I found the key difference: when TCP timestamps (RFC 7323) are enabled on the
TCP sender, the RX HW timestamp of LROed packets on the DPDK PMD on middlebox
machine becomes inconsistent or invalid. Is there a known limitation or erratum
in the mlx5 driver or CX7 firmware regarding this?
Sincerely,
Junghan Yoon
On Jul 24, 2025, 12:26 AM +0900, Ivan Malov <ivan.ma...@arknetworks.am>, wrote:
> On Thu, 24 Jul 2025, Yoon Junghan wrote:
>
> > I'm not sure I did well. All interface show the same result.
> >
> > current settings:
> > tx_type 0
> > rx_filter 0
>
> But that should mean.. no timestamping?
>
> 1) May be also check 'sudo ethtool -T <ifname>'.
> 2) May be try to enable 'sudo hwstamp_ctl -i <ifname> -r 1'.
>
> Thank you.
>
> >
> > Sincerely,
> > Junghan Yoon
> > On 2025년 7월 23일 PM 11:19 +0900, Ivan Malov <ivan.ma...@arknetworks.am>,
> > wrote:
> > On Wed, 23 Jul 2025, Yoon Junghan wrote:
> >
> > I isolated port 1 using -a option for EAL parameter and got the similar
> > result.
> >
> > Note that port 1 becomes port 0 in this time.
> > [port 0] RX HW timestamp: 0x3eac4214bc574368 (LROed)
> > [port 0] RX HW timestamp: 0x3eac4214bc574368 (LROed)
> > [port 0] RX HW timestamp: 0x3eac4214bc574368 (LROed)
> > [port 0] RX HW timestamp: 0x3eac4214bc574368 (LROed)
> > [port 0] RX HW timestamp: 0x00042819272fad (not LROed)
> > [port 0] RX HW timestamp: 0x000428192e6e77 (not LROed)
> > [port 0] RX HW timestamp: 0x000428192e7f01 (not LROed)
> > [port 0] RX HW timestamp: 0x000428192e833d (not LROed)
> >
> > FYI, I have 4 CX-7 on the same machine. (eth0 = port 0, ... eth3 = port 3
> > in DPDK)
> > pci@0000:16:00.0 eth0 network MT2910 Family [ConnectX-7]
> > pci@0000:40:00.0 eth1 network MT2910 Family [ConnectX-7]
> > pci@0000:6a:00.0 eth2 network MT2910 Family [ConnectX-7]
> > pci@0000:94:00.0 eth3 network MT2910 Family [ConnectX-7]
> >
> > Among them, only the first CX-7 shows consistent timestamp regardless of
> > LRO.
> >
> >
> > Does 'sudo hwstamp_ctl -i <ifname>' show consistent results across all the
> > NICs?
> >
> > Thank you.
> >
> >
> > Sincerely,
> > Junghan Yoon
> > On 2025년 7월 23일 PM 10:28 +0900, Ivan Malov <ivan.ma...@arknetworks.am>,
> > wrote:
> > On Wed, 23 Jul 2025, Yoon Junghan wrote:
> >
> > Thank you for quick response.
> >
> > 1) They are different NICs. Not in the same board. Separate adapters in
> > different PCIe slots.
> > 2) My DPDK app uses 4 separate ports; port 0, port 1, port 2, and port 3.
> > They are all on different boards. Thus, they are running at the same time.
> >
> >
> > Excellent. I apologise for one more dumb question, but does isolating the
> > very
> > specific NIC (so that DPDK does not grab the other ones) that is known to
> > give
> > strange timestamps, result in the same/unexpected behaviour? Just to make
> > sure.
> >
> > Thank you.
> >
> >
> > Sincerely,
> > Junghan Yoon
> > On 2025년 7월 23일 PM 10:09 +0900, Ivan Malov <ivan.ma...@arknetworks.am>,
> > wrote:
> >
> > Hello,
> >
> > On Wed, 23 Jul 2025, Yoon Junghan wrote:
> >
> >
> > Hello,
> > As advised, I tested hardware timestamps with LRO enabled on our ConnectX-7
> > NICs. However, the timestamps of LROed packets still show inconsistent and
> > abnormally
> > large
> > gaps from normal
> > packets.
> >
> > Interestingly, I found this issue does not appear on all CX-7 NICs. Even
> > with identical DPDK code, firmware version (28.43.2566), and hardware
> > models from the
> > same
> > manufacturer, only
> > specific NICs exhibit this inconsistency.
> > I have confirmed that:
> > * All NICs use the same driver and firmware version.
> > * All NICs are of the same model (MCX75310AAS-NEA_Ax).
> >
> >
> >
> > 1) Do the two "NICs" ('port 0' and 'port 1' from below printout) represent
> > two
> > different ports/PFs of the same physical 'board'/'adapter card' in fact?
> >
> > 2) If (1) is true, were the results obtained by running the application on
> > both
> > ports simultaneously (both managed by the DPDK at the same time)?
> >
> > (just to clarify, -- I'm confused by the fact that the NIC driver itself
> > seems
> > to invoke 'rte_mbuf_dyn_rx_timestamp_register' for each new RxQ rather than
> > call
> > it once and then look-up and reuse the existing offsets for more
> > ports/queue ).
> >
> > Thank you.
> >
> >
> > * The issue occurs only when LRO is enabled together with RX hardware
> > timestamping.
> > * Disabling LRO eliminates the issue.
> > I would appreciate any insight into how this behavior can occur on only
> > some ports despite same software and hardware setup.
> >
> > Below is my code snippet.
> >
> > ```c
> > /*----------------------------------------------------------------------------*/
> > static inline int
> > is_timestamp_enabled(struct rte_mbuf *mbuf)
> > {
> > static uint64_t timestamp_rx_dynflag = 0;
> > int timestamp_rx_dynflag_offset;
> >
> > if (!timestamp_rx_dynflag)
> > {
> > timestamp_rx_dynflag_offset =
> > rte_mbuf_dynflag_lookup(RTE_MBUF_DYNFLAG_RX_TIMESTAMP_NAME,
> > NULL);
> > if (timestamp_rx_dynflag_offset < 0)
> > {
> > return 0;
> > }
> > timestamp_rx_dynflag = RTE_BIT64(timestamp_rx_dynflag_offset);
> > }
> >
> > return mbuf->ol_flags & timestamp_rx_dynflag;
> > }
> > /*----------------------------------------------------------------------------*/
> > static inline rte_mbuf_timestamp_t *
> > get_timestamp(struct rte_mbuf *mbuf)
> > {
> > static int timestamp_dynfield_offset = -1;
> >
> > if (timestamp_dynfield_offset < 0)
> > {
> > timestamp_dynfield_offset =
> > rte_mbuf_dynfield_lookup(RTE_MBUF_DYNFIELD_TIMESTAMP_NAME, NULL);
> > if (timestamp_dynfield_offset < 0)
> > {
> > return 0;
> > }
> > }
> >
> > return RTE_MBUF_DYNFIELD(mbuf,
> > timestamp_dynfield_offset,
> > rte_mbuf_timestamp_t *);
> > }
> > /*----------------------------------------------------------------------------*/
> > static inline rte_mbuf_timestamp_t *
> > get_rx_hw_timestamp(struct rte_mbuf *pkt)
> > {
> > if (!is_timestamp_enabled(pkt))
> > {
> > printf("rx_hw_timestamp not enabled in mbuf!\n");
> > return NULL;
> > }
> >
> > return get_timestamp(pkt);
> > }
> > ```
> >
> > My DPDK application prints logs as below.
> >
> > ```c
> > /* parse HW timestamp */
> > rte_mbuf_timestamp_t *rx_timestamp = get_rx_hw_timestamp(pkt);
> > printf("[port %d] RX HW timestamp: %#016lx %s\n",
> > pctx->port_id,
> > *rx_timestamp,
> > pkt->ol_flags & PKT_RX_LRO ? "(LROed)" : "(not LROed)");
> > ```
> >
> > Below are observations from two CX-7 ports under identical conditions.
> >
> > Normal NIC (port 0):
> > [port 0] RX HW timestamp: 0x00007dcd2d185b (LROed)
> > [port 0] RX HW timestamp: 0x00007dcd2d1911 (LROed)
> > [port 0] RX HW timestamp: 0x00007dcd2d19c9 (LROed)
> > [port 0] RX HW timestamp: 0x00007dcd2d37ca (LROed)
> > [port 0] RX HW timestamp: 0x00007dcd2d4cb3 (not LROed)
> > [port 0] RX HW timestamp: 0x00007dcd2d4cb3 (not LROed)
> > [port 0] RX HW timestamp: 0x00007dcd30e019 (not LROed)
> > [port 0] RX HW timestamp: 0x00007dcd3280bb (not LROed)
> >
> > Erroneous NIC (port 1):
> > Below is erroneous NIC's timestamp.
> > [port 1] RX HW timestamp: 0x3e6eef91bc19f0fd (LROed)
> > [port 1] RX HW timestamp: 0x3e6eef91bc19f0fd (LROed)
> > [port 1] RX HW timestamp: 0x3e6eef91bc19f0fd (LROed)
> > [port 1] RX HW timestamp: 0x3e6eef91bc19f0fd (LROed)
> > [port 1] RX HW timestamp: 0x000080691b7557 (not LROed)
> > [port 1] RX HW timestamp: 0x000080691e2311 (not LROed)
> > [port 1] RX HW timestamp: 0x00008069357553 (not LROed)
> > [port 1] RX HW timestamp: 0x0000806936e8c1 (not LROed)
> >
> > As shown above, non-LRO packets consistently have normal hardware
> > timestamps on both NICs. However, on port 1, all LROed packets return a
> > fixed, invalid timestamp
> > (0x3e6eef91bc19f0fd),
> > which is clearly inconsistent.
> > I have also confirmed that other dynfields (rather than dynfield[1] and
> > dynfield[2]) are unused.
> >
> >
> > Sincerely,
> > Junghan Yoon
> > On Jul 22, 2025, 5:31 PM +0900, Ivan Malov <ivan.ma...@arknetworks.am>,
> > wrote:
> > Hello,
> >
> > On Tue, 22 Jul 2025, Yoon Junghan wrote:
> >
> > Hello,
> >
> > I'm currently using DPDK 20.11 with a ConnectX-7 NIC, and I'm trying to
> > retrieve RX hardware timestamps using
> > `rte_mbuf_dyn_rx_timestamp_register()`.
> >
> >
> > Does the application invoke 'rte_mbuf_dyn_rx_timestamp_register' on its
> > own? If
> > yes, consider to replace this with invocations of APIs [1] (with field name
> > [2])
> > and [3] (with flag name [4]). For an example, please refer to [5] and [6].
> >
> > This is because, as per [7], the driver in question might 'register' the
> > field
> > and the flag on its own, in response to 'DEV_RX_OFFLOAD_TIMESTAMP' request,
> > so,
> > the user application should look up the field/flag, not 'register' it
> > afresh.
> >
> > If this does not help, then consider to clarify whether the timestamps are
> > accurate (and whether the flag is seen in the mbufs) when LRO is not
> > enabled.
> >
> > [1]
> > https://doc.dpdk.org/api-20.11/rte__mbuf__dyn_8h.html#a6adf9b352a83e7d521fd6aa04e305b1c
> > [2]
> > https://doc.dpdk.org/api-20.11/rte__mbuf__dyn_8h.html#a5159b2d34fa801d171ed0ccce451121b
> > [3]
> > https://doc.dpdk.org/api-20.11/rte__mbuf__dyn_8h.html#a89d835027034f76a27eb2afe7987ae35
> > [4]
> > https://doc.dpdk.org/api-20.11/rte__mbuf__dyn_8h.html#a831d7066c7193788351797a65186848a
> > [5]
> > https://github.com/DPDK/dpdk/blob/d69724b1dcc69784bcef00b96597469b7f6e6207/app/test-pmd/util.c#L44
> > [6]
> > https://github.com/DPDK/dpdk/blob/d69724b1dcc69784bcef00b96597469b7f6e6207/app/test-pmd/util.c#L60
> > [7]
> > https://github.com/DPDK/dpdk/blob/d69724b1dcc69784bcef00b96597469b7f6e6207/drivers/net/mlx5/mlx5_rxq.c#L1743
> >
> > Thank you.
> >
> >
> > When LRO is enabled, I notice that LROed mbufs seem to share identical
> > timestamp values, and the timestamps are unexpectedly large or
> > inconsistent. This raises
> > the question of whether
> > LRO is interfering with the correctness of the RX HW timestamps.
> >
> > I’d appreciate any clarification on whether HW RX timestamping is reliable
> > when LRO is enabled on this platform, or if LRO should be just disabled for
> > accurate
> > per-packet timestamping.
> >
> >
> > Sincerely,
> > Junghan Yoon
> >
> >
> >
> >
> >
> >
> >
