[PATCH v3 36/39] ubifs: implement ubifs_get_qsize to get quota size in ubifs

[Jan Kara]

On Mon 21-09-15 17:16:49, Dongsheng Yang wrote:
> On 09/21/2015 05:13 PM, Jan Kara wrote:
> >On Mon 21-09-15 12:35:01, Dongsheng Yang wrote:
> >>On 09/18/2015 07:20 PM, Jan Kara wrote:
> >>>>.... TBH, there is a little different with other filesystems. I did not
> >>>>use the "disk" space, but the file space in ubifs quota, although dquot
> >>>>means disk quota. Same with btrfs quota. If we use disk space for quota,
> >>>>the most common problem from user is that: why I did not reach the limit
> >>>>but I can not write any byte. COW in btrfs or out-place-updating in
> >>>>ubifs makes this problem much worse.
> >>>>
> >>>>So I choose file space here for ubifs.
> >>>
> >>>OK, so these are really two separate questions. I understand your choice of
> >>>using file space as the amount of space to account for quota purposes and
> >>>I'm fine with that choice. Another thing is that regardless of how you
> >>>decide to do quota accounting, you must maintain i_blocks / i_bytes to
> >>>contain proper value because dquot_transfer() uses that information to update
> >>>quota usage when inode owner is changed.
> >>
> >>But if we don't use i_blocks to get qsize, what we care only in
> >>dquot_transter() is dquot->dq_dqb. That means, even if the i_blocks
> >>is not correct in dquot_transfer() in ubifs, that's okey, because we
> >>will never use this value, right?
> >
> >dquot_transfer() will use the value - when file F changes owner from user A
> >to user B, then you need to decrement amount of space used by F from A's
> >quota usage and add that amount to B's quota usage. And the amount of space
> >is obtained via inode_get_bytes() which uses i_blocks and i_bytes. See
> >__dquot_transfer() in fs/quota/dquot.c for details.
> 
> Yes, I see it. But if ubifs doesn't use i_blocks and i_bytes to stand
> for quota size, as I mentioned I want to use i_size. Then we will never
> use i_blocks. So we only need a to transfer dquot->dq_dqb values. That
> means, even the i_blocks in dquot_transfer() is not correct, we don't
> care it. we only need to make sure values in dquot_dq_dqb are
> transfered, such as dquot->dq_dqb->curspace for user B is equal with
> i_size.

I think you are missing one thing:

Assume user A has files F1 (1MB) F2 (2MB) F3 (3MB).
User B has one file G1 (4MB).

So user A uses 6 MB in total (stored in dquot for user A in dqb_curspace
field). User B uses 4MB.

Now you do "chown B F1"

You need to change dq_dqb of F1 to point to user B quota structure as you
explain, that is correct. However you also need to subtract 1MB from user
A's total usage (which is stored in struct dquot for user A) since file F1
no longer belongs to him. And you need to add 1 MB to user B's total usage.

After this user A will have total usage of 5 MB and user B has usage of 5
MB as well. You need to properly update the total usage of each user so
that you can check the usage against quota limits.

								Honza
-- 
Jan Kara <jack at suse.com>
SUSE Labs, CR