Re: users Digest, Vol 38, Issue 13

Re: users Digest, Vol 38, Issue 13

[Michael Niederle]

I'm evaluating nilfs as a replacement for the typical layered filesystem
approach (unionfs, aufs) used in live-systems.

nilfs performs incredibly well! Its sequential write / random read philosophy
is great for accessing pen-drives.

I have only one big problem: I twice corrupted the nilfs-superblock (due to
inproper shutdown) leading to a complete data loss each time.

The todo-list mentions that more copies of the superblock are needed. I
strongly agree! I think this should be one of the top items on the todo-list.

Greetings, Michael

Re: users Digest, Vol 38, Issue 13

[Ryusuke Konishi]

Hi!
On Thu, 14 May 2009 19:39:16 +0200, "Michael Niederle" wrote:
> I'm evaluating nilfs as a replacement for the typical layered filesystem
> approach (unionfs, aufs) used in live-systems.
> 
> nilfs performs incredibly well! Its sequential write / random read philosophy
> is great for accessing pen-drives.

Thanks, but I suspect partially the writes are not done on your
pendrive ;-)
 
> I have only one big problem: I twice corrupted the nilfs-superblock (due to
> inproper shutdown) leading to a complete data loss each time.

Yeah, that is a headache.  Milos Jana has been reporting this issue
for long periods, but we haven't solved this yet.

What is your pendrive product?

I'd like to buy a pendrive to clarify the root cause of this problem.

BTW, could you try switching IO scheduler to "noop" ?
It can be done as follows:

 # cd /sys/block/sda/queue    
   ('sda' should be the device name of the pendrive)
 # cat scheduler
 anticipatory deadline [cfq] noop
 # echo noop > scheduler
 # cat scheduler
 anticipatory deadline cfq [noop]

I think this would help to narrow down the problem.

> The todo-list mentions that more copies of the superblock are needed. I
> strongly agree! I think this should be one of the top items on the todo-list.
> 
> Greetings, Michael

The latest nilfs already has secondary super block in tail of
partition.  But it looks insufficient for this pendrive problem.

I think ensuring write on unmount should come before that.

Regards,
Ryusuke Konishi

Re: users Digest, Vol 38, Issue 13

[Ryusuke Konishi]

On Fri, 15 May 2009 09:21:43 +0900 (JST), Ryusuke Konishi wrote:
> On Thu, 14 May 2009 19:39:16 +0200, "Michael Niederle" wrote:
> > The todo-list mentions that more copies of the superblock are needed. I
> > strongly agree! I think this should be one of the top items on the todo-list.
> > 
> > Greetings, Michael
> 
> The latest nilfs already has secondary super block in tail of
> partition.  But it looks insufficient for this pendrive problem.
> 
> I think ensuring write on unmount should come before that.

Just a quick note.  The write on unmount is already _DONE_ on the
source code.  I here mean some enhanced operation like diskcache
flushing.

Ryusuke Konishi