UBIFS synchronisation questions about file create/delete

UBIFS synchronisation questions about file create/delete

[Ronny Dierckx]

Hi,

I'm using UBIFS (git://git.infradead.org/~dedekind/ubifs-v2.6.27.git)
on an ARM device with OneNAND flash.

I've been reading the documentation, FAQ, ... about sync/fsync/fdatasync and
the
write-back buffers, but there are still some things I don't understand.

I create a file with vi, write some data in it, close it, don't call sync.
Then 
after waiting 10 minutes I cut the power. When the system is restarted, the
file is
not there anymore?! Is this normal behavior? Shouldn't there be at least
a file of 0 bytes because "UBIFS operations which change directory entries
are synchronous"?
Doesn't this mean that creating or deleting a file entry doesn't require a
sync ?

Thanks,

Ronny Dierckx

Re: UBIFS synchronisation questions about file create/delete

[Adrian Hunter]

Ronny Dierckx wrote:
> Hi,
> 
> I'm using UBIFS (git://git.infradead.org/~dedekind/ubifs-v2.6.27.git)
> on an ARM device with OneNAND flash.
> 
> I've been reading the documentation, FAQ, ... about sync/fsync/fdatasync and
> the
> write-back buffers, but there are still some things I don't understand.
> 
> I create a file with vi, write some data in it, close it, don't call sync.
> Then 
> after waiting 10 minutes I cut the power. When the system is restarted, the
> file is
> not there anymore?! Is this normal behavior? Shouldn't there be at least
> a file of 0 bytes because "UBIFS operations which change directory entries
> are synchronous"?
> Doesn't this mean that creating or deleting a file entry doesn't require a
> sync ?

No it is not normal behaviour.

10 minutes is long enough for everything to be written and flushed.  The file
should be completely intact.

No UBIFS operations which change directory entries are not entirely
synchronous.  There is a NAND-page-sized buffer which must be flushed.
It is flushed every 5 seconds or so.  Otherwise you must do some kind of
sync.

I can only guess that that version of UBIFS is somehow broken.

Please send a bug report as described here:
http://www.linux-mtd.infradead.org/doc/ubifs.html#L_how_send_bugreport
but set the Default message level to 1.

Re: UBIFS synchronisation questions about file create/delete

[Jamie Lokier]

Adrian Hunter wrote:
> 10 minutes is long enough for everything to be written and flushed.  The 
> file should be completely intact.

Unless laptop_mode is used.  It is normally for letting disks spin
down by not flushing writes quickly.

-- Jamie

RE: UBIFS synchronisation questions about file create/delete

[Ronny Dierckx]

>No it is not normal behaviour.
>
>10 minutes is long enough for everything to be written and flushed.  The
file
>should be completely intact.
>
>No UBIFS operations which change directory entries are not entirely
>synchronous.  There is a NAND-page-sized buffer which must be flushed.
>It is flushed every 5 seconds or so.  Otherwise you must do some kind of
>sync.
>
>I can only guess that that version of UBIFS is somehow broken.

Hi Adrian,

I tried some different versions from
git://git.infradead.org/~dedekind/ubifs-v2.6.27.git
and the problem is caused by commit " UBIFS: fix 64-bit divisions in debug
print " 
9d016dc4b3006ef3493b21442394028614e26da5 :

dbg_io("set timer for jhead %d, %llu millisecs", wbuf->jhead,
-       ktime_to_ns(wbuf->hardlimit)/USEC_PER_SEC);
-       hrtimer_start(&wbuf->timer, wbuf->hardlimit, HRTIMER_MODE_REL);
+       div_u64(ktime_to_ns(wbuf->hardlimit), USEC_PER_SEC)); 
}

As you can see the timer is no longer started. Could you please fix this ?

Thanks,

Ronny Dierckx

Re: UBIFS synchronisation questions about file create/delete

[Adrian Hunter]

Ronny Dierckx wrote:
>> No it is not normal behaviour.
>>
>> 10 minutes is long enough for everything to be written and flushed.  The
> file
>> should be completely intact.
>>
>> No UBIFS operations which change directory entries are not entirely
>> synchronous.  There is a NAND-page-sized buffer which must be flushed.
>> It is flushed every 5 seconds or so.  Otherwise you must do some kind of
>> sync.
>>
>> I can only guess that that version of UBIFS is somehow broken.
> 
> Hi Adrian,
> 
> I tried some different versions from
> git://git.infradead.org/~dedekind/ubifs-v2.6.27.git
> and the problem is caused by commit " UBIFS: fix 64-bit divisions in debug
> print " 
> 9d016dc4b3006ef3493b21442394028614e26da5 :
> 
> dbg_io("set timer for jhead %d, %llu millisecs", wbuf->jhead,
> -       ktime_to_ns(wbuf->hardlimit)/USEC_PER_SEC);
> -       hrtimer_start(&wbuf->timer, wbuf->hardlimit, HRTIMER_MODE_REL);
> +       div_u64(ktime_to_ns(wbuf->hardlimit), USEC_PER_SEC)); 
> }
> 
> As you can see the timer is no longer started. Could you please fix this ?

It is Artem's git tree.

Artem, can you look at this?

Re: UBIFS synchronisation questions about file create/delete

[Artem Bityutskiy]

On Fri, 2009-11-27 at 15:44 +0200, Adrian Hunter wrote:
> > I tried some different versions from
> > git://git.infradead.org/~dedekind/ubifs-v2.6.27.git
> > and the problem is caused by commit " UBIFS: fix 64-bit divisions in debug
> > print " 
> > 9d016dc4b3006ef3493b21442394028614e26da5 :
> > 
> > dbg_io("set timer for jhead %d, %llu millisecs", wbuf->jhead,
> > -       ktime_to_ns(wbuf->hardlimit)/USEC_PER_SEC);
> > -       hrtimer_start(&wbuf->timer, wbuf->hardlimit, HRTIMER_MODE_REL);
> > +       div_u64(ktime_to_ns(wbuf->hardlimit), USEC_PER_SEC)); 
> > }
> > 
> > As you can see the timer is no longer started. Could you please fix this ?
> 
> It is Artem's git tree.
> 
> Artem, can you look at this?

Done.

-- 
Best Regards,
Artem Bityutskiy (Артём Битюцкий)