Re: RFC: x86: Fall back to lazy binding for unrelocated IFUNC symbol [BZ #23240]

[Florian Weimer <fweimer at redhat dot com>]

On 06/06/2018 05:38 PM, H.J. Lu wrote:
> I pushed hjl/pr23240/fw, which is fw/bug21242 + tests for [BZ #23176] and
> [BZ #23240].
>
> 1. On x86-64:
>
> FAIL: elf/reldep6a

The crash happens because “weak” is undefined in reldep6mod1.so.

The “weak” function is defined in reldep6mod4.so.  But reldep6mod1.so is 
loaded before reldep6mod4.so:

     14289:     calling init: 
/home/fweimer/src/gnu/glibc/build/elf/reldep6mod1.so
…
     14289:     opening 
file=/home/fweimer/src/gnu/glibc/build/elf/reldep6mod4.so [0]; 
direct_opencount=1
     14289:
     14289:     binding file 
/home/fweimer/src/gnu/glibc/build/elf/reldep6mod4.so [0] to 
/home/fweimer/src/gnu/glibc/build/elf/reldep6mod1.so [0]: normal symbol 
`baz'

I think this is simply invalid.  I think that if you had a “weak == 
NULL” check in reldep6mod1.so, it would be false even after loading 
reldep6mod4.so because the address is not lazily bound.

Is this derived from the X server use case, where they load additional 
DSOs which provide definitions of functions which are lazily bound? 
That's just not compatible with BIND_NOW, and I don't see a compelling 
reason why it should be.  If it's just about fixing X, then we need to 
fix X, and not enhance the dynamic linker.

> 2. On i686:
>
> FAIL: elf/ifuncpreload1
> FAIL: elf/reldep6a

There's no delayed processing for i686 yet.  I wanted to get consensus 
on the overall approach first.

Thanks,
Florian