Re: [PATCH v5 40/40] 9p: Use netfslib read/write_iter

[David Howells <dhowells@redhat.com>]

Dominique Martinet <asmadeus@codewreck.org> wrote:

> > +static void v9fs_upload_to_server(struct netfs_io_subrequest *subreq)
> > +{
> > +	struct inode *inode = subreq->rreq->inode;
> > +	struct v9fs_inode __maybe_unused *v9inode = V9FS_I(inode);
> 
> Any reason to have this variable assignment at all?

I'll just remove it.  The __maybe_unused suppressed the warning, otherwise I'd
have removed it already.

> p9_client_write return value should always be subreq->len, but I believe
> we should use it unless err is set.
> (It's also possible for partial writes to happen, e.g. p9_client_write
> looped a few times and then failed, at which point the size returned
> would be the amount that actually got through -- we probably should do
> something with that?)

How about something like:

-	int err;
+	int err, len;
 
 	trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
-	p9_client_write(fid, subreq->start, &subreq->io_iter, &err);
-	netfs_write_subrequest_terminated(subreq, err < 0 ? err : subreq->len,
-					  false);
+	len = p9_client_write(fid, subreq->start, &subreq->io_iter, &err);
+	netfs_write_subrequest_terminated(subreq, len ?: err, false);

> > +	total = p9_client_read(fid, subreq->start + subreq->transferred,
> > +			       &subreq->io_iter, &err);
> 
> Just to clarify: subreq->io_iter didn't exist (or some conditions to use
> it weren't cleared) before?

Correct.  It's added in the netfs-lib patches.  I've provided a way to
separate the user-side iterator from the I/O-side iterator to allow the use of
a bounce buffer for the purpose of content crypto, compression or just having
to deal with RMW cycles to a larger block size on the server.

> > +	if (file) {
> > +		fid = file->private_data;
> > +		BUG_ON(!fid);
> 
> This probably should be WARN + return EINVAL like find by inode?
> It's certainly a huge problem, but we should avoid BUG if possible...

Sure.  The BUG_ON() was already there, but I can turn it into a WARN+error.

> nit: not sure what's cleaner?
> Since there's a message that makes for a bit awkward if...
> 
> if (WARN_ONCE(!fid, "folio expected an open fid inode->i_private=%p\n",
> 	      rreq->inode->i_private))
> 	return -EINVAL;
> 
> (as a side note, I'm not sure what to make of this i_private pointer
> here, but if that'll help you figure something out sure..)

Um.  9p is using i_private.  But perhaps i_ino would be a better choice:

	if (file) {
		fid = file->private_data;
		if (!fid)
			goto no_fid;
		p9_fid_get(fid);
	} else {
		fid = v9fs_fid_find_inode(rreq->inode, writing, INVALID_UID, true);
		if (!fid)
			goto no_fid;
	}

	...

no_fid:
	WARN_ONCE(1, "folio expected an open fid inode->i_ino=%lx\n",
		  rreq->inode->i_ino);
	return -EINVAL;

> This is as follow on your netfs-lib branch:
> -       WARN_ON(rreq->origin == NETFS_READ_FOR_WRITE &&
> -                       !(fid->mode & P9_ORDWR));
> -
> -       p9_fid_get(fid);
> +       WARN_ON(rreq->origin == NETFS_READ_FOR_WRITE && !(fid->mode & P9_ORDWR));
> 
> So the WARN_ON has been reverted back with only indentation changed;
> I guess there were patterns that were writing despite the fid not having
> been open as RDWR?
> Do you still have details about these?

The condition in the WARN_ON() here got changed.  It was:

	WARN_ON(writing && ...

at one point, but that caused a bunch of incorrect warning to appear because
only NETFS_READ_FOR_WRITE requires read-access as well as write-access.  All
the others:

	bool writing = (rreq->origin == NETFS_READ_FOR_WRITE ||
			rreq->origin == NETFS_WRITEBACK ||
			rreq->origin == NETFS_WRITETHROUGH ||
			rreq->origin == NETFS_LAUNDER_WRITE ||
			rreq->origin == NETFS_UNBUFFERED_WRITE ||
			rreq->origin == NETFS_DIO_WRITE);

only require write-access.

There will be an additional one if we roll out content crypto to 9p as we may
need to do RMW cycles occasionally - but that's off to one side just for the
moment.

> If a file has been open without the write bit it might not go through,
> and it's incredibly difficult to get such users back to userspace in
> async cases (e.g. mmap flushes), so would like to understand that.

The VFS/VM should prevent writing to files that aren't open O_WRONLY or
O_RDWR, so I don't think we should be called in otherwise.

Note that I'm intending to change the way fscache is driven when we fetch
cacheable data from the server so that I can free up the PG_fscache bit and
return it to the MM folks.  Instead of marking the page PG_fscache, I mark it
PG_dirty and set page->private with a special value to indicate it should only
be written to the cache - then the writepages sees that and just writes these
pages to the cache.  I have a patch to do this and it seems to work, but I
need to make ceph and cifs use netfslib before I can apply it.

> > +	p9_debug(P9_DEBUG_VFS, "(cached)\n");
> 
> (Not a new problem so no need to address here, but having just
> "(cached)" on a split line is a bit weird.. We first compute cached or
> not as a bool and make it %s + cached ? " (cached)" : "" or
> something... I'll send a patch after this gets in to avoid conflicts)

Okay.

> > +	return netfs_page_mkwrite(vmf, NULL);
> 
> (I guess there's no helper that could be used directly in .page_mkwrite
> op?)

I could provide a helper that just supplies NULL as the second argument.  I
think only 9p will use it, but that's fine.

David