On 7/16/19 5:07 PM, Xiaoming Ni wrote:
> On 2019/7/16 18:20, Vasily Averin wrote:
>> On 7/16/19 5:00 AM, Xiaoming Ni wrote:
>>> On 2019/7/15 13:38, Vasily Averin wrote:
>>>> On 7/14/19 5:45 AM, Xiaoming Ni wrote:
>>>>> On 2019/7/12 22:07, gregkh@xxxxxxxxxxxxxxxxxxx wrote:
>>>>>> On Fri, Jul 12, 2019 at 09:11:57PM +0800, Xiaoming Ni wrote:
>>>>>>> On 2019/7/11 21:57, Vasily Averin wrote:
>>>>>>>> On 7/11/19 4:55 AM, Nixiaoming wrote:
>>>>>>>>> On Wed, July 10, 2019 1:49 PM Vasily Averin wrote:
>>>>>>>>>> On 7/10/19 6:09 AM, Xiaoming Ni wrote:
>>>>>>>>>>> Registering the same notifier to a hook repeatedly can cause the hook
>>>>>>>>>>> list to form a ring or lose other members of the list.
>>>>>>>>>>
>>>>>>>>>> I think is not enough to _prevent_ 2nd register attempt,
>>>>>>>>>> it's enough to detect just attempt and generate warning to mark host in bad state.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Duplicate registration is prevented in my patch, not just "mark host in bad state"
>>>>>>>>>
>>>>>>>>> Duplicate registration is checked and exited in notifier_chain_cond_register()
>>>>>>>>>
>>>>>>>>> Duplicate registration was checked in notifier_chain_register() but only
>>>>>>>>> the alarm was triggered without exiting. added by commit 831246570d34692e
>>>>>>>>> ("kernel/notifier.c: double register detection")
>>>>>>>>>
>>>>>>>>> My patch is like a combination of 831246570d34692e and notifier_chain_cond_register(),
>>>>>>>>> which triggers an alarm and exits when a duplicate registration is detected.
>>>>>>>>>
>>>>>>>>>> Unexpected 2nd register of the same hook most likely will lead to 2nd unregister,
>>>>>>>>>> and it can lead to host crash in any time:
>>>>>>>>>> you can unregister notifier on first attempt it can be too early, it can be still in use.
>>>>>>>>>> on the other hand you can never call 2nd unregister at all.
>>>>>>>>>
>>>>>>>>> Since the member was not added to the linked list at the time of the second registration,
>>>>>>>>> no linked list ring was formed.
>>>>>>>>> The member is released on the first unregistration and -ENOENT on the second unregistration.
>>>>>>>>> After patching, the fault has been alleviated
>>>>>>>>
>>>>>>>> You are wrong here.
>>>>>>>> 2nd notifier's registration is a pure bug, this should never happen.
>>>>>>>> If you know the way to reproduce this situation -- you need to fix it.
>>>>>>>>
>>>>>>>> 2nd registration can happen in 2 cases:
>>>>>>>> 1) missed rollback, when someone forget to call unregister after successfull registration,
>>>>>>>> and then tried to call register again. It can lead to crash for example when according module will be unloaded.
>>>>>>>> 2) some subsystem is registered twice, for example from different namespaces.
>>>>>>>> in this case unregister called during sybsystem cleanup in first namespace will incorrectly remove notifier used
>>>>>>>> in second namespace, it also can lead to unexpacted behaviour.
>>>>>>>>
>>>>>>> So in these two cases, is it more reasonable to trigger BUG() directly when checking for duplicate registration ?
>>>>>>> But why does current notifier_chain_register() just trigger WARN() without exiting ?
>>>>>>> notifier_chain_cond_register() direct exit without triggering WARN() ?
>>>>>>
>>>>>> It should recover from this, if it can be detected. The main point is
>>>>>> that not all apis have to be this "robust" when used within the kernel
>>>>>> as we do allow for the callers to know what they are doing :)
>>>>>>
>>>>> In the notifier_chain_register(), the condition ( (*nl) == n) is the same registration of the same hook.
>>>>> We can intercept this situation and avoid forming a linked list ring to make the API more rob
>>>>
>>>> Once again -- yes, you CAN prevent list corruption, but you CANNOT recover the host and return it back to safe state.
>>>> If double register event was detected -- it means you have bug in kernel.
>>>>
>>>> Yes, you can add BUG here and crash the host immediately, but I prefer to use warning in such situations.
>>>>
>>>>>> If this does not cause any additional problems or slow downs, it's
>>>>>> probably fine to add.
>>>>>>
>>>>> Notifier_chain_register() is not a system hotspot function.
>>>>> At the same time, there is already a WARN_ONCE judgment. There is no new judgment in the new patch.
>>>>> It only changes the processing under the condition of (*nl) == n, which will not cause performance problems.
>>>>> At the same time, avoiding the formation of a link ring can make the system more robust.
>>>>
>>>> I disagree,
>>>> yes, node will have correct list, but anyway node will work wrong and can crash the host in any time.
>>>
>>> Sorry, my description is not accurate.
>>>
>>> My patch feature does not prevent users from repeatedly registering hooks.
>>> But avoiding the chain ring caused by the user repeatedly registering the hook
>>>
>>> There are no modules for duplicate registration hooks in the current system.
>>> But considering that not all modules are in the kernel source tree,
>>> In order to improve the robustness of the kernel API, we should avoid the linked list ring caused by repeated registration.
>>> Or in order to improve the efficiency of problem location, when the duplicate registration is checked, the system crashes directly.
>>
>> Detect of duplicate registration means an unrecoverable error,
>> from this point of view it makes sense to replace WARN_ONCE by BUG_ON.
>>
>>> On the other hand, the difference between notifier_chain_register() and notifier_chain_cond_register() for duplicate registrations is confusing:
>>> Blocking the formation of the linked list ring in notifier_chain_cond_register()
>>> There is no interception of the linked list ring in notifier_chain_register(), just an alarm.
>>> Give me the illusion: Isn't notifier_chain_register() allowed to create a linked list ring?
>>
>> I'm not sure that I understood your question correctly but will try to answer.
>> As far as I see all callers of notifier_chain_cond_register checks return value, expect possible failure and handle it somehow.
>> On the other hand callers of notifier_chain_register() in many cases do not check return value and always expect success.
>> The goal of original WARN_ONCE -- to detect possible misuse of notifiers and it seems for me it correctly handles this task.
>>
> Notifier_chain_cond_register() has only one return value: 0
It looks wrong for me.
> At the same time, it is only called by blocking_notifier_chain_cond_register().
> In the function comment of blocking_notifier_chain_cond_register there is " Currently always returns zero."
> Therefore, the user cannot check whether the hook has duplicate registration or other errors by checking the return value.
I think notifier_chain_cond_register can be changed to return error.
It is safe now, all its in-tree callers checks return value and can properly react on such error.
On the other hand, in all cases notifier_chain_cond_register are __init functions,
they are called once only and double registration seems is impossible here:
even if some old notifier was lost and was not properly unregistered,
new one will have another address.
And even if these addresses was equal -- it is critical error
and I prefer to generate warning instead of silent failure of module load.
> If the interceptor list ring is added to notifier_chain_register(), notifier_chain_register()
> And notifier_chain_cond_register() becomes redundant code, we can delete one of them
Yes, I'm agree, at present there are no difference between
notifier_chain_cond_register() and notifier_chain_register()
Question is -- how to improve it.
You propose to remove notifier_chain_cond_register() by some way.
Another option is return an error, for some abstract callers who expect possible double registration.
Frankly speaking I prefer second one,
however because of kernel do not have any such callers right now seems you are right,
and we can delete notifier_chain_cond_register().
So let me finally accept your patch-set.
Thank you,
Vasily Averin
