Re: [PATCH v8 7/8] livepatch: Correctly handle atomic replace for not yet loaded modules

[Joe Lawrence]

On 03/01/2018 05:28 AM, Petr Mladek wrote:
> On Thu 2018-02-22 22:00:28, Miroslav Benes wrote:
>> On Wed, 21 Feb 2018, Petr Mladek wrote:
>>> This patch allows the late initialization.
>>>
>>> diff --git a/kernel/livepatch/core.c b/kernel/livepatch/core.c
>>> index ad508a86b2f9..da1438d47d83 100644
>>> --- a/kernel/livepatch/core.c
>>> +++ b/kernel/livepatch/core.c
>>> @@ -984,7 +988,12 @@ static void klp_free_patch(struct klp_patch *patch)
>>>  
>>>  static int klp_init_func(struct klp_object *obj, struct klp_func *func)
>>>  {
>>> -	if (!func->old_name || !func->new_func)
>>> +	if (!func->old_name)
>>> +		return -EINVAL;
>>> +
>>> +	/* NOPs do not know the address until the patched module is loaded. */
>>> +	if (!func->new_func &&
>>> +	    (!klp_is_func_type(func, KLP_FUNC_NOP) || klp_is_object_loaded(obj)))
>>>  		return -EINVAL;
>>
>> If we changed the order of klp_init_func() and klp_init_object_loaded() 
>> calls in klp_init_object(), the hunk would not be needed. Is that correct? 
> 
> Not really. klp_init_object_loaded() would set func->new_func only
> when the object was loaded. But we want to proceed here and create
> the kobject for NOPs even when it was not loaded.
> 
> 
>>>  	INIT_LIST_HEAD(&func->stack_node);
>>> @@ -1039,6 +1048,9 @@ static int klp_init_object_loaded(struct klp_patch *patch,
>>>  			return -ENOENT;
>>>  		}
>>>  
>>> +		if (klp_is_func_type(func, KLP_FUNC_NOP))
>>> +			func->new_func = (void *)func->old_addr;
>>
>> Is there a reason why you left the same assignment in 
>> klp_alloc_func_nop()? This one is enough, no?
> 
> Good point! I am going to replace the obsolete assignment
> with a comment in v8.

Hi Petr, Miroslav,

I don't think the assignment in klp_alloc_func_nop() was necessarily
redundant.  It was removed in v9 and that breaks my atomic replace
sample module when I try to load it.  (Perhaps the sample patch has
issues, but here are my debug notes):

To recap:

  patch 1 - modifies cmdline_proc_show()
  patch 2 - modifies only meminfo_proc_show()

when I load patch 2 with .replace=1, klp_add_nops() is called and this
adds a nop function to patch 2 so it reverts cmdline_proc_show():

klp_init_patch()
  if (patch->replace)
    klp_add_nops()
      list_for_each_entry(old_patch, &klp_patches, list) {
        klp_for_each_object(old_patch, old_obj) {
          klp_add_object_nops()
            klp_add_func_nop()
              klp_alloc_func_nop()

the patch continues initialization and I hit the second -EINVAL
condition on that nop function in klp_init_func():

  klp_init_object
    klp_init_func

(from added debug):
[   48.456980] livepatch: func->old_name=cmdline_proc_show
[   48.457620] livepatch: func->new_func=          (null)
[   48.458042] livepatch: klp_is_func_type(func, KLP_FUNC_NOP)=1
[   48.458573] livepatch: klp_is_object_loaded(obj)=1

If I restore the assignment of func->new_func to klp_alloc_func_nop()
then the replacement patch properly loads.  (Reordering the code may
have similar effect?)

I think this problem is contained to only replacement patches that need
the nop-revert feature... if the replacement patch provides a new
function definition, then it shouldn't be affected.

Man, we need a regression test suite for all these cases :)

-- Joe