Re: [PATCH v3 5/6] ext4: introduce direct IO write path using iomap infrastructure

[Matthew Bobrowski <mbobrowski@mbobrowski.org>]

On Tue, Sep 17, 2019 at 02:02:33AM -0700, Christoph Hellwig wrote:
> On Tue, Sep 17, 2019 at 02:30:15PM +0530, Ritesh Harjani wrote:
> > So if we have a delayed buffered write to a file,
> > in that case we first only update inode->i_size and update
> > i_disksize at writeback time
> > (i.e. during block allocation).
> > In that case when we call for ext4_dio_write_iter
> > since offset + len > i_disksize, we call for ext4_update_i_disksize().
> > 
> > Now if writeback for some reason failed. And the system crashes, during the
> > DIO writes, after the blocks are allocated. Then during reboot we may have
> > an inconsistent inode, since we did not add the inode into the
> > orphan list before we updated the inode->i_disksize. And journal replay
> > may not succeed.
> > 
> > 1. Can above actually happen? I am still not able to figure out the
> >    race/inconsistency completely.
> > 2. Can you please help explain under what other cases
> >    it was necessary to call ext4_update_i_disksize() in DIO write paths?
> > 3. When will i_disksize be out-of-sync with i_size during DIO writes?
> 
> None of the above seems new in this patchset, does it?

That's correct.

*Ritesh - FWIW, I think you'll find the answers to your questions above by
 referring to the following commits:

 1) 73fdad00b208b
 2) 45d8ec4d9fd54

If you drop the check (offset + count > EXT4_I(inode)->i_disksize) and the
call to ext4_update_i_disksize(), under some workloads i.e. "generic/475"
you'll generally end up with metadata corruption.

> That being said I found the early size update odd. XFS updates the on-disk
> size only at I/O completion time to deal with various races including the
> potential exposure of stale data.

Indeed a little odd. But, I think delalloc/writeback implementation is
possibly to blame here (based on what's detailed in 45d8ec4d9fd54)? Ideally, I
had the same approach as XFS in mind, but I couldn't do it.

--<M>--