> On Nov 9, 2020, at 3:10 PM, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
>
>
>
> On 11/9/20 8:31 PM, Chuck Lever wrote:
>>
>>
>>> On Nov 9, 2020, at 1:16 PM, Trond Myklebust <trondmy@xxxxxxxxxxxxxxx> wrote:
>>>
>>> On Mon, 2020-11-09 at 12:36 -0500, Chuck Lever wrote:
>>>>
>>>>
>>>>> On Nov 9, 2020, at 12:32 PM, Trond Myklebust <
>>>>> trondmy@xxxxxxxxxxxxxxx> wrote:
>>>>>
>>>>> On Mon, 2020-11-09 at 12:12 -0500, Chuck Lever wrote:
>>>>>>
>>>>>>
>>>>>>> On Nov 9, 2020, at 12:08 PM, Trond Myklebust
>>>>>>> <trondmy@xxxxxxxxxxxxxxx> wrote:
>>>>>>>
>>>>>>> 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.
>>>>>>>>
>>>>>>>> Investigation showed that tcp_sendmsg() was not using zero-
>>>>>>>> copy
>>>>>>>> to
>>>>>>>> send the xdr_buf's bvec pages, but instead was relying on
>>>>>>>> memcpy.
>>>>>>>>
>>>>>>>> Set up the socket and each msghdr that bears bvec pages to
>>>>>>>> use
>>>>>>>> the
>>>>>>>> zero-copy mechanism in tcp_sendmsg.
>>>>>>>>
>>>>>>>> Reported-by: Daire Byrne <daire@xxxxxxxx>
>>>>>>>> BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=209439
>>>>>>>> Fixes: da1661b93bf4 ("SUNRPC: Teach server to use
>>>>>>>> xprt_sock_sendmsg
>>>>>>>> for socket sends")
>>>>>>>> Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
>>>>>>>> ---
>>>>>>>> net/sunrpc/socklib.c | 5 ++++-
>>>>>>>> net/sunrpc/svcsock.c | 1 +
>>>>>>>> net/sunrpc/xprtsock.c | 1 +
>>>>>>>> 3 files changed, 6 insertions(+), 1 deletion(-)
>>>>>>>>
>>>>>>>> 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.
>>>>>>>>
>>>>>>>>
>>>>>>>> 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;
>>>>>>>>
>>>>>>>> + msg->msg_flags |= 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 = xprt_sendmsg(sock, msg, base + xdr->page_base);
>>>>>>>> + msg->msg_flags &= ~MSG_ZEROCOPY;
>>>>>>>> + return err;
>>>>>>>> }
>>>>>>>>
>>>>>>>> /* 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 = 0;
>>>>>>>> memset(&svsk->sk_pages[0], 0, sizeof(svsk-
>>>>>>>>> sk_pages));
>>>>>>>>
>>>>>>>> + sock_set_flag(sk, SOCK_ZEROCOPY);
>>>>>>>> tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
>>>>>>>>
>>>>>>>> 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)
>>>>>>>>
>>>>>>>> /* socket options */
>>>>>>>> sock_reset_flag(sk, SOCK_LINGER);
>>>>>>>> + sock_set_flag(sk, SOCK_ZEROCOPY);
>>>>>>>> tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
>>>>>>>>
>>>>>>>> xprt_clear_connected(xprt);
>>>>>>>>
>>>>>>>>
>>>>>>> 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.
>>>>>>
>>>>>> 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.
>>>>>>
>>>>>> However, the client loses the benefits of sending a page at a
>>>>>> time.
>>>>>> Is there a desire to remedy that somehow?
>>>>>
>>>>> What about splice reads on the server side?
>>>>
>>>> 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?
>>>
>>> It copies the data. 🙂
>>
>> 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() ?
>>
>
>
> MSG_ZEROCOPY is only accepted if sock_flag(sk, SOCK_ZEROCOPY) is true,
> ie if SO_ZEROCOPY socket option has been set earlier.
The patch does set both SO_ZEROCOPY, and MSG_ZEROCOPY when appropriate.
--
Chuck Lever
