Re: [PATCH AUTOSEL 5.10 03/45] SUNRPC: Handle TCP socket sends with kernel_sendpage() again

[Chuck Lever <chuck.lever@oracle.com>]

> On Feb 8, 2021, at 3:12 PM, Trond Myklebust <trondmy@hammerspace.com> wrote:
> 
> On Mon, 2021-02-08 at 19:48 +0000, Chuck Lever wrote:
>> 
>> 
>>> On Feb 8, 2021, at 2:34 PM, Trond Myklebust <
>>> trondmy@hammerspace.com> wrote:
>>> 
>>> On Tue, 2021-01-19 at 20:25 -0500, Sasha Levin wrote:
>>>> From: Chuck Lever <chuck.lever@oracle.com>
>>>> 
>>>> [ Upstream commit 4a85a6a3320b4a622315d2e0ea91a1d2b013bce4 ]
>>>> 
>>>> 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.
>>>> This means copying every byte of a large NFS READ payload.
>>>> 
>>>> It looks like TLS sockets do indeed support a ->sendpage method,
>>>> so it's really not necessary to use xprt_sock_sendmsg() to
>>>> support
>>>> TLS fully on the server. A mechanical reversion of da1661b93bf4
>>>> is
>>>> not possible at this point, but we can re-implement the server's
>>>> TCP socket sendmsg path using kernel_sendpage().
>>>> 
>>>> Reported-by: Daire Byrne <daire@dneg.com>
>>>> BugLink: https://bugzilla.kernel.org/show_bug.cgi?id=209439
>>>> Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
>>>> Signed-off-by: Sasha Levin <sashal@kernel.org>
>>>> ---
>>>>  net/sunrpc/svcsock.c | 86
>>>> +++++++++++++++++++++++++++++++++++++++++++-
>>>>  1 file changed, 85 insertions(+), 1 deletion(-)
>>>> 
>>>> diff --git a/net/sunrpc/svcsock.c b/net/sunrpc/svcsock.c
>>>> index c2752e2b9ce34..4404c491eb388 100644
>>>> --- a/net/sunrpc/svcsock.c
>>>> +++ b/net/sunrpc/svcsock.c
>>>> @@ -1062,6 +1062,90 @@ static int svc_tcp_recvfrom(struct
>>>> svc_rqst
>>>> *rqstp)
>>>>         return 0;       /* record not complete */
>>>>  }
>>>>  
>>>> +static int svc_tcp_send_kvec(struct socket *sock, const struct
>>>> kvec
>>>> *vec,
>>>> +                             int flags)
>>>> +{
>>>> +       return kernel_sendpage(sock, virt_to_page(vec->iov_base),
>>>> +                              offset_in_page(vec->iov_base),
>>>> +                              vec->iov_len, flags);
>> 
>> Thanks for your review!
>> 
>>> I'm having trouble with this line. This looks like it is trying to
>>> push
>>> a slab page into kernel_sendpage().
>> 
>> The head and tail kvec's in rq_res are not kmalloc'd, they are
>> backed by pages in rqstp->rq_pages[].
>> 
>> 
>>> What guarantees that the nfsd
>>> thread won't call kfree() before the socket layer is done
>>> transmitting
>>> the page?
>> 
>> If I understand correctly what Neil told us last week, the page
>> reference count on those pages is set up so that one of
>> svc_xprt_release() or the network layer does the final put_page(),
>> in a safe fashion.
>> 
>> Before da1661b93bf4 ("SUNRPC: Teach server to use xprt_sock_sendmsg
>> for socket sends"), the original svc_send_common() code did this:
>> 
>> -       /* send head */
>> -       if (slen == xdr->head[0].iov_len)
>> -               flags = 0;
>> -       len = kernel_sendpage(sock, headpage, headoffset,
>> -                                 xdr->head[0].iov_len, flags);
>> -       if (len != xdr->head[0].iov_len)
>> -               goto out;
>> -       slen -= xdr->head[0].iov_len;
>> -       if (slen == 0)
>> -               goto out;
>> 
>> 
>> 
> 
> OK, so then only the argument kvec can be allocated on the slab (thanks
> to  svc_deferred_recv)? Is that correct?

The RPC/RDMA Receive buffer is kmalloc'd, that would be used for
rq_arg.head/tail. But for TCP, I believe the head kvec is always
pulled out of rq_pages[].

svc_process() sets up rq_res.head this way:

1508         resv->iov_base = page_address(rqstp->rq_respages[0]);
1509         resv->iov_len = 0;

I would need to audit the code to confirm that rq_res.tail is never
kmalloc'd.


--
Chuck Lever