[parted-devel] Failure with O_DIRECT

[Jim Meyering]

Michael Brennan <brennan.brisad at gmail.com> wrote:
> I would like to start to contribute to a free software project, and
> I've found that this one seems pretty nice.  I find filesystems
> and partitions to be an interesting subject and I really hope to be
> able to help in the future.  I'm now digging through the code trying
> to understand it and now I wonder if you could clarify a couple of
> things for me.
>
> First of all, parted won't work on my main computer, I have
> experimented a little and concluded that it is because of O_DIRECT.
> I run linux on all my computers and on all the others parted works
> fine, but on this one (running kernel 2.4.34.1) the read and writes
> only fails with Invalid argument.  If I remove O_DIRECT it works fine,

Thanks for the report.
FYI, O_DIRECT works fine with ext3 at least on recent linux-2.6.
But there is definitely a file-system-dependent bit here.
O_DIRECT does not work on "tmpfs", and there are some restrictions
even with recent reiserfs.  See the comments in m4/o-direct.m4.

> or if I align the memory with posix_memalign() in linux_read() to
> be aligned to 1024 instead of the dev->sector_size value of 512 the
> read will succeed.  I picked 1024 because I read somewhere that
> and O_DIRECT read/write must be aligned to the blocksize of the
> filesystem, 1024 on ext3 in my case.  But that requirement
> is obviously not needed on the other machines since they work
> fine with 512 aligned memory.  Do you know what the real alignment
> requirement is for reading/writing with O_DIRECT?

According to open(2), the alignment requirement is 512 for linux-2.6 and
FS logical block size for linux-2.4.  But the worst is that for 2.4,
the _transfer size_ (i.e., all read and write sizes) must also be a
multiple of the logical block size of the underlying file system.

I've been working, off and on, on a patch to make parted work
with arbitrary "sector_size".  That will be required for newer
disks with larger-than-512-byte physical sector sizes.
With a little luck, it'll be in shape to post soon.

> And two other questions:
> 1. While phys_sector_size is pretty clear, what does sector_size
> (logical sector size) in PedDevice represent?  Could it be so that
> this variable should be the same as the block size of the
> filesystem?

I wondered the same thing, and concluded "Yes", but
maybe Andrew or Leslie (Cc'd) can confirm that.
If so, I'll document it, so no one else has to wonder :)

> 2. How should the problem with O_DIRECT on my computer be solved?
> Open without O_DIRECT or find out the correct alignment in some way?

Another possibility is to use posix_fadvise, but
that has its own problems.  Some of them were discussed
when we added oflag=direct support to coreutils' dd:

  http://thread.gmane.org/gmane.comp.gnu.coreutils.bugs/9968/focus=9982