Re: Problem with aufs on top of hfsplus(ro) and ext3(rw): any client process hangs when performing a file rename

[Ananda T. <an...@ma...>]

Hello J. R.,

Thanks a lot for your work to improve hfsplus support in aufs.
Concerning not supporting r/w hfsplus branches, I don't think it's a
problem at all, and even, it's better not to support it.
Indeed mounting hsplus partition in r/w mode on Linux is not advised
because the partition journal is not supported yet, which is not safe.

I will definitely upgrade to you last version on monday.
Also, I hope your latest version will be integrated into next Ubuntu
10.04, that would be great.

Also I have another process freeze problem to report.
Everything was working properly since applying your patch of March the
3rd, over the old version 20090126. But I have just encountered another
process freeze again just now, which is different.
This time, it follows a 'sys_write' system call.
The hanged process is the 'bzr' version control utility I use here.
The uninterruptible call still occurs in hfsplus_file_release function.

FYI, please find the call trace attached.
May be this problem does not exist anymore in your last version, so I
will upgrade on monday, and inform whenever I come accross any other
problem (I currently use aufs everyday).

Kind regards

Anand

Le vendredi 05 mars 2010 à 07:05 +0900, sf...@us... a
écrit : 
> Ananda Tallur:
> > I had to backport it to old version of aufs sources provided in Ubuntu
> > karmic: 20090126.
> 
> Hmm... That is about one year old version. Roughly speaking, I had
> released the newer version 40 or 50 times since then.
> 
> 
> > I would like to use the latest aufs source in the git standalone
> > repository and I will try to do it soon. For this I will have to
> 	:::
> 
> That is the most likely way, I believe.
> 
> 
> > Please find attached the backported patch I have applied to aufs source
> > version 20090126. There were also two small typo fixes:
> 
> Ah, sorry.
> Now I refined the patch and applied the similar approach to the internal
> copy-up (several places), handling chmod/chown or any other attribute
> changes, link(2) and rename(2).
> The patch grew up and it will be included in next Monday release. While
> I don't attach the latest patch to this mail, I'd suggest you to upgrade
> aufs on next Monday.
> 
> Here is a list which I've found about hfsplus.
> These things never mean that hfsplus is bad. It just differs from other
> popular linux filesystems. Additionally you are using it as RO. They
> won't be problem on your system generally.
> - hfsplus does not set limit to the link count in link(2). Potentially
>   or theoritically it can be a problem when the link count overflows (I
>   don't think it can happen easily though).
> - it doesn't seem to be available to be remounted RO --> RW.
> - it doesn't seem to handle the block count of a file which has a hole
>   in it (sparse file) and a symlink. Or its blockc count looks uncommon.
> - it doesn't support splice(2) for write, and doesn't update atime for
>   splice read (or stat/fstat cannot get the updated atime).
> - it may return a positive number as the link count of an unlinked file,
>   and its size became 0 (incorrect).
> - sgid bit of the parent dir of a newly created file doesn't seem to be
>   inherited (sysv:bsd behaviour).
> - finally, I gave up supporting writable hfsplus branch in aufs.
> 
> 
> J. R. Okajima