On Tue, 12 Aug 2008 13:05:43 -0700 (PDT)
Linus Torvalds [off-list ref] wrote:

On Wed, 13 Aug 2008, OGAWA Hirofumi wrote:

quoted

However, there are some similar stuff: ->st_size, ->st_nlink and
->st_ino in stat() (cp_old_stat()). Maybe EOVERFLOW is the reason for
consistency...

.. I actually had an old binary that triggered that case (actually, it was 
O_LARGEFILE in open()), and didn't work as a result. I only needed to run 
it once, so I literally hacked up a once-time-use kernel that just removed 
the EOVERFLOW in open.

So no, I'm not a fan of EOVERFLOW at all. Not in readdir(), not really 
anywhere else. 

In a lot of places the standard mandates it to ensure you don't get nasty
suprises. There are places where during the LFS work people found apps
whose large file response was unpleasant - things like "get the file
length, reduce it to an exact number of records long with ftruncate in
case we got a partial record write somewhere" do horrible things when the
length wraps.

That said there is nothing that says we can't have a 'posix_me_harder'
sysfs control for such things. The standards say what should occur in the
normal situation not what should occur if you intentionally move out of
the standard definition.

Alan