[lxc-devel] Memory Resources

[Daniel Lezcano]

Serge E. Hallyn wrote:
> Quoting Daniel Lezcano (daniel.lezcano at free.fr):
>   
>> Serge E. Hallyn wrote:
>>     
>>> Quoting Daniel Lezcano (daniel.lezcano at free.fr):
>>>   
>>>       
>>>> Krzysztof Taraszka wrote:
>>>>     
>>>>         
>>>>> Okey.
>>>>> I made few tests and this two ways work:
>>>>>
>>>>> First way:
>>>>> =======
>>>>> lxc. smack enabled, policy loaded. cgroup not labeled.
>>>>>
>>>>> a) start container
>>>>> b) mount cgroup inside container
>>>>> c) mount --bind /cgroup/foo/memory.meminfo /proc/meminfo
>>>>> d) secure the /cgroup on the host (ie: attr -S -s SMACK64 -V host /cgroup).
>>>>>
>>>>> this step can be done inside lxc tools ;)
>>>>>
>>>>> Second way:
>>>>> ==========
>>>>> lxc. smack enabled, policy loaded. cgroup not labeled.
>>>>>
>>>>> a) do not label whole /cgrop directory (DO NOT DO: attr -S -s SMACK64 -V
>>>>> host /cgroup). Label dedicate files only (for example: /cgroup/cpuset.cpus,
>>>>> /cgroup/vs1/cpuset.cpus, etc). Do not label the /cgrop/vs1 directory. Label
>>>>> with vs1 label only /cgroup/vs1/memory.meminfo. All other files label with
>>>>> host label to do not allow read them.
>>>>> b) start container
>>>>> c) mount cgroup inside container
>>>>> d) mount --bind /cgroup/foo/memory.meminfo /proc/meminfo
>>>>>
>>>>> steps: b, c, d can be done inside lxc tools. step a can't and it is base on
>>>>> the admin policy.
>>>>>
>>>>> I think that the first solution is more automatic and can be done by lxc
>>>>> tools (maybe command line switch? I can prepare a patch for that.
>>>>>         
>>>>>           
>>>> I do not know smack, what does smack here ? Will this solution avoid 
>>>> the container to overwrite /proc/meminfo by remounting /proc ?
>>>>     
>>>>         
>>> Right, in the first way he is labeling the whole cgroupfs with a label
>>> which prevents the container from mounting it.  In the second way,
>>> the specific files are labeled.
>>>   
>>>       
>> Ah, got it ! :)
>>
>> The idea of Kamezawa-san to use a fuse proc is maybe a good idea in this  
>> case. So we can address the entire /proc specific informations. For  
>>     
>
> I agree, nice idea.  And hopefully pretty simple to whip up for the
> meminfo and cpuinfo files as an example.
>
> Are you thinking a fuse fs which takes a config file, holds an open
> ref to its ancestor /proc, and for each file looks in a config file to
> decide whether to show userspace:
> 	1. nothing
> 	2. the underlying file, unprocessed
> 	3. a simple ascii file instead
> 	4. the underlying file, processed?
>   

Yes, exactly :)
But, I am not sure how to retrieve the container context, I mean how to 
pick and return the right information.
eg: in the container foo, when looking at /proc/meminfo, fuse-lxcfs 
should process /cgroup/foo/(somefiles), how to know the request is 
coming from 'foo' without doing multiple mount, one in each container ?

>> example, like the /proc/meminfo, there is the /proc/cpuinfo. If you  
>> restrict the usage to a subset of your cpus with cpuset and you look at  
>> /proc/cpuinfo, you see all the cpus; it is not a big problem until a  
>> computation application looks at this file and choose to fork(n cpus)  
>> and set the affinity of each process to each cpu ... AFAIR, this is the  
>> case for HPC applications.
>>