On Thu, 1 Oct 2026 13:02:54 +0530 Prashant Gupta <[email protected]> wrote:
> This is the second of four series upstreaming the missing NXP dpaa2 > driver changes. It does not depend on series 1 and can be applied > independently. It adds the net/dpaa2 flow, metering and parser features: > > - fix an integer overflow in the CCSR region mapping, > - set Tx confirmation on device init and support a larger burst size, > - support MPLS and PPPoE flow distribution, meter and policing, and the > flow drop action, with a per-device default flow miss action, > - identify Rx mbuf hash information by FLC and add minimum key size > support, > - restructure the parser processing, parse tunnel and fragmented packet > types, remove the unused soft parser driver and rename the Rx queue > flags. > > Note that patch 12 removes the soft parser driver and, with it, the > rte_flow VXLAN and eCPRI pattern items that were only accepted while an > SP image was believed to be loaded. Both items are dropped from the dpaa2 > feature matrix and a release note is added. > > Every commit builds cleanly (including the aarch64 DPAA cross build with > -Werror) and the series is bisectable. Still have several open issues in AI review. [PATCH v5-S2 00/13] net/dpaa2: flow, meter and parser updates Applied to main (062b389). Patches 5 and 12 conflict in release_26_11.rst; applied with release notes excluded. Rebase. Per-commit build with -Dwerror=true passes for all 13 patches (x86, gcc, dpaa2 drivers only). Series-level: The v4 review comments for patches 5, 7, 8 and 10 were pasted into the commit messages as "Known limitations and follow-up items" instead of being fixed. Known regressions and bugs in new code are fixed before merge, not listed as TODOs. Several of the pasted items describe code no longer in the patch. Remove these sections and fix the issues. Patch 1/13 net/dpaa2: fix integer overflow in CCSR region mapping Info: the commit message says the old round-down did not account for offset + len crossing a page boundary, but "len = len & page_mask" is kept. All callers pass page-aligned addresses and touch only the first page, so nothing faults today. Map RTE_ALIGN_CEIL(offset + len, page_size) or drop the sentence. Patch 3/13 net/dpaa2: support larger burst size Info: on non-LX2160A or MC < 8.7 the accepted burst drops from 65536 to 0xF7FF. Say so in the commit message. Patch 5/13 net/dpaa2: support meter and policing Error: nothing in the tree calls dpni_set_rx_tc_policing() after this series. Every rte_mtr op returns success and only edits the software lists. Profile mode, unit, rates and policy red_drop are stored and never read. dpaa2.rst and the release note advertise policing that does not happen. Implement it or drop the patch. Error: the "Known limitations" section describes code that is not in this patch (the /1000 conversion, the meter_create HW error sign, meter_destroy ignoring the dpni return). Warning: profile_add and policy_add have no duplicate-ID check. Warning: the profiles, policies and meters lists are never freed on dpaa2_dev_close(). Warning: s_dpaa2_mtr_capa is file-scope and written under the per-device meter_lock, so two ports race on it. s_err_msg is a shared static buffer returned through error->message. Use a local capa and constant strings. Warning: capabilities still advertise color_aware_trtcm_rfc2698 and color_aware_trtcm_rfc4115. Warning: dpaa2.ini adds "Rate limitation = Y", which the commit message itself says is wrong. Warning: dpni_set_rx_tc_policing_v1() is declared and never defined. DPNI_POLICER_OPT_DO_NOT_RESET_COUNTERS is unused. Info: of the four release note bullets, only metering belongs to this patch. Flow classification belongs with patch 4. Patch 7/13 net/dpaa2: set default flow miss action per device Error: default_flow = RTE_MIN(fs_entries, dist_queues) - 1 sends FS misses to DPNI flow dist_queues - 1. Rx queues map flow_id = dist_idx % num_rxqueue_per_tc over all DPNI queues, so that flow is ethdev Rx queue dist_queues - 1 of the TC. An application that configures fewer Rx queues never polls it. Once a QUEUE or DROP rule exists, unmatched traffic is lost. Previously it went to flow 0. fs_entries is a table size, not a queue count. Warning: the commit message says default_flow is "initialised to 0". Warning: the commit message says a release note "should be added". Add it in this patch. Patch 8/13 net/dpaa2: identify Rx mbuf hash information by FLC Warning: "Known limitations" describes RX_FDIR_ID and a raw FLC word in hash.fdir.hi. The code sets RTE_MBUF_F_RX_FDIR and writes hash.sched. Remove the section. Info: hash.sched is the rte_sched Tx field. The TC/flow it carries is the Rx queue the application already polled. Patch 9/13 net/dpaa2: add minimum key size support Warning: the removed comment says MC only supports 56-byte entries. State which MC version supports 24 bytes and gate on dpaa2_dev_cmp_dpni_ver() as patch 3 does. Patch 10/13 net/dpaa2: restructure dpaa2 parser processing Error: Rx timestamping is broken. The dynfield write is removed. dpaa2_enable_ts[] is set in dpaa2_eth_dev_configure() and never read. dpaa2_dev_rx_common() reads a dynfield nothing writes. Listing it as a known limitation does not make it acceptable. Error: on non-LX2160A, dpaa2_dev_rx_parse() no longer sets L3/L4 checksum flags for the common IPv4/IPv6 TCP/UDP cases. The switch returns directly, and only parse_slow sets the flags. Before, the flags were set at function entry for every packet. Bad checksums are no longer reported. The commit message claim that the observable result is the same is false. Error: dpaa2_dev_rx_parse_frc() starts with packet_type = RTE_PTYPE_L4_NONFRAG (0x600) and ORs in L4 values. L4_TCP and L4_ICMP become 0x700 (L4_IGMP). L4_UDP and L4_SCTP collapse to NONFRAG. Per the commit message, ICMP and SCTP now always take this path on LX2160A, so every ICMP packet is reported as IGMP. GTPU/GTPC report L4 NONFRAG instead of UDP. Non-VLAN frames get no RTE_PTYPE_L2_ETHER. Assign the L4 type rather than OR it. Warning: VLAN-tagged frames on LX2160A now take the fast switch with the VLAN bit stripped and report RTE_PTYPE_L2_ETHER. Before, they went to parse_slow and reported L2_ETHER_VLAN/QINQ. This is an undescribed functional change. Warning: dpaa2_dev_rx_mbuf_sched_set() now also runs on the non-LX2160A path, so those SoCs report RX_RSS_HASH from flc_hi for the first time. This is not described. Warning: the parser dump is dead. The env var sets dpaa2_print_parser_result, which is never read. The helper tests DPAA2_RX_PRINT_PSR_RESULT_FLAG, which nothing sets. Once wired up, its RX_FDIR branch walks rx_vq[0..MAX_RX_QUEUES) and dereferences rxq without a NULL check, but only nb_rx_queues entries exist. Warning: psr_dynfield_offset is never registered, so dpaa2_dev_rx_annot_prefetch(), dpaa2_dev_rx_parse_offset() and struct dpaa2_dyn_rx_protocol_pos are dead. sp_protocol is never assigned, so the GENEVE block is dead. dpaa2_parser_ecpri_extract() and dpaa2_protocol_psr_bit_offset() are unused. Patch 12/13 net/dpaa2: remove unused soft parser driver Warning: rte_flow_create() does not call dpaa2_dev_verify_patterns(). In dpaa2_generic_flow_set() the default case sets ret = -ENOTSUP and breaks only out of the switch. The loop continues, and later item and action calls overwrite ret. With the VXLAN case gone, a pattern eth/ipv4/udp/vxlan/eth is installed with VXLAN ignored and the inner eth handled as in_tunnel. Return from the default case. Info: soft parser leftovers remain: priv->ss_offset, ss_iova, ss_param_iova, sp_protocol, enum dpaa2_ecpri_fafe_type and DPAA2_ECPRI_MAX_EXTRACT_NB.
