Re: [PATCH] Return short read or 0 at end of a raw device, not EIO

[David Jeffery <djeffery@redhat.com>]

On 09/30/2014 11:28 AM, Christoph Hellwig wrote:
> On Mon, Sep 29, 2014 at 06:08:11PM -0400, David Jeffery wrote:
>> On 09/29/2014 03:05 PM, Christoph Hellwig wrote:
>>> Seems like this should be changed in the generic code, or is there some
>>> reason why it would return EIO only for devices, but not for regular
>>> files in this case?
>>>
>>
>> Regular files shouldn't be returning EIO and don't in my tests. The file
>> systems manage direct I/O EOF handling in their own block or direct_IO
>> callbacks.  Block devices do not and instead do the size checks up
>> front.  Raw devices were bypassing the block device check, so only the
>> raw driver should be having this issue.
> 
> So I guess the problem is commit
> 
> "blkdev_aio_read(): switch to generic_file_read_iter(), get rid of iov_shorten()"
> 
> which removes the iov_shorten call in blkdev_aio_read?  This should be
> mentioned in the changelog.
> 
> But maybe we should instead make block devices behave more similar to
> regular files in this respect?

The issue's origin should be with bbec0270bdd8 "blkdev_max_block: make
private to fs/buffer.c".  This change intentionally removed the size
checks from blkdev_get_block() to avoid issues with the size variables
changing underneath the calls.  Commit 684c9aaebbb0 is the initial
commit to add the size checks with blkdev_aio_read() since
blkdev_get_block() no longer checks size.  As the raw driver was also
dependent on the size checks in blkdev_get_block(), it was broken by
bbec0270bdd8 but not fixed at all by 684c9aaebbb0.

> 
> Also did you make sure to add your regression test somewhere, e.g. ltp?
> 

No I have not.  I'll see about an ltp test.