On Tue, 22 Sep 2026 20:37:33 +0800 Zaiyu Wang <[email protected]> wrote:
> This series addresses link-related issues on Wangxun Amber-lite 25G/40G NICs > (CR/KR training, hot-plug, 10G link state), and additionally refines UDP > offload handling. > > --- These AI comments look like the need addressing Review: [PATCH v4 00/16] net/txgbe: Amber-Lite link and offload fixes Author: Zaiyu Wang <[email protected]> Series applies to main; every commit builds with -Dwerror=true. Fixes: hashes resolve except the one noted in 13/16. Patch 10/16: net/txgbe: fix SFP hot-plug when auto-negotiation is on Warning: AN73 watchdog can be re-armed after dev_stop txgbe_dev_detect_sfp() now calls rte_eal_alarm_set() for txgbe_dev_e56_check_bp_event with no started check. dev_stop cancels check_bp_event first (txgbe_ethdev.c ~2121) and detect_sfp second (~2126). A detect_sfp instance running in the alarm thread between the two cancels, or one armed by a GPIO interrupt before txgbe_disable_intr(), re-arms the watchdog on a stopped port. check_bp_event re-arms itself unconditionally at "out:", so it keeps running, and after dev_close it dereferences freed port data. Patch 11 gates its own re-arm on dev->data->dev_started, but that flag is only cleared after dev_stop returns, so it does not close this window either. Cancel txgbe_dev_detect_sfp before txgbe_dev_e56_check_bp_event in dev_stop, and skip the re-arm in detect_sfp when hw->adapter_stopped is set. Patch 09/16: net/txgbe: add offload support for tunnel type UDP Warning: RTE_MBUF_F_TX_TUNNEL_UDP resolved by destination port The generic UDP tunnel flag is defined for tunnels that are not VXLAN or GENEVE. txgbe_get_tun_len() maps port 6081 to GENEVE and everything else to VXLAN, i.e. a fixed 16 byte UDP+VXLAN header. Any other UDP encapsulation gets a wrong tunnel length, so the inner checksum and TSO offsets are wrong, and GENEVE on a non-default port loses its options length. Since the driver advertises RTE_ETH_TX_OFFLOAD_UDP_TNL_TSO, either reject tunnels it cannot size in txgbe_prep_pkts() or stop advertising the capability. Info: the new rte_pktmbuf_read() NULL check is right, but the GRE and GENEVE cases in the same switch still dereference grh and gh without one. Same fix, same function. Patch 13/16: net/txgbe: fix unset pre2 FFE tap and backplane capability Warning: Fixes: 6104fd11089d does not exist. The 25G commit is 6104fd11086e, as used in 05/16 and 12/16. Warning: this is a feature, not a fix The recommended pre2 values are S25G_TX_FFE_CFG_DAC_PRE2 = 0x0 and S40G_TX_FFE_CFG_PRE2 = 0x0, so programming ffe_pre2 = 0 was never wrong. bp_capa = 0 is the existing KR4+CR4 behaviour. What remains is two new devargs, which do not belong in stable. Drop the Fixes tags and Cc: stable. The patch also changes the ffe_main/pre/post defaults on the 25G MAC from 27/8/44 to 0x2a/0x03/0x11. Neither the commit message nor txgbe.rst mentions this; the guide still documents 27, 8 and 44. Info: bp_capa is not range checked. Any value above 2 advertises no 40G ability at all on the 40G backplane. Patch 16/16: net/txgbe: align link capabilities and DAC classification Warning: get_link_capabilities_aml40() clears hw->devarg.auto_neg For a DAC whose fiber_suppport_speed is 10G only, the new DAC branch writes hw->devarg.auto_neg = false. Nothing sets it back. With the hot-plug support in 10/16 and 11/16, replacing that module with a 40G QSFP DAC leaves txgbe_xpcs_an_enabled() false, so AN73 never runs until the port is re-probed. The aml function has the same pattern, but hot-plug now makes it reachable on aml40. Return *autoneg = false without touching the devarg. Info: PMD_DRV_LOG in base/txgbe_aml40.c; base code uses DEBUGOUT. txgbe_is_40g_fiber_qsfp() and txgbe_is_10g_fiber_sfp() return int with true/false; make them bool. Patch 02/16: net/txgbe: use the requested speed in E56 AN setup Info: default AN advertisement changes. With dev_start passing 10G|25G (aml) or 10G|40G (aml40 after 07/16), the backplane branch now advertises 10GBASE-KR on the 25G and 40G parts, where previously it advertised only 25G or only 40G. Say so in the commit message. Patch 07/16: net/txgbe: fix link speed display info for 10G mode Info: adding 10G to the aml40 autoneg speed mask is not needed for the reported speed; that comes from the new PORTSTAT check. It changes what is advertised (see 02/16) and belongs with 01/16. Patch 15/16: net/txgbe: fix CR/KR link training and recovery Info: the CL72 poll (400 x 1 ms) and the page exchange (up to 200 x 1 ms) busy-wait inside txgbe_dev_e56_check_bp_event(), which runs on the EAL interrupt thread. This stalls interrupt and alarm handling for all ports. Info: BP_LOG expands to RTE_LOG, so its arguments are evaluated even when the log type is disabled. txgbe_e56_get_txffe() and the new "an_int" arguments do a dozen PHY reads per call only for logging. The BP_LOG("%s %d\n", __func__, __LINE__) in check_bp_event is a leftover debug line.
