Re: [PATCH v14 13/21] xfs: use read ioend for fsverity data verification

["Darrick J. Wong" <djwong@kernel.org>]

On Tue, Aug 11, 2026 at 10:18:05AM +0200, Andrey Albershteyn wrote:
> > > > "If we have fsverity and block device integrity attached to this bio,
> > > > we need to run both validations from the separate fsverity workqueue
> > > > to avoid deadlocking due to fsverity issuing its own reads."
> > > > 
> > > > (Assuming I understand the fsverity && pi case correctly.)
> > > > 
> > > > One thing I'm not clear about -- why is it safe to do the fsverity
> > > > validation here if PI isn't enabled?  Can't that also issue IO to pull
> > > > in merkle tree blocks?
> > > 
> > > Without PI, fsverity metadata is read without XFS bio completion
> > > path, we don't get here for the descriptor/metadata reads
> > > (see xfs_get_iomap_read_ops()). So, we won't block the queue, as
> > > data ioends won't be mixed with metadata ioends.
> > > 
> > > With PI, all fsverity reads goes through this path. We could get a
> > > case that data ioend is waiting for metadata IO to be completed which
> > > in turn is pending for data ioend to be finished (due to batch
> > > processing of multiple BIOs in the bio_complete wq).
> > > 
> > > So, this will issue more IO, but this IO will not get onto this
> > > queue (it will go through iomap_bio_submit_read()).
> > 
> > Ah, ok.  Maybe add to that comment:
> > 
> > "If we have fsverity enabled but block device integrity is not enabled,
> > completion of the fsverity metadata reads does not require a workqueue
> > so there is no deadlock potential."
> > 
> > then?
> > 
> > (Just echoing you to make sure I understand completely.)
> > 
> 
> yes,
> 
> I've changed it to:
> 
> +       /*
> +        * If we have fsverity and block device integrity attached to this bio,
> +        * we need to run fsverity verification of data folios from a separate
> +        * fsverity workqueue. This is necessary to avoid deadlocking due to
> +        * fsverity issuing more reads of fsverity metadata which would be
> +        * processed by the same worker in the BIO completion workqueue.
> +        *
> +        * Without device integrity, fsverity metadata IO will not use ioends for

I would like to nitpick this to "Without block device integrity..." but
otherwise this comment looks good to me :)

With that fixed,
Reviewed-by: "Darrick J. Wong" <djwong@kernel.org>

--D

> +        * completion.
> +        */
> 
> 
> -- 
> - Andrey
> 
>