CHECK_STATIC_TLS in R_X86_64_TPOFF64 type relocation

CHECK_STATIC_TLS in R_X86_64_TPOFF64 type relocation

[Fengkai Sun]

Hi list,

I noticed that when resolving R_X86_64_TPOFF64 type relocations, the
runtime dynamic linker would call the macro CHECK_STATIC_TLS, which can be
expanded into a call to _dl_try_allocate_static_tls. (in
sysdeps/x86_64/dl-machine.h)

The comments of _dl_try_allocate_static_tls says:

> We are trying to perform a static TLS relocation in MAP, but it was
dynamically loaded.

To the best of my knowledge, I found that R_X86_64_TPOFF64 is for
Initial-exec TLS model, which only appears in executables.

Since glibc has refused to dlopen PIEs since 2.29, why does it not see it
as an error, but still try to allocate static TLS for PIEs?

---
Thanks,
Fengkai

Re: CHECK_STATIC_TLS in R_X86_64_TPOFF64 type relocation

[Florian Weimer]

* Fengkai Sun via Libc-help:

>> We are trying to perform a static TLS relocation in MAP, but it was
> dynamically loaded.
>
> To the best of my knowledge, I found that R_X86_64_TPOFF64 is for
> Initial-exec TLS model, which only appears in executables.

No, initial-exec in shared objects is fine (as long as static TLS
space is available).

Local-exec TLS in executables does not need relocations, and the link
editor typically relaxes initial-exec TLS reference in main programs
to local-exec TLS, leaving no of them trace (except for the PT_TLS
segment).

Re: CHECK_STATIC_TLS in R_X86_64_TPOFF64 type relocation

[Fengkai Sun]

Hi Florian,

Thanks for your reply!

I figure out that dlopened DSOs could use the initial-exec model: as long
as they can have access to static TLS, they are the same as DSOs introduced
by the main executable.

However, I am confused why (and when) the link editor emits initial-exec
for DSOs, because this model is not guaranteed to work. If the TLS segment
in a dlopened DSO is too large to fit in static space, it will fail the
main program and the user does not have a way to fix it. In contrast, the
two dynamic models always work.

Sorry if this sounds deviated and like a linker-related question.


---
Thanks,
Fengkai

On Sun, Dec 26, 2021 at 11:48 PM Florian Weimer <fw@deneb.enyo.de> wrote:

> * Fengkai Sun via Libc-help:
>
> >> We are trying to perform a static TLS relocation in MAP, but it was
> > dynamically loaded.
> >
> > To the best of my knowledge, I found that R_X86_64_TPOFF64 is for
> > Initial-exec TLS model, which only appears in executables.
>
> No, initial-exec in shared objects is fine (as long as static TLS
> space is available).
>
> Local-exec TLS in executables does not need relocations, and the link
> editor typically relaxes initial-exec TLS reference in main programs
> to local-exec TLS, leaving no of them trace (except for the PT_TLS
> segment).
>