Re: [PATCH v2 1/2] udp: UDP socket send queue repair

[Willem de Bruijn]

On Fri, Aug 13, 2021 at 4:52 AM Bui Quang Minh <minhquangbui99@xxxxxxxxx> wrote:
>
>
>
> On 8/12/2021 10:51 PM, Eric Dumazet wrote:
> >
> >
> > On 8/12/21 3:46 PM, Bui Quang Minh wrote:
> >>
> >>
> >> On 8/11/2021 11:14 PM, Eric Dumazet wrote:
> >>>
> >>>
> >>> On 8/11/21 5:45 PM, Bui Quang Minh wrote:
> >>>> In this patch, I implement UDP_REPAIR sockoption and a new path in
> >>>> udp_recvmsg for dumping the corked packet in UDP socket's send queue.
> >>>>
> >>>> A userspace program can use recvmsg syscall to get the packet's data and
> >>>> the msg_name information of the packet. Currently, other related
> >>>> information in inet_cork that are set in cmsg are not dumped.
> >>>>
> >>>> While working on this, I was aware of Lese Doru Calin's patch and got some
> >>>> ideas from it.
> >>>
> >>>
> >>> What is the use case for this feature, adding a test in UDP fast path ?
> >>
> >> This feature is used to help CRIU to dump CORKed UDP packet in send queue. I'm sorry for being not aware of the performance perspective here.
> >
> > UDP is not reliable.
> >
> > I find a bit strange we add so many lines of code
> > for a feature trying very hard to to drop _one_ packet.
> >
> > I think a much better changelog would be welcomed.
>
> The reason we want to dump the packet in send queue is to make to state of the
> application consistent. The scenario is that when an application sends UDP
> packets via UDP_CORK socket or with MSG_MORE, CRIU comes and checkpoints the
> application. If we drop the data in send queue, when application restores, it
> sends some more data then turns off the cork and actually sends a packet. The
> receiving side may get that packet but it's unusual that the first part of that
> packet is missing because we drop it. So we try to solve this problem with some
> help from the Linux kernel.

Instead of checkpointing the state, how about making the kernel drop
the next packet.

For instance by setting up->pending to something else than AF_UNSPEC,
AF_INET, AF_INET6 from a new setsockopt and testing for this case in
the udp_sendmsg up->pending slowpath.

udp_sendmsg already calls udp_v6_flush_pending_frames on error when
appending to a pending packet, so returning an error on the next call
after restore and have that imply a flush is acceptable. I would
introduce a new error code.

The state can perhaps be inferred in other ways, e.g., from
up->pending && !up->len or up->pending && !skb_peek_tail(queue). But
an explicit up->pending mode will be easier to grasp.