Gold Linker Patch: Introduce the "retpoline" x86 mitigation technique for variant #2 of the speculative execution vulnerabilities disclosed today, specifically identified by CVE-2017-5715 and in some places called "spectre".

[H.J. Lu]

On Mon, Jan 8, 2018 at 1:09 PM, Sriraman Tallam <tmsriram@google.com> wrote:
> On Mon, Jan 8, 2018 at 12:22 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>> On Mon, Jan 8, 2018 at 12:17 PM, Sriraman Tallam via binutils
>> <binutils@sourceware.org> wrote:
>>> On Mon, Jan 8, 2018 at 11:01 AM, Florian Weimer <fweimer@redhat.com> wrote:
>>>> On 01/08/2018 07:51 PM, Rui Ueyama wrote:
>>>>>
>>>>> A drawback of using BIND_NOW is that an application that has a PLT entry
>>>>> that cannot be resolved but not used fails to start with that option.
>>>>
>>>>
>>>> That can be a good or bad thing, depending on your perspective.  With more
>>>> and more use of symbol versioning, the point is increasingly moot because
>>>> the set of symbol versions is not checked lazily.
>>>
>>> Ok, my attempt to summarize the discussions around this patch:
>>>
>>> a)  We don't need this patch.
>>>       * We could deploy fno-plt and now binding and remove PLTs
>>> altogether. We have to fix correctness issues related to these, like
>>> the one Rui pointed out.
>>
>> What correctness issue?
>
> If a lazy bound symbol cannot be resolved and is not resolved at
> run-time, now binding will expose the issue.

Is this the missing definition issue?  I won't call it as "correctness".  I
consider it as "abuse".  It can even be a security issue when a program
crashes unexpectedly due to the missing definition.

>>
>>>       * One other pain point is we do have internally is we use a
>>> configuration for tests where we build a number of shared objects and
>>> keep the main binary pretty thin.  We have explicitly disabled now
>>> binding for this due to performance reasons, huge increase in the
>>> number of dynamic relocations putting unacceptable overheads on our
>>> distributed build system.  We need to find a solution here.
>>
>> Have you measured performance impact of -fno-plt?
>
> I have conducted some experiments with fno-plt for binaries that
> mostly statically linked with some hot calls to libc.  fno-plt did
> gives us 0.5 %- 1% improvements here and we have plans to turn this on
> for performance sensitive binaries.  fno-plt seems to help in reducing
>  iTLB misses when used in conjunction with kernel huge pages.

So removing PLT isn't that bad for performance.

>>
>>>       * The compiler is eliminating indirect branches and calls
>>> anyway, might as well do it with fno-plt also.  With
>>> -mindirect-branch=think this might also be unnecessary but LLVM
>>> atleast does not support this yet.
>>
>> Shouldn't LLVM be fixed?
>
> Yes,  I can take a look at fixing LLVM fo this.
>

I think you should experiment all options before changing PLT.


-- 
H.J.