Hotpatching a C Function on x86

In this post I’m going to do a silly, but interesting, exercise that should never be done in any program that actually matters. I’m going write a program that changes one of its function definitions while it’s actively running and using that function. Unlike last time, this won’t involve shared libraries, but it will require x86_64 and GCC. Most of the time it will work with Clang, too, but it’s missing an important compiler option that makes it stable.

If you want to see it all up front, here’s the full source: hotpatch.c

Here’s the function that I’m going to change:

void
hello(void)
{
    puts("hello");
}

It’s dead simple, but that’s just for demonstration purposes. This will work with any function of arbitrary complexity. The definition will be changed to this:

void
hello(void)
{
    static int x;
    printf("goodbye %d\n", x++);
}

I was only going change the string, but I figured I should make it a little more interesting.

Here’s how it’s going to work: I’m going to overwrite the beginning of the function with an unconditional jump that immediately moves control to the new definition of the function. It’s vital that the function prototype does not change, since that would be a far more complex problem.

But first there’s some preparation to be done. The target needs to be augmented with some GCC function attributes to prepare it for its redefinition. As is, there are three possible problems that need to be dealt with:

The solution is the ms_hook_prologue function attribute. This tells GCC to put a hotpatch prologue on the function: a big, fat, 8-byte NOP that I can safely clobber. This idea originated in Microsoft’s Win32 API, hence the “ms” in the name.

The solution is the aligned function attribute, ensuring the hotpatch prologue is properly aligned.

As you might have guessed, this is primarily fixed with the noinline function attribute. Since GCC may also clone the function and call that instead, so it also needs the noclone attribute.

Even further, if GCC determines there are no side effects, it may cache the return value and only ever call the function once. To convince GCC that there’s a side effect, I added an empty inline assembly string (__asm("")). Since puts() has a side effect (output), this isn’t truly necessary for this particular example, but I’m being thorough.

What does the function look like now?

__attribute__ ((ms_hook_prologue))
__attribute__ ((aligned(8)))
__attribute__ ((noinline))
__attribute__ ((noclone))
void
hello(void)
{
    __asm("");
    puts("hello");
}

And what does the assembly look like?

$ objdump -Mintel -d hotpatch
0000000000400848 <hello>:
  400848:       48 8d a4 24 00 00 00    lea    rsp,[rsp+0x0]
  40084f:       00
  400850:       bf d4 09 40 00          mov    edi,0x4009d4
  400855:       e9 06 fe ff ff          jmp    400660 <puts@plt>

It’s 8-byte aligned and it has the 8-byte NOP: that lea instruction does nothing. It copies rsp into itself and changes no flags. Why not 8 1-byte NOPs? I need to replace exactly one instruction with exactly one other instruction. I can’t have another thread in between those NOPs.

Hotpatching

Next, let’s take a look at the function that will perform the hotpatch. I’ve written a generic patching function for this purpose. This part is entirely specific to x86.

void
hotpatch(void *target, void *replacement)
{
    assert(((uintptr_t)target & 0x07) == 0); // 8-byte aligned?
    void *page = (void *)((uintptr_t)target & ~0xfff);
    mprotect(page, 4096, PROT_WRITE | PROT_EXEC);
    uint32_t rel = (char *)replacement - (char *)target - 5;
    union {
        uint8_t bytes[8];
        uint64_t value;
    } instruction = { {0xe9, rel >> 0, rel >> 8, rel >> 16, rel >> 24} };
    *(uint64_t *)target = instruction.value;
    mprotect(page, 4096, PROT_EXEC);
}

It takes the address of the function to be patched and the address of the function to replace it. As mentioned, the target must be 8-byte aligned (enforced by the assert). It’s also important this function is only called by one thread at a time, even on different targets. If that was a concern, I’d wrap it in a mutex to create a critical section.

There are a number of things going on here, so let’s go through them one at a time:

Make the function writeable

The .text segment will not be writeable by default. This is for both security and safety. Before I can hotpatch the function I need to make the function writeable. To make the function writeable, I need to make its page writable. To make its page writeable I need to call mprotect(). If there was another thread monkeying with the page attributes of this page at the same time (another thread calling hotpatch()) I’d be in trouble.

It finds the page by rounding the target address down to the nearest 4096, the assumed page size (sorry hugepages). Warning: I’m being a bad programmer and not checking the result of mprotect(). If it fails, the program will crash and burn. It will always fail systems with W^X enforcement, which will likely become the standard in the future. Under W^X (“write XOR execute”), memory can either be writeable or executable, but never both at the same time.

What if the function straddles pages? Well, I’m only patching the first 8 bytes, which, thanks to alignment, will sit entirely inside the page I just found. It’s not an issue.

At the end of the function, I mprotect() the page back to non-writeable.

Create the instruction

I’m assuming the replacement function is within 2GB of the original in virtual memory, so I’ll use a 32-bit relative jmp instruction. There’s no 64-bit relative jump, and I only have 8 bytes to work within anyway. Looking that up in the Intel manual, I see this:

Fortunately it’s a really simple instruction. It’s opcode 0xE9 and it’s followed immediately by the 32-bit displacement. The instruction is 5 bytes wide.

To compute the relative jump, I take the difference between the functions, minus 5. Why the 5? The jump address is computed from the position after the jump instruction and, as I said, it’s 5 bytes wide.

I put 0xE9 in a byte array, followed by the little endian displacement. The astute may notice that the displacement is signed (it can go “up” or “down”) and I used an unsigned integer. That’s because it will overflow nicely to the right value and make those shifts clean.

Finally, the instruction byte array I just computed is written over the hotpatch NOP as a single, atomic, 64-bit store.

    *(uint64_t *)target = instruction.value;

Other threads will see either the NOP or the jump, nothing in between. There’s no synchronization, so other threads may continue to execute the NOP for a brief moment even through I’ve clobbered it, but that’s fine.

Trying it out

Here’s what my test program looks like:

void *
worker(void *arg)
{
    (void)arg;
    for (;;) {
        hello();
        usleep(100000);
    }
    return NULL;
}

int
main(void)
{
    pthread_t thread;
    pthread_create(&thread, NULL, worker, NULL);
    getchar();
    hotpatch(hello, new_hello);
    pthread_join(thread, NULL);
    return 0;
}

I fire off the other thread to keep it pinging at hello(). In the main thread, it waits until I hit enter to give the program input, after which it calls hotpatch() and changes the function called by the “worker” thread. I’ve now changed the behavior of the worker thread without its knowledge. In a more practical situation, this could be used to update parts of a running program without restarting or even synchronizing.

Further Reading

These related articles have been shared with me since publishing this article:

Have a comment on this article? Start a discussion in my public inbox by sending an email to ~skeeto/public-inbox@lists.sr.ht [mailing list etiquette] , or see existing discussions.

null program

Chris Wellons

wellons@nullprogram.com (PGP)
~skeeto/public-inbox@lists.sr.ht (view)