Re: New TLS usage in libgcc_s.so.1, compatibility impact

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

The 01/13/2024 13:49, Florian Weimer wrote:
> This commit
> 
> commit 8abddb187b33480d8827f44ec655f45734a1749d
> Author: Andrew Burgess <andrew.burgess@embecosm.com>
> Date:   Sat Aug 5 14:31:06 2023 +0200
> 
>     libgcc: support heap-based trampolines
>     
>     Add support for heap-based trampolines on x86_64-linux, aarch64-linux,
>     and x86_64-darwin. Implement the __builtin_nested_func_ptr_created and
>     __builtin_nested_func_ptr_deleted functions for these targets.
>     
>     Co-Authored-By: Maxim Blinov <maxim.blinov@embecosm.com>
>     Co-Authored-By: Iain Sandoe <iain@sandoe.co.uk>
>     Co-Authored-By: Francois-Xavier Coudert <fxcoudert@gcc.gnu.org>
> 
> added TLS usage to libgcc_s.so.1.  The way that libgcc_s is currently
> built, it ends up using a dynamic TLS variant on the Linux targets.
> This means that there is no up-front TLS allocation with glibc (but
> there would be one with musl).
> 
> There is still a compatibility impact because glibc assigns a TLS module
> ID upfront.  This seems to be what causes the
> ust/libc-wrapper/test_libc-wrapper test in lttng-tools to fail.  We end
> up with an infinite regress during process termination because
> libgcc_s.so.1 has been loaded, resulting in a DTV update.  When this
> happens, the bottom of the stack looks like this:
> 
> #4447 0x00007ffff7f288f0 in free () from /lib64/liblttng-ust-libc-wrapper.so.1
> #4448 0x00007ffff7fdb142 in free (ptr=<optimized out>)
>     at ../include/rtld-malloc.h:50
> #4449 _dl_update_slotinfo (req_modid=3, new_gen=2) at ../elf/dl-tls.c:822
> #4450 0x00007ffff7fdb214 in update_get_addr (ti=0x7ffff7f2bfc0, 
>     gen=<optimized out>) at ../elf/dl-tls.c:916
> #4451 0x00007ffff7fddccc in __tls_get_addr ()
>     at ../sysdeps/x86_64/tls_get_addr.S:55
> #4452 0x00007ffff7f288f0 in free () from /lib64/liblttng-ust-libc-wrapper.so.1
> #4453 0x00007ffff7fdb142 in free (ptr=<optimized out>)
>     at ../include/rtld-malloc.h:50
> #4454 _dl_update_slotinfo (req_modid=2, new_gen=2) at ../elf/dl-tls.c:822
> #4455 0x00007ffff7fdb214 in update_get_addr (ti=0x7ffff7f39fa0, 
>     gen=<optimized out>) at ../elf/dl-tls.c:916
> #4456 0x00007ffff7fddccc in __tls_get_addr ()
>     at ../sysdeps/x86_64/tls_get_addr.S:55
> #4457 0x00007ffff7f36113 in lttng_ust_cancelstate_disable_push ()
>    from /lib64/liblttng-ust-common.so.1
> #4458 0x00007ffff7f4c2e8 in ust_lock_nocheck () from /lib64/liblttng-ust.so.1
> #4459 0x00007ffff7f5175a in lttng_ust_cleanup () from /lib64/liblttng-ust.so.1
> #4460 0x00007ffff7fca0f2 in _dl_call_fini (
>     closure_map=closure_map@entry=0x7ffff7fbe000) at dl-call_fini.c:43
> #4461 0x00007ffff7fce06e in _dl_fini () at dl-fini.c:114
> #4462 0x00007ffff7d82fe6 in __run_exit_handlers () from /lib64/libc.so.6
> 
> Cc:ing <lttng-dev@lists.lttng.org> for awareness.
> 
> The issue also requires a recent glibc with changes to DTV management:
> commit d2123d68275acc0f061e73d5f86ca504e0d5a344 ("elf: Fix slow tls
> access after dlopen [BZ #19924]").  If I understand things correctly,
> before this glibc change, we didn't deallocate the old DTV, so there was
> no call to the free function.

with 19924 fixed, after a dlopen or dlclose every thread updates
its dtv on the next dynamic tls access.

before that, dtv was only updated up to the generation of the
module being accessed for a particular tls access.

so hitting the free in the dtv update path is now more likely
but the free is not new, it was there before.

also note that this is unlikely to happen on aarch64 since
tlsdesc only does dynamic tls access after a 512byte static
tls reservation runs out.

> 
> On the glibc side, we should recommend that intercepting mallocs and its
> dependencies use initial-exec TLS because that kind of TLS does not use
> malloc.  If intercepting mallocs using dynamic TLS work at all, that's
> totally by accident, and was in the past helped by glibc bug 19924.  (I

right.

> don't think there is anything special about libgcc_s.so.1 that triggers
> the test failure above, it is just an object with dynamic TLS that is
> implicitly loaded via dlopen at the right stage of the test.)  In this
> particular case, we can also paper over the test failure in glibc by not
> call free at all because the argument is a null pointer:
> 
> diff --git a/elf/dl-tls.c b/elf/dl-tls.c
> index 7b3dd9ab60..14c71cbd06 100644
> --- a/elf/dl-tls.c
> +++ b/elf/dl-tls.c
> @@ -819,7 +819,8 @@ _dl_update_slotinfo (unsigned long int req_modid, size_t new_gen)
>  		 dtv entry free it.  Note: this is not AS-safe.  */
>  	      /* XXX Ideally we will at some point create a memory
>  		 pool.  */
> -	      free (dtv[modid].pointer.to_free);
> +	      if (dtv[modid].pointer.to_free != NULL)
> +		free (dtv[modid].pointer.to_free);
>  	      dtv[modid].pointer.val = TLS_DTV_UNALLOCATED;
>  	      dtv[modid].pointer.to_free = NULL;

can be done, but !=NULL is more likely since we do modid reuse
after dlclose.

there is also a realloc in dtv resizing which happens when more
than 16 modules with tls are loaded after thread creation
(DTV_SURPLUS).

i'm not sure if it's worth supporting malloc interposers that
only work sometimes.

> 
> As the comment hints, we shouldn't be using malloc for TLS memory at all
> because it is not AS-safe, but that's a long-term change.  This change
> seems rather specific to this particular test case failure because it
> relies on libgcc_s.so.1 never using TLS before it gets unloaded.
> 
> Regarding the libgcc_s side, I'm not sure if the TLS usage there should
> be considered a real problem, although I'm a bit nervous about it.
> However, the current implementation caches one page of trampolines past
> the outermost nested function pointer deallocation (otherwise creating
> one function pointer per thread in a loop would be really expensive).
> It looks to me that is never freed, so if the thread exits even with
> proper unwinding (e.g., on glibc with code compiled with -fexceptions),
> there is a memory leak.  Integration with glibc could avoid this issue,
> and also help with the longjmp problem, and fix setcontext/swapcontext,
> too.
> 
> Thanks,
> Florian
>