Re: [PATCH 08/10] s390/dasd: Display FC Endpoint Security information via sysfs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Am 08.10.20 um 09:03 schrieb Cornelia Huck:
> On Wed, 7 Oct 2020 22:10:11 +0200
> Jan Höppner <hoeppner@xxxxxxxxxxxxx> wrote:
>
>> On 10/7/20 6:40 PM, Cornelia Huck wrote:
>>> On Wed, 7 Oct 2020 16:33:37 +0200
>>> Jan Höppner <hoeppner@xxxxxxxxxxxxx> wrote:
>>>   
>>>>>>>> +static inline void dasd_path_release(struct kobject *kobj)
>>>>>>>> +{
>>>>>>>> +/* Memory for the dasd_path kobject is freed when dasd_free_device() is called */
>>>>>>>> +}
>>>>>>>> +      
>>>>>>> As already said, I don't think that's a correct way to implement this.
>>>>>>>       
>>>>>> As you correctly pointed out, our release function doesn't do anything.
>>>>>> This is because our path data is a (static) part of our device.
>>>>>> This data is critical to keep our devices operational.
>>>>>> We can't simply rely on allocated memory if systems are under stress.     
>>>>> Yes, avoiding freeing and reallocating memory certainly makes sense.
>>>>>     
>>>>>> Having this data dynamically allocated involves a lot of rework of our
>>>>>> path handling as well. There are a few things that are subject to improvement
>>>>>> and evaluating whether our dasd_path structures can be dynamic is one of
>>>>>> these things. However, even then, the above concern persists and I
>>>>>> highly doubt that dynamic dasd_paths objects are doable for us at this
>>>>>> moment.
>>>>>>
>>>>>> I do understand the concerns, however, we release the memory for dasd_path
>>>>>> structures eventually when dasd_free_device() is called. Until that point,
>>>>>> the data has to be kept alive. The rest is taking care of by the kobject
>>>>>> library.    
>>>>> Yes, there doesn't seem to be any memory leakage.
>>>>>     
>>>>>> In our path handling we also make sure that we can always verify/validate
>>>>>> paths information even if a system is under high memory pressure. Another
>>>>>> reason why it would contradictory for dasd_path objects to be dynamic.
>>>>>>
>>>>>> I hope this explains the reasoning behind the release function.    
>>>>> I understand where you're coming from.
>>>>>
>>>>> However, "static" kobjects (in the sense of "we may re-register the
>>>>> same kobject") are still problematic. Is there any way to simply
>>>>> "disappear" path objects that are not valid at the moment, or mark them
>>>>> as not valid?    
>>>> You could use kobject_del(), but it is rather intended to be used for
>>>> a two-stage removal of the kobject.
>>>>  
>>>>> Also, the simple act of registering/unregistering a kobject already
>>>>> creates stress from its sysfs interactions... it seems you should try
>>>>> to avoid that as well?
>>>>>     
>>>> We don't re-register kobjects over and over again. The kobjects are
>>>> infact initialized and created only _once_. This is done either during
>>>> device initialization (after dasd_eckd_read_conf() in
>>>> dasd_eckd_check_characteristics()) or when a path is newly added
>>>> (in the path event handler).
>>>> The kobject will stay until the memory for the whole device is being
>>>> freed. This is also the reason why the kobject can stay initialized and
>>>> we track ourselves whether we did the initialization/creation already
>>>> (which we check e.g. when a path is removed and added again).
>>>> So, instead of the release function freeing the kobject data,
>>>> it is done by our dasd_free_device() (same thing, different function IMHO).
>>>>
>>>> I think the concerns would be more worrisome if we'd remove/add
>>>> the kobjects every time. And then I agree, we'd run into trouble.
>>>>  
>>> The thing that tripped me is
>>>
>>> +void dasd_path_remove_kobj(struct dasd_device *device, int chp)
>>> +{
>>> +	if (device->path[chp].in_sysfs) {
>>> +		kobject_put(&device->path[chp].kobj);
>>> +		device->path[chp].in_sysfs = false;
>>> +	}
>>> +}
>>>
>>> As an exported function, it is not clear where this may be called from.
>>> Given your explanation above (and some more code reading on my side),
>>> the code looks ok in its current incarnation (but non-idiomatic).
>>>
>>> Is there a way to check that indeed nobody re-adds a previously removed
>>> path object due to a (future) programming error? And maybe add a
>>> comment that you must never re-register a path? "The path is gone,
>>> let's remove the object" looks quite tempting.
>>>   
>> A comment is the minimum I can think of at the moment and
>> I'll prepare a fixup patch or a new version of this patch that adds
>> a proper comment for this function.
>> Other ways to protect the usage must be investigated. 
>> I have to discuss with Stefan what the best approach would be as the patchset
>> is supposed to be ready for upstream integration.
>>
>> I'd prefer a fixup patch that we could send with at least one more fixup patch
>> that we have in the pipe already. Let's see. I hope that's fine with you
>> (and Jens obviously) so far.
> Fine with me. I don't really have a horse in that race; I just wanted
> to look at this from a vfio-ccw perspective and then stumbled over the
> kobject handling...
>

Thanks for this. I will send a v2 shortly.




[Index of Archives]     [Linux RAID]     [Linux SCSI]     [Linux ATA RAID]     [IDE]     [Linux Wireless]     [Linux Kernel]     [ATH6KL]     [Linux Bluetooth]     [Linux Netdev]     [Kernel Newbies]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Device Mapper]

  Powered by Linux