Re: [PATCH] elf: Remove fallback to the start of DT_STRTAB for dladdr

[Fangrui Song <maskray@google.com>]

On 2022-05-02, Florian Weimer wrote:
>* Fangrui Song:
>
>> On 2022-05-02, Florian Weimer wrote:
>>>* Fangrui Song via Libc-alpha:
>>>
>>>> When neither DT_HASH nor DT_GNU_HASH is present, the code scans
>>>> [DT_SYMTAB, DT_STRTAB). However, there is no guarantee that .dynstr
>>>> immediately follows .dynsym (e.g. lld typically places .gnu.version
>>>> after .dynsym).
>>>
>>>The code is compatible with lld because it always generates a hash
>>>table.  Maybe it was added to support old binaries without a hash table.
>>>So we would have to check if such binaries exist from the early
>>>libc.so.6 days (or check if binutils every generated ELF binaries
>>>without a hash table).  The glibc comment isn't clear if this was added
>>>because it was required at the time, or just because it seemed a good
>>>idea.
>>>
>>>I couldn't find any binaries with DT_SYMTAB, but without DT_HASH or
>>>DT_GNU_HASH in my collection, but doesn't mean that they don't exist.
>>>
>>>Thanks,
>>>Florian
>>
>> With a linker script .hash and .gnu.hash can be removed.
>> But such an object behaves as if it has no symbol: symbol search will fail.
>> It makes sense for dladdr to not return a symbol for it.
>>
>> % bmake
>> cc -pipe -g   -fuse-ld=bfd -g -fpic -shared -Wl,--version-script=b.ver b.c -o b.so
>> cc -pipe -g   -fuse-ld=bfd -g a.c -Wl,--no-as-needed -fno-pie -no-pie -Wl,-rpath=/tmp/d b.so -ldl -o a
>> % ./a
>> 42
>>
>> % cat b.lds   # GNU ld doesn't have this yet: https://sourceware.org/bugzilla/show_bug.cgi?id=26404
>> OVERWRITE_SECTIONS {
>>   /DISCARD/ : { *(.hash) *(.gnu.hash) }
>> }
>> % clang -fpic -fuse-ld=lld -shared b.c -Wl,b.lds -o b.so
>> % ./a
>> ./a: symbol lookup error: ./a: undefined symbol: var
>
>This looks like it might be an lld bug.  DT_HASH is mandatory in the ELF
>specification.  We ignore that requirement in the GNU ABI and use
>DT_GNU_HASH instead, even for static PIE binaries.

I am not sure supporting discarding .hash/.gnu.hash is a bug.
The linker can even discard .dynsym/.dynstr/.dynamic .

In the Linux kernel, there is something like

         /DISCARD/ : {
                 *(.interp .dynamic)
                 *(.dynsym .dynstr .hash .gnu.hash)
         }

If we argue that DT_HASH is mandatory, then --hash-style=gnu breaks the ELF
specification as well.

My point is that a normal dynamic linking object always has either .hash or .gnu.hash .
If neither is present, it seems pretty fair for dladdr to not give a symbol,
just like symbol search (including relocation resolving and dl(v)sym) won't give any result.

>Do you want to drop symbol tables from static PIE binaries?

I don't have the plan.