Re: [PATCH] aarch64: Remove ld.so __tls_get_addr plt usage

[Palmer Dabbelt <palmer@dabbelt.com>]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 4015 bytes --]

On Tue, 09 Apr 2024 07:03:45 PDT (-0700), adhemerval.zanella@linaro.org wrote:
>
>
> On 09/04/24 05:30, Szabolcs Nagy wrote:
>> The 04/08/2024 13:57, Adhemerval Zanella Netto wrote:
>>> On 08/04/24 04:26, Szabolcs Nagy wrote:
>>>> The 04/07/2024 16:29, Cristian Rodríguez wrote:
>>>>> On Fri, Apr 5, 2024 at 11:59 AM Szabolcs Nagy <szabolcs.nagy@arm.com> wrote:
>>>>>> The 04/05/2024 09:35, Adhemerval Zanella wrote:
>>>>>>> Use the hidden alias instead.
>>>>>>>
>>>>>>> Checked on aarch64-linux-gnu.
>>>>>>
>>>>>> does this change behaviour in case __tls_get_addr is interposed?
>>>>>
>>>>> Wut ? is that really supported.. I mean.. isn't that symbol prefix
>>>>> reserved for the implementation and any assumption about it is either
>>>>> ID or UB?
>>>>
>>>> a behaviour can change even if it's not supported.
>>>> i did not try to imply that it should be supported.
>>>>
>>>> i know sanitizers interpose __tls_get_addr, because
>>>> https://sourceware.org/bugzilla/show_bug.cgi?id=16291
>>>> i don't know if that hack works at all now for tlsdesc
>>>> (where the ld.so calls __tls_get_addr, not user code)
>>>>
>>>> my question was if we investigated this issue since it
>>>> is useful to document then in the commit msg (or news
>>>> entry if this affects users)
>>>
>>> This change 'breaks' the sanitizer trick to get the dynamic TLS, with
>>> this patch I now see:
>>>
>>>   MemorySanitizer-AARCH64 :: dtls_test.c
>>>   SanitizerCommon-asan-aarch64-Linux :: Linux/resize_tls_dynamic.cpp
>>>   SanitizerCommon-msan-aarch64-Linux :: Linux/resize_tls_dynamic.cpp
>>>   SanitizerCommon-tsan-aarch64-Linux :: Linux/resize_tls_dynamic.cpp
>>>
>>> And it does not fail on x86 only because it uses -mtls=gnu as default
>>> (the same tests fail on x86 with -mtls=gnu2).
>>>
>>> Now that GCC and distributions are aiming to use GNU2/DESC as the
>>> default TLS, this hack will also break on x86.  So the question is
>>> whether we revert 050f7298e1ecc39887c329037575ccd972071255 and
>>> document that __tls_get_addr should be interposable, or move with this
>>> change and try to come up with a possible solution for BZ#16291.
>>>
>>> I bringing this because we will have another two ABIs with tlsdesc
>>> support (loongarch and riscv).
>>
>> adding some sanitizer committers to cc.
>>
>> tl;dr: in the next glibc release tlsdesc will not call
>> __tls_get_addr in an interposable way in the dynamic tls
>> allocation case, unless somebody screems that this is needed.
>> (affects targets that may default to tlsdesc, but note that
>> the dynamic case only triggers with tlsdesc when a lot of
>> dlopened tls is used, otherwise static tls area is used)
>
> Just a note that this already true for x86 with -mtls=gnu2 since
> 2.21.  And now that distro are aiming to make it default, this issues
> will happen more often.
>
>>
>> i think it is also possible that we will use custom malloc
>> in ld.so which may be just as big change for the sanitizers.
>> (this can make tls access signal safe)
>>
>> i'm not against the change, but if we plan to add several
>> interposable hooks as in
>> https://sourceware.org/glibc/wiki/ThreadPropertiesAPI
>> then we might as well keep __tls_get_addr PLT for now.
>>
>
> I don't have a strong opinion, but what I really want is to have
> consistency over the architectures.  Meaning that if we want to keep
> the __tls_get_addr PLT for sanitizer/runtime hooks, it would be good
> to revert the x86 change.
>
> It also means to document it properly somewhere and make the new
> RISC-V and loongarch follow the same guidelines.

I also don't have a strong opinion on whether __tls_get_addr should be 
interposable, but I'm happy to try and make the RISC-V port match 
arm64/x86.  I guess we're kind of safe for now as we don't have TLSDESC 
merged, though I think we were getting pretty close there so we should 
probbaly decide before we accidentally commit to an ABI.

> I will take a look again on the ThreadPropertiesAPI, since it is has
> been more and more a demanding issue.