Re: [PATCH V5 7/9] dax: read holder_ops once in dax_holder_notify_failure()

[John Groves <John@groves.net>]

On 26/06/12 11:02AM, Richard Cheng wrote:
> On Thu, Jun 11, 2026 at 05:32:45PM +0800, John Groves wrote:
> > From: John Groves <John@Groves.net>
> > 
> > dax_holder_notify_failure() reads dax_dev->holder_ops twice without
> > READ_ONCE() -- once for the NULL check and once for the indirect
> > notify_failure() call. A concurrent fs_put_dax() or kill_dax() can clear
> > holder_ops between the two reads, so the check can observe a non-NULL
> > pointer while the call dereferences NULL.
> > 
> 
> Hello John,
> 
> Thanks for this.
> 
> Reviewed-by: Richard Cheng <icheng@nvidia.com>
> 
> Small message nit, kill_dax() isn't a racing clearer.

Right -- kill_dax() clears holder_ops only after synchronize_srcu(), so it
can't race a reader under dax_read_lock(). Reworded the commit message and
the code comment for V6 to name only fs_put_dax().

> Plus I think this only fix holder_ops double-fetch, the fs_put_dax()
> race issue is separate and pre-existing.

Yes, intentionally -- this patch is just the reader-side double-fetch. The
fs_put_dax() race is the separate, pre-existing one handled by the next
patch in the series ("dax: fix holder_ops race in fs_put_dax()").

> 
> Best regards,
> Richard Cheng.

Thanks,
John

<snip>