Re: [PATCH v6 21/34] 9p: Pin pages rather than ref'ing if appropriate

[David Howells]

Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:

> Wait a sec; just how would that work for ITER_KVEC?  AFAICS, in your
> tree that would blow with -EFAULT...

You're right.  I wonder if I should handle ITER_KVEC in
iov_iter_extract_pages(), though I'm sure I've been told that a kvec might
point to data that doesn't have a matching page struct.  Or maybe it's that
the refcount shouldn't be changed on it.

A question for the 9p devs:

Looking more into p9_virtio_zc_request(), it might be better to use
netfs_extract_iter_to_sg(), since the page list is going to get turned into
one, instead of calling p9_get_mapped_pages() and pack_sg_list().

This would, however, require that chan->sg[] be populated outside of the
spinlock'd section - is there any reason that this can't be the case?  There's
nothing inside the locked section that makes sure the chan can be used before
it launches into loading up the scatterlist.  There is a wait afterwards, but
it has to drop the lock first, so wouldn't stop a parallel op from clobbering
chan->sg[] anyway.

Further, if virtqueue_add_sgs() fails with -ENOSPC and we go round again to
req_retry_pinned, do we actually need to reload chan->sg[]?

David