Re: [RFC PATCH 27/28] sunrpc: Use sendmsg(MSG_SPLICE_PAGES) rather then sendpage

From: Chuck Lever III <chuck.lever@oracle.com>
To: David Howells <dhowells@redhat.com>
Cc: Trond Myklebust <trondmy@hammerspace.com>,
	Matthew Wilcox <willy@infradead.org>,
	"David S. Miller" <davem@davemloft.net>,
	Eric Dumazet <edumazet@google.com>,
	Jakub Kicinski <kuba@kernel.org>, Paolo Abeni <pabeni@redhat.com>,
	Al Viro <viro@zeniv.linux.org.uk>,
	Christoph Hellwig <hch@infradead.org>,
	Jens Axboe <axboe@kernel.dk>, Jeffrey Layton <jlayton@kernel.org>,
	Christian Brauner <brauner@kernel.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	"netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"linux-fsdevel@vger.kernel.org" <linux-fsdevel@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"linux-mm@kvack.org" <linux-mm@kvack.org>,
	Anna Schumaker <anna@kernel.org>,
	Linux NFS Mailing List <linux-nfs@vger.kernel.org>
Subject: Re: [RFC PATCH 27/28] sunrpc: Use sendmsg(MSG_SPLICE_PAGES) rather then sendpage
Date: Thu, 16 Mar 2023 17:41:44 +0000	[thread overview]
Message-ID: <3B4C4D62-E80E-4921-BA7A-1A157942159F@oracle.com> (raw)
In-Reply-To: <782776.1678987682@warthog.procyon.org.uk>

> On Mar 16, 2023, at 1:28 PM, David Howells <dhowells@redhat.com> wrote:
> 
> Chuck Lever III <chuck.lever@oracle.com> wrote:
> 
>> That means I haven't seen the cover letter and do not have any
>> context for this proposed change.
> 
> https://lore.kernel.org/linux-fsdevel/20230316152618.711970-1-dhowells@redhat.com/
> 
>> We've tried combining the sendpages calls in here before. It
>> results in a significant and measurable performance regression.
>> See:
>> 
>> da1661b93bf4 ("SUNRPC: Teach server to use xprt_sock_sendmsg for socket sends")
> 
> The commit replaced the use of sendpage with sendmsg, but that took away the
> zerocopy aspect of sendpage.  The idea behind MSG_SPLICE_PAGES is that it
> allows you to do keep that.  I'll have to try reapplying this commit and
> adding the MSG_SPLICE_PAGES flag.

Note that, as Trond point out, NFSD can handle an NFS READ
request with either a splice actor or by copying through a
vector, depending on what the underlying filesystem can
support and whether we are using a security flavor that
requires stable pages. Grep for RQ_SPLICE_OK.

Eventually we want to make use of iomaps to ensure that
reading areas of a file that are not allocated on disk
does not trigger an extent allocation. Anna is working on
that, but I have no idea what it will look like. We can
talk more at LSF, if you'll both be around.

Also... I find I have to put back the use of MSG_MORE and
friends in here, otherwise kTLS will split each of these
kernel_sendsomething() calls into its own TLS record. This
code is likely going to look different after support for
RPC-with-TLS goes in.

>> Therefore, this kind of change needs to be accompanied by both
>> benchmark results and some field testing to convince me it won't
>> cause harm.
> 
> Yep.
> 
>> And, we have to make certain that this doesn't break operation
>> with kTLS sockets... do they support MSG_SPLICE_PAGES ?
> 
> I haven't yet tackled AF_TLS, AF_KCM or AF_SMC as they seem significantly more
> complex than TCP and UDP.  I thought I'd get some feedback on what I have
> before I tried my hand at those.

OK, I didn't mean AF_TLS, I meant the stuff under net/tls,
which is AF_INET[6] and TCP, but with a ULP in place. It's
got its own sendpage and sendmsg methods that choke when
an unrecognized MSG_ flag is present.

But OK, you're just asking for feedback, so I'll put my red
pencil down.

--
Chuck Lever