Re: [PATCH 1/2] task_struct cleanup: move binfmt field to mm_struct

[Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>]

Oleg Nesterov wrote:
> (add Rusty)
> 
> On 07/27, Hiroshi Shimamoto wrote:
>>  void set_binfmt(struct linux_binfmt *new)
>>  {
>> -	if (current->binfmt)
>> -		module_put(current->binfmt->module);
>> +	struct mm_struct *mm = current->mm;
>> +
>> +	BUG_ON(!mm);
>> +	if (mm->binfmt)
> 
> I am not sure we need this BUG_ON() above. If mm == NULL we will
> have the same debug info after the crash.

Ah, right. It will crash at accessing mm->binfmt and we'll get the
same information. I didn't think seriously.
BTW, now I noticed that it looks similar the security issue with
NULL page mapping. If the kernel has the page mapping to NULL,
it causes unstable behavior, no?

> 
>> @@ -619,6 +621,9 @@ struct mm_struct *dup_mm(struct task_struct *tsk)
>>  	mm->hiwater_rss = get_mm_rss(mm);
>>  	mm->hiwater_vm = mm->total_vm;
>>
>> +	if (mm->binfmt && !try_module_get(mm->binfmt->module))
>> +		goto free_pt;
> 
> free_pt does mmput() which calls module_put(mm->binfmt->module). This
> is not right when try_module_get() fails. Hmm, the same if dup_mmap()
> fails, we are doing an extra module_put().

Good catch, should avoid over putting binfmt->module.

> 
> Perhaps we can use __module_get() again, current->mm->binfmt holds a
> reference. But this is a user-visible change, currently fork() can't
> succeed if ->module->state == MODULE_STATE_GOING.
> 
> So I think we need another small change:
> 
> 	 free_pt:
> 	+	mm->binfmt = NULL;
> 	 	mmput(mm);

looks good for me.
Will send an update.

thanks,
Hiroshi