[parted-devel] Failure with O_DIRECT

[Michael Brennan]

Sorry I accidentally mailed you instead of the list first,
I've CC'd this one.

On Thu, Jul 05, 2007 at 01:59:28PM -0300, Otavio Salvador wrote:
> "Michael Brennan" <brennan.brisad at gmail.com> writes:
> 
> >    On 7/4/07, Otavio Salvador <otavio at debian.org> wrote:
> >      > Have you tested with lastest 2.4 series too?
> >
> >    Yes, I have tried the latest kernels from both 2.4 and 2.6 now and
> >    I found that 2.4 does not work while 2.6 works fine.
> 
> Ok, right.
> 
> >      I did some research about it and looks like we're not suppose to use
> >      O_DIRECT but madvice or posix_fadvice.
> >
> >    What is the main reason this non-buffering mechanism is used in parted?
> >    Is it for performance?  Or something other, like reducing the risk of
> >    corrupting data?
> 
> Probably to avoid possible corruption due a poweroff and like.

I'm probably missing something, but, if a power failure happens in the
middle of an unbuffered write, wouldn't that do more damage than if
the write is still in the buffers and haven't been synced yet?

> 
> >    The other partitioning tools I've used seem to use normal write
> >    operations and then sync the disks right afterwards, on the other
> >    hand, they only write the partition table to disk and does not
> >    have the advanced features parted has.
> 
> The only risk I can think is when a freeze or a poweroff happens but
> we might output an warning on those cases and then move to
> posix_fadvice.

Output a warning on which cases?

> 
> It need to be checked ...
>