Weird behavior in struct destruction

问题

I am trying to write a struct to a file and read it back. The code to do so is here:

#include <fstream>
#include <iostream>
#include <cstring>

using namespace std;

struct info {
  int id;
  string name;
};

int main(void) {
  info adam;
  adam.id = 50;
  adam.name = "adam";

  ofstream file("student_info.dat", ios::binary);
  file.write((char*)&adam, sizeof(info));
  file.close();

  info student;
  ifstream file2("student_info.dat", ios::binary);
  file2.read((char*)&student, sizeof(student));
  cout << "ID =" << student.id << " Name = " << student.name << endl;

  file2.close();
  return 0;
}

However I get a strange segmentation fault in the end.

The output is :

ID =50 Name = adam
Segmentation fault (core dumped)

On looking at the core dump, I see that there is something weird happening in the destruction of the struct info.

(gdb) bt
#0  0x00007f035330595c in ?? ()
#1  0x00000000004014d8 in info::~info() () at binio.cc:7
#2  0x00000000004013c9 in main () at binio.cc:21

I suspect that something weird is happening in the string destruction but I am not able to figure out the exact problem. Any help will be great.

I am using gcc 8.2.0.

回答1:

You can't serialize/deserialize like that. On this line here:

file2.read((char*)&student, sizeof(student));

You're just writing 1:1 over an instance of info, which includes an std::string. Those aren't just arrays of characters - they dynamically allocate their storage on the heap and manage that using pointers. Thus the string becomes invalid if you overwrite it like that, it's undefined behavior because its pointers aren't pointing to a valid place anymore.

Instead you should save the actual characters, not the string object, and make a new string with that content on load.

Generally, you can do a copy like that with trivial objects. You can test it like this:

std::cout << std::is_trivially_copyable<std::string>::value << '\n';

回答2:

To add to the accepted answer, because the asker is still confused about "why does it crash on the deletion of the first object?":

Let's look at the diassembly, because it cannot lie, even in the face of an incorrect program that exhibits UB (unlike the debugger).

https://godbolt.org/z/pstZu5

(Take note that rsp - our stack pointer - is never changed aside from the adjustment at the beginning and end of main).

Here is the initialization of adam:

    lea     rax, [rsp+24]
    // ...
    mov     QWORD PTR [rsp+16], 0
    mov     QWORD PTR [rsp+8], rax
    mov     BYTE PTR [rsp+24], 0

It seems [rsp+16] and [rsp+24] hold size and capacity of the string, while [rsp+8] holds the pointer to the internal buffer. That pointer is set up to point into the string object itself.

Then adam.name is overwritten with "adam":

   call    std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >::_M_replace(unsigned long, unsigned long, char const*, unsigned long)

Due to small string optimization, the buffer pointer at [rsp+8] probably still points to the same place (rsp+24) to indicate the string that we have a small buffer and no memory allocation (that's my guess to be clear).

Later on we initialize student much in the same way:

    lea     rax, [rsp+72]
    // ...
    mov     QWORD PTR [rsp+64], 0
    // ...
    mov     QWORD PTR [rsp+56], rax
    mov     BYTE PTR [rsp+72], 0

Note how student's buffer pointer points into student to signify a small buffer.

Now you brutally replace the internals of student with those of adam. And suddenly, student's buffer pointer doesn't point to the expected place anymore. Is that a problem?

    mov     rdi, QWORD PTR [rsp+56]
    lea     rax, [rsp+72]
    cmp     rdi, rax
    je      .L90
    call    operator delete(void*)

Yep! If the internal buffer of student points anywhere else than where we initially set it to (rsp+72), it will delete that pointer. At this point we don't know where exactly adam's buffer pointer (that you copied into student) points to, but it's certainly the wrong place. As explained above, "adam" is likely still covered by small string optimization, so adam's buffer pointer was likely in the exact same place as before: rsp+24. Since we copied that into student and it is different from rsp+72, we call delete(rsp+24) - which is in the middle of our own stack. The environment doesn't think that's very funny and you get a segfault right there, in the first deallocation (the second one wouldn't even delete anything because the world would still be fine over there - adam was unharmed by you).

Bottom line: Don't try to outclever the compiler ("it can't segfault because it'll be on the same heap!"). You will lose. Follow the rules of the language and nobody gets hurt. ;)

Side note: This design in gcc might even be intentional. I believe they could just as easily store a nullptr instead of pointing into the string object to denote a small string buffer. But in that case you wouldn't segfault from this malpractice.

回答3:

Briefly and thinking conceptually, when adam.name = "adam"; is done, appropriate memory is allocated internally for adam.name.

When file2.read((char*)&student, sizeof(student)); is done, you are writing at memory location i.e at address &student which is not yet allocated properly to accommodate the data which is being read. student.adam doesn't have enough valid memory allocated to it. On doing such read into student object's location is actually causing memory corruption.

来源：https://stackoverflow.com/questions/55181063/weird-behavior-in-struct-destruction

标签

c++

serialization

struct

file-writing