Re: [RFC 4/5] elf: Enable RTLD_NODELETE on __libc_unwind_link_get

[Florian Weimer <fweimer@redhat.com>]

* Adhemerval Zanella Netto:

> On 12/06/24 06:54, Florian Weimer wrote:
>> * Adhemerval Zanella:
>> 
>>> The libgcc_s.so can also be sealed.  The library is loaded once
>>> and not unloaded during process execution (only for memory debug
>>> with __libc_unwind_link_freeres).
>> 
>> The unwind-link change to use RTLD_NODELETE could go in separately.
>
> Ok, I will send a separate patch.
>
>> 
>>> diff --git a/misc/unwind-link.c b/misc/unwind-link.c
>>> index 213a0162a4..7267ecbec3 100644
>>> --- a/misc/unwind-link.c
>>> +++ b/misc/unwind-link.c
>> 
>>> @@ -100,7 +100,8 @@ __libc_unwind_link_get (void)
>>>  
>>>    __libc_lock_lock (lock);
>>>    if (atomic_load_relaxed (&global_libgcc_handle) != NULL)
>>> -    /* This thread lost the race.  Clean up.  */
>>> +    /* This thread lost the race.  Drop the l_direct_opencount and issue
>>> +       the debug log.  */
>>>      __libc_dlclose (local_libgcc_handle);
>>>    else
>>>      {
>> 
>> I don't understand what “debug log” means in this context.
>
> Sorry, I meant the __libc_unwind_link_freeres usually triggered by
> memory profilers.  With sealing the dlclose won't unmap the memory.

I still don't understand the comment.

We can still deallocate the helper data structures.  With the switch to
read-only link maps, most of the allocations will be hidden from malloc
tracing anyway and no longer appear as leaks.

Thanks,
Florian