Date: Tue, 22 Nov 2005 12:37:23 -0500
To: "Theodore Ts'o" <tytso@mit.edu>, Christoph Hellwig <hch@infradead.org>,
       J?rn Engel <joern@wohnheim.fh-wedel.de>,
       Alfred Brons <alfredbrons@yahoo.com>, pocm@sat.inesc-id.pt,
       linux-kernel@vger.kernel.org
Subject: Re: what is our answer to ZFS?
Message-ID: <20051122173723.GY12760@delft.aura.cs.cmu.edu>
Mail-Followup-To: Theodore Ts'o <tytso@mit.edu>,
	Christoph Hellwig <hch@infradead.org>,
	J?rn Engel <joern@wohnheim.fh-wedel.de>,
	Alfred Brons <alfredbrons@yahoo.com>, pocm@sat.inesc-id.pt,
	linux-kernel@vger.kernel.org
References: <11b141710511210144h666d2edfi@mail.gmail.com> <20051121095915.83230.qmail@web36406.mail.mud.yahoo.com> <20051121101959.GB13927@wohnheim.fh-wedel.de> <20051122075148.GB20476@infradead.org> <20051122145047.GB29179@thunk.org> <20051122152531.GU12760@delft.aura.cs.cmu.edu> <20051122162836.GA31444@thunk.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20051122162836.GA31444@thunk.org>
User-Agent: Mutt/1.5.9i
From: Jan Harkes <jaharkes@cs.cmu.edu>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 4425
Lines: 86

On Tue, Nov 22, 2005 at 11:28:36AM -0500, Theodore Ts'o wrote:
> On Tue, Nov 22, 2005 at 10:25:31AM -0500, Jan Harkes wrote:
> > On Tue, Nov 22, 2005 at 09:50:47AM -0500, Theodore Ts'o wrote:
> > > I will note though that there are people who are asking for 64-bit
> > > inode numbers on 32-bit platforms, since 2**32 inodes are not enough
> > > for certain distributed/clustered filesystems.  And this is something
> > > we don't yet support today, and probably will need to think about much
> > > sooner than 128-bit filesystems....
> > 
> > As far as the kernel is concerned this hasn't been a problem in a while
> > (2.4.early). The iget4 operation that was introduced by reiserfs (now
> > iget5) pretty much makes it possible for a filesystem to use anything to
> > identify it's inodes. The 32-bit inode numbers are simply used as a hash
> > index.
> 
> iget4 wasn't even strictly necessary, unless you want to use the inode
> cache (which has always been strictly optional for filesystems, even
> inode-based ones) --- Linux's VFS is dentry-based, not inode-based, so
> we don't use inode numbers to index much of anything inside the
> kernel, other than the aforementioned optional inode cache.

Ah yes, you're right.

> The main issue is the lack of a 64-bit interface to extract inode
> numbers, which is needed as you point out for userspace archiving
> tools like tar. There are also other programs or protocols that in the
> past have broken as a result of inode number collisions.

64-bit? Coda has been using 128-bit file identifiers for a while now.
And I can imagine someone trying to plug something like git into the VFS
might want to use 168-bits. Or even more for a CAS-based storage that
identifies objects by their SHA256 or SHA512 checksum.

On the other hand, any large scale distributed/cluster based file system
probably will have some sort of snapshot based backup strategy as part
of the file system design. Using tar to back up a couple of tera/peta
bytes just seems like asking for trouble, even keeping track of the
possible hardlinks by remembering previously seen inode numbers over
vast amounts of files will become difficult at some point.

> As another example, a quick google search indicates that the some mail
> programs can use inode numbers as a part of a technique to create
> unique filenames in maildir directories.  One could easily also

Hopefully it is only part of the technique. Like combining it with
grabbing a timestamp, the hostname/MAC address where the operation
occurred, etc.

> imagine using inode numbers as part of creating unique ids returned by
> an IMAP server --- not something I would recommend, but it's an
> example of what some people might have done, since everybody _knows_
> they can count on inode numbers on Unix systems, right?  POSIX
> promises that they won't break!

Under limited conditions. Not sure how stable/unique 32-bit inode
numbers are on NFS clients, taking into account client-reboots, failing
disks that are restored from tape, or when the file system reuses inode
numbers of recently deleted files, etc. It doesn't matter how much
stability and uniqueness POSIX demands, I simply can't see how it can be
guaranteed in all cases.

> > The only thing that tends to break are userspace archiving tools like
> > tar, which assume that 2 objects with the same 32-bit st_ino value are
> > identical. I think that by now several actually double check that the
> > inode linkcount is larger than 1.
> 
> Um, that's not good enough to avoid failure modes; consider what might
> happen if you have two inodes that have hardlinks, so that st_nlink >
> 1, but whose inode numbers are the same if you only look at the low 32
> bits?  Oops.
>
> It's not a bad hueristic, if you don't have that many hard-linked
> files on your system, but if you have a huge number of hard-linked
> trees (such as you might find on a kernel developer with tons of
> hard-linked trees), I wouldn't want to count on this always working.

Yeah, bad example for the typical case. But there must be some check to
at least avoid problems when files are removed/created and the inode
numbers are reused during a backup run.

Jan
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/