On Mon, 2 Oct 2023 at 11:09, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> BTW, is there any reason not to have synchronize_rcu() in delete_module(2),
> just before calling ->exit()?
>
> It's not a hot path, unless something really weird is going on, and it
> would get rid of the need to delay unload_nls() calls...
We already have one - it's hidden in free_module().
It's done after the kallsyms list removal. Is that too late?
Note that module *loading* actually has a few other synchronize_rcu()
calls in the failure path. In fact, load_module() has *two*
"synchronize_rcu()" calls:
- after ftrace_release_mod(mod)
- before releasing module_mutex
for the failure path, but there's only one for module unloading.
Adding one to after ftrace_release_mod() - where we have that
async_synchronize_full() - would make module unloading match the
loading error path.
But the synchronize_rcu calls do seem to be a bit randomly sprinkled
about. Maybe the one in free_module() is already sufficient for nls?
Linus
