Re: [PATCH] Add EXT4_STATE_MAY_INLINE_DATA check in ext4_readdir

On Thu, Aug 13, 2026 at 11:31:03PM -0400, Theodore Tso wrote:
> On Thu, Aug 13, 2026 at 04:33:57PM -0500, pipishuo wrote:
> > After a failure in the route ext4_restore_inline_data ->
> > ext4_create_inline_data -> ext4_mark_iloc_dirty, the inode will still
> > be considered as having inline data.
> 
> How did you run across this?  Was this something that was actually
> happening?  If so, what was triggering the error?
> 
> Or was this something that was found via AI?
I didn't have a reproducer. I was reviewing the inline data error paths 
and noticed the interaction with ext4_readdir looked fragile
so I attempted to make it more robust.
> 
> I'm fairly sure this patch was created using AI.  This useless
> inclusion was a bit of a tip-off:
> 
> > +#include <asm-generic/errno-base.h>
It was automatically inserted by my editor's include suggestion feature,
and I missed it in review. That's entirely my fault.
> 
> As far as the fix is concerned, there's a much better way of
> implementing the functionality in ext4_restore_inline_data(), which is
> to just capture a copy of the inode (since the inline data is stored
> in the inode), and if we need to back out a failed conversion in
> ext4_convert_inline_data_nolock(), we can just copy the old contents
> of the inode back in place.
Because ext4_convert_inline_data_nolock() has already called
ext4_destroy_inline_data_nolock(), the inode is no longer in the inline
state. So we need ext4_create_inline_data() to rebuild the inline state.
> 
> We can also do the conversion in a much safer way, by allocating the
> data block first, and writing a copy of the data to the data block,
> and only if that is successful, do we update the extent tree or
> indirect block mapping in the inode.  Of course, this would require a
> rewrite of ext4_convert_inline_data_nolock(), but this would allow us
> to fix up the other aspects of the error handling which is not quite
> right. 
I agree with your direction. I'll take some time to study the code 
and work on a proper v2 that rewrites ext4_convert_inline_data_nolock()
following that model.
> 
> 						- Ted