On Mon, Jul 27, 2020 at 10:44:59AM -0700, Eric Dumazet wrote:
> On Mon, Jul 27, 2020 at 10:35 AM Jonathan Lemon
> <jonathan.le...@gmail.com> wrote:
> >
> > On Mon, Jul 27, 2020 at 09:09:48AM -0700, Eric Dumazet wrote:
> > > On Mon, Jul 27, 2020 at 8:56 AM Jonathan Lemon <jonathan.le...@gmail.com>
> > > wrote:
> > > >
> > > > On Mon, Jul 27, 2020 at 08:19:43AM -0700, Eric Dumazet wrote:
> > > > > On Mon, Jul 27, 2020 at 12:51 AM Jonathan Lemon
> > > > > <jonathan.le...@gmail.com> wrote:
> > > > > >
> > > > > > This flag indicates that the attached data is a zero-copy send,
> > > > > > and the pages should be retrieved from the netgpu module. The
> > > > > > socket should should already have been attached to a netgpu queue.
> > > > > >
> > > > > > Signed-off-by: Jonathan Lemon <jonathan.le...@gmail.com>
> > > > > > ---
> > > > > > include/linux/socket.h | 1 +
> > > > > > net/ipv4/tcp.c | 8 ++++++++
> > > > > > 2 files changed, 9 insertions(+)
> > > > > >
> > > > > > diff --git a/include/linux/socket.h b/include/linux/socket.h
> > > > > > index 04d2bc97f497..63816cc25dee 100644
> > > > > > --- a/include/linux/socket.h
> > > > > > +++ b/include/linux/socket.h
> > > > > > @@ -310,6 +310,7 @@ struct ucred {
> > > > > > */
> > > > > >
> > > > > > #define MSG_ZEROCOPY 0x4000000 /* Use user data in kernel
> > > > > > path */
> > > > > > +#define MSG_NETDMA 0x8000000
> > > > > > #define MSG_FASTOPEN 0x20000000 /* Send data in TCP SYN */
> > > > > > #define MSG_CMSG_CLOEXEC 0x40000000 /* Set close_on_exec for
> > > > > > file
> > > > > > descriptor received
> > > > > > through
> > > > > > diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> > > > > > index 261c28ccc8f6..340ce319edc9 100644
> > > > > > --- a/net/ipv4/tcp.c
> > > > > > +++ b/net/ipv4/tcp.c
> > > > > > @@ -1214,6 +1214,14 @@ int tcp_sendmsg_locked(struct sock *sk,
> > > > > > struct msghdr *msg, size_t size)
> > > > > > uarg->zerocopy = 0;
> > > > > > }
> > > > > >
> > > > > > + if (flags & MSG_NETDMA && size && sock_flag(sk,
> > > > > > SOCK_ZEROCOPY)) {
> > > > > > + zc = sk->sk_route_caps & NETIF_F_SG;
> > > > > > + if (!zc) {
> > > > > > + err = -EFAULT;
> > > > > > + goto out_err;
> > > > > > + }
> > > > > > + }
> > > > > >
> > > > >
> > > > > Sorry, no, we can not allow adding yet another branch into TCP fast
> > > > > path for yet another variant of zero copy.
> > > >
> > > > I'm not in disagreement with that statement, but the existing zerocopy
> > > > work makes some assumptions that aren't suitable. I take it that you'd
> > > > rather have things folded together so the old/new code works together?
> > >
> > > Exact. Forcing users to use MSG_NETDMA, yet reusing SOCK_ZEROCOPY is
> > > silly.
> > >
> > > SOCK_ZEROCOPY has been added to that user space and kernel would agree
> > > on MSG_ZEROCOPY being not a nop (as it was on old kernels)
> > >
> > > >
> > > > Allocating an extra structure for every skbuff isn't ideal in my book.
> > > >
> > >
> > > We do not allocate a structure for every skbuff. Please look again.
> >
> > I'm looking here:
> >
> > uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
> >
> > Doesn't sock_zerocopy_realloc() allocate a new structure if the skb
> > doesn't have one already?
> >
> >
> > > > > Overall, I think your patch series desperately tries to add changes in
> > > > > TCP stack, while there is yet no proof
> > > > > that you have to use TCP transport between the peers.
> > > >
> > > > The goal is having a reliable transport without resorting to RDMA.
> > >
> > > And why should it be TCP ?
> > >
> > > Are you dealing with lost packets, retransmits, timers, and al ?
> >
> > Yes? If there was a true lossless medium, RDMA would have taken over by
> > now. Or are you suggesting that the transport protocol reliability
> > should be performed in userspace? (not all the world is QUIC yet)
> >
>
> The thing is : this patch series is a monster thing adding stuff that
> is going to impact 100% % of TCP flows,
> even if not used in this NETDMA context.
>
> So you need to convince us you are really desperate to get this in
> upstream linux.
>
> I have implemented TCP RX zero copy without adding a single line in
> standard TCP code.
That's a bit of an exaggeration, as I see skb_zcopy_*() calls scattered
around the normal TCP code path. I also haven't changed the normal TCP
path either, other than doing some of the same things as skb_zcopy_*().
(ignoring the ugly moron about padding out the TCP header, which I'll
put under a static_branch_unlikely).
The thing is, the existing zero copy code is zero-copy to /host/ memory,
which is not the same thing as zero-copy to other memory areas.
--
Jonathan
