Re: [Ext2-devel] Re: [RFC/RFT] [PATCH] EXT3: Retry allocation after journal commit

["Theodore Ts'o" <tytso@mit.edu>]

On Thu, May 13, 2004 at 07:53:10PM -0700, Andrew Morton wrote:
> "Theodore Ts'o" <tytso@mit.edu> wrote:
> >
> > It is possible for block allocation to fail, even if there is space in
> >  the filesystem, because all of the free blocks were recently deleted and
> >  so could not be allocated until after the currently running transaction
> >  is committed.   This can result in a very strange and surprising result
> >  where a system call such as a mkdir() will fail even though there is
> >  plenty of disk space apparently available.
> 
> I merged a little patch for this into post-2.6.6, but that only addresses
> prepare_write().

Oh, sorry, I didn't see that patch.  The specific bug I was trying to
fix was actually in mkdir though (the regression test suite did the
equivalent of rm -rf /mntpt/*, followed by mkdir's which failed).

I can respin the patch versus BK-latest.

> I wonder if there's much value in having ext3_has_free_blocks()?  We could
> just retry three times even if the fs is really full?

We could; it's an error path, after all.  On the other hand, the cost
of ext3_has_free_blocks() is pretty small, and if we know that there's
no point doing the retry, why not skip the work.

> I'd be inclined to pass a pointer to the `retry' counter into
> ext3_should_retry_alloc() and implement the counting in there.  It'll tidy
> things up a bit and consolidates that piece of policy in a single place.

Yes, good point.

						- Ted