Re: [PATCH net-next] xfrm: policy: Remove obsolete WARN while xfrm policy inserting

[Yuehaibing]

On 2020/4/15 15:14, Steffen Klassert wrote:
> On Thu, Apr 09, 2020 at 04:19:37PM +0800, Yuehaibing wrote:
>>
>>
>> On 2020/4/6 17:03, Steffen Klassert wrote:
>>> On Mon, Mar 30, 2020 at 10:05:32PM +0800, Yuehaibing wrote:
>>>> On 2020/3/28 19:23, Steffen Klassert wrote:
>>>>> On Fri, Mar 27, 2020 at 08:34:43PM +0800, YueHaibing wrote:
>>>>>> Since commit 7cb8a93968e3 ("xfrm: Allow inserting policies with matching
>>>>>> mark and different priorities"), we allow duplicate policies with
>>>>>> different priority, this WARN is not needed any more.
>>>>>
>>>>> Can you please describe a bit more detailed why this warning
>>>>> can't trigger anymore?
>>>>
>>>> No, this warning is triggered while detect a duplicate entry in the policy list
>>>>
>>>> regardless of the priority. If we insert policy like this:
>>>>
>>>> policy A (mark.v = 3475289, mark.m = 0, priority = 1)	//A is inserted
>>>> policy B (mark.v = 0, mark.m = 0, priority = 0) 	//B is inserted
>>>> policy C (mark.v = 3475289, mark.m = 0, priority = 0)	//C is inserted and B is deleted
>>>
>>> The codepath that replaces a policy by another should just trigger
>>> on policy updates (XFRM_MSG_UPDPOLICY). Is that the case in your
>>> test?
>>
>> Yes, this is triggered by XFRM_MSG_UPDPOLICY
>>
>>>
>>> It should not be possible to add policy C with XFRM_MSG_NEWPOLICY
>>> as long as you have policy B inserted.
>>>
>>> The update replaces an old policy by a new one, the lookup keys of
>>> the old policy must match the lookup keys of the new one. But policy
>>> B has not the same lookup keys as C, the mark is different. So B should
>>> not be replaced with C.
>>
>> 1436 static bool xfrm_policy_mark_match(struct xfrm_policy *policy,
>> 1437                                    struct xfrm_policy *pol)
>> 1438 {
>> 1439         u32 mark = policy->mark.v & policy->mark.m;
>> 1440
>> 1441         if (policy->mark.v == pol->mark.v && policy->mark.m == pol->mark.m)
>> 1442                 return true;
>> 1443
>> 1444         if ((mark & pol->mark.m) == pol->mark.v &&    //policy is C, pol is B, so mark is 0, pol->mark.m is 0, pol->mark.v is 0
>> 1445             policy->priority == pol->priority)	   //priority is same zero, so return true, B is replaced with C
>> 1446                 return true;
>> 1447
>> 1448         return false;
>> 1449 }
>>
>> Should xfrm_policy_mark_match be fixedï
> 
> Yes, xfrm_policy_mark_match should only replace if the found
> policy has the same lookup keys.

I'm wonder that lookup keys means association of mark.v and mark.m, or the mark (mark.v & mark.m).

In above my case, policy B and C has the same mark (that is 0), if the lookup keys is mark, replacement is permitted.

If lookup keys is association of mark.v and mark.m, then:

policy E (mark.v = 0x1, mark.m = 0x3, priority = 1)
policy F (mark.v = 0x1, mark.m = 0x5, priority = 1)

E should not be replaced by F, but this is permitted now.

> 
> .
>