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

[Szabolcs Nagy <szabolcs.nagy@arm.com>]

The 04/09/2024 10:50, Fangrui Song wrote:
> On Tue, Apr 9, 2024 at 7:03 AM Adhemerval Zanella Netto
> <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:
> > >>> 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).
> 
> Last time I analyzed the __tls_get_addr interceptor in sanitizers, I
> have made quite some notes at
> https://maskray.me/blog/2021-02-14-all-about-thread-local-storage#why-does-compiler-rt-need-to-know-tls-blocks

says

> Note: aarch64 uses TLSDESC by default and there is no interposable symbol.

not quite true: the tlsdesc call cannot be interposed, but then
ld.so internally calls __tls_get_addr via a PLT that can be.
(only for first time dynamic tls setup, but that's enough to
get dynamic tls boundaries)

you need another way to detect the static tls area, but that's
true for non-tlsdesc too.

> Current TLSDESC implementations not providing an interposable symbol
> is indeed an issue, but I haven't carefully analyzed it on an aarch64
> machine.

we are planning to remove the interposable symbol in the
next glibc version, so the current __tls_get_addr trick
won't work for x86_64, aarch64, loongarch, riscv.