From: Bjorn Ketelaars Subject: Re: igc(4) tso To: Moritz Buhl Cc: tech@openbsd.org, j.klemkow@wemelug.de Date: Fri, 1 Mar 2024 07:42:59 +0100 On Tue 09/01/2024 23:22, Moritz Buhl wrote: > Here is a diff that implements tso for igc(4). > It looks very similar to the other intel cards but it was missing > VLAN_HWTAGGING. > > I am not too happy how olinfo_status is initialized now so I am > open to suggestions. > > I tested the diff on I225-V and I225-LM in various configurations. > More testing welcome. As privately discussed, I rebased your diff on top of recent changes to igc(4). Initially the interface locked up after about an hour, which was fixed by copying a recent change to ix(4), em(4), and others (from bluhm@ [0]). With this I have been running your diff for the last couple of days without any issues. $ dmesg | grep igc0 igc0 at pci2 dev 0 function 0 "Intel I225-V" rev 0x03, msix, 4 queues, address 00:e2:69:5b:a6:aa $ ifconfig igc0 hwfeatures igc0: flags=8a43 mtu 1500 hwfeatures=31b7 hardmtu 9216 lladdr 00:e2:69:5b:a6:aa index 1 priority 0 llprio 3 media: Ethernet autoselect (2500baseT full-duplex) status: active I refactored your diff to make it look more similar to the approach in ix(4), and split it up in 1.) a vlan hwtagging part and 2.) a tso part, as has been suggested by jan@. Included with this mail is the hwtagging part, which has been tested successfully on a system with multiple igc(4) interfaces. I will send the tso part later. BTW, the issue with locking that has been resolved by picking a change from [0] is in the tso part. [0] https://codeberg.org/OpenBSD/src/commit/e78a66e5775d3e9cf59b16f193be327ae6d56426 diff --git sys/dev/pci/if_igc.c sys/dev/pci/if_igc.c index b88644ade21..159051d8124 100644 --- sys/dev/pci/if_igc.c +++ sys/dev/pci/if_igc.c @@ -117,7 +117,8 @@ int igc_media_change(struct ifnet *); void igc_iff(struct igc_softc *); void igc_update_link_status(struct igc_softc *); int igc_get_buf(struct rx_ring *, int); -int igc_tx_ctx_setup(struct tx_ring *, struct mbuf *, int, uint32_t *); +int igc_tx_ctx_setup(struct tx_ring *, struct mbuf *, int, uint32_t *, + uint32_t *); void igc_configure_queues(struct igc_softc *); void igc_set_queues(struct igc_softc *, uint32_t, uint32_t, int); @@ -790,10 +791,8 @@ igc_setup_interface(struct igc_softc *sc) ifp->if_capabilities = IFCAP_VLAN_MTU; -#ifdef notyet #if NVLAN > 0 ifp->if_capabilities |= IFCAP_VLAN_HWTAGGING; -#endif #endif ifp->if_capabilities |= IFCAP_CSUM_IPv4; @@ -1001,7 +1000,11 @@ igc_start(struct ifqueue *ifq) bus_dmamap_sync(txr->txdma.dma_tag, map, 0, map->dm_mapsize, BUS_DMASYNC_PREWRITE); - if (igc_tx_ctx_setup(txr, m, prod, &olinfo_status)) { + cmd_type_len = IGC_ADVTXD_DCMD_IFCS | IGC_ADVTXD_DTYP_DATA | + IGC_ADVTXD_DCMD_DEXT; + + if (igc_tx_ctx_setup(txr, m, prod, &cmd_type_len, + &olinfo_status)) { /* Consume the first descriptor */ prod++; prod &= mask; @@ -1011,14 +1014,11 @@ igc_start(struct ifqueue *ifq) for (i = 0; i < map->dm_nsegs; i++) { txdesc = &txr->tx_base[prod]; - cmd_type_len = IGC_ADVTXD_DCMD_IFCS | IGC_ADVTXD_DTYP_DATA | - IGC_ADVTXD_DCMD_DEXT | map->dm_segs[i].ds_len; - if (i == map->dm_nsegs - 1) - cmd_type_len |= IGC_ADVTXD_DCMD_EOP | - IGC_ADVTXD_DCMD_RS; - - htolem64(&txdesc->read.buffer_addr, map->dm_segs[i].ds_addr); - htolem32(&txdesc->read.cmd_type_len, cmd_type_len); + htolem64(&txdesc->read.buffer_addr, + map->dm_segs[i].ds_addr); + htolem32(&txdesc->read.cmd_type_len, cmd_type_len | + map->dm_segs[i].ds_len | ((i == map->dm_nsegs - 1) ? + IGC_ADVTXD_DCMD_EOP | IGC_ADVTXD_DCMD_RS : 0)); htolem32(&txdesc->read.olinfo_status, olinfo_status); last = prod; @@ -2019,7 +2019,7 @@ igc_free_transmit_buffers(struct tx_ring *txr) int igc_tx_ctx_setup(struct tx_ring *txr, struct mbuf *mp, int prod, - uint32_t *olinfo_status) + uint32_t *cmd_type_len, uint32_t *olinfo_status) { struct ether_extracted ext; struct igc_adv_tx_context_desc *txdesc; @@ -2027,6 +2027,7 @@ igc_tx_ctx_setup(struct tx_ring *txr, struct mbuf *mp, int prod, uint32_t vlan_macip_lens = 0; int off = 0; + ether_extract_headers(mp, &ext); vlan_macip_lens |= (sizeof(*ext.eh) << IGC_ADVTXD_MACLEN_SHIFT); /* @@ -2034,18 +2035,15 @@ igc_tx_ctx_setup(struct tx_ring *txr, struct mbuf *mp, int prod, * be placed into the context descriptor. Hence * we need to make one even if not doing offloads. */ -#ifdef notyet #if NVLAN > 0 if (ISSET(mp->m_flags, M_VLANTAG)) { uint32_t vtag = mp->m_pkthdr.ether_vtag; vlan_macip_lens |= (vtag << IGC_ADVTXD_VLAN_SHIFT); + *cmd_type_len |= IGC_ADVTXD_DCMD_VLE; off = 1; } -#endif #endif - ether_extract_headers(mp, &ext); - if (ext.ip4) { type_tucmd_mlhl |= IGC_ADVTXD_TUCMD_IPV4; if (ISSET(mp->m_pkthdr.csum_flags, M_IPV4_CSUM_OUT)) { @@ -2056,8 +2054,6 @@ igc_tx_ctx_setup(struct tx_ring *txr, struct mbuf *mp, int prod, } else if (ext.ip6) { type_tucmd_mlhl |= IGC_ADVTXD_TUCMD_IPV6; #endif - } else { - return 0; } vlan_macip_lens |= ext.iphlen;