Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Content-Type: text/plain;
        charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 13.4 \(3608.120.23.2.4\))
Subject: Re: [PATCH RFC] SUNRPC: Use zero-copy to perform socket send
 operations
From:   Chuck Lever <chuck.lever@oracle.com>
In-Reply-To: <3194609c525610dc502d69f11c09cff1c9b21f2d.camel@hammerspace.com>
Date:   Mon, 9 Nov 2020 14:31:26 -0500
Cc:     Linux NFS Mailing List <linux-nfs@vger.kernel.org>,
        "netdev@vger.kernel.org" <netdev@vger.kernel.org>
Content-Transfer-Encoding: quoted-printable
Message-Id: <A3D0FF41-D88F-4116-AD47-AF9C94B1D984@oracle.com>
References: <160493771006.15633.8524084764848931537.stgit@klimt.1015granger.net>
 <9ce015245c916b2c90de72440a22f801142f2c6e.camel@hammerspace.com>
 <0313136F-6801-434F-8304-72B9EADD389E@oracle.com>
 <f03dae6d36c0f008796ae01bbb6de3673e783571.camel@hammerspace.com>
 <5056C7C7-7B26-4667-9691-D2F634C02FB1@oracle.com>
 <3194609c525610dc502d69f11c09cff1c9b21f2d.camel@hammerspace.com>
To:     Trond Myklebust <trondmy@hammerspace.com>
Precedence: bulk


> On Nov 9, 2020, at 1:16 PM, Trond Myklebust <trondmy@hammerspace.com> =
wrote:
>=20
> On Mon, 2020-11-09 at 12:36 -0500, Chuck Lever wrote:
>>=20
>>=20
>>> On Nov 9, 2020, at 12:32 PM, Trond Myklebust <
>>> trondmy@hammerspace.com> wrote:
>>>=20
>>> On Mon, 2020-11-09 at 12:12 -0500, Chuck Lever wrote:
>>>>=20
>>>>=20
>>>>> On Nov 9, 2020, at 12:08 PM, Trond Myklebust
>>>>> <trondmy@hammerspace.com> wrote:
>>>>>=20
>>>>> On Mon, 2020-11-09 at 11:03 -0500, Chuck Lever wrote:
>>>>>> Daire Byrne reports a ~50% aggregrate throughput regression
>>>>>> on
>>>>>> his
>>>>>> Linux NFS server after commit da1661b93bf4 ("SUNRPC: Teach
>>>>>> server
>>>>>> to
>>>>>> use xprt_sock_sendmsg for socket sends"), which replaced
>>>>>> kernel_send_page() calls in NFSD's socket send path with
>>>>>> calls to
>>>>>> sock_sendmsg() using iov_iter.
>>>>>>=20
>>>>>> Investigation showed that tcp_sendmsg() was not using zero-
>>>>>> copy
>>>>>> to
>>>>>> send the xdr_buf's bvec pages, but instead was relying on
>>>>>> memcpy.
>>>>>>=20
>>>>>> Set up the socket and each msghdr that bears bvec pages to
>>>>>> use
>>>>>> the
>>>>>> zero-copy mechanism in tcp_sendmsg.
>>>>>>=20
>>>>>> Reported-by: Daire Byrne <daire@dneg.com>
>>>>>> BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=3D209439
>>>>>> Fixes: da1661b93bf4 ("SUNRPC: Teach server to use
>>>>>> xprt_sock_sendmsg
>>>>>> for socket sends")
>>>>>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>>>>>> ---
>>>>>>  net/sunrpc/socklib.c  |    5 ++++-
>>>>>>  net/sunrpc/svcsock.c  |    1 +
>>>>>>  net/sunrpc/xprtsock.c |    1 +
>>>>>>  3 files changed, 6 insertions(+), 1 deletion(-)
>>>>>>=20
>>>>>> This patch does not fully resolve the issue. Daire reports
>>>>>> high
>>>>>> softIRQ activity after the patch is applied, and this
>>>>>> activity
>>>>>> seems to prevent full restoration of previous performance.
>>>>>>=20
>>>>>>=20
>>>>>> diff --git a/net/sunrpc/socklib.c b/net/sunrpc/socklib.c
>>>>>> index d52313af82bc..af47596a7bdd 100644
>>>>>> --- a/net/sunrpc/socklib.c
>>>>>> +++ b/net/sunrpc/socklib.c
>>>>>> @@ -226,9 +226,12 @@ static int xprt_send_pagedata(struct
>>>>>> socket
>>>>>> *sock, struct msghdr *msg,
>>>>>>         if (err < 0)
>>>>>>                 return err;
>>>>>> =20
>>>>>> +       msg->msg_flags |=3D MSG_ZEROCOPY;
>>>>>>         iov_iter_bvec(&msg->msg_iter, WRITE, xdr->bvec,
>>>>>> xdr_buf_pagecount(xdr),
>>>>>>                       xdr->page_len + xdr->page_base);
>>>>>> -       return xprt_sendmsg(sock, msg, base + xdr-
>>>>>>> page_base);
>>>>>> +       err =3D xprt_sendmsg(sock, msg, base + xdr->page_base);
>>>>>> +       msg->msg_flags &=3D ~MSG_ZEROCOPY;
>>>>>> +       return err;
>>>>>>  }
>>>>>> =20
>>>>>>  /* Common case:
>>>>>> diff --git a/net/sunrpc/svcsock.c b/net/sunrpc/svcsock.c
>>>>>> index c2752e2b9ce3..c814b4953b15 100644
>>>>>> --- a/net/sunrpc/svcsock.c
>>>>>> +++ b/net/sunrpc/svcsock.c
>>>>>> @@ -1176,6 +1176,7 @@ static void svc_tcp_init(struct
>>>>>> svc_sock
>>>>>> *svsk,
>>>>>> struct svc_serv *serv)
>>>>>>                 svsk->sk_datalen =3D 0;
>>>>>>                 memset(&svsk->sk_pages[0], 0, sizeof(svsk-
>>>>>>> sk_pages));
>>>>>> =20
>>>>>> +               sock_set_flag(sk, SOCK_ZEROCOPY);
>>>>>>                 tcp_sk(sk)->nonagle |=3D TCP_NAGLE_OFF;
>>>>>> =20
>>>>>>                 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
>>>>>> diff --git a/net/sunrpc/xprtsock.c b/net/sunrpc/xprtsock.c
>>>>>> index 7090bbee0ec5..343c6396b297 100644
>>>>>> --- a/net/sunrpc/xprtsock.c
>>>>>> +++ b/net/sunrpc/xprtsock.c
>>>>>> @@ -2175,6 +2175,7 @@ static int
>>>>>> xs_tcp_finish_connecting(struct
>>>>>> rpc_xprt *xprt, struct socket *sock)
>>>>>> =20
>>>>>>                 /* socket options */
>>>>>>                 sock_reset_flag(sk, SOCK_LINGER);
>>>>>> +               sock_set_flag(sk, SOCK_ZEROCOPY);
>>>>>>                 tcp_sk(sk)->nonagle |=3D TCP_NAGLE_OFF;
>>>>>> =20
>>>>>>                 xprt_clear_connected(xprt);
>>>>>>=20
>>>>>>=20
>>>>> I'm thinking we are not really allowed to do that here. The
>>>>> pages
>>>>> we
>>>>> pass in to the RPC layer are not guaranteed to contain stable
>>>>> data
>>>>> since they include unlocked page cache pages as well as
>>>>> O_DIRECT
>>>>> pages.
>>>>=20
>>>> I assume you mean the client side only. Those issues aren't a
>>>> factor
>>>> on the server. Not setting SOCK_ZEROCOPY here should be enough to
>>>> prevent the use of zero-copy on the client.
>>>>=20
>>>> However, the client loses the benefits of sending a page at a
>>>> time.
>>>> Is there a desire to remedy that somehow?
>>>=20
>>> What about splice reads on the server side?
>>=20
>> On the server, this path formerly used kernel_sendpages(), which I
>> assumed is similar to the sendmsg zero-copy mechanism. How does
>> kernel_sendpages() mitigate against page instability?
>=20
> It copies the data. =F0=9F=99=82

tcp_sendmsg_locked() invokes skb_copy_to_page_nocache(), which is
where Daire's performance-robbing memcpy occurs.

do_tcp_sendpages() has no such call site. Therefore the legacy
sendpage-based path has at least one fewer data copy operations.

What is the appropriate way to make tcp_sendmsg() treat a bvec-bearing
msghdr like an array of struct page pointers passed to kernel_sendpage() =
?


--
Chuck Lever