invalid fastbin entry (free)

Question

I am trying to find the cause for:

*** glibc detected *** ...: invalid fastbin entry (free): 0x00007fc384ced120 ***

The program dumped core, so I was able to trace this back to a destructor of a very simple class similar to this:

class foo : public foo_base
{
    ...
    ...
    std::vector<boost::weak_ptr<bar> > vec;
}

The destructor is virtual in foo_base and not implemented in foo

The vector vec is only assigned to in the constructor and not modified thereafter.

The address mentioned by the glibc error is identical to vec._M_impl._M_start

• Where could I start searching for the cause?

• Knowing what a fastbin is, how can it be invalid?

• Could this be a double free situation, or would glibc definitely raise a double free in this case?

Answer 1

To "answer" my own question:

I was able to rule out a double free situation, because it turned out that all foo instances were always (correctly) kept in smart pointers.

A memory corruption bug has recently been found. It is impossible to confirm this to have been the cause for the original problem, but it seems reasonable.

The problem was never reproduced.

Answer 2

---

This may be due to a bug in glibc. The RedHat Advisories provide additional details: https://rhn.redhat.com/errata/RHBA-2014-0480.html

To identify if you are affected by this bug:

rpm -qa | grep glibc

If your version of glibc is 2.12 and doesn't have a .149 or later suffix, then your server may be affected by this issue.

---

Answer 3

For me this error came out of pytorch torch api for c++

I'm a noob still at the subject, but error rises when:

torch::Tensor var = torch::zeros({1, 1}, torch::kFloat32);
torch::Tensor foo = torch::full({1, 1}, val, torch::kFloat32);

It went away when:

torch::Tensor var = torch::zeros({1, 1}, torch::kFloat64);
torch::Tensor foo = torch::full({1, 1}, val, torch::kFloat64);