Re: TLS issue on aarch64

[Szabolcs Nagy <nsz@port70.net>]

* Rich Felker <dalias@libc.org> [2018-05-31 20:52:00 -0400]:
> On Fri, Jun 01, 2018 at 02:11:02AM +0200, Szabolcs Nagy wrote:
> > 	aarch64: tp + alignup(16, align) must be aligned == tp must be aligned
> 
> OK, I see two possible solutions here:
> 
> 1. tp==self+sizeof(struct pthread). In this case we'll waste some
> space (vs the current approach) when no extra alignment is needed, but
> it's simple and clean because all the alignments match up naturally.
> 
> 2. tp==self+sizeof(struct pthread)-16 (or rather -reserved in
> general). This preserves the current memory usage, but requires
> complex new alignment logic since self will no longer be aligned mod
> tls_align when tls_align>reserved.
> 
> I pretty strongly prefer option 1.
> 

ok.

> In either case, the main_tls.offset/app.tls.offset value needs to
> correctly reflect the offset of the TLS from TP, so it either needs to
> be alignup(reserved,tls_align) or alignup(reserved,tls_align)-reserved
> depending on option 1 or 2. After that change is made, we need to make
> sure the storage needs (libc.tls_size) are computed correctly and
> account for the extra space due to the initial positive offset.
> 
> No change is then needed in __copy_tls.
> 
> Changes to TP_ADJ and __pthread_self are needed to get reserved out of
> them, and the value of reserved needs to be provided somewhere else
> for computing main_tls.offset.
> 

ok.
i'll try to prepare a patch.

> > for initial-exec to work:
> > 	tp + *got - add must be aligned (i.e. *got has to be set up to meet
> > 	the alignment requirement of the module, this does not seem to require
> > 	realignment of tp so even runtime loading of initial-exec tls should
> > 	be possible assuming there is enough space etc...)
> 
> There's never space so it's not even a question, but even if there
> were, no, it can't be done because tp will not be aligned mod some
> possibly-larger alignment than the alignment in effect at the time the
> thread was created.
> 

ah right because there are many threads with different tp
so tp + *got can only be aligned if tp modulo alignment
is the same in all of them.