Re: UDP implementation and the MSG_MORE flag

[Oliver Graute <oliver.graute@gmail.com>]

On 27/01/21, Willem de Bruijn wrote:
> On Wed, Jan 27, 2021 at 9:53 PM Willem de Bruijn
> <willemdebruijn.kernel@gmail.com> wrote:
> >
> > On Tue, Jan 26, 2021 at 10:25 PM Willem de Bruijn
> > <willemdebruijn.kernel@gmail.com> wrote:
> > >
> > > On Tue, Jan 26, 2021 at 5:00 PM Willem de Bruijn
> > > <willemdebruijn.kernel@gmail.com> wrote:
> > > >
> > > > On Tue, Jan 26, 2021 at 4:54 PM Willem de Bruijn
> > > > <willemdebruijn.kernel@gmail.com> wrote:
> > > > >
> > > > > On Tue, Jan 26, 2021 at 9:58 AM Oliver Graute <oliver.graute@gmail.com> wrote:
> > > > > >
> > > > > > Hello,
> > > > > >
> > > > > > we observe some unexpected behavior in the UDP implementation of the
> > > > > > linux kernel.
> > > > > >
> > > > > > Some UDP packets send via the loopback interface are dropped in the
> > > > > > kernel on the receive side when using sendto with the MSG_MORE flag.
> > > > > > Every drop increases the InCsumErrors in /proc/self/net/snmp. Some
> > > > > > example code to reproduce it is appended below.
> > > > > >
> > > > > > In the code we tracked it down to this code section. ( Even a little
> > > > > > further but its unclear to me wy the csum() is wrong in the bad case)
> > > > > >
> > > > > > udpv6_recvmsg()
> > > > > > ...
> > > > > > if (checksum_valid || udp_skb_csum_unnecessary(skb)) {
> > > > > >                 if (udp_skb_is_linear(skb))
> > > > > >                         err = copy_linear_skb(skb, copied, off, &msg->msg_iter);
> > > > > >                 else
> > > > > >                         err = skb_copy_datagram_msg(skb, off, msg, copied);
> > > > > >         } else {
> > > > > >                 err = skb_copy_and_csum_datagram_msg(skb, off, msg);
> > > > > >                 if (err == -EINVAL) {
> > > > > >                         goto csum_copy_err;
> > > > > >                 }
> > > > > >         }
> > > > > > ...
> > > > > >
> > > > >
> > > > > Thanks for the report with a full reproducer.
> > > > >
> > > > > I don't have a full answer yet, but can reproduce this easily.
> > > > >
> > > > > The third program, without MSG_MORE, builds an skb with
> > > > > CHECKSUM_PARTIAL in __ip_append_data. When looped to the receive path
> > > > > that ip_summed means no additional validation is needed. As encoded in
> > > > > skb_csum_unnecessary.
> > > > >
> > > > > The first and second programs are essentially the same, bar for a
> > > > > slight difference in length. In both cases packet length is very short
> > > > > compared to the loopback device MTU. Because of MSG_MORE, these
> > > > > packets have CHECKSUM_NONE.
> > > > >
> > > > > On receive in
> > > > >
> > > > >   __udp4_lib_rcv()
> > > > >     udp4_csum_init()
> > > > >       err = skb_checksum_init_zero_check()
> > > > >
> > > > > The second program validates and sets ip_summed = CHECKSUM_COMPLETE
> > > > > and csum_valid = 1.
> > > > > The first does not, though err == 0.
> > > > >
> > > > > This appears to succeed consistently for packets <= 68B of payload,
> > > > > fail consistently otherwise. It is not clear to me yet what causes
> > > > > this distinction.
> > > >
> > > > This is from
> > > >
> > > > "
> > > > /* For small packets <= CHECKSUM_BREAK perform checksum complete directly
> > > >  * in checksum_init.
> > > >  */
> > > > #define CHECKSUM_BREAK 76
> > > > "
> > > >
> > > > So the small packet gets checksummed immediately in
> > > > __skb_checksum_validate_complete, but the larger one does not.
> > > >
> > > > Question is why the copy_and_checksum you pointed to seems to fail checksum.
> > >
> > > Manually calling __skb_checksum_complete(skb) in
> > > skb_copy_and_csum_datagram_msg succeeds, so it is the
> > > skb_copy_and_csum_datagram that returns an incorrect csum.
> > >
> > > Bisection shows that this is a regression in 5.0, between
> > >
> > > 65d69e2505bb datagram: introduce skb_copy_and_hash_datagram_iter helper (fail)
> > > d05f443554b3 iov_iter: introduce hash_and_copy_to_iter helper
> > > 950fcaecd5cc datagram: consolidate datagram copy to iter helpers
> > > cb002d074dab iov_iter: pass void csum pointer to csum_and_copy_to_iter (pass)
> > >
> > > That's a significant amount of code change. I'll take a closer look,
> > > but checkpointing state for now..
> >
> > Key difference is the csum_block_add when handling frags, and the
> > removal of temporary csum2.
> >
> > In the reproducer, there is one 13B csum_and_copy_to_iter from
> > skb->data + offset, followed by a 73B csum_and_copy_to_iter from the
> > first frag. So the second one passes pos 13 to csum_block_add.
> >
> > The original implementation of skb_copy_and_csum_datagram similarly
> > fails the test, if we fail to account for the position
> >
> > -                       *csump = csum_block_add(*csump, csum2, pos);
> > +                       *csump = csum_block_add(*csump, csum2, 0);
> 
> One possible approach:

very thx for your analysis and your patch proposal. After a first quick
test this patch proposal solves the problem.

Best regards,

Oliver