vDSO parsing bug?

vDSO parsing bug?

[Andy Lutomirski]

Hi-

It looks like musl can't handle vDSO images with more than one PT_LOAD
segment.  Both glibc and the reference code in
Documentation/vDSO/parse_vdso.c can handle this.

It's not currently an issue, since I don't think that any
architectures provide such a vDSO, but they might want to in the
future.

Also, using the name "linux-gate.so.1" is sort of wrong on x86_64,
where the convention is "linux-vdso.so.1".  Wouldn't it be better to
just read the name out of the SONAME header?

--Andy

Re: vDSO parsing bug?

[Rich Felker]

On Thu, Mar 20, 2014 at 08:23:24AM -0700, Andy Lutomirski wrote:
> Hi-
> 
> It looks like musl can't handle vDSO images with more than one PT_LOAD
> segment.  Both glibc and the reference code in
> Documentation/vDSO/parse_vdso.c can handle this.

I'm not clear on what you think is wrong. The computation of
vdso->base? That's all musl uses the PT_LOAD header for.

> It's not currently an issue, since I don't think that any
> architectures provide such a vDSO, but they might want to in the
> future.
> 
> Also, using the name "linux-gate.so.1" is sort of wrong on x86_64,
> where the convention is "linux-vdso.so.1".  Wouldn't it be better to
> just read the name out of the SONAME header?

Yes, that sounds like a better solution. I just didn't realize it was
an option.

BTW the vdso usage is currently broken as far as I can tell. Due to
the way musl binds function references at link time, the weak-alias
setup that's intended to allow the vdso functions to interpose does
not actually work.

I'm thinking it might be better to avoid accessing the vdso through
the dynamic linker like this and simply parsing it directly in code
that can be used in static-linked programs too. Then it's not clear to
me whether the dynamic linker would even need to report the vdso in
the link map, but maybe it still should to assist gdb...?

I'd welcome further input on this topic.

Rich

Re: vDSO parsing bug?

[Andy Lutomirski]

On 03/20/2014 10:02 AM, Rich Felker wrote:
> On Thu, Mar 20, 2014 at 08:23:24AM -0700, Andy Lutomirski wrote:
>> Hi-
>>
>> It looks like musl can't handle vDSO images with more than one PT_LOAD
>> segment.  Both glibc and the reference code in
>> Documentation/vDSO/parse_vdso.c can handle this.
> 
> I'm not clear on what you think is wrong. The computation of
> vdso->base? That's all musl uses the PT_LOAD header for.

Hmm.  I guess if all the PT_LOAD segments have the same offset-vaddr,
then all is well.  I'm not sure what happens if there's a filesz=0
PT_LOAD segment, though.

> 
>> It's not currently an issue, since I don't think that any
>> architectures provide such a vDSO, but they might want to in the
>> future.
>>
>> Also, using the name "linux-gate.so.1" is sort of wrong on x86_64,
>> where the convention is "linux-vdso.so.1".  Wouldn't it be better to
>> just read the name out of the SONAME header?
> 
> Yes, that sounds like a better solution. I just didn't realize it was
> an option.
> 
> BTW the vdso usage is currently broken as far as I can tell. Due to
> the way musl binds function references at link time, the weak-alias
> setup that's intended to allow the vdso functions to interpose does
> not actually work.
> 
> I'm thinking it might be better to avoid accessing the vdso through
> the dynamic linker like this and simply parsing it directly in code
> that can be used in static-linked programs too. Then it's not clear to
> me whether the dynamic linker would even need to report the vdso in
> the link map, but maybe it still should to assist gdb...?
> 
> I'd welcome further input on this topic.

It's almost certainly a bad idea to expose vdso symbols as normal
symbols (in the RTLD_GLOBAL sense).  For example, the 64-bit vdso
defines gettimeofday, and that version of gettimeofday doesn't set errno.

That being said, it is useful to be able to fish out symbols with dlsym.
 And gdb may care.

But statically-linked programs should be able to use the vdso, too.  Go
can do this.

--Andy