Re: [EXTERNAL] Re: Hitting g dwfl->lookup_elts limit in report_r_debug, so not all modules show up and backtracing fails

[Luke Diamand <ldiamand@roku.com>]

On 08/05/2023 17:35, Mark Wielaard wrote:
> Hi Florian, Hi Luke,
> 
> On Tue, 2023-05-02 at 09:57 +0200, Florian Weimer via Elfutils-devel
> wrote:
>> * Luke Diamand via Elfutils-devel:
>>
>>> I've got a few cores where report_r_debug() in link_map.c fails to
>>> find all of the modules - for example I had libc.so missing. This
>>> obviously meant that elfutils could not backtrace my core.
>>>
>>> It seems to be related to this code:
>>>
>>>    /* There can't be more elements in the link_map list than there are
>>>       segments.  DWFL->lookup_elts is probably twice that number, so it
>>>       is certainly above the upper bound.  If we iterate too many times,
>>>       there must be a loop in the pointers due to link_map clobberation.  */
>>>    size_t iterations = 0;
>>>
>>>    while (next != 0 && ++iterations < dwfl->lookup_elts)
>>>
>>> I've changed this to just keep going until it reaches
>>> dwfl->lookup_elts*5, which seems to "fix" it, but I feel there must be
>>> a better fix!
>>>
>>> The most recent core I saw with this had lookup_elts=36, and hit 109
>>> iterations of the loop and then backtraced just fine.
>>
>> It's probably another fallout from -z separate-code, which tends to
>> create four LOAD segments.  The magic number 5 sounds about right, as
>> gold also has -z text-unlikely-segment, which might result in creating
>> that number of load segments (but I haven't tried).
> 
> Wow, that had never occurred to me. Thanks.
> 
> Luke does the binary/libraries from which your core file was generated
> contain multiple PT_LOAD segments?
> 
> We could add something like:
> 
> diff --git a/libdwfl/link_map.c b/libdwfl/link_map.c
> index 06d85eb6..76f23354 100644
> --- a/libdwfl/link_map.c
> +++ b/libdwfl/link_map.c
> @@ -331,11 +331,17 @@ report_r_debug (uint_fast8_t elfclass, uint_fast8_t elfdata,
>     int result = 0;
>   
>     /* There can't be more elements in the link_map list than there are
> -     segments.  DWFL->lookup_elts is probably twice that number, so it
> -     is certainly above the upper bound.  If we iterate too many times,
> -     there must be a loop in the pointers due to link_map clobberation.  */
> +     segments.  A segment is created for each PT_LOAD and there can be
> +     up to 5 per module (-z separate-code, tends to create four LOAD
> +     segments, gold has -z text-unlikely-segment, which might result
> +     in creating that number of load segments) DWFL->lookup_elts is
> +     probably twice the number of modules, so that multiplied by max
> +     PT_LOADs is certainly above the upper bound.  If we iterate too
> +     many times, there must be a loop in the pointers due to link_map
> +     clobberation.  */
> +#define MAX_PT_LOAD 5
>     size_t iterations = 0;
> -  while (next != 0 && ++iterations < dwfl->lookup_elts)
> +  while (next != 0 && ++iterations < dwfl->lookup_elts * MAX_PT_LOAD)
>       {
>         if (read_addrs (&memory_closure, elfclass, elfdata,
>                        &buffer, &buffer_available, next, &read_vaddr,
> 
> Does that sound reasonable?

Sorry - I did not see this until just after sending in my patch!

Let me try it with this change and I will re-roll it.

Luke