Re: State of dump utility

[Lukas Czerner <lczerner@redhat.com>]

On Tue, Sep 29, 2020 at 10:06:46PM -0400, Theodore Y. Ts'o wrote:
> On Tue, Sep 29, 2020 at 04:37:13PM +0200, Lukas Czerner wrote:
> > 
> > lately we've had couple of bugs against dump utility and a after a quick
> > look at the code I realized that it is very much outdated at least on
> > the extN side of things and would need some work and attention to make it
> > work reliably with modern ext4 features.
> > 
> > However the code has been neglected for a while and talking to the
> > maintainer he is pretty much done with it. At this point I am ready to
> > pull the plug on dump/restore in Fedora, but before I do I was wondering
> > whether there is any interest in moving dump/restore, or part of it, into
> > e2fsprogs ?
> > 
> > I have not looked at the code close enought to say whether it's worth it
> > or whether it would be better to write something from scratch. There is
> > also a question about what to do with the tape code - that's not
> > something I have any interest in digging into.
> > 
> > In my eyes dump had a good run and I would be happy just dumping it, but
> > it is worth asking here on the list. Is there anyone interested in
> > maintaining dump/restore, or is there interest in or objections agains
> > merging it into e2fsprogs ?
> 
> One of the interesting questions is how reliable the dump utility
> really is; that's because it works by reading the metadata directly
> --- while the file system is mounted.  So it's quite possible for the
> metadata to be changing out from under the dump/restore process.
> Especially with metadata checksums, I suspect dump/restore is going
> much more unreliable in terms of the libext2fs returning checksum
> failures.

Hi Ted,

this is a very good point. I have not even thought about checksums, but
that is just one example where it is likely to fail miserably. Granted
that's a relatively new feature, as well as inline data however it can't
even handle uninitialized extents correctly and hardly anyone is
noticing.

> 
> In the future, if we ever try to bypass the use of the buffer cache,
> and instead have jbd2 write out directly to the bio layer so we cant
> get better write error codes.  There was a discussion about this
> recently, and there are two problems.  First, we need to worry about
> programs like tune2fs and e2label that need to be able to read and
> modify the superblock while the file system is modified.  We'd want to
> add ioctl's to set and get the superblock, and update e2fsprogs to try
> to use those system calls first.  And then.... there is dump/restore.i
> 
> I could imagine adding ioctl's which allow safe read-only access to
> all metadata blocks, and not just for the superblock.  The question,
> though is... is it worth it, especially if it's only to make
> dump/restore work?

I have a feeling that the answer is no, it's not worth it for
dump/restore alone.

> 
> On the other hand, if we want to try to implement some kind of on-line
> fsck work, then -perhaps safe metadata reading would be part of that
> interface.  So I'd never say never, but I do wonder if it's time to
> pull the plug on dump/restore --- especially if we want to allow it to
> support not just inline files/directories, but also things like
> extended attributes and ACL's.
> 
> 						- Ted
> 

Thanks Ted, you made some good points and while there are some good
ideas for the future, there is no place for dump there. I think we're in
agreement to pull the plug on dump.

Thanks!
-Lukas