Message-ID: <7781.1235374526@jrobl>
Date:	Mon, 23 Feb 2009 16:35:26 +0900
From:	hooanon05@...oo.co.jp
To:	linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: [RFC 3/8] Aufs2: lookup


> This is my second trial to ask incorporating aufs into mainline.


Lookup in a Branch
----------------------------------------------------------------------
Since aufs has a character of sub-VFS (see Introduction), it operates
lookup for branches as VFS does. It may be a heavy work. Generally
speaking struct nameidata is a bigger structure and includes many
information. But almost all lookup operation in aufs is the simplest
case, ie. lookup only an entry directly connected to its parent. Digging
down the directory hierarchy is unnecessary.

VFS has a function lookup_one_len() for that use, but it is not usable
for a branch filesystem which requires struct nameidata. So aufs
implements a simple lookup wrapper function. When a branch filesystem
allows NULL as nameidata, it calls lookup_one_len(). Otherwise it builds
a simplest nameidata and calls lookup_hash().
Here aufs applies "a principle in NFSD", ie. if the filesystem supports
NFS-export, then it has to support NULL as a nameidata parameter for
->create(), ->lookup() and ->d_revalidate(). So the lookup wrapper in
aufs tests if ->s_export_op in the branch is NULL or not.

When a branch is a remote filesystem, aufs trusts its ->d_revalidate().
For d_revalidate, aufs implements three levels of revalidate tests. See
"Revalidate Dentry and UDBA" in detail.


Loopback Mount
----------------------------------------------------------------------
Basically aufs supports any type of filesystem and block device for a
branch (actually there are some exceptions). But it is prohibited to add
a loopback mounted one whose backend file exists in a filesystem which is
already added to aufs. The reason is to protect aufs from a recursive
lookup. If it was allowed, the aufs lookup operation might re-enter a
lookup for the loopback mounted branch in the same context, and will
cause a deadlock.


Revalidate Dentry and UDBA (User's Direct Branch Access)
----------------------------------------------------------------------
Generally VFS helpers re-validate a dentry as a part of lookup.
0. digging down the directory hierarchy.
1. lock the parent dir by its i_mutex.
2. lookup the final (child) entry.
3. revalidate it.
4. call the actual operation (create, unlink, etc.)
5. unlock the parent dir

If the filesystem implements its ->d_revalidate() (step 3), then it is
called. Actually aufs implements it and checks the dentry on a branch is
still valid.
But it is not enough. Because aufs has to release the lock for the
parent dir on a branch at the end of ->lookup() (step 2) and
->d_revalidate() (step 3) while the i_mutex of the aufs dir is still
held by VFS.
If the file on a branch is changed directly, eg. bypassing aufs, after
aufs released the lock, then the subsequent operation may cause
something unpleasant result.

This situation is a result of VFS architecture, ->lookup() and
->d_revalidate() is separated. But I never say it is wrong. It is a good
design from VFS's point of view. It is just not suitable for sub-VFS
character in aufs.

Aufs supports such case by three level of revalidation which is
selectable by user.
1. Simple Revalidate
   Addition to the native flow in VFS's, confirm the child-parent
   relationship on the branch just after locking the parent dir on the
   branch in the "actual operation" (step 4). When this validation
   fails, aufs returns EBUSY. ->d_revalidate() (step 3) in aufs still
   checks the validation of the dentry on branches.
2. Monitor Changes Internally by Inotify
   Addition to above, in the "actual operation" (step 4) aufs re-lookup
   the dentry on the branch, and returns EBUSY if it finds different
   dentry.
   Additionally, aufs sets the inotify watch for every dir on branches
   during it is in cache. When the event is notified, aufs registers a
   function to kernel 'events' thread by schedule_work(). And the
   function sets some special status to the cached aufs dentry and inode
   private data. If they are not cached, then aufs has nothing to
   do. When the same file is accessed through aufs (step 0-3) later,
   aufs will detect the status and refresh all necessary data.
   In this mode, aufs has to ignore the event which is fired by aufs
   itself.
3. No Extra Validation
   This is the simplest test and doesn't add any additional revalidation
   test, and skip therevalidatin in step 4. It is useful and improves
   aufs performance when system surely hide the aufs branches from user,
   by over-mounting something (or another method).
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists