexes not addresses

Mitch Frazier's picture

The register "addresses" you're referring to (EAX, EBX, etc) are indexes:

#define EBX 0
#define ECX 1
#define EDX 2
#define ESI 3
#define EDI 4

into the user data:

struct user_regs_struct
  long int ebx;
  long int ecx;
  long int edx;
  long int esi;
  long int edi;

So to get an address, multiply by 4 (the "word" size on a x86 system).

Mitch Frazier is an Associate Editor for Linux Journal.

ptrace and threads

Anonymous's picture

I'm posting this hoping the next fellow who encounters this gotcha can save a little time...

ptrace only works in the base thread of the parent process. ptrace(PTRACE_CONT, pid) will fail with ESRCH (process not found) if issued in a child thread on Linux.

If you are thinking of using a debugger thread to watch each child thread, give it up. It won't work. And unless you find this message or have a sudden epiphany, you are liable to spend a great deal of time bashing your poor head against the wall.

Google, take it from here!

line numbers ptrace

Karan Verma's picture


If we want to put a breakpoint at a particular line number, how would we accomplish that using ptrace?

Hi Pradeep, Thanks a

Patel's picture

Hi Pradeep,

Thanks a lot for this article, this is the one which helped me to write my own tool, to capture all the system calls(including calls from forked/child processes too). Thanks again!!!