Re: [PATCH] elf/dl-deps.c: Make _dl_build_local_scope breadth first

[Mark Hatle <mark.hatle@kernel.crashing.org>]

On 1/12/22 2:12 PM, Adhemerval Zanella wrote:
> 
> 
> On 12/01/2022 16:08, Mark Hatle wrote:
>>
>>
>> On 1/11/22 1:26 PM, Adhemerval Zanella wrote:
>>>
>>>
>>> On 09/12/2021 20:53, Khem Raj via Libc-alpha wrote:
>>>> From: Mark Hatle <mark.hatle@windriver.com>
>>>>
>>>> According to the ELF specification:
>>>>
>>>> When resolving symbolic references, the dynamic linker examines the symbol
>>>> tables with a breadth-first search.
>>>>
>>>> This function was using a depth first search.  By doing so the conflict
>>>> resolution reported to the prelinker (when LD_TRACE_PRELINKING=1 is set)
>>>> was incorrect.  This caused problems when their were various circular
>>>> dependencies between libraries.  The problem usually manifested itself by
>>>> the wrong IFUNC being executed.
>>>>
>>>> Similar issue has been reported here [1]
>>>>
>>>> [BZ# 20488]
>>>>
>>>> [1] https://sourceware.org/legacy-ml/libc-alpha/2016-05/msg00034.html
>>>>
>>>> Signed-off-by: Mark Hatle <mark.hatle@kernel.crashing.org>
>>>> Signed-off-by: Khem Raj <raj.khem@gmail.com>
>>>
>>> I am trying to understand why it this only an issue for LD_TRACE_PRELINKING=1,
>>> do we have a testcase that stress it for a default usercase?
>>
>> The underlying issue here is that resolution is happening depth first and not breadth first.  According to the ELF spec, all resolution should be breadth-first.
>>
>>
>> As noted in item in above, the prelinker just happens to be a way to actually show that the behavior is incorrect.  (There even appears to be a related defect with a reproducer.)
>>
>>
>> When taking the values from LD_TRACE_PRELINKING=1, various addresses and conflict resolutions are specified.  When you compare what is reported, vs what happens, vs what the spec says they don't align as they should.
> 
> That was pretty clear from bug report, what I am trying to understand is
> why it seems to not being reported before in default non-prelinked usage.
> 
> Also, the patch only changes the semantic to prelinked binaries which is
> at least troublesome: this semantic difference at symbol resolution is
> a potential source of issues where user see different behavior depending
> whether prelinked was used or not.
> 
> I am also worried because making such change might potentially trigger
> some hard to diagnostic breakage.
> 
>>
>>
>>> When you say the 'wrong IFUNC being executed' what exactly you mean here?
>>> Could we use a testcase based on this?
>>
>>
>> The prelinker (and possibly just in general), the IFUNC address used is the one from the wrong library scope.  I personally have never tried to reproduce this outside of the prelinking use-case, but based on the referenced report and the code at the time of the change, it is believed this could happen (but rarely) without a prelinked system in a very complex case with multiple libraries providing the same functions.
>>
>> The main issue (my memory is sketchy sorry this might be wrong) is that if one or more functions is an IFUNC and one or more functions is NOT, then the you can get into a situation with a conflict of the wrong function being called using the wrong mechanism.
>>
>> Definitely in the prelink case, the resolver gave the address of a regular function which was then placed into an IFUNC (or maybe it was the other way around) triggering the runtime segfault.
> 
> I asked about ifunc because it should not really be dependent where ifunc
> is used or not, what might be happening is this issue triggers the long
> standing ifunc ordering issue more often.

IFUNC is the only place we ever reproduced it.  At least one of the functions 
had to be an ifunc.  (Or maybe a better way to say it is at least one function 
was an ifunc and one wasn't, but they both had the same name.)

> To summarize, although this change only affects prelinked binaries I think
> it would be better to change for default semantic as well.  We can make
> it a tunable for the transition, like new DSO sorting algorithm with
> glibc.rtld.dynamic_sort, and evaluate on Fedora rawhide, and then hit
> the switch to make it the default.

It absolutely affects prelinked binaries.  At the time we believed it COULD 
affect regular binaries, but we had never seen a failure in the wild.

> We will also need regression testcases, although not sure what kind
> of coverage we will need to provide (our own testsuite currently does not
> trigger any issue though).

This is the test case that was developed at the time for including inside of the 
prelinker:

https://git.yoctoproject.org/prelink-cross/commit/testsuite?id=8f55afd84b3580b1f1d6af904e8c2a39221055b7

It essentially makes three libraries, with two functions (two libraries have the 
same function).  The resulting value expects the breadth-first loading behavior 
in order to generate the correct output.

The main purpose of the prelink test cases was to ensure the same behavior 
before and after the prelinking.  The same should be true with this change to 
glibc, everything should work the same before and after the change, unless the 
stuff before was an error (wrong function used).

We have also been using this patch in the Yocto Project since 2016.  And we've 
never had a report of an incompatibility/failure.  So it really is quite low 
risk, but I'll never say it's "no risk".

>>
>> --Mark
>>
>>>> ---
>>>>    elf/dl-deps.c | 14 ++++++++++----
>>>>    1 file changed, 10 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/elf/dl-deps.c b/elf/dl-deps.c
>>>> index 237d9636c5..e15f7f83d8 100644
>>>> --- a/elf/dl-deps.c
>>>> +++ b/elf/dl-deps.c
>>>> @@ -73,13 +73,19 @@ _dl_build_local_scope (struct link_map **list, struct link_map *map)
>>>>    {
>>>>      struct link_map **p = list;
>>>>      struct link_map **q;
>>>> +  struct link_map **r;
>>>>        *p++ = map;
>>>>      map->l_reserved = 1;
>>>> -  if (map->l_initfini)
>>>> -    for (q = map->l_initfini + 1; *q; ++q)
>>>> -      if (! (*q)->l_reserved)
>>>> -    p += _dl_build_local_scope (p, *q);
>>>> +
>>>> +  for (r = list; r < p; ++r)
>>>> +    if ((*r)->l_initfini)
>>>> +      for (q = (*r)->l_initfini + 1; *q; ++q)
>>>> +    if (! (*q)->l_reserved)
>>>> +      {
>>>> +        *p++ = *q;
>>>> +        (*q)->l_reserved = 1;
>>>> +      }
>>>>      return p - list;
>>>>    }
>>>>