[Gluster-devel] Quota-2

Venky Shankar vshankar at redhat.com
Mon Dec 21 15:53:15 UTC 2015

Venky Shankar wrote:
> Shyam wrote:
>> On 12/19/2015 01:13 AM, Venky Shankar wrote:
>>> Shyam wrote:
>>>> <first off the cuff response>
>>>> On 12/18/2015 04:00 AM, Vijaikumar Mallikarjuna wrote:
>>>>> Hi All,
>>>>> Here is the summary of discussion we had today on Quota-v2
>>>>> Project, user and group quotas will use the same logic of accounting
>>>>> the
>>>>> usage
>>>>> Quota-v2 should be compatible with both DHT-v1 and DHT-v2
>>>>> Project quotas will use gfid of the given path as the project-id.
>>>>> each inode will have a associated project-id which needs to be
>>>>> embedded within inode.
>>>>> when creating new file, it will inherit project-id from its parent
>>>>> inode.
>>>>> In case if parent inode doesn't contain project-id, then there
>>>>> should be a mechanism to
>>>>> find the project-id given a gfid, so back-pointers can solve this
>>>>> problem
>>>> Why would the parent not contain a project-id? and if it does not
>>>> contain a project-id, why would you need to find a project-id?
>>> Updating project-id's for all children for a given directory on which a
>>> project-id is set would required scanning. If we could figure
>>> out the project-id for a given inode using back-pointers (closest
>>> ancestor for which a project-id is set) then we'd not need the crawl and
>>> update.
>> For existing directories when a quota is set, we need to do the scan and
>> set/update the quota information, right? This is unavoidable as far as I
>> can see.
> I was more into trying to save the sub-tree update when a project-id is
> set on a directory. If (with current dht) we could figure out the
> project-id for the closest project-id set ancestor using back-pointers
> (or whatever..) then we could possibly avoid the scan + update.

Vijaikumar mentioned that XFS does the scan + update during project-id 
set, however it's pretty fast - however, that's not the case with Gluster.

Correct me if I'm wrong please.

>> For a new entry, as responded in the other mail on this thread, if the
>> parent is not yet marked with a project-id then the child will inherit
>> the same when the scanner updates the project-id for the parent, till
>> then there is no need to account.
>>> With DHT2, such a mechanism would query each MDS. For current DHT, it
>>> might still be good enough.
>> With DHT2 the scanner (for updating quota information for pre-existing
>> directories) may need to split the work up and perform the operation. A
>> crude thought at the moment would be,
>> - Set the project-id for the grandparent where the quota enforcement
>> starts
>> - Set the pj-id for the leaf objects for this directory, and set the
>> pj-id for the directories under this grandparent, but take up marking
>> leaves of these sub-directories as tasks in the other MDS stack
>> (wherever they belong, i.e same MDS or a different one).
>>>> What I can visualize is that, the parent is not yet part of a quota
>>>> project, hence it does not contain a project-id. When this becomes a
>>>> part of some quota restriction it would start getting a quota. This
>>>> seems to be one case where the project-id would be missing, but by
>>>> choice.
>>>> What am I missing here?
>>>>> which xlator DHT/marker will set the project-id in the inode? this
>>>>> needs to be decided
>>>>> User and group quotas will use uid and gid to account the usage
>>>>> Quota has below set of operation which should be performed as a single
>>>>> transaction
>>>>> read current size
>>>>> read current contribution
>>>>> update delta to the contribution
>>>>> update size to user/project inode
>>>>> In the current directory quota, to make this transaction crash
>>>>> consistent we use a mechanism of setting dirty flag in a parent inode.
>>>>> This mechanism will not work efficiently with project quotas.
>>>>> Journaling
>>>>> will be the good solution for crash-consistency.
>>>>> Quota 2 depends on:
>>>>> back-pointers: to find project-id from gfid
>>>>> journaling: for crash consistency of accounting
>>>>> Thanks,
>>>>> Vijay

More information about the Gluster-devel mailing list