direct-IO writes strange behavior

[Alberto Nava <beto@xxxxxxxxxxx>]

Hi,

I'm seeing a very strange behavior when I write to a file using 512k
direct-io writes. The filesystem is XFS and the kernel is 2.6.0-test9.

The first time I do it the 512k writes are broken down into 4k requests
and the drives only see these small requests ( a lot of them).

The second time (after the data is been written once) the 512k writes
are preserved all the way to the drives.

Preallocating the space for the file have little impact on this
behavior.

Here is a more detail description of the two situations


first time:
        - rm /content/file
        - touch /content/file
        - xfs_io -c 'resvsp 0 1048576000' /content/file
        - xfs_bmap /content/file
                /content/file:
                0: [0..2047999]: 348456960..350504959

        - perform 2000 sequential direct IO writes onto the file 		 
starting from 0
        - all the BIOs that are passed to the io-scheduler are 4k long

second time:
       - xfs_bmap /content/file
                /content/file:
                0: [0..2047999]: 348456960..350504959

       - perform 2000 sequential direct IO writes onto the file 
starting 	from 0
       - all the BIOs that are passed to the io-scheduler are 512k long

Is this behavior expected on XFS or is this a generic direct-IO related 
issued?

Any information or suggestions would be greatly appreciated?
Thanks
beto