Re: [dpdk-users] Why is udp packet, sent by dpdk, received with zero'd payload?

[Cliff Burdick <shaklee3@gmail.com>]

Another option is to use rte_pktmbuf_chain(), and chain the buffers
together. That sounds more like what you were intending to do in the first
place. If you do that, make sure your driver supports it (mlx5 does), and
set the data/pkt_len appropriately for each part of the chain. This also is
zero-copy.

On Wed, Jul 1, 2020 at 9:48 AM David Aldrich <david.aldrich.ntml@gmail.com>
wrote:

> Hi Matt
> Thanks very much for your help. I was about to write out in more detail
> what our code does but I've just noticed that it uses a zero-copy method to
> attach a data buffer to the mbuf using a call to:
>
> rte_pktmbuf_attach()
>
> The docs says for this function: "Right now, not supported:"
>
> So maybe it doesn't work. Anyway, I need to revisit this code and take the
> simpler approach of just copying the payload into the mbuf. Maybe I can get
> the zero copy method working later. I'll see how this goes and come back if
> I need more help.
>
> David
>
> On Wed, Jul 1, 2020 at 4:56 PM Matt Laswell <laswell@infinite.io> wrote:
>
> > OK, this makes sense.  Understand that the mbuf data structure in DPDK
> > contains an offset that tells other consumers of the mbuf where to find
> > content.  When you call rte_pktmbuf_prepend(), you're adjusting this
> > offset, in this case moving it forward by 42 bytes.
> >
> > Based on what I've seen you say here and on Stack Overflow, I think
> > perhaps your code is doing roughly this:
> > 1. You have an MBuf that contains your payload
> > 2. You create a second MBuf that contains your headers, and link it to
> the
> > payload mbuf via the ->next pointer and nb_segs
> > 3. But you then call rte_pktmbuf_prepend() on the payload MBuf, making
> > space in the payload MBuf for headers.
> > 4. You send the header payload, implicitly also sending the payload one
> > (since they're linked)
> >
> > If that's an accurate description of what you're doing, step 3 is the
> > problem.  You're telling the transmitting NIC to start pulling 15 bytes
> of
> > payload from a spot that's 42 bytes ahead of where the payload actually
> > lives.  Hence, you get whatever happens to be in the buffer at that
> point.
> > In this case, it's zeroes.  It could be whatever garbage was lying around
> > from the last user of the MBuf.
> >
> > Let me know if that's an accurate description of what your application is
> > doing, or if I've misunderstood.
> >
> > On Wed, Jul 1, 2020 at 10:07 AM David Aldrich <
> > david.aldrich.ntml@gmail.com> wrote:
> >
> >> >If you look at the code for  rte_pktmbuf_prepend  it appears to be just
> >> >incrementing data_len and pkt_len by the same amount.
> >>
> >> I'm sorry, but I don't understand. My calls to rte_pktmbuf_prepend are:
> >>
> >>
> >> p_udp_hdr = (struct udp_hdr*)rte_pktmbuf_prepend(ptMbuf,
> >> (uint16_t)sizeof(struct udp_hdr));
> >> p_ip_hdr  = (struct ipv4_hdr*)rte_pktmbuf_prepend(ptMbuf,
> >> (uint16_t)sizeof(struct ipv4_hdr));
> >>
> >> are you saying that those calls are wrong?  When you say 'if you look
> >> at the code' are you referring to this code?
> >>
> >> (I thought it best that I should ask for clarification rather than
> guess).
> >>
> >> Thanks
> >>
> >>
> >> On Wed, Jul 1, 2020 at 3:59 PM Cliff Burdick <shaklee3@gmail.com>
> wrote:
> >>
> >> > If you look at the code for  rte_pktmbuf_prepend  it appears to be
> just
> >> > incrementing data_len and pkt_len by the same amount. My guess is that
> >> > those fields were not set to the correct values before you called
> >> > rte_pktmbuf_prepend().
> >> >
> >> > On Wed, Jul 1, 2020 at 6:41 AM David Aldrich <
> >> david.aldrich.ntml@gmail.com>
> >> > wrote:
> >> >
> >> >> I think this is where my lack of understanding is my problem.
> Looking
> >> at
> >> >> the mbuf dump:
> >> >>
> >> >> dump mbuf at 0x53e4e92c0, iova=73e4e9340, buf_len=2176
> >> >>   pkt_len=57, ol_flags=c0000000000000, nb_segs=2, in_port=65535
> >> >>   segment at 0x53e4e92c0, data=0x53e4e9396, data_len=42
> >> >>   Dump data at [0x53e4e9396], len=42
> >> >> 00000000: 94 40 C9 1F C4 DB 94 40 C9 1E 13 7D 08 00 45 B8 |
> >> .@.....@...}..E.
> >> >> 00000010: 00 2B 00 00 00 00 40 11 F5 E8 C0 A8 01 69 C0 A8 |
> >> .+....@......i..
> >> >> 00000020: 01 68 0B B9 0B B8 00 17 64 2D |  |  |  |  |  |  |
> .h......d-
> >> >>   segment at 0x53e146e40, data=0x53e4fbbc0, data_len=15
> >> >>   Dump data at [0x53e4fbbc0], len=15
> >> >> 00000000: 48 65 6C 6C 6F 20 66 72 6F 6D 20 64 70 64 6B |  | Hello
> from
> >> dpdk
> >> >>
> >> >> It consists of two parts (segments?), of lengths 42 and 15. This
> makes
> >> sense to me as I first
> >> >> put the payload in the mbuf (15 bytes) and then added the Ethernet
> and
> >> L3 headers (42 bytes) by calling rte_pktmbuf_prepend().
> >> >>
> >> >> I guess only the first segment is getting transmitted?
> >> >>
> >> >>
> >> >> On Wed, Jul 1, 2020 at 1:57 PM Cliff Burdick <shaklee3@gmail.com>
> >> wrote:
> >> >>
> >> >>> Are you setting data_len and packet_len in the mbuf before sending?
> >> >>>
> >> >>>
> >> >>> On Wed, Jul 1, 2020, 03:23 David Aldrich <
> >> david.aldrich.ntml@gmail.com>
> >> >>> wrote:
> >> >>>
> >> >>>>  Hi,
> >> >>>>
> >> >>>> I have a problem transmitting udp packets with
> dpdk-stable-18.11.8. I
> >> >>>> have
> >> >>>> posted a question about it on stackoverflow:
> >> >>>>
> >> >>>>
> >> >>>>
> >>
> https://stackoverflow.com/questions/62674596/why-is-udp-packet-sent-by-dpdk-received-with-zerod-payload
> >> >>>>
> >> >>>> I would be grateful for any comments either there or in this group
> >> >>>> please.
> >> >>>>
> >> >>>> Best regards
> >> >>>>
> >> >>>> David
> >> >>>>
> >> >>>
> >>
> >
>