Home Original page

[0/3] Revert to using version index 1

Jan Beulich jbeulich@suse.com
Fri Nov 21 08:15:57 GMT 2025 
On 20.11.2025 16:29, Michael Matz wrote:
> Hello,
> 
> [here's three patches that try to fix some of the recent version index 
> breakage.  I'd like to get approval for 1+2, but see below for why there 
> are three.]
> 
> Defined unversioned symbols certainly can use VER_NDX_GLOBAL (1) as their 
> version index.  There are different opinions if VER_NDX_LOCAL (0) should 
> be usable for those as well, but many tools don't expect that, as in: link 
> editors don't resolve against those (ld.bfd being an exception).
> 
> So, no matter the opinions, the change from NDX_GLOBAL to NDX_LOCAL for 
> defined symbols causes real-world breakage.  It also has no precedent, as 
> the reason for this change was a bug report about inconsistency for 
> _undefined_ symbols vis Solaris.
> 
> IMO it's reasonable to continue using the value 0 for undefined symbols 
> (when the reference itself wasn't versioned); it increases consistency 
> with other OSes and doesn't hurt existing infrastructure.  (I do think 
> value 1 is reasonable as well, FWIW)
> 
> So, to that end the first patch makes it so that NDX_GLOBAL is again 
> generated for defined, global unversioned symbols.  It leaves alone the 
> NDX_LOCAL for undefined syms.
> 
> In connection to all the above readelf was also changed to print 'foo@@' 
> for normal unversioned defined exported symbols (those using NDX_GLOBAL).  
> I think that's harmful, as the symbol to resolve against (and hence what 
> users will try to look up when investigating stuff) is named 'foo', not 
> anything else.  It's also very peculiar and not what other readelfs do.
> 
> To that end the second patch reverts that part of 2be0f2da, but again only 
> for NDX_GLOBAL.  I.e. for symbols with a hidden unnamed version we 
> continue printing 'foo@', which seems useful as an indication of something 
> special going on (in particular a reference to 'foo' can not be resolve by 
> such symbol).
> 
> I really really want to have that second patch in as well.  But just in 
> case the decision is that readelf should indeed continue to print 'foo@@' 
> for normal symbols then patch three is there to fix testsuite fallout from 
> the first patch.
> 
> I've tested the combination 1+2 and 1+2+3 on x86_64-linux on 155 targets 
> without regressions.  Okay for 1+2, please? :)

Okay, perhaps - as you indicated in a reply on the 3/3 sub-thread, in the
order 2, then 1.

Jan
More information about the Binutils mailing list