Re: [RFC PATCH 2/2] ceph: quota: fix quota subdir mounts

[Luis Henriques]

"Yan, Zheng" <ukernel@xxxxxxxxx> writes:

> On Sat, Mar 2, 2019 at 3:13 AM Luis Henriques <lhenriques@xxxxxxxx> wrote:
>>
>> The CephFS kernel client doesn't enforce quotas that are set in a
>> directory that isn't visible in the mount point.  For example, given the
>> path '/dir1/dir2', if quotas are set in 'dir1' and the mount is done in with
>>
>>   mount -t ceph <server>:<port>:/dir1/ /mnt
>>
>> then the client can't access the 'dir1' inode from the quota realm dir2
>> belongs to.
>>
>> This patch fixes this by simply doing an MDS LOOKUPINO Op and grabbing a
>> reference to it (so that it doesn't disappear again).  This also requires an
>> extra field in ceph_snap_realm so that we know we have to release that
>> reference when destroying the realm.
>>
>
> This may cause circle reference if somehow an inode owned by snaprealm
> get moved into mount subdir (other clients do rename).  how about
> holding these inodes in mds_client?

Ok, before proceeded any further I wanted to make sure that what you
were suggesting was something like the patch below.  It simply keeps a
list of inodes in ceph_mds_client until the filesystem is umounted,
iput()ing them at that point.

I'm sure I'm missing another place where the reference should be
dropped, but I couldn't figure it out yet.  It can't be
ceph_destroy_inode; drop_inode_snap_realm is a possibility, but what if
the inode becomes visible in the meantime?  Well, I'll continue thinking
about it.

Function get_quota_realm() should have a similar construct to lookup
inodes.  But it's a bit more tricky, because ceph_quota_is_same_realm()
requires snap_rwsem to be held for the 2 invocations of
get_quota_realm.

Cheers,
-- 
Luis