[PATCH v9 3/5] Fix assert during static startup (BZ 33326)

[Adhemerval Zanella Netto]

On 08/06/26 11:26, Adhemerval Zanella Netto wrote:
> 
> 
> On 08/06/26 09:14, Adhemerval Zanella Netto wrote:
>>
>>
>> On 08/06/26 07:46, Florian Weimer wrote:
>>> * Adhemerval Zanella:
>>>
>>> I think the general issue is not right: During startup, we should not
>>> use mmap to allocate the backing store.  We can preserve the assert
>>> message and file name by copying them to on-stack buffer.  As no user
>>> code should be running at this point, we do not need to be prepared for
>>> arbitrarily-long assert messages (but we still should have a length
>>> check, of course).
>>>
>>> Can we redirect to a different assert implementation during early
>>> startup?
>>
>> I think I can reinstate the assert wrapper [1]. I kept the assert used 
>> simplify the the implementation and use the same code path.  We can use
>> a stack allocated buffer and cap the message string, for glibc own usage we
>> would know when we need to increase it.
>>
>> But it also raises the question whether we do copy the buffer in this case,
>> this assert will be only used during process startup and it would be triggered
>> in cases that can be considered glibc issues.
>>
>> [1] https://sourceware.org/pipermail/libc-alpha/2026-March/176039.html
> It turned out that for the rtld it is already done by elf/dl-minimal.c
> __assert_fail/__assert_perror_fail which only calls _dl_fatal_printf instead
> of __libc_assert_fail (which in turn calls __libc_message_impl).
> 
> Now changing this to *static-pie* means making static-pie act more like
> rtld and less lime a static binary: we will need to add some logic on
> the libc __libc_message_impl to handle static-pie and/or adding some
> mechanism to override assert (similar how it is done for rtld). I don't
> think either worth the trouble.

I am planning to send a new version with your remarks fixes, do you still have
reservations for this assert on static-pie Florian? This patch does not really
change any semantic, loader still uses the minimal implementation.