Re: [PATCH] 9p/virtio: restrict page pinning to user_backed_iter() iovec

[Christoph Hellwig <hch@infradead.org>]

On Mon, Dec 15, 2025 at 04:34:12PM +0900, Dominique Martinet wrote:
> Christoph Hellwig wrote on Sun, Dec 14, 2025 at 09:55:12PM -0800:
> > > Ok, I don't understand why the current code locks everything down and
> > > wants to use a single scatterlist shared for the whole channel (and
> > > capped to 128 pages?), it should only need to lock around the
> > > virtqueue_add_sg() call, I'll need to play with that some more.
> > 
> > What do you mean with "lock down"?
> 
> Just the odd (to me) use of the chan->lock around basically all of
> p9_virtio_request() and most of p9_virtio_zc_request() -- I'm not pretty
> sure this was just the author trying to avoid an allocation by recycling
> the chan->sg array around though, so ignore this.

Oh, ok.  This seems unrelated to the handling of the iov_iters and
I'm sorry that I don't really know anything about that part.

> 
> > > Looking at other virtio drivers I could probably use a sg_table and
> > > have extract_iter_to_sg() do all the work for us...
> > 
> > Looking at the code I'm actually really confused.  Both because I
> > actually though we were talking about the 9fs direct I/O code, but
> > that has actually been removed / converted to netfs a long time ago.
> >
> > But even more so what the net/9p code is actually doing..  How do
> > we even end up with user addresses here at all?
> 
> FWIW I tried logging and saw ITER_BVEC, ITER_KVEC and ITER_FOLIOQ --
> O_DIRECT writes are seen as BVEC so I guess it's not as direct as I
> expected them to be -- that code could very well be leftovers from
> the switch to iov_iter back in 2015...

Oh right, I think this from Dave's netfs_extract_user_iter.

> (waiting for David's answer here, but as far as I see the contract
> between the transport and the vfs is that the transport should handle
> whatever it's being fed, so it doesn't really matter if it's a bio_vec
> or an iov_iter -- ultimately virtio or whatever backend that wants to
> handle zc likely won't handle bio_vec any better so it'll need
> converting anyway)

Yeah.  Looking at what the code does with the pages, I think all this
should go away in favor of using extract_iter_to_sg and build the
scatterlists directly from the iters, without an extra page indirection.

(and of course one day virtio should migrate away from scatterlists,
but that's for another time).