Re: [PATCH 18/22] xfs: use BIO_COMPLETE_IN_TASK for bounce buffered read I/Os

[Christoph Hellwig <hch@lst.de>]

On Fri, Jul 24, 2026 at 11:38:53AM +0200, Andrey Albershteyn wrote:
> On 2026-07-24 08:21:20, Christoph Hellwig wrote:
> > On Thu, Jul 23, 2026 at 05:58:49PM +0200, Andrey Albershteyn wrote:
> > > I would probably need something like this for fsverity.
> > > 
> > > When using integrity checksums with fsverity together, both file data
> > > ioends and fsverity metadata ioends get onto the inode queue
> > > (without checksums there's no metadata ioends). This way worker could
> > > self-deadlock if data ioend processed first in xfs_end_io(). It will
> > > call verify_bio to read merkle pages which could be already waiting
> > > in queue.
> > 
> > Yes.  Same as the situation we have (or had as this is changing
> > for 7.3) with fsverity and fscrypt.
> > 
> > > I initially considered changing xfs_end_io, for read ioends, to just
> > > schedule them instead of adding to the queue, but decided just sort
> > > metadata ioends first in the queue [1] as a bit simpler fix.
> > > 
> > > The difference is that work won't be scheduled on the high-priority
> > > fsverity's workqueue. Not sure how critical this is as with
> > > checksums all reads would be the same priority.
> > 
> > I'll need to go back to your series, but just sorting isn't going
> > to fix the problem, as the reading of the verity metadata could
> > be kicked off from the completion workqueue and you could be blocking
> > the rescuer thread.  So either verity needs it's own WQ, or we'd need
> > some way for a currently running I/O completion to exit and get
> > restarted.
> 
> Oh I see, then, the data reads could be queue further with
> fsverity_enqueue_verify_work(ioend). This way bio completion queue
> can continue process incoming metadata reads while data ioend is
> waiting for verification.


fsverity_enqueue_verify_work would actually really benefit from doing
a list (percpu or otherwise) so that the caller only needs a single
or doubly linked list entry instead of adding another struct.

But if we don't want to do that work yet, adding a verity context
of some sort and using fsverity_enqueue_verify_work could work.