Re: [PATCH] ftrace: Fix missing synchronize_rcu() removing trampoline from kallsyms

[Adrian Hunter]

On 11/09/20 2:41 pm, peterz@xxxxxxxxxxxxx wrote:
> On Tue, Sep 01, 2020 at 12:16:17PM +0300, Adrian Hunter wrote:
>> Add synchronize_rcu() after list_del_rcu() in
>> ftrace_remove_trampoline_from_kallsyms() to protect readers of
>> ftrace_ops_trampoline_list (in ftrace_get_trampoline_kallsym)
>> which is used when kallsyms is read.
>>
>> Fixes: fc0ea795f53c8d ("ftrace: Add symbols for ftrace trampolines")
>> Signed-off-by: Adrian Hunter <adrian.hunter@xxxxxxxxx>
>> ---
>>  kernel/trace/ftrace.c | 1 +
>>  1 file changed, 1 insertion(+)
>>
>> diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
>> index 275441254bb5..4e64367c9774 100644
>> --- a/kernel/trace/ftrace.c
>> +++ b/kernel/trace/ftrace.c
>> @@ -2782,6 +2782,7 @@ static void ftrace_remove_trampoline_from_kallsyms(struct ftrace_ops *ops)
>>  {
>>  	lockdep_assert_held(&ftrace_lock);
>>  	list_del_rcu(&ops->list);
>> +	synchronize_rcu();
>>  }
> 
> 
> Hurmph, we've just done a ton of that:
> 
> 
> 	ftrace_shutdown()
> 	  synchronize_rcu_tasks_rude()
> 	  ftrace_trampoline_free()
> 	    ftrace_remove_trampoline_from_kallsyms()
> 
> 
> So would it not be better to move that call before the existing
> synchronize_rcu_tasks stuff rather than adding another synchronize_rcu()
> call?

Doesn't that mean removing the symbol while the trampoline is potentially
still in use?

Could follow up the fix with a patch to allocate list nodes instead, and use
call_rcu() to free it, so another synchronize_rcu() is not needed.