On 4/21/16 8:43 AM, Jan Tulak wrote:
> On Thu, Apr 21, 2016 at 11:39 AM, Jan Tulak <jtulak@redhat.com <mailto:jtulak@redhat.com>>wrote:
>
>     From: Dave Chinner <dchinner@redhat.com <mailto:dchinner@redhat.com>>
>
>     THIS PATCH HAS KNOWN ISSUES - it fails xfs/206 and xfs/216 tests, as it
>     shrinks a file instead just not using it entirely, when -d size is used.
>
>
> ​So the shrinking is happening here:
> 3127        /*
> 3128         * If the data area is a file, then grow it out to its final size
> 3129         * so that the reads for the end of the device in the mount code
> 3130         * will succeed.
> 3131         */
> 3132        if (xi.disfile && ftruncate64(xi.dfd, dblocks * blocksize) < 0) {​
>
>
> Before the patch, xi.disfile was 0 and so it didn't shrink the file
> to the size of the new FS. Now, what is the correct solve to this?
> Tests are written for the old behaviour, but this shrinking seems to
> be an intentional thing. It seems that the FS works ok even when this
> truncating is not applied, so I think that I should remove this chunk
> (or change it to xi.dcreate=1 only), and keep the old behaviour.
>
> What do you think about it, guys?

Can't remove it; that would break the other side of things, if you try
to mkfs.xfs -d size=2g on an existing 1g file... mount tries to do
IO to the last block, and if it's not truncated out, that will fail
(as the comment says).

I suppose the simple way to fix it is to only truncate up, never down.

i.e. truncate to max(dblocks * blocksize, st_size) or
if (xi.disfile && st_size <  dblocks * blocksize) { truncate ... }