Search code examples
cunixcrashpowerpc

How can dereferencing a NULL pointer in C not crash a program?


I need help of a real C guru to analyze a crash in my code. Not for fixing the crash; I can easily fix it, but before doing so I'd like to understand how this crash is even possible, as it seems totally impossible to me.

This crash only happens on a customer machine and I cannot reproduce it locally (so I cannot step through the code using a debugger), as I cannot obtain a copy of this user's database. My company also won't allow me to just change a few lines in the code and make a custom build for this customer (so I cannot add some printf lines and have him run the code again) and of course the customer has a build without debug symbols. In other words, my debbuging abilities are very limited. Nonetheless I could nail down the crash and get some debugging information. However when I look at that information and then at the code I cannot understand how the program flow could ever reach the line in question. The code should have crashed long before getting to that line. I'm totally lost here.

Let's start with the relevant code. It's very little code:

// ... code above skipped, not relevant ...

if (data == NULL) return -1;

information = parseData(data);

if (information == NULL) return -1;

/* Check if name has been correctly \0 terminated */
if (information->kind.name->data[information->kind.name->length] != '\0') {
    freeParsedData(information);
    return -1;
}

/* Copy the name */
realLength = information->kind.name->length + 1;
*result = malloc(realLength);
if (*result == NULL) {
    freeParsedData(information);
    return -1;
}
strlcpy(*result, (char *)information->kind.name->data, realLength);

// ... code below skipped, not relevant ...

That's already it. It crashes in strlcpy. I can tell you even how strlcpy is really called at runtime. strlcpy is actually called with the following paramaters:

strlcpy ( 0x341000, 0x0, 0x1 );

Knowing this it is rather obvious why strlcpy crashes. It tries to read one character from a NULL pointer and that will of course crash. And since the last parameter has a value of 1, the original length must have been 0. My code clearly has a bug here, it fails to check for the name data being NULL. I can fix this, no problem.

My question is:
How can this code ever get to the strlcpy in the first place?
Why does this code not crash at the if-statement?

I tried it locally on my machine:

int main (
    int argc,
    char ** argv
) {
    char * nullString = malloc(10);
    free(nullString);
    nullString = NULL;
    
    if (nullString[0] != '\0') {
        printf("Not terminated\n");
        exit(1);
    }
    printf("Can get past the if-clause\n");
    
    char xxx[10];
    strlcpy(xxx, nullString, 1);
    return 0;   
}

This code never gets passed the if statement. It crashes in the if statement and that is definitely expected.

So can anyone think of any reason why the first code can get passed that if-statement without crashing if name->data is really NULL? This is totally mysterious to me. It doesn't seem deterministic.

Important extra information:
The code between the two comments is really complete, nothing has been left out. Further the application is single threaded, so there is no other thread that could unexpectedly alter any memory in the background. The platform where this happens is a PPC CPU (a G4, in case that could play any role). And in case someone wonders about "kind.", this is because "information" contains a "union" named "kind" and name is a struct again (kind is a union, every possible union value is a different type of struct); but this all shouldn't really matter here.

I'm grateful for any idea here. I'm even more grateful if it's not just a theory, but if there is a way I can verify that this theory really holds true for the customer.

Solution

I accepted the right answer already, but just in case anyone finds this question on Google, here's what really happened:

The pointers were pointing to memory, that has already been freed. Freeing memory won't make it all zero or cause the process to give it back to the system at once. So even though the memory has been erroneously freed, it was containing the correct values. The pointer in question is not NULL at the time the "if check" is performed.

After that check I allocate some new memory, calling malloc. Not sure what exactly malloc does here, but every call to malloc or free can have far-reaching consequences to all dynamic memory of the virtual address space of a process. After the malloc call, the pointer is in fact NULL. Somehow malloc (or some system call malloc uses) zeros the already freed memory where the pointer itself is located (not the data it points to, the pointer itself is in dynamic memory). Zeroing that memory, the pointer now has a value of 0x0, which is equal to NULL on my system and when strlcpy is called, it will of course crash.

So the real bug causing this strange behavior was at a completely different location in my code. Never forget: Freed memory keeps it values, but it is beyond your control for how long. To check if your app has a memory bug of accessing already freed memory, just make sure the freed memory is always zeroed before it is freed. In OS X you can do this by setting an environment variable at runtime (no need to recompile anything). Of course this slows down the program quite a bit, but you will catch those bugs much earlier.


Solution

  • It is possible that the structure is located in memory that has been free()'d, or the heap is corrupted. In that case, malloc() could be modifying the memory, thinking that it is free.

    You might try running your program under a memory checker. One memory checker that supports Mac OS X is valgrind, although it supports Mac OS X only on Intel, not on PowerPC.