Bug 2251557

Upstream bug report (from 2022-04 but nothing was done): https://sourceware.org/bugzilla/show_bug.cgi?id=29039

Here's the sequence of events:

- Early on during dynamic module load, `_dl_assign_tls_modid` is called to find a free module ID for the module. It finds a free slotinfo entry, and sets its map to point to the module, marking it as used. At this point the generation ID is the default zero.
- Relocations are processed. Dynamic TLS relocations create their associated data by calling `_dl_make_tlsdesc_dynamic`. This calls `map_generation` to identify the generation for this module. That goes through the slot info, looking for an entry that already has the current map (module) assigned and a valid generation. We do have an entry, but the generation is zero, so that fails. The fallback path then assigns the current global generation + 1.
- Once all relocations have been processed, the linker calls `_dl_add_to_slotinfo`. This finally sets up the slot info to the current map (which was already set anyway), and the current global generation + 1.
- Whenever a TLS symbol is used for this module, it will call `_dl_tlsdesc_dynamic`. The first time, that will check the DTV generation against the symbol generation. Since it will be too old, it calls into the slow path via `__tls_get_addr`, which checks the generation and calls `_dl_update_slotinfo`
- `_dl_update_slotinfo` finally expands the DTV if needed, and initializes the DTV entry with the pointer `TLS_DTV_UNALLOCATED`. The caller then checks that, and finally allocates the TLS block for this thread/module.
- When a module is unloaded, `remove_slotinfo` will find its slot info entry, and set the map to NULL and the generation to, again, the current generation + 1 (which will have incremented by now if a DTV update happened in the interim).

This all goes wrong when a modid is reused:

- A module with TLS is dynamically loaded. It gets assigned a slot, a generation, everything. So far so good.
- That module **never uses its TLS symbols**, so the DTV resize/allocation never happens.
- The module is unloaded, which replaces its slot pointer with NULL and bumps up the generation
- A bunch of other stuff happens involving TLS, which causes the generation to increase and the DTV to be updated. Since the module was unloaded, its DTV entry is no longer in use and is not initialized.
- Another module is loaded. It is assigned the same module ID. **`_dl_assign_tls_modid` assigns it to the same slot, but leaves the generation untouched, which is now nonzero since it was never cleared.**
- Since the generation is nonzero, relocation processing goes ahead and **puts that old generation into the TLS relocation entries**. This generation is now obsolete, since the global generation increased in the interim.
- After relocations, `_dl_add_to_slotinfo` is finally called, which bumps up the generation in the slot info to the current one + 1. **But the relocations were already processed with an old generation!**
- On first access to TLS data, `_dl_tlsdesc_dynamic` checks the generation. Since the DTV generation is newer than the symbol generation, it checks the pointer against `TLS_DTV_UNALLOCATED`, **which is (void*)-1**. Since the DTV entry was never initialized, its value is zero, so it returns from the happy path.
- Some pointer arithmetic nonsense later, that NULL pointer is dereferenced.

This is currently breaking GNOME in Fedora Asahi.

Reproducible: Always

Steps to Reproduce:
1. Install Fedora Asahi GNOME
2. Try to start an X11 GNOME session
Actual Results:  
Mutter crashloops

Expected Results:  
Mutter does not crashloop