Re: [PATCHv1 3/6] rdmacg: implements rdma cgroup
From: Parav Pandit
Date: Thu Jan 07 2016 - 15:39:30 EST
On Fri, Jan 8, 2016 at 1:58 AM, Tejun Heo <tj@xxxxxxxxxx> wrote:
> Hello, Parav.
>
> On Fri, Jan 08, 2016 at 01:55:09AM +0530, Parav Pandit wrote:
> ...
>> Above functionality is achieved, by maintaining the map this tgid and
>> associated original cgroup at try_charge(), uncharge() time.
>
> Hmmm, what happens after the following?
>
> 1. A process allocates some rdma resources and get registered on the
> hash table.
>
> 2. The process gets migrated to a different cgroup.
>
> 3. The process allocates more rdma resources.
>
> Which cgroup would the resources from #3 be attributed to?
Since the pid/tgid of the process doesn't change in step_3, it
allocates from original cgroup of step_1.
However in next patch V2, as described below, since IB resource will
store rdma_cg pointer,
in step_3, new resource will be allocated from new cgroup.
old resource will be freed from older cgroup.
This is what you are expecting, right?
>
>> In alternate method,
>> Its simple to store the pointer of rdma_cgroup structure in the IB
>> resource structure and hold on reference count to rdma_cgroup.
>> so that when its freed, uncharge_resource can accept rdma_cgroup
>> structure pointer.
>
> That'd be a lot more in line with how other controllers behave.
>
> Thanks.
>
> --
> tejun