[DPDK/testpmd Bug 1479] mlx5: Not able to create rte_flows to match head fragments and sub fragments

[email protected]

author

2024-07-02 14:21 UTC #0 lore raw

https://bugs.dpdk.org/show_bug.cgi?id=1479

Bug ID: 1479
Summary: mlx5: Not able to create rte_flows to match head
fragments and sub fragments
Product: DPDK
Version: 21.11
Hardware: All
OS: All
Status: UNCONFIRMED
Severity: major
Priority: Normal
Component: testpmd
Assignee: [email protected]
Reporter: [email protected]
Target Milestone: ---

I am trying to create an RTE flow rule to match head and non-head fragments to compute NIC RSS based on 5tuple/3tuple respectively on connectX-6 DX NIC and mlx5 driver.

As part of it when trying to install RTE flow rule using testpmd on dpdk 21.11/23.07 version, the following errors are thrown.

Flow rule to match head fragment
=================================
testpmd> flow create 0 ingress pattern eth / ipv4 fragment_offset spec 0x2000 fragment_offset mask 0x3fff / end actions drop / count / end port_flow_complain(): Caught PMD error type 13 (specific pattern item): cause:
0x7ffd1954c548, match on first fragment not supported: Operation not supported

Flow rule to match non-head fragments
======================================
testpmd> flow validate 0 ingress pattern eth / ipv4 fragment_offset is 0x2001 fragment_offset last 0x1fff / end actions drop / end port_flow_complain(): Caught PMD error type 11 (item specification range):
cause: 0x7ffc6f629534, specified range not supported: Operation not supported

When I browsed mlx5_flow_dv.c driver file, there are set of conditions implemented to block this configurations.

Could you kindly help is there a way to compute different RSS hash for fragments and non-fragments using RTE flow rules on mellanox?

Thanks! in advance.

/*
* Match on fragment_offset 0x2000 means MF is 1 and frag-offset is 0,
* indicating this is 1st fragment of fragmented packet.
* This is not yet supported in MLX5, return appropriate error message. */
if (fragment_offset_spec == RTE_BE16(RTE_IPV4_HDR_MF_FLAG))
return rte_flow_error_set(error, ENOTSUP,
RTE_FLOW_ERROR_TYPE_ITEM, item, "match on first fragment not " "supported");
if (fragment_offset_spec && !last)
return rte_flow_error_set(error, ENOTSUP,
RTE_FLOW_ERROR_TYPE_ITEM, item, "specified value not supported");
/*
* Match on fragment_offset spec 0x2001 and last 0x3fff
* means MF is 1 and frag-offset is > 0.
* This packet is fragment 2nd and onward, excluding last.
* This is not yet supported in MLX5, return appropriate
* error message.
*/
if (fragment_offset_spec == RTE_BE16(RTE_IPV4_HDR_MF_FLAG + 1) &&
fragment_offset_last == RTE_BE16(MLX5_IPV4_FRAG_OFFSET_MASK))
return rte_flow_error_set(error, ENOTSUP,
RTE_FLOW_ERROR_TYPE_ITEM_LAST, last, "match on following " "fragments not supported");
/*
* Match on fragment_offset spec 0x0001 and last 0x1fff
* means MF is 0 and frag-offset is > 0.
* This packet is last fragment of fragmented packet.
* This is not yet supported in MLX5, return appropriate
* error message.
*/
if (fragment_offset_spec == RTE_BE16(1) &&
fragment_offset_last == RTE_BE16(RTE_IPV4_HDR_OFFSET_MASK))
return rte_flow_error_set(error, ENOTSUP,
RTE_FLOW_ERROR_TYPE_ITEM_LAST, last, "match on last " "fragment not supported");

Signature · 2 lines

You are receiving this mail because:
You are the assignee for the bug.

Asaf Penso

2024-07-03 06:19 UTC #1 lore raw

Hello

Mlx5 pmd supports only these two modes:

1. fragment_offset=0x0000 with mask 0xffff means MF is 0 and frag-offset is 0.
This is an unfragmented packet.
2. fragment_offset spec 0x0001 and last 0x3fff with mask 0xffff, means MF and/or frag-offset is not 0.
This is a fragmented packet.
To answer your questions, mlx5 doesn’t support match on head fragment and to match on non-head fragment, try mode 2 above.

Regards,
Asaf Penso

From: [email protected] <[email protected]>
Sent: Tuesday, 2 July 2024 17:21
To: [email protected]
Subject: [DPDK/testpmd Bug 1479] mlx5: Not able to create rte_flows to match head fragments and sub fragments

Bug ID
1479<https://bugs.dpdk.org/show_bug.cgi?id=1479>
Summary
mlx5: Not able to create rte_flows to match head fragments and sub fragments Product
DPDK
Version
21.11
Hardware
All
OS
All
Status
UNCONFIRMED
Severity
major
Priority
Normal
Component
testpmd
Assignee
[email protected]<mailto:[email protected]>
Reporter
[email protected]<mailto:[email protected]>
Target Milestone
---

I am trying to create an RTE flow rule to match head and non-head fragments to

compute NIC RSS based on 5tuple/3tuple respectively on connectX-6 DX NIC and

mlx5 driver.

As part of it when trying to install RTE flow rule using testpmd on dpdk

21.11/23.07 version, the following errors are thrown.

Flow rule to match head fragment

=================================

testpmd> flow create 0 ingress pattern eth / ipv4 fragment_offset spec 0x2000

fragment_offset mask 0x3fff / end actions drop / count / end

port_flow_complain(): Caught PMD error type 13 (specific pattern item): cause:

0x7ffd1954c548, match on first fragment not supported: Operation not supported

Flow rule to match non-head fragments

======================================

testpmd> flow validate 0 ingress pattern eth / ipv4 fragment_offset is 0x2001

fragment_offset last 0x1fff / end actions drop / end

port_flow_complain(): Caught PMD error type 11 (item specification range):

cause: 0x7ffc6f629534, specified range not supported: Operation not supported

When I browsed mlx5_flow_dv.c driver file, there are set of conditions

implemented to block this configurations.

Could you kindly help is there a way to compute different RSS hash for

fragments and non-fragments using RTE flow rules on mellanox?

Thanks! in advance.

/*

* Match on fragment_offset 0x2000 means MF is 1 and frag-offset is 0,

* indicating this is 1st fragment of fragmented packet.

* This is not yet supported in MLX5, return appropriate error message.

*/

if (fragment_offset_spec == RTE_BE16(RTE_IPV4_HDR_MF_FLAG))

return rte_flow_error_set(error, ENOTSUP,

RTE_FLOW_ERROR_TYPE_ITEM, item,

"match on first fragment not "

"supported");

if (fragment_offset_spec && !last)

return rte_flow_error_set(error, ENOTSUP,

RTE_FLOW_ERROR_TYPE_ITEM, item,

"specified value not supported");

/*

* Match on fragment_offset spec 0x2001 and last 0x3fff

* means MF is 1 and frag-offset is > 0.

* This packet is fragment 2nd and onward, excluding last.

* This is not yet supported in MLX5, return appropriate

* error message.

*/

if (fragment_offset_spec == RTE_BE16(RTE_IPV4_HDR_MF_FLAG + 1) &&

fragment_offset_last == RTE_BE16(MLX5_IPV4_FRAG_OFFSET_MASK))

return rte_flow_error_set(error, ENOTSUP,

RTE_FLOW_ERROR_TYPE_ITEM_LAST,

last, "match on following "

"fragments not supported");

/*

* Match on fragment_offset spec 0x0001 and last 0x1fff

* means MF is 0 and frag-offset is > 0.

* This packet is last fragment of fragmented packet.

* This is not yet supported in MLX5, return appropriate

* error message.

*/

if (fragment_offset_spec == RTE_BE16(1) &&

fragment_offset_last == RTE_BE16(RTE_IPV4_HDR_OFFSET_MASK))

return rte_flow_error_set(error, ENOTSUP,

RTE_FLOW_ERROR_TYPE_ITEM_LAST,

last, "match on last "

"fragment not supported");

________________________________
You are receiving this mail because:

* You are the assignee for the bug.

[email protected]

author

2025-10-27 09:17 UTC #2 lore raw

In reply to [email protected]

http://bugs.dpdk.org/show_bug.cgi?id=1479

[email protected] ([email protected]) changed:

What |Removed |Added
----------------------------------------------------------------------------

                 CC|                            |[email protected]
             Status|CONFIRMED                   |RESOLVED
         Resolution|---                         |FIXED

--- Comment #1 from [email protected] ([email protected]) --- Hi, Siva,
Sorry for the late response,

You're correct that the mlx5 driver explicitly blocks the fragment matching patterns you're trying to use.

For your use case (computing different RSS based on 5-tuple vs 3-tuple):
Create two flow rules:
One for ANY fragmented packets → RSS with 3-tuple (ipv4-frag)
One for non-fragmented packets → RSS with 5-tuple (ipv4-tcp, ipv4-udp) The NIC will automatically:
Use 3-tuple for fragments (only IP src/dst/protocol available) Use 5-tuple for non-fragments (full 5-tuple available)
You cannot differentiate between first, middle, and last fragments in flow rules on ConnectX-6 DX.
The hardware RSS engine already handles this correctly by design - fragmented packets can only use 3-tuple because port information is only in the first fragment, and may not be available to all fragments.

Signature · 2 lines

You are receiving this mail because:
You are the assignee for the bug.