I attempted fixing this issue before at 43fbaf99. Although it did work,
it worked wrong, and it was just consistently working (but in wrong way)
until it didn't.

Since this is (hopefully) the real fix, I will try to explain exactly
what is going on.

This is explaination by example:

our TLS is memory of size 0x1000 starting at 0x7ffff6c50000,
but the real size is 0x000068 so we have padding stored at master.offset
= 0xf98

Now our symbol looks as follows

  Offset          Type                Sym. Value    Name
000000432b20  R_X86_64_DTPOFF64   0000000000000058 errno

The old code did 0x7ffff6c50000 + 0xf98 + 000000432b20 which is
obviosly overflowing the memory and wrong.

The right way 0x7ffff6c50000 + 0xf98 + 0000000000000058.

THe Tls base part and offset are added at __tls_get_addr function.
What is left is storing the 0x58 at the relocation address. The problem
is that we don't have 0x58, but we have (binary base + 0x58) in global
symbol table and binary base so what we store is the (binarybase + 0x58
- binary base).

I hope this does turn out to be wrong.