Re: assertion failed: last_size == new_size, file: fs/btrfs/inode.c

[Liu Bo <bo.li.liu@oracle.com>]

On Thu, Mar 09, 2017 at 03:24:21PM +0100, David Sterba wrote:
> On Thu, Mar 02, 2017 at 06:04:33PM -0800, Liu Bo wrote:
> > On Thu, Mar 02, 2017 at 07:58:01AM -0800, Liu Bo wrote:
> > > On Wed, Mar 01, 2017 at 03:03:19PM -0500, Dave Jones wrote:
> > > > On Tue, Feb 28, 2017 at 05:12:01PM -0800, Liu Bo wrote:
> > > >  > On Mon, Feb 27, 2017 at 11:23:42AM -0500, Dave Jones wrote:
> > > >  > > On Mon, Feb 27, 2017 at 07:53:48AM -0800, Liu Bo wrote:
> > > >  > >  > On Sun, Feb 26, 2017 at 07:18:42PM -0500, Dave Jones wrote:
> > > >  > >  > > Hitting this fairly frequently.. I'm not sure if this is the same bug I've
> > > >  > >  > > been hitting occasionally since 4.9. The assertion looks new to me at least.
> > > >  > >  > >
> > > >  > >  > 
> > > >  > >  > It was recently introduced by my commit and used to catch data loss at truncate.
> > > >  > >  > 
> > > >  > >  > Were you running the test with a mkfs.btrfs -O NO_HOLES?
> > > >  > >  > (We just queued a fix for the NO_HOLES case in btrfs-next.)
> > > >  > > 
> > > >  > > No, a fs created with default mkfs.btrfs options.
> > > >  > 
> > > >  > I have this patch[1] to fix a bug which results in file hole extent, and this
> > > >  > bug could lead us to hit the assertion.
> > > >  > 
> > > >  > Would you try to run the test w/ it, please?
> > > >  > 
> > > >  > [1]: https://patchwork.kernel.org/patch/9597281/
> > > > 
> > > > Made no difference. Still see the same trace & assertion.
> > > 
> > > Some updates here, I've got it reproduced, somehow a corner case ends up
> > > with a inline file extent following by some pre-alloc extents, along the
> > > way, isize also got updated unexpectedly.  Will try to narrow it down.
> > >
> > 
> > I realized that btrfs now could tolerate files that mix inline extents with
> > regular extents,
> 
> Where did that change? I thought that inline extent can represent only
> the entire file, so any mixing of extents does not make sense to me.
> Do you refer to some intermediate state where the file is being
> converted from inline to non-inline, thus we could see both types of
> extents at some point?
>

We could get that by doing the following step,

xfs_io -f -c "pwrite 0 8" foo
xfs_io -f -c "falloc 4 8188" foo

offset 4 is rounded down to offset 0 and before allocating blocks we wait for
any dirty stuff starting at offset 0, since the isize is not yet updated, the
inline extent is created as is again.  Now we have an inline extent from 0 to 8
and a prealloc extent from 4096 to 8192, AND its isize is 8192.

Thanks,

-liubo