Suspected bug in DW_OP_addr handling

Suspected bug in DW_OP_addr handling

[Yichao Yu]

I noticed that gdb unwind failed with read of invalid memory address
when I used `DW_OP_addr` in my unwind info. Upon checking, it seems
that the handling of this operation in the dwarf interpreter is very
suspicious.

The interpreter for the op code has a comment[1] sayijng,

> Some versions of GCC emit DW_OP_addr before
> DW_OP_GNU_push_tls_address. In this case the value is an
> index, not an address.

However, the code appears to check for exactly the opposite condition
`op_ptr >= op_end || *op_ptr != DW_OP_GNU_push_tls_address` to decide
whether the result should be used as index. This was added 12 years
ago in commit ac56253ddece [2].

Am I missing something or is this a long-standing bug?

Yichao


[1] https://github.com/bminor/binutils-gdb/blob/2392dc0f8e243a4f5f9d0db6cee64a1698886e56/gdb/dwarf2/expr.c#L1574-L1585
[2] https://github.com/bminor/binutils-gdb/commit/ac56253ddece35aff4402b848f88ba40856102b1#diff-9c28445d6bd87cff9dcb11dc4b58ff0e605a078b08ceb37da22d8d2c7b866f6aR406

Re: Suspected bug in DW_OP_addr handling

[Simon Marchi]

On 2022-05-06 23:54, Yichao Yu via Gdb wrote:
> I noticed that gdb unwind failed with read of invalid memory address
> when I used `DW_OP_addr` in my unwind info. Upon checking, it seems
> that the handling of this operation in the dwarf interpreter is very
> suspicious.
> 
> The interpreter for the op code has a comment[1] sayijng,
> 
>> Some versions of GCC emit DW_OP_addr before
>> DW_OP_GNU_push_tls_address. In this case the value is an
>> index, not an address.
> 
> However, the code appears to check for exactly the opposite condition
> `op_ptr >= op_end || *op_ptr != DW_OP_GNU_push_tls_address` to decide
> whether the result should be used as index. This was added 12 years
> ago in commit ac56253ddece [2].
> 
> Am I missing something or is this a long-standing bug?

My understanding is:

 - normally, a value given by DW_OP_addr is an address that should be
   relocated with the base address of the objfile
 - the DW_OP_GNU_push_tls_address operation (now called
   DW_OP_form_tls_address in DWARF 5) expects some kind of value on top
   of the stack, identifying the TLS variable to fetch.  It's
   implementation-defined what this value means, but in practice it
   means it's some value that should not be relocated.
 - a contemporary version of GCC produces something like this for a TLS
   variable's DW_AT_location:

     DW_OP_const8u 0x4, DW_OP_form_tls_address

 - A version of gcc in the past must have used this instead:

     DW_OP_addr 0x4, DW_OP_form_tls_address

The usage of DW_OP_addr was wrong, and it made so GDB had to avoid
relocating the DW_OP_addr value in this particular case.  So it is
checking, if we have DW_OP_addr followed by DW_OP_form_tls_address, then
we don't relocate, because we are in this buggy situation.

And so the condition:

	  if (op_ptr >= op_end || *op_ptr != DW_OP_GNU_push_tls_address)
	    result += this->m_per_objfile->objfile->text_section_offset ();

looks right to me.  It says, only relocate if:

 - DW_OP_addr is the last operation of the sequence
 - the following op is not DW_OP_form_tls_address / DW_OP_GNU_push_tls_address

Simon

Re: Suspected bug in DW_OP_addr handling

[Yichao Yu]

> My understanding is:

Thanks for the detailed explanation!

>
>  - normally, a value given by DW_OP_addr is an address that should be
>    relocated with the base address of the objfile

OK, this is my confusion I guess. I guess I wasn't sure what "address"
refers to in the dwarf spec.
I now see that dwarf spec has a section dedicated to relocation (7.3)
and it mentions that dw_op_addr should be relocated. I was previously
confused by the "constant address" in table 7.9 in 7.7.1 and assumed
that it is the full address.

The following then all makes sense.

>  - the DW_OP_GNU_push_tls_address operation (now called
>    DW_OP_form_tls_address in DWARF 5) expects some kind of value on top
>    of the stack, identifying the TLS variable to fetch.  It's
>    implementation-defined what this value means, but in practice it
>    means it's some value that should not be relocated.
>  - a contemporary version of GCC produces something like this for a TLS
>    variable's DW_AT_location:
>
>      DW_OP_const8u 0x4, DW_OP_form_tls_address
>
>  - A version of gcc in the past must have used this instead:
>
>      DW_OP_addr 0x4, DW_OP_form_tls_address
>
> The usage of DW_OP_addr was wrong, and it made so GDB had to avoid
> relocating the DW_OP_addr value in this particular case.  So it is
> checking, if we have DW_OP_addr followed by DW_OP_form_tls_address, then
> we don't relocate, because we are in this buggy situation.
>
> And so the condition:
>
>           if (op_ptr >= op_end || *op_ptr != DW_OP_GNU_push_tls_address)
>             result += this->m_per_objfile->objfile->text_section_offset ();
>
> looks right to me.  It says, only relocate if:
>
>  - DW_OP_addr is the last operation of the sequence
>  - the following op is not DW_OP_form_tls_address / DW_OP_GNU_push_tls_address
>
> Simon