[PATCH v15 0/8] Add rseq extensible ABI support

[Michael Jeanson]

On 2025-01-08 06:05, Florian Weimer wrote:
> * Michael Jeanson:
> 
>> Introduced in Linux v6.3 the rseq extensible ABI [1] will allow adding
>> rseq features past the initial 32 bytes of the original ABI.
> 
> I've completed my review.  I think we are getting really close.
> 
> Andreas, do you think we can still merge this for glibc 2.41?  Ideally I
> would give this some soak time in Fedora rawhide, maybe a week or two.
> So if you want to release next Monday, merging this later this the week
> would not be a good idea.
> 
> The pointer/size arithmetic looks in the extra TLS setup looks correct
> to me (except for two nits, one of them not relevant to Linux).  But
> incorrect TLS allocation overlap does not always lead to immediate
> issues.  And the other problem is that we do not have a kernel that
> truly exercises the extension bits (as seen by the open question of the
> relationship between the size/alignment values).

I've tested earlier revisions of this patchset with custom kernels that
had different combinations of AT_RSEQ_FEATURE_SIZE and AT_RSEQ_ALIGN.
Once we are done with your reviews, I can redo those tests on a few
architectures. It wouldn't guarantee we don't have any TLS overlap
issues but it could give us little but more confidence.

> 
> If we want to merge this, we also need the TLS_TP_OFFSET series:
> 
>   [PATCH v2 0/3] Define TLS_TP_OFFSET on all architectures
>   <https://inbox.sourceware.org/libc-alpha/cover.1736275068.git.fweimer@redhat.com/>
> 
> Thanks,
> Florian
>