xfs
[Top] [All Lists]

Re: fsync, rename, O_ATOMIC/O_PONIES

To: Dave Chinner <david@xxxxxxxxxxxxx>
Subject: Re: fsync, rename, O_ATOMIC/O_PONIES
From: Olaf van der Spek <olaf@xxxxxxxx>
Date: Mon, 05 Mar 2012 16:44:26 +0100
Cc: Christoph Hellwig <hch@xxxxxxxxxxxxx>, xfs@xxxxxxxxxxx
In-reply-to: <20120305010203.GK5091@dastard>
References: <4F50BF89.7020909@xxxxxxxx> <20120302131240.GA14186@xxxxxxxxxxxxx> <4F53A2FF.3000305@xxxxxxxx> <20120305010203.GK5091@dastard>
User-agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:10.0.2) Gecko/20120216 Thunderbird/10.0.2
On 5-3-2012 2:02, Dave Chinner wrote:
Argh, come on.
That's not real and it's not complete. tmpfile is undefined, errors
aren't handled and you have lots of unlisted assumptions or
regressions.

The above is perfectly reasonable psuedo code for quickly describing
how to safely overwriting a file. If you want to know about error
handling and assumptions, read the man pages for operation.

But I don't have a psuedo code compiler. Using psuedo code hides complexity and bugs. Even the code from Jeff Moyer in the article you're refering too contained bugs.

Don't you think it's quite strange there's no real code available to handle this widespread problem?

To make it easy for you, I'll just point out this has already been
dealt with early in in the LWN thread for that article, here's the
link to the comment and the relevant LWN article providing you with
all the information you want:

http://lwn.net/Articles/476606/
https://lwn.net/Articles/457667/

It also doesn't handle the assumptions / regressions: http://lwn.net/Articles/485182/
Assuming the target is not a symlink to a different volume
Assuming you are allowed to create the tmp file
Assuming you are allowed to overwrite an existing file having the same name as your tmp file Assuming it's ok to reset meta-data, like file owner, permissions, acls, creation timestamp, etc. Assuming the performance regression due to fsync is ok (request was for atomic, not durable)

And that my stance on this atomic rename subject is simply this: If
we want to change reanme behaviour, then the filesystem is not the
right place to do it - atomic rename semantics need to be defined
and enforced at the VFS.  See here:

http://lwn.net/Articles/479152/

I don't think rename is part of the solution.

Olaf

<Prev in Thread] Current Thread [Next in Thread>