Re: [RFC 08/13] cifs: Add a function to read into an iter from a socket

From: David Howells <dhowells@redhat.com>
To: David Laight <David.Laight@ACULAB.COM>
Cc: dhowells@redhat.com, Steve French <smfrench@gmail.com>,
	Al Viro <viro@zeniv.linux.org.uk>,
	Shyam Prasad N <nspmangalore@gmail.com>,
	Rohith Surabattula <rohiths.msft@gmail.com>,
	"Tom Talpey" <tom@talpey.com>,
	Stefan Metzmacher <metze@samba.org>,
	"Christoph Hellwig" <hch@infradead.org>,
	Matthew Wilcox <willy@infradead.org>,
	"Jeff Layton" <jlayton@kernel.org>,
	"linux-cifs@vger.kernel.org" <linux-cifs@vger.kernel.org>,
	"linux-fsdevel@vger.kernel.org" <linux-fsdevel@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	Steve French <sfrench@samba.org>
Subject: Re: [RFC 08/13] cifs: Add a function to read into an iter from a socket
Date: Thu, 26 Jan 2023 15:44:01 +0000	[thread overview]
Message-ID: <2862713.1674747841@warthog.procyon.org.uk> (raw)
In-Reply-To: <0d53a3cc9f9448298bba04d06f51b23d@AcuMS.aculab.com>

David Laight <David.Laight@ACULAB.COM> wrote:

> On the face of it that passes a largely uninitialised 'struct msghdr'
> to cifs_readv_from_socket() in order to pass an iov_iter.
> That seems to be asking for trouble.
> 
> If cifs_readv_from_socket() only needs the iov_iter then wouldn't
> it be better to do the wrapper the other way around?
> (Probably as an inline function)
> Something like:
> 
> int
> cifs_readv_from_socket(struct TCP_Server_Info *server, struct msghdr *smb_msg)
> {
> 	return cifs_read_iter_from_socket(server, &smb_msg->msg_iter, smb_msg->msg_iter.count);
> }
> 
> and then changing cifs_readv_from_socket() to just use the iov_iter.

Yeah.  And smbd_recv() only cares about the iterator too.

> I'm also not 100% sure that taking a copy of an iov_iter is a good idea.

It shouldn't matter as the only problematic iterator is ITER_PIPE (advancing
that has side effects) - and splice_read is handled specially by patch 4.  The
problem with splice_read with the way cifs works is that it likes to subdivide
its read/write requests across multiple reqs and then subsubdivide them if
certain types of failure occur.  But you can't do that with ITER_PIPE.

I build an ITER_BVEC from ITER_PIPE, ITER_UBUF and ITER_IOVEC in the top
levels with pins inserted as appropriate and hand the ITER_BVEC down.  For
user-backed iterators it has to be done this way because the I/O may get
shuffled off to a different thread.

Reqs can then just copy the BVEC/XARRAY/KVEC and narrow the region because the
master request at the top does holds the vector list and the top cifs level or
the caller above the vfs (eg. sys_execve) does what is necessary to retain the
pages.

David