[PATCH v2] elf: Support DT_RELR relative relocation format [BZ #27924]

[Joseph Myers]

On Mon, 18 Oct 2021, H.J. Lu via Libc-alpha wrote:

> On Mon, Oct 18, 2021 at 3:30 PM Joseph Myers <joseph@codesourcery.com> wrote:
> >
> > On Mon, 18 Oct 2021, H.J. Lu via Libc-alpha wrote:
> >
> > > If a binary requires a new glibc to run, running it with the old glibc will
> > > lead to an ld.so diagnostic message.   DT_RELR shouldn't be different.
> >
> > My point is that in many cases (for example, a binary using a new flag as
> > an argument to a function taking a flags argument) it *won't* lead to an
> > ld.so diagnostic message, just to the relevant function call misbehaving
> > at runtime.
> 
> Do you have such a testcase?  It sounds like a glibc bug to me.

We could take my recently submitted patch 
<https://sourceware.org/pipermail/libc-alpha/2021-October/131905.html> for 
printf %b and %B as an example of the sort of thing we haven't considered 
to need new symbol versions in the past.  If someone builds a program 
using any of about 70 printf-like functions (three times as many on 
powerpc64le because of long double variants) with a %b or %B format, it 
won't work correctly with older glibc versions, but the symbol versions 
won't prevent it from running on older glibc versions.  We haven't 
considered new printf/scanf/strftime/strfmon formats to need new symbol 
versions before (and if we did, there would be a lot of such versions in 
the printf case because there are a lot of printf-like functions).  (The 
strtol-like and scanf-like functions will need new symbol versions for the 
binary input changes because they will also affect the handling of base 2 
/ base 0 input starting with 0b, i.e. change semantics of code that was 
already valid with previous glibc and standard versions.  But there are 
fewer such functions.)

-- 
Joseph S. Myers
joseph@codesourcery.com