Index: stable/11/sys/dev/sfxge/sfxge.c =================================================================== --- stable/11/sys/dev/sfxge/sfxge.c (revision 342453) +++ stable/11/sys/dev/sfxge/sfxge.c (revision 342454) @@ -1,1202 +1,1202 @@ /*- * Copyright (c) 2010-2016 Solarflare Communications Inc. * All rights reserved. * * This software was developed in part by Philip Paeps under contract for * Solarflare Communications, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * The views and conclusions contained in the software and documentation are * those of the authors and should not be interpreted as representing official * policies, either expressed or implied, of the FreeBSD Project. */ #include __FBSDID("$FreeBSD$"); #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef RSS #include #endif #include "common/efx.h" #include "sfxge.h" #include "sfxge_rx.h" #include "sfxge_ioc.h" #include "sfxge_version.h" #define SFXGE_CAP (IFCAP_VLAN_MTU | IFCAP_VLAN_HWCSUM | \ IFCAP_RXCSUM | IFCAP_TXCSUM | \ IFCAP_RXCSUM_IPV6 | IFCAP_TXCSUM_IPV6 | \ IFCAP_TSO4 | IFCAP_TSO6 | \ IFCAP_JUMBO_MTU | \ IFCAP_VLAN_HWTSO | IFCAP_LINKSTATE | IFCAP_HWSTATS) #define SFXGE_CAP_ENABLE SFXGE_CAP #define SFXGE_CAP_FIXED (IFCAP_VLAN_MTU | \ IFCAP_JUMBO_MTU | IFCAP_LINKSTATE | IFCAP_HWSTATS) MALLOC_DEFINE(M_SFXGE, "sfxge", "Solarflare 10GigE driver"); SYSCTL_NODE(_hw, OID_AUTO, sfxge, CTLFLAG_RD, 0, "SFXGE driver parameters"); #define SFXGE_PARAM_RX_RING SFXGE_PARAM(rx_ring) static int sfxge_rx_ring_entries = SFXGE_NDESCS; TUNABLE_INT(SFXGE_PARAM_RX_RING, &sfxge_rx_ring_entries); SYSCTL_INT(_hw_sfxge, OID_AUTO, rx_ring, CTLFLAG_RDTUN, &sfxge_rx_ring_entries, 0, "Maximum number of descriptors in a receive ring"); #define SFXGE_PARAM_TX_RING SFXGE_PARAM(tx_ring) static int sfxge_tx_ring_entries = SFXGE_NDESCS; TUNABLE_INT(SFXGE_PARAM_TX_RING, &sfxge_tx_ring_entries); SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_ring, CTLFLAG_RDTUN, &sfxge_tx_ring_entries, 0, "Maximum number of descriptors in a transmit ring"); #define SFXGE_PARAM_RESTART_ATTEMPTS SFXGE_PARAM(restart_attempts) static int sfxge_restart_attempts = 3; TUNABLE_INT(SFXGE_PARAM_RESTART_ATTEMPTS, &sfxge_restart_attempts); SYSCTL_INT(_hw_sfxge, OID_AUTO, restart_attempts, CTLFLAG_RDTUN, &sfxge_restart_attempts, 0, "Maximum number of attempts to bring interface up after reset"); #if EFSYS_OPT_MCDI_LOGGING #define SFXGE_PARAM_MCDI_LOGGING SFXGE_PARAM(mcdi_logging) static int sfxge_mcdi_logging = 0; TUNABLE_INT(SFXGE_PARAM_MCDI_LOGGING, &sfxge_mcdi_logging); #endif static void sfxge_reset(void *arg, int npending); static int sfxge_estimate_rsrc_limits(struct sfxge_softc *sc) { efx_drv_limits_t limits; int rc; unsigned int evq_max; uint32_t evq_allocated; uint32_t rxq_allocated; uint32_t txq_allocated; /* * Limit the number of event queues to: * - number of CPUs * - hardwire maximum RSS channels * - administratively specified maximum RSS channels */ #ifdef RSS /* * Avoid extra limitations so that the number of queues * may be configured at administrator's will */ evq_max = MIN(MAX(rss_getnumbuckets(), 1), EFX_MAXRSS); #else evq_max = MIN(mp_ncpus, EFX_MAXRSS); #endif if (sc->max_rss_channels > 0) evq_max = MIN(evq_max, sc->max_rss_channels); memset(&limits, 0, sizeof(limits)); limits.edl_min_evq_count = 1; limits.edl_max_evq_count = evq_max; - limits.edl_min_txq_count = SFXGE_TXQ_NTYPES; - limits.edl_max_txq_count = evq_max + SFXGE_TXQ_NTYPES - 1; + limits.edl_min_txq_count = SFXGE_EVQ0_N_TXQ(sc); + limits.edl_max_txq_count = evq_max + SFXGE_EVQ0_N_TXQ(sc) - 1; limits.edl_min_rxq_count = 1; limits.edl_max_rxq_count = evq_max; efx_nic_set_drv_limits(sc->enp, &limits); if ((rc = efx_nic_init(sc->enp)) != 0) return (rc); rc = efx_nic_get_vi_pool(sc->enp, &evq_allocated, &rxq_allocated, &txq_allocated); if (rc != 0) { efx_nic_fini(sc->enp); return (rc); } - KASSERT(txq_allocated >= SFXGE_TXQ_NTYPES, - ("txq_allocated < SFXGE_TXQ_NTYPES")); + KASSERT(txq_allocated >= SFXGE_EVQ0_N_TXQ(sc), + ("txq_allocated < %u", SFXGE_EVQ0_N_TXQ(sc))); sc->evq_max = MIN(evq_allocated, evq_max); sc->evq_max = MIN(rxq_allocated, sc->evq_max); - sc->evq_max = MIN(txq_allocated - (SFXGE_TXQ_NTYPES - 1), + sc->evq_max = MIN(txq_allocated - (SFXGE_EVQ0_N_TXQ(sc) - 1), sc->evq_max); KASSERT(sc->evq_max <= evq_max, ("allocated more than maximum requested")); #ifdef RSS if (sc->evq_max < rss_getnumbuckets()) device_printf(sc->dev, "The number of allocated queues (%u) " "is less than the number of RSS buckets (%u); " "performance degradation might be observed", sc->evq_max, rss_getnumbuckets()); #endif /* * NIC is kept initialized in the case of success to be able to * initialize port to find out media types. */ return (0); } static int sfxge_set_drv_limits(struct sfxge_softc *sc) { efx_drv_limits_t limits; memset(&limits, 0, sizeof(limits)); /* Limits are strict since take into account initial estimation */ limits.edl_min_evq_count = limits.edl_max_evq_count = sc->intr.n_alloc; limits.edl_min_txq_count = limits.edl_max_txq_count = - sc->intr.n_alloc + SFXGE_TXQ_NTYPES - 1; + sc->intr.n_alloc + SFXGE_EVQ0_N_TXQ(sc) - 1; limits.edl_min_rxq_count = limits.edl_max_rxq_count = sc->intr.n_alloc; return (efx_nic_set_drv_limits(sc->enp, &limits)); } static int sfxge_start(struct sfxge_softc *sc) { int rc; SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); if (sc->init_state == SFXGE_STARTED) return (0); if (sc->init_state != SFXGE_REGISTERED) { rc = EINVAL; goto fail; } /* Set required resource limits */ if ((rc = sfxge_set_drv_limits(sc)) != 0) goto fail; if ((rc = efx_nic_init(sc->enp)) != 0) goto fail; /* Start processing interrupts. */ if ((rc = sfxge_intr_start(sc)) != 0) goto fail2; /* Start processing events. */ if ((rc = sfxge_ev_start(sc)) != 0) goto fail3; /* Fire up the port. */ if ((rc = sfxge_port_start(sc)) != 0) goto fail4; /* Start the receiver side. */ if ((rc = sfxge_rx_start(sc)) != 0) goto fail5; /* Start the transmitter side. */ if ((rc = sfxge_tx_start(sc)) != 0) goto fail6; sc->init_state = SFXGE_STARTED; /* Tell the stack we're running. */ sc->ifnet->if_drv_flags |= IFF_DRV_RUNNING; sc->ifnet->if_drv_flags &= ~IFF_DRV_OACTIVE; return (0); fail6: sfxge_rx_stop(sc); fail5: sfxge_port_stop(sc); fail4: sfxge_ev_stop(sc); fail3: sfxge_intr_stop(sc); fail2: efx_nic_fini(sc->enp); fail: device_printf(sc->dev, "sfxge_start: %d\n", rc); return (rc); } static void sfxge_if_init(void *arg) { struct sfxge_softc *sc; sc = (struct sfxge_softc *)arg; SFXGE_ADAPTER_LOCK(sc); (void)sfxge_start(sc); SFXGE_ADAPTER_UNLOCK(sc); } static void sfxge_stop(struct sfxge_softc *sc) { SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); if (sc->init_state != SFXGE_STARTED) return; sc->init_state = SFXGE_REGISTERED; /* Stop the transmitter. */ sfxge_tx_stop(sc); /* Stop the receiver. */ sfxge_rx_stop(sc); /* Stop the port. */ sfxge_port_stop(sc); /* Stop processing events. */ sfxge_ev_stop(sc); /* Stop processing interrupts. */ sfxge_intr_stop(sc); efx_nic_fini(sc->enp); sc->ifnet->if_drv_flags &= ~IFF_DRV_RUNNING; } static int sfxge_vpd_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc) { efx_vpd_value_t value; int rc = 0; switch (ioc->u.vpd.op) { case SFXGE_VPD_OP_GET_KEYWORD: value.evv_tag = ioc->u.vpd.tag; value.evv_keyword = ioc->u.vpd.keyword; rc = efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value); if (rc != 0) break; ioc->u.vpd.len = MIN(ioc->u.vpd.len, value.evv_length); if (ioc->u.vpd.payload != 0) { rc = copyout(value.evv_value, ioc->u.vpd.payload, ioc->u.vpd.len); } break; case SFXGE_VPD_OP_SET_KEYWORD: if (ioc->u.vpd.len > sizeof(value.evv_value)) return (EINVAL); value.evv_tag = ioc->u.vpd.tag; value.evv_keyword = ioc->u.vpd.keyword; value.evv_length = ioc->u.vpd.len; rc = copyin(ioc->u.vpd.payload, value.evv_value, value.evv_length); if (rc != 0) break; rc = efx_vpd_set(sc->enp, sc->vpd_data, sc->vpd_size, &value); if (rc != 0) break; rc = efx_vpd_verify(sc->enp, sc->vpd_data, sc->vpd_size); if (rc != 0) break; rc = efx_vpd_write(sc->enp, sc->vpd_data, sc->vpd_size); break; default: rc = EOPNOTSUPP; break; } return (rc); } static int sfxge_private_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc) { switch (ioc->op) { case SFXGE_MCDI_IOC: return (sfxge_mcdi_ioctl(sc, ioc)); case SFXGE_NVRAM_IOC: return (sfxge_nvram_ioctl(sc, ioc)); case SFXGE_VPD_IOC: return (sfxge_vpd_ioctl(sc, ioc)); default: return (EOPNOTSUPP); } } static int sfxge_if_ioctl(struct ifnet *ifp, unsigned long command, caddr_t data) { struct sfxge_softc *sc; struct ifreq *ifr; sfxge_ioc_t ioc; int error; ifr = (struct ifreq *)data; sc = ifp->if_softc; error = 0; switch (command) { case SIOCSIFFLAGS: SFXGE_ADAPTER_LOCK(sc); if (ifp->if_flags & IFF_UP) { if (ifp->if_drv_flags & IFF_DRV_RUNNING) { if ((ifp->if_flags ^ sc->if_flags) & (IFF_PROMISC | IFF_ALLMULTI)) { sfxge_mac_filter_set(sc); } } else sfxge_start(sc); } else if (ifp->if_drv_flags & IFF_DRV_RUNNING) sfxge_stop(sc); sc->if_flags = ifp->if_flags; SFXGE_ADAPTER_UNLOCK(sc); break; case SIOCSIFMTU: if (ifr->ifr_mtu == ifp->if_mtu) { /* Nothing to do */ error = 0; } else if (ifr->ifr_mtu > SFXGE_MAX_MTU) { error = EINVAL; } else if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) { ifp->if_mtu = ifr->ifr_mtu; error = 0; } else { /* Restart required */ SFXGE_ADAPTER_LOCK(sc); sfxge_stop(sc); ifp->if_mtu = ifr->ifr_mtu; error = sfxge_start(sc); SFXGE_ADAPTER_UNLOCK(sc); if (error != 0) { ifp->if_flags &= ~IFF_UP; ifp->if_drv_flags &= ~IFF_DRV_RUNNING; if_down(ifp); } } break; case SIOCADDMULTI: case SIOCDELMULTI: if (ifp->if_drv_flags & IFF_DRV_RUNNING) sfxge_mac_filter_set(sc); break; case SIOCSIFCAP: { int reqcap = ifr->ifr_reqcap; int capchg_mask; SFXGE_ADAPTER_LOCK(sc); /* Capabilities to be changed in accordance with request */ capchg_mask = ifp->if_capenable ^ reqcap; /* * The networking core already rejects attempts to * enable capabilities we don't have. We still have * to reject attempts to disable capabilities that we * can't (yet) disable. */ KASSERT((reqcap & ~ifp->if_capabilities) == 0, ("Unsupported capabilities 0x%x requested 0x%x vs " "supported 0x%x", reqcap & ~ifp->if_capabilities, reqcap , ifp->if_capabilities)); if (capchg_mask & SFXGE_CAP_FIXED) { error = EINVAL; SFXGE_ADAPTER_UNLOCK(sc); break; } /* Check request before any changes */ if ((capchg_mask & IFCAP_TSO4) && (reqcap & (IFCAP_TSO4 | IFCAP_TXCSUM)) == IFCAP_TSO4) { error = EAGAIN; SFXGE_ADAPTER_UNLOCK(sc); if_printf(ifp, "enable txcsum before tso4\n"); break; } if ((capchg_mask & IFCAP_TSO6) && (reqcap & (IFCAP_TSO6 | IFCAP_TXCSUM_IPV6)) == IFCAP_TSO6) { error = EAGAIN; SFXGE_ADAPTER_UNLOCK(sc); if_printf(ifp, "enable txcsum6 before tso6\n"); break; } if (reqcap & IFCAP_TXCSUM) { ifp->if_hwassist |= (CSUM_IP | CSUM_TCP | CSUM_UDP); } else { ifp->if_hwassist &= ~(CSUM_IP | CSUM_TCP | CSUM_UDP); if (reqcap & IFCAP_TSO4) { reqcap &= ~IFCAP_TSO4; if_printf(ifp, "tso4 disabled due to -txcsum\n"); } } if (reqcap & IFCAP_TXCSUM_IPV6) { ifp->if_hwassist |= (CSUM_TCP_IPV6 | CSUM_UDP_IPV6); } else { ifp->if_hwassist &= ~(CSUM_TCP_IPV6 | CSUM_UDP_IPV6); if (reqcap & IFCAP_TSO6) { reqcap &= ~IFCAP_TSO6; if_printf(ifp, "tso6 disabled due to -txcsum6\n"); } } /* * The kernel takes both IFCAP_TSOx and CSUM_TSO into * account before using TSO. So, we do not touch * checksum flags when IFCAP_TSOx is modified. * Note that CSUM_TSO is (CSUM_IP_TSO|CSUM_IP6_TSO), * but both bits are set in IPv4 and IPv6 mbufs. */ ifp->if_capenable = reqcap; SFXGE_ADAPTER_UNLOCK(sc); break; } case SIOCSIFMEDIA: case SIOCGIFMEDIA: error = ifmedia_ioctl(ifp, ifr, &sc->media, command); break; #ifdef SIOCGI2C case SIOCGI2C: { struct ifi2creq i2c; error = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c)); if (error != 0) break; if (i2c.len > sizeof(i2c.data)) { error = EINVAL; break; } SFXGE_ADAPTER_LOCK(sc); error = efx_phy_module_get_info(sc->enp, i2c.dev_addr, i2c.offset, i2c.len, &i2c.data[0]); SFXGE_ADAPTER_UNLOCK(sc); if (error == 0) error = copyout(&i2c, ifr_data_get_ptr(ifr), sizeof(i2c)); break; } #endif case SIOCGPRIVATE_0: error = priv_check(curthread, PRIV_DRIVER); if (error != 0) break; error = copyin(ifr_data_get_ptr(ifr), &ioc, sizeof(ioc)); if (error != 0) return (error); error = sfxge_private_ioctl(sc, &ioc); if (error == 0) { error = copyout(&ioc, ifr_data_get_ptr(ifr), sizeof(ioc)); } break; default: error = ether_ioctl(ifp, command, data); } return (error); } static void sfxge_ifnet_fini(struct ifnet *ifp) { struct sfxge_softc *sc = ifp->if_softc; SFXGE_ADAPTER_LOCK(sc); sfxge_stop(sc); SFXGE_ADAPTER_UNLOCK(sc); ifmedia_removeall(&sc->media); ether_ifdetach(ifp); if_free(ifp); } static int sfxge_ifnet_init(struct ifnet *ifp, struct sfxge_softc *sc) { const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp); device_t dev; int rc; dev = sc->dev; sc->ifnet = ifp; if_initname(ifp, device_get_name(dev), device_get_unit(dev)); ifp->if_init = sfxge_if_init; ifp->if_softc = sc; ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; ifp->if_ioctl = sfxge_if_ioctl; ifp->if_capabilities = SFXGE_CAP; ifp->if_capenable = SFXGE_CAP_ENABLE; ifp->if_hw_tsomax = SFXGE_TSO_MAX_SIZE; ifp->if_hw_tsomaxsegcount = SFXGE_TX_MAPPING_MAX_SEG; ifp->if_hw_tsomaxsegsize = PAGE_SIZE; #ifdef SFXGE_LRO ifp->if_capabilities |= IFCAP_LRO; ifp->if_capenable |= IFCAP_LRO; #endif if (encp->enc_hw_tx_insert_vlan_enabled) { ifp->if_capabilities |= IFCAP_VLAN_HWTAGGING; ifp->if_capenable |= IFCAP_VLAN_HWTAGGING; } ifp->if_hwassist = CSUM_TCP | CSUM_UDP | CSUM_IP | CSUM_TSO | CSUM_TCP_IPV6 | CSUM_UDP_IPV6; ether_ifattach(ifp, encp->enc_mac_addr); ifp->if_transmit = sfxge_if_transmit; ifp->if_qflush = sfxge_if_qflush; ifp->if_get_counter = sfxge_get_counter; DBGPRINT(sc->dev, "ifmedia_init"); if ((rc = sfxge_port_ifmedia_init(sc)) != 0) goto fail; return (0); fail: ether_ifdetach(sc->ifnet); return (rc); } void sfxge_sram_buf_tbl_alloc(struct sfxge_softc *sc, size_t n, uint32_t *idp) { KASSERT(sc->buffer_table_next + n <= efx_nic_cfg_get(sc->enp)->enc_buftbl_limit, ("buffer table full")); *idp = sc->buffer_table_next; sc->buffer_table_next += n; } static int sfxge_bar_init(struct sfxge_softc *sc) { efsys_bar_t *esbp = &sc->bar; esbp->esb_rid = PCIR_BAR(EFX_MEM_BAR); if ((esbp->esb_res = bus_alloc_resource_any(sc->dev, SYS_RES_MEMORY, &esbp->esb_rid, RF_ACTIVE)) == NULL) { device_printf(sc->dev, "Cannot allocate BAR region %d\n", EFX_MEM_BAR); return (ENXIO); } esbp->esb_tag = rman_get_bustag(esbp->esb_res); esbp->esb_handle = rman_get_bushandle(esbp->esb_res); SFXGE_BAR_LOCK_INIT(esbp, device_get_nameunit(sc->dev)); return (0); } static void sfxge_bar_fini(struct sfxge_softc *sc) { efsys_bar_t *esbp = &sc->bar; bus_release_resource(sc->dev, SYS_RES_MEMORY, esbp->esb_rid, esbp->esb_res); SFXGE_BAR_LOCK_DESTROY(esbp); } static int sfxge_create(struct sfxge_softc *sc) { device_t dev; efx_nic_t *enp; int error; char rss_param_name[sizeof(SFXGE_PARAM(%d.max_rss_channels))]; #if EFSYS_OPT_MCDI_LOGGING char mcdi_log_param_name[sizeof(SFXGE_PARAM(%d.mcdi_logging))]; #endif dev = sc->dev; SFXGE_ADAPTER_LOCK_INIT(sc, device_get_nameunit(sc->dev)); sc->max_rss_channels = 0; snprintf(rss_param_name, sizeof(rss_param_name), SFXGE_PARAM(%d.max_rss_channels), (int)device_get_unit(dev)); TUNABLE_INT_FETCH(rss_param_name, &sc->max_rss_channels); #if EFSYS_OPT_MCDI_LOGGING sc->mcdi_logging = sfxge_mcdi_logging; snprintf(mcdi_log_param_name, sizeof(mcdi_log_param_name), SFXGE_PARAM(%d.mcdi_logging), (int)device_get_unit(dev)); TUNABLE_INT_FETCH(mcdi_log_param_name, &sc->mcdi_logging); #endif sc->stats_node = SYSCTL_ADD_NODE( device_get_sysctl_ctx(dev), SYSCTL_CHILDREN(device_get_sysctl_tree(dev)), OID_AUTO, "stats", CTLFLAG_RD, NULL, "Statistics"); if (sc->stats_node == NULL) { error = ENOMEM; goto fail; } TASK_INIT(&sc->task_reset, 0, sfxge_reset, sc); (void) pci_enable_busmaster(dev); /* Initialize DMA mappings. */ DBGPRINT(sc->dev, "dma_init..."); if ((error = sfxge_dma_init(sc)) != 0) goto fail; /* Map the device registers. */ DBGPRINT(sc->dev, "bar_init..."); if ((error = sfxge_bar_init(sc)) != 0) goto fail; error = efx_family(pci_get_vendor(dev), pci_get_device(dev), &sc->family); KASSERT(error == 0, ("Family should be filtered by sfxge_probe()")); DBGPRINT(sc->dev, "nic_create..."); /* Create the common code nic object. */ SFXGE_EFSYS_LOCK_INIT(&sc->enp_lock, device_get_nameunit(sc->dev), "nic"); if ((error = efx_nic_create(sc->family, (efsys_identifier_t *)sc, &sc->bar, &sc->enp_lock, &enp)) != 0) goto fail3; sc->enp = enp; /* Initialize MCDI to talk to the microcontroller. */ DBGPRINT(sc->dev, "mcdi_init..."); if ((error = sfxge_mcdi_init(sc)) != 0) goto fail4; /* Probe the NIC and build the configuration data area. */ DBGPRINT(sc->dev, "nic_probe..."); if ((error = efx_nic_probe(enp)) != 0) goto fail5; if (!ISP2(sfxge_rx_ring_entries) || (sfxge_rx_ring_entries < EFX_RXQ_MINNDESCS) || (sfxge_rx_ring_entries > EFX_RXQ_MAXNDESCS)) { log(LOG_ERR, "%s=%d must be power of 2 from %u to %u", SFXGE_PARAM_RX_RING, sfxge_rx_ring_entries, EFX_RXQ_MINNDESCS, EFX_RXQ_MAXNDESCS); error = EINVAL; goto fail_rx_ring_entries; } sc->rxq_entries = sfxge_rx_ring_entries; if (!ISP2(sfxge_tx_ring_entries) || (sfxge_tx_ring_entries < EFX_TXQ_MINNDESCS) || (sfxge_tx_ring_entries > efx_nic_cfg_get(enp)->enc_txq_max_ndescs)) { log(LOG_ERR, "%s=%d must be power of 2 from %u to %u", SFXGE_PARAM_TX_RING, sfxge_tx_ring_entries, EFX_TXQ_MINNDESCS, efx_nic_cfg_get(enp)->enc_txq_max_ndescs); error = EINVAL; goto fail_tx_ring_entries; } sc->txq_entries = sfxge_tx_ring_entries; SYSCTL_ADD_STRING(device_get_sysctl_ctx(dev), SYSCTL_CHILDREN(device_get_sysctl_tree(dev)), OID_AUTO, "version", CTLFLAG_RD, SFXGE_VERSION_STRING, 0, "Driver version"); SYSCTL_ADD_UINT(device_get_sysctl_ctx(dev), SYSCTL_CHILDREN(device_get_sysctl_tree(dev)), OID_AUTO, "phy_type", CTLFLAG_RD, NULL, efx_nic_cfg_get(enp)->enc_phy_type, "PHY type"); /* Initialize the NVRAM. */ DBGPRINT(sc->dev, "nvram_init..."); if ((error = efx_nvram_init(enp)) != 0) goto fail6; /* Initialize the VPD. */ DBGPRINT(sc->dev, "vpd_init..."); if ((error = efx_vpd_init(enp)) != 0) goto fail7; efx_mcdi_new_epoch(enp); /* Reset the NIC. */ DBGPRINT(sc->dev, "nic_reset..."); if ((error = efx_nic_reset(enp)) != 0) goto fail8; /* Initialize buffer table allocation. */ sc->buffer_table_next = 0; /* * Guarantee minimum and estimate maximum number of event queues * to take it into account when MSI-X interrupts are allocated. * It initializes NIC and keeps it initialized on success. */ if ((error = sfxge_estimate_rsrc_limits(sc)) != 0) goto fail8; /* Set up interrupts. */ DBGPRINT(sc->dev, "intr_init..."); if ((error = sfxge_intr_init(sc)) != 0) goto fail9; /* Initialize event processing state. */ DBGPRINT(sc->dev, "ev_init..."); if ((error = sfxge_ev_init(sc)) != 0) goto fail11; /* Initialize port state. */ DBGPRINT(sc->dev, "port_init..."); if ((error = sfxge_port_init(sc)) != 0) goto fail12; /* Initialize receive state. */ DBGPRINT(sc->dev, "rx_init..."); if ((error = sfxge_rx_init(sc)) != 0) goto fail13; /* Initialize transmit state. */ DBGPRINT(sc->dev, "tx_init..."); if ((error = sfxge_tx_init(sc)) != 0) goto fail14; sc->init_state = SFXGE_INITIALIZED; DBGPRINT(sc->dev, "success"); return (0); fail14: sfxge_rx_fini(sc); fail13: sfxge_port_fini(sc); fail12: sfxge_ev_fini(sc); fail11: sfxge_intr_fini(sc); fail9: efx_nic_fini(sc->enp); fail8: efx_vpd_fini(enp); fail7: efx_nvram_fini(enp); fail6: fail_tx_ring_entries: fail_rx_ring_entries: efx_nic_unprobe(enp); fail5: sfxge_mcdi_fini(sc); fail4: sc->enp = NULL; efx_nic_destroy(enp); SFXGE_EFSYS_LOCK_DESTROY(&sc->enp_lock); fail3: sfxge_bar_fini(sc); (void) pci_disable_busmaster(sc->dev); fail: DBGPRINT(sc->dev, "failed %d", error); sc->dev = NULL; SFXGE_ADAPTER_LOCK_DESTROY(sc); return (error); } static void sfxge_destroy(struct sfxge_softc *sc) { efx_nic_t *enp; /* Clean up transmit state. */ sfxge_tx_fini(sc); /* Clean up receive state. */ sfxge_rx_fini(sc); /* Clean up port state. */ sfxge_port_fini(sc); /* Clean up event processing state. */ sfxge_ev_fini(sc); /* Clean up interrupts. */ sfxge_intr_fini(sc); /* Tear down common code subsystems. */ efx_nic_reset(sc->enp); efx_vpd_fini(sc->enp); efx_nvram_fini(sc->enp); efx_nic_unprobe(sc->enp); /* Tear down MCDI. */ sfxge_mcdi_fini(sc); /* Destroy common code context. */ enp = sc->enp; sc->enp = NULL; efx_nic_destroy(enp); /* Free DMA memory. */ sfxge_dma_fini(sc); /* Free mapped BARs. */ sfxge_bar_fini(sc); (void) pci_disable_busmaster(sc->dev); taskqueue_drain(taskqueue_thread, &sc->task_reset); /* Destroy the softc lock. */ SFXGE_ADAPTER_LOCK_DESTROY(sc); } static int sfxge_vpd_handler(SYSCTL_HANDLER_ARGS) { struct sfxge_softc *sc = arg1; efx_vpd_value_t value; int rc; value.evv_tag = arg2 >> 16; value.evv_keyword = arg2 & 0xffff; if ((rc = efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value)) != 0) return (rc); return (SYSCTL_OUT(req, value.evv_value, value.evv_length)); } static void sfxge_vpd_try_add(struct sfxge_softc *sc, struct sysctl_oid_list *list, efx_vpd_tag_t tag, const char *keyword) { struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); efx_vpd_value_t value; /* Check whether VPD tag/keyword is present */ value.evv_tag = tag; value.evv_keyword = EFX_VPD_KEYWORD(keyword[0], keyword[1]); if (efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value) != 0) return; SYSCTL_ADD_PROC( ctx, list, OID_AUTO, keyword, CTLTYPE_STRING|CTLFLAG_RD, sc, tag << 16 | EFX_VPD_KEYWORD(keyword[0], keyword[1]), sfxge_vpd_handler, "A", ""); } static int sfxge_vpd_init(struct sfxge_softc *sc) { struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); struct sysctl_oid *vpd_node; struct sysctl_oid_list *vpd_list; char keyword[3]; efx_vpd_value_t value; int rc; if ((rc = efx_vpd_size(sc->enp, &sc->vpd_size)) != 0) { /* * Unpriviledged functions deny VPD access. * Simply skip VPD in this case. */ if (rc == EACCES) goto done; goto fail; } sc->vpd_data = malloc(sc->vpd_size, M_SFXGE, M_WAITOK); if ((rc = efx_vpd_read(sc->enp, sc->vpd_data, sc->vpd_size)) != 0) goto fail2; /* Copy ID (product name) into device description, and log it. */ value.evv_tag = EFX_VPD_ID; if (efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value) == 0) { value.evv_value[value.evv_length] = 0; device_set_desc_copy(sc->dev, value.evv_value); device_printf(sc->dev, "%s\n", value.evv_value); } vpd_node = SYSCTL_ADD_NODE( ctx, SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), OID_AUTO, "vpd", CTLFLAG_RD, NULL, "Vital Product Data"); vpd_list = SYSCTL_CHILDREN(vpd_node); /* Add sysctls for all expected and any vendor-defined keywords. */ sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "PN"); sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "EC"); sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "SN"); keyword[0] = 'V'; keyword[2] = 0; for (keyword[1] = '0'; keyword[1] <= '9'; keyword[1]++) sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, keyword); for (keyword[1] = 'A'; keyword[1] <= 'Z'; keyword[1]++) sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, keyword); done: return (0); fail2: free(sc->vpd_data, M_SFXGE); fail: return (rc); } static void sfxge_vpd_fini(struct sfxge_softc *sc) { free(sc->vpd_data, M_SFXGE); } static void sfxge_reset(void *arg, int npending) { struct sfxge_softc *sc; int rc; unsigned attempt; (void)npending; sc = (struct sfxge_softc *)arg; SFXGE_ADAPTER_LOCK(sc); if (sc->init_state != SFXGE_STARTED) goto done; sfxge_stop(sc); efx_nic_reset(sc->enp); for (attempt = 0; attempt < sfxge_restart_attempts; ++attempt) { if ((rc = sfxge_start(sc)) == 0) goto done; device_printf(sc->dev, "start on reset failed (%d)\n", rc); DELAY(100000); } device_printf(sc->dev, "reset failed; interface is now stopped\n"); done: SFXGE_ADAPTER_UNLOCK(sc); } void sfxge_schedule_reset(struct sfxge_softc *sc) { taskqueue_enqueue(taskqueue_thread, &sc->task_reset); } static int sfxge_attach(device_t dev) { struct sfxge_softc *sc; struct ifnet *ifp; int error; sc = device_get_softc(dev); sc->dev = dev; /* Allocate ifnet. */ ifp = if_alloc(IFT_ETHER); if (ifp == NULL) { device_printf(dev, "Couldn't allocate ifnet\n"); error = ENOMEM; goto fail; } sc->ifnet = ifp; /* Initialize hardware. */ DBGPRINT(sc->dev, "create nic"); if ((error = sfxge_create(sc)) != 0) goto fail2; /* Create the ifnet for the port. */ DBGPRINT(sc->dev, "init ifnet"); if ((error = sfxge_ifnet_init(ifp, sc)) != 0) goto fail3; DBGPRINT(sc->dev, "init vpd"); if ((error = sfxge_vpd_init(sc)) != 0) goto fail4; /* * NIC is initialized inside sfxge_create() and kept inialized * to be able to initialize port to discover media types in * sfxge_ifnet_init(). */ efx_nic_fini(sc->enp); sc->init_state = SFXGE_REGISTERED; DBGPRINT(sc->dev, "success"); return (0); fail4: sfxge_ifnet_fini(ifp); fail3: efx_nic_fini(sc->enp); sfxge_destroy(sc); fail2: if_free(sc->ifnet); fail: DBGPRINT(sc->dev, "failed %d", error); return (error); } static int sfxge_detach(device_t dev) { struct sfxge_softc *sc; sc = device_get_softc(dev); sfxge_vpd_fini(sc); /* Destroy the ifnet. */ sfxge_ifnet_fini(sc->ifnet); /* Tear down hardware. */ sfxge_destroy(sc); return (0); } static int sfxge_probe(device_t dev) { uint16_t pci_vendor_id; uint16_t pci_device_id; efx_family_t family; int rc; pci_vendor_id = pci_get_vendor(dev); pci_device_id = pci_get_device(dev); DBGPRINT(dev, "PCI ID %04x:%04x", pci_vendor_id, pci_device_id); rc = efx_family(pci_vendor_id, pci_device_id, &family); if (rc != 0) { DBGPRINT(dev, "efx_family fail %d", rc); return (ENXIO); } if (family == EFX_FAMILY_SIENA) { device_set_desc(dev, "Solarflare SFC9000 family"); return (0); } if (family == EFX_FAMILY_HUNTINGTON) { device_set_desc(dev, "Solarflare SFC9100 family"); return (0); } if (family == EFX_FAMILY_MEDFORD) { device_set_desc(dev, "Solarflare SFC9200 family"); return (0); } DBGPRINT(dev, "impossible controller family %d", family); return (ENXIO); } static device_method_t sfxge_methods[] = { DEVMETHOD(device_probe, sfxge_probe), DEVMETHOD(device_attach, sfxge_attach), DEVMETHOD(device_detach, sfxge_detach), DEVMETHOD_END }; static devclass_t sfxge_devclass; static driver_t sfxge_driver = { "sfxge", sfxge_methods, sizeof(struct sfxge_softc) }; DRIVER_MODULE(sfxge, pci, sfxge_driver, sfxge_devclass, 0, 0); Index: stable/11/sys/dev/sfxge/sfxge_ev.c =================================================================== --- stable/11/sys/dev/sfxge/sfxge_ev.c (revision 342453) +++ stable/11/sys/dev/sfxge/sfxge_ev.c (revision 342454) @@ -1,1038 +1,1039 @@ /*- * Copyright (c) 2010-2016 Solarflare Communications Inc. * All rights reserved. * * This software was developed in part by Philip Paeps under contract for * Solarflare Communications, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * The views and conclusions contained in the software and documentation are * those of the authors and should not be interpreted as representing official * policies, either expressed or implied, of the FreeBSD Project. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include "common/efx.h" #include "sfxge.h" static void sfxge_ev_qcomplete(struct sfxge_evq *evq, boolean_t eop) { struct sfxge_softc *sc; unsigned int index; struct sfxge_rxq *rxq; struct sfxge_txq *txq; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; index = evq->index; rxq = sc->rxq[index]; if ((txq = evq->txq) != NULL) { evq->txq = NULL; evq->txqs = &(evq->txq); do { struct sfxge_txq *next; next = txq->next; txq->next = NULL; KASSERT(txq->evq_index == index, ("txq->evq_index != index")); if (txq->pending != txq->completed) sfxge_tx_qcomplete(txq, evq); txq = next; } while (txq != NULL); } if (rxq->pending != rxq->completed) sfxge_rx_qcomplete(rxq, eop); } static struct sfxge_rxq * sfxge_get_rxq_by_label(struct sfxge_evq *evq, uint32_t label) { struct sfxge_rxq *rxq; KASSERT(label == 0, ("unexpected rxq label != 0")); rxq = evq->sc->rxq[evq->index]; KASSERT(rxq != NULL, ("rxq == NULL")); KASSERT(evq->index == rxq->index, ("evq->index != rxq->index")); return (rxq); } static boolean_t sfxge_ev_rx(void *arg, uint32_t label, uint32_t id, uint32_t size, uint16_t flags) { struct sfxge_evq *evq; struct sfxge_softc *sc; struct sfxge_rxq *rxq; unsigned int stop; unsigned int delta; struct sfxge_rx_sw_desc *rx_desc; evq = arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; if (evq->exception) goto done; rxq = sfxge_get_rxq_by_label(evq, label); if (__predict_false(rxq->init_state != SFXGE_RXQ_STARTED)) goto done; stop = (id + 1) & rxq->ptr_mask; id = rxq->pending & rxq->ptr_mask; delta = (stop >= id) ? (stop - id) : (rxq->entries - id + stop); rxq->pending += delta; if (delta != 1) { if ((delta <= 0) || (delta > efx_nic_cfg_get(sc->enp)->enc_rx_batch_max)) { evq->exception = B_TRUE; device_printf(sc->dev, "RX completion out of order" " (id=%#x delta=%u flags=%#x); resetting\n", id, delta, flags); sfxge_schedule_reset(sc); goto done; } } rx_desc = &rxq->queue[id]; prefetch_read_many(rx_desc->mbuf); for (; id != stop; id = (id + 1) & rxq->ptr_mask) { rx_desc = &rxq->queue[id]; KASSERT(rx_desc->flags == EFX_DISCARD, ("rx_desc->flags != EFX_DISCARD")); rx_desc->flags = flags; KASSERT(size < (1 << 16), ("size > (1 << 16)")); rx_desc->size = (uint16_t)size; } evq->rx_done++; if (rxq->pending - rxq->completed >= SFXGE_RX_BATCH) sfxge_ev_qcomplete(evq, B_FALSE); done: return (evq->rx_done >= SFXGE_EV_BATCH); } static boolean_t sfxge_ev_exception(void *arg, uint32_t code, uint32_t data) { struct sfxge_evq *evq; struct sfxge_softc *sc; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; DBGPRINT(sc->dev, "[%d] %s", evq->index, (code == EFX_EXCEPTION_RX_RECOVERY) ? "RX_RECOVERY" : (code == EFX_EXCEPTION_RX_DSC_ERROR) ? "RX_DSC_ERROR" : (code == EFX_EXCEPTION_TX_DSC_ERROR) ? "TX_DSC_ERROR" : (code == EFX_EXCEPTION_UNKNOWN_SENSOREVT) ? "UNKNOWN_SENSOREVT" : (code == EFX_EXCEPTION_FWALERT_SRAM) ? "FWALERT_SRAM" : (code == EFX_EXCEPTION_UNKNOWN_FWALERT) ? "UNKNOWN_FWALERT" : (code == EFX_EXCEPTION_RX_ERROR) ? "RX_ERROR" : (code == EFX_EXCEPTION_TX_ERROR) ? "TX_ERROR" : (code == EFX_EXCEPTION_EV_ERROR) ? "EV_ERROR" : "UNKNOWN"); evq->exception = B_TRUE; if (code != EFX_EXCEPTION_UNKNOWN_SENSOREVT) { device_printf(sc->dev, "hardware exception (code=%u); resetting\n", code); sfxge_schedule_reset(sc); } return (B_FALSE); } static boolean_t sfxge_ev_rxq_flush_done(void *arg, uint32_t rxq_index) { struct sfxge_evq *evq; struct sfxge_softc *sc; struct sfxge_rxq *rxq; unsigned int index; uint16_t magic; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; rxq = sc->rxq[rxq_index]; KASSERT(rxq != NULL, ("rxq == NULL")); /* Resend a software event on the correct queue */ index = rxq->index; if (index == evq->index) { sfxge_rx_qflush_done(rxq); return (B_FALSE); } evq = sc->evq[index]; magic = sfxge_sw_ev_rxq_magic(SFXGE_SW_EV_RX_QFLUSH_DONE, rxq); KASSERT(evq->init_state == SFXGE_EVQ_STARTED, ("evq not started")); efx_ev_qpost(evq->common, magic); return (B_FALSE); } static boolean_t sfxge_ev_rxq_flush_failed(void *arg, uint32_t rxq_index) { struct sfxge_evq *evq; struct sfxge_softc *sc; struct sfxge_rxq *rxq; unsigned int index; uint16_t magic; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; rxq = sc->rxq[rxq_index]; KASSERT(rxq != NULL, ("rxq == NULL")); /* Resend a software event on the correct queue */ index = rxq->index; evq = sc->evq[index]; magic = sfxge_sw_ev_rxq_magic(SFXGE_SW_EV_RX_QFLUSH_FAILED, rxq); KASSERT(evq->init_state == SFXGE_EVQ_STARTED, ("evq not started")); efx_ev_qpost(evq->common, magic); return (B_FALSE); } static struct sfxge_txq * sfxge_get_txq_by_label(struct sfxge_evq *evq, enum sfxge_txq_type label) { unsigned int index; - KASSERT((evq->index == 0 && label < SFXGE_TXQ_NTYPES) || + KASSERT((evq->index == 0 && label < SFXGE_EVQ0_N_TXQ(evq->sc)) || (label == SFXGE_TXQ_IP_TCP_UDP_CKSUM), ("unexpected txq label")); - index = (evq->index == 0) ? label : (evq->index - 1 + SFXGE_TXQ_NTYPES); + index = (evq->index == 0) ? + label : (evq->index - 1 + SFXGE_EVQ0_N_TXQ(evq->sc)); return (evq->sc->txq[index]); } static boolean_t sfxge_ev_tx(void *arg, uint32_t label, uint32_t id) { struct sfxge_evq *evq; struct sfxge_txq *txq; unsigned int stop; unsigned int delta; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); txq = sfxge_get_txq_by_label(evq, label); KASSERT(txq != NULL, ("txq == NULL")); KASSERT(evq->index == txq->evq_index, ("evq->index != txq->evq_index")); if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) goto done; stop = (id + 1) & txq->ptr_mask; id = txq->pending & txq->ptr_mask; delta = (stop >= id) ? (stop - id) : (txq->entries - id + stop); txq->pending += delta; evq->tx_done++; if (txq->next == NULL && evq->txqs != &(txq->next)) { *(evq->txqs) = txq; evq->txqs = &(txq->next); } if (txq->pending - txq->completed >= SFXGE_TX_BATCH) sfxge_tx_qcomplete(txq, evq); done: return (evq->tx_done >= SFXGE_EV_BATCH); } static boolean_t sfxge_ev_txq_flush_done(void *arg, uint32_t txq_index) { struct sfxge_evq *evq; struct sfxge_softc *sc; struct sfxge_txq *txq; uint16_t magic; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; txq = sc->txq[txq_index]; KASSERT(txq != NULL, ("txq == NULL")); KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, ("txq not initialized")); if (txq->evq_index == evq->index) { sfxge_tx_qflush_done(txq); return (B_FALSE); } /* Resend a software event on the correct queue */ evq = sc->evq[txq->evq_index]; magic = sfxge_sw_ev_txq_magic(SFXGE_SW_EV_TX_QFLUSH_DONE, txq); KASSERT(evq->init_state == SFXGE_EVQ_STARTED, ("evq not started")); efx_ev_qpost(evq->common, magic); return (B_FALSE); } static boolean_t sfxge_ev_software(void *arg, uint16_t magic) { struct sfxge_evq *evq; struct sfxge_softc *sc; unsigned int label; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; label = magic & SFXGE_MAGIC_DMAQ_LABEL_MASK; magic &= ~SFXGE_MAGIC_DMAQ_LABEL_MASK; switch (magic) { case SFXGE_SW_EV_MAGIC(SFXGE_SW_EV_RX_QFLUSH_DONE): sfxge_rx_qflush_done(sfxge_get_rxq_by_label(evq, label)); break; case SFXGE_SW_EV_MAGIC(SFXGE_SW_EV_RX_QFLUSH_FAILED): sfxge_rx_qflush_failed(sfxge_get_rxq_by_label(evq, label)); break; case SFXGE_SW_EV_MAGIC(SFXGE_SW_EV_RX_QREFILL): sfxge_rx_qrefill(sfxge_get_rxq_by_label(evq, label)); break; case SFXGE_SW_EV_MAGIC(SFXGE_SW_EV_TX_QFLUSH_DONE): { struct sfxge_txq *txq = sfxge_get_txq_by_label(evq, label); KASSERT(txq != NULL, ("txq == NULL")); KASSERT(evq->index == txq->evq_index, ("evq->index != txq->evq_index")); sfxge_tx_qflush_done(txq); break; } default: break; } return (B_FALSE); } static boolean_t sfxge_ev_sram(void *arg, uint32_t code) { (void)arg; (void)code; switch (code) { case EFX_SRAM_UPDATE: EFSYS_PROBE(sram_update); break; case EFX_SRAM_CLEAR: EFSYS_PROBE(sram_clear); break; case EFX_SRAM_ILLEGAL_CLEAR: EFSYS_PROBE(sram_illegal_clear); break; default: KASSERT(B_FALSE, ("Impossible SRAM event")); break; } return (B_FALSE); } static boolean_t sfxge_ev_timer(void *arg, uint32_t index) { (void)arg; (void)index; return (B_FALSE); } static boolean_t sfxge_ev_wake_up(void *arg, uint32_t index) { (void)arg; (void)index; return (B_FALSE); } #if EFSYS_OPT_QSTATS static void sfxge_evq_stat_update(struct sfxge_evq *evq) { clock_t now; SFXGE_EVQ_LOCK(evq); if (__predict_false(evq->init_state != SFXGE_EVQ_STARTED)) goto out; now = ticks; if ((unsigned int)(now - evq->stats_update_time) < (unsigned int)hz) goto out; evq->stats_update_time = now; efx_ev_qstats_update(evq->common, evq->stats); out: SFXGE_EVQ_UNLOCK(evq); } static int sfxge_evq_stat_handler(SYSCTL_HANDLER_ARGS) { struct sfxge_evq *evq = arg1; struct sfxge_softc *sc = evq->sc; unsigned int id = arg2; SFXGE_ADAPTER_LOCK(sc); sfxge_evq_stat_update(evq); SFXGE_ADAPTER_UNLOCK(sc); return (SYSCTL_OUT(req, &evq->stats[id], sizeof(evq->stats[id]))); } static int sfxge_evq_stat_init(struct sfxge_evq *evq) { struct sfxge_softc *sc = evq->sc; struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); char name[16]; struct sysctl_oid *evq_stats_node; unsigned int id; snprintf(name, sizeof(name), "%u", evq->index); evq_stats_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->evqs_stats_node), OID_AUTO, name, CTLFLAG_RD, NULL, ""); if (evq_stats_node == NULL) return (ENOMEM); for (id = 0; id < EV_NQSTATS; id++) { SYSCTL_ADD_PROC( ctx, SYSCTL_CHILDREN(evq_stats_node), OID_AUTO, efx_ev_qstat_name(sc->enp, id), CTLTYPE_U64|CTLFLAG_RD, evq, id, sfxge_evq_stat_handler, "Q", ""); } return (0); } static void sfxge_ev_stat_update(struct sfxge_softc *sc) { struct sfxge_evq *evq; unsigned int index; clock_t now; unsigned int id; SFXGE_ADAPTER_LOCK(sc); now = ticks; if ((unsigned int)(now - sc->ev_stats_update_time) < (unsigned int)hz) goto out; sc->ev_stats_update_time = now; memset(sc->ev_stats, 0, sizeof(sc->ev_stats)); /* Update and add event counts from each event queue in turn */ for (index = 0; index < sc->evq_count; index++) { evq = sc->evq[index]; sfxge_evq_stat_update(evq); for (id = 0; id < EV_NQSTATS; id++) sc->ev_stats[id] += evq->stats[id]; } out: SFXGE_ADAPTER_UNLOCK(sc); } static int sfxge_ev_stat_handler(SYSCTL_HANDLER_ARGS) { struct sfxge_softc *sc = arg1; unsigned int id = arg2; sfxge_ev_stat_update(sc); return (SYSCTL_OUT(req, &sc->ev_stats[id], sizeof(sc->ev_stats[id]))); } static void sfxge_ev_stat_init(struct sfxge_softc *sc) { struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); struct sysctl_oid_list *stat_list; unsigned int id; char name[40]; stat_list = SYSCTL_CHILDREN(sc->stats_node); for (id = 0; id < EV_NQSTATS; id++) { snprintf(name, sizeof(name), "ev_%s", efx_ev_qstat_name(sc->enp, id)); SYSCTL_ADD_PROC( ctx, stat_list, OID_AUTO, name, CTLTYPE_U64|CTLFLAG_RD, sc, id, sfxge_ev_stat_handler, "Q", ""); } } #endif /* EFSYS_OPT_QSTATS */ static void sfxge_ev_qmoderate(struct sfxge_softc *sc, unsigned int idx, unsigned int us) { struct sfxge_evq *evq; efx_evq_t *eep; evq = sc->evq[idx]; eep = evq->common; KASSERT(evq->init_state == SFXGE_EVQ_STARTED, ("evq->init_state != SFXGE_EVQ_STARTED")); (void)efx_ev_qmoderate(eep, us); } static int sfxge_int_mod_handler(SYSCTL_HANDLER_ARGS) { struct sfxge_softc *sc = arg1; struct sfxge_intr *intr = &sc->intr; unsigned int moderation; int error; unsigned int index; SFXGE_ADAPTER_LOCK(sc); if (req->newptr != NULL) { if ((error = SYSCTL_IN(req, &moderation, sizeof(moderation))) != 0) goto out; /* We may not be calling efx_ev_qmoderate() now, * so we have to range-check the value ourselves. */ if (moderation > efx_nic_cfg_get(sc->enp)->enc_evq_timer_max_us) { error = EINVAL; goto out; } sc->ev_moderation = moderation; if (intr->state == SFXGE_INTR_STARTED) { for (index = 0; index < sc->evq_count; index++) sfxge_ev_qmoderate(sc, index, moderation); } } else { error = SYSCTL_OUT(req, &sc->ev_moderation, sizeof(sc->ev_moderation)); } out: SFXGE_ADAPTER_UNLOCK(sc); return (error); } static boolean_t sfxge_ev_initialized(void *arg) { struct sfxge_evq *evq; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); /* Init done events may be duplicated on 7xxx */ KASSERT(evq->init_state == SFXGE_EVQ_STARTING || evq->init_state == SFXGE_EVQ_STARTED, ("evq not starting")); evq->init_state = SFXGE_EVQ_STARTED; return (0); } static boolean_t sfxge_ev_link_change(void *arg, efx_link_mode_t link_mode) { struct sfxge_evq *evq; struct sfxge_softc *sc; evq = (struct sfxge_evq *)arg; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); sc = evq->sc; sfxge_mac_link_update(sc, link_mode); return (0); } static const efx_ev_callbacks_t sfxge_ev_callbacks = { .eec_initialized = sfxge_ev_initialized, .eec_rx = sfxge_ev_rx, .eec_tx = sfxge_ev_tx, .eec_exception = sfxge_ev_exception, .eec_rxq_flush_done = sfxge_ev_rxq_flush_done, .eec_rxq_flush_failed = sfxge_ev_rxq_flush_failed, .eec_txq_flush_done = sfxge_ev_txq_flush_done, .eec_software = sfxge_ev_software, .eec_sram = sfxge_ev_sram, .eec_wake_up = sfxge_ev_wake_up, .eec_timer = sfxge_ev_timer, .eec_link_change = sfxge_ev_link_change, }; int sfxge_ev_qpoll(struct sfxge_evq *evq) { int rc; SFXGE_EVQ_LOCK(evq); if (__predict_false(evq->init_state != SFXGE_EVQ_STARTING && evq->init_state != SFXGE_EVQ_STARTED)) { rc = EINVAL; goto fail; } /* Synchronize the DMA memory for reading */ bus_dmamap_sync(evq->mem.esm_tag, evq->mem.esm_map, BUS_DMASYNC_POSTREAD); KASSERT(evq->rx_done == 0, ("evq->rx_done != 0")); KASSERT(evq->tx_done == 0, ("evq->tx_done != 0")); KASSERT(evq->txq == NULL, ("evq->txq != NULL")); KASSERT(evq->txqs == &evq->txq, ("evq->txqs != &evq->txq")); /* Poll the queue */ efx_ev_qpoll(evq->common, &evq->read_ptr, &sfxge_ev_callbacks, evq); evq->rx_done = 0; evq->tx_done = 0; /* Perform any pending completion processing */ sfxge_ev_qcomplete(evq, B_TRUE); /* Re-prime the event queue for interrupts */ if ((rc = efx_ev_qprime(evq->common, evq->read_ptr)) != 0) goto fail; SFXGE_EVQ_UNLOCK(evq); return (0); fail: SFXGE_EVQ_UNLOCK(evq); return (rc); } static void sfxge_ev_qstop(struct sfxge_softc *sc, unsigned int index) { struct sfxge_evq *evq; evq = sc->evq[index]; KASSERT(evq->init_state == SFXGE_EVQ_STARTED, ("evq->init_state != SFXGE_EVQ_STARTED")); SFXGE_EVQ_LOCK(evq); evq->init_state = SFXGE_EVQ_INITIALIZED; evq->read_ptr = 0; evq->exception = B_FALSE; #if EFSYS_OPT_QSTATS /* Add event counts before discarding the common evq state */ efx_ev_qstats_update(evq->common, evq->stats); #endif efx_ev_qdestroy(evq->common); efx_sram_buf_tbl_clear(sc->enp, evq->buf_base_id, EFX_EVQ_NBUFS(evq->entries)); SFXGE_EVQ_UNLOCK(evq); } static int sfxge_ev_qstart(struct sfxge_softc *sc, unsigned int index) { struct sfxge_evq *evq; efsys_mem_t *esmp; int count; int rc; evq = sc->evq[index]; esmp = &evq->mem; KASSERT(evq->init_state == SFXGE_EVQ_INITIALIZED, ("evq->init_state != SFXGE_EVQ_INITIALIZED")); /* Clear all events. */ (void)memset(esmp->esm_base, 0xff, EFX_EVQ_SIZE(evq->entries)); /* Program the buffer table. */ if ((rc = efx_sram_buf_tbl_set(sc->enp, evq->buf_base_id, esmp, EFX_EVQ_NBUFS(evq->entries))) != 0) return (rc); /* Create the common code event queue. */ if ((rc = efx_ev_qcreate(sc->enp, index, esmp, evq->entries, evq->buf_base_id, sc->ev_moderation, EFX_EVQ_FLAGS_TYPE_AUTO, &evq->common)) != 0) goto fail; SFXGE_EVQ_LOCK(evq); /* Prime the event queue for interrupts */ if ((rc = efx_ev_qprime(evq->common, evq->read_ptr)) != 0) goto fail2; evq->init_state = SFXGE_EVQ_STARTING; SFXGE_EVQ_UNLOCK(evq); /* Wait for the initialization event */ count = 0; do { /* Pause for 100 ms */ pause("sfxge evq init", hz / 10); /* Check to see if the test event has been processed */ if (evq->init_state == SFXGE_EVQ_STARTED) goto done; } while (++count < 20); rc = ETIMEDOUT; goto fail3; done: return (0); fail3: SFXGE_EVQ_LOCK(evq); evq->init_state = SFXGE_EVQ_INITIALIZED; fail2: SFXGE_EVQ_UNLOCK(evq); efx_ev_qdestroy(evq->common); fail: efx_sram_buf_tbl_clear(sc->enp, evq->buf_base_id, EFX_EVQ_NBUFS(evq->entries)); return (rc); } void sfxge_ev_stop(struct sfxge_softc *sc) { struct sfxge_intr *intr; efx_nic_t *enp; int index; intr = &sc->intr; enp = sc->enp; KASSERT(intr->state == SFXGE_INTR_STARTED, ("Interrupts not started")); /* Stop the event queue(s) */ index = sc->evq_count; while (--index >= 0) sfxge_ev_qstop(sc, index); /* Tear down the event module */ efx_ev_fini(enp); } int sfxge_ev_start(struct sfxge_softc *sc) { struct sfxge_intr *intr; int index; int rc; intr = &sc->intr; KASSERT(intr->state == SFXGE_INTR_STARTED, ("intr->state != SFXGE_INTR_STARTED")); /* Initialize the event module */ if ((rc = efx_ev_init(sc->enp)) != 0) return (rc); /* Start the event queues */ for (index = 0; index < sc->evq_count; index++) { if ((rc = sfxge_ev_qstart(sc, index)) != 0) goto fail; } return (0); fail: /* Stop the event queue(s) */ while (--index >= 0) sfxge_ev_qstop(sc, index); /* Tear down the event module */ efx_ev_fini(sc->enp); return (rc); } static void sfxge_ev_qfini(struct sfxge_softc *sc, unsigned int index) { struct sfxge_evq *evq; evq = sc->evq[index]; KASSERT(evq->init_state == SFXGE_EVQ_INITIALIZED, ("evq->init_state != SFXGE_EVQ_INITIALIZED")); KASSERT(evq->txqs == &evq->txq, ("evq->txqs != &evq->txq")); sfxge_dma_free(&evq->mem); sc->evq[index] = NULL; SFXGE_EVQ_LOCK_DESTROY(evq); free(evq, M_SFXGE); } static int sfxge_ev_qinit(struct sfxge_softc *sc, unsigned int index) { struct sfxge_evq *evq; efsys_mem_t *esmp; int rc; KASSERT(index < SFXGE_RX_SCALE_MAX, ("index >= SFXGE_RX_SCALE_MAX")); evq = malloc(sizeof(struct sfxge_evq), M_SFXGE, M_ZERO | M_WAITOK); evq->sc = sc; evq->index = index; sc->evq[index] = evq; esmp = &evq->mem; /* Build an event queue with room for one event per tx and rx buffer, * plus some extra for link state events and MCDI completions. * There are three tx queues in the first event queue and one in * other. */ if (index == 0) evq->entries = ROUNDUP_POW_OF_TWO(sc->rxq_entries + 3 * sc->txq_entries + 128); else evq->entries = ROUNDUP_POW_OF_TWO(sc->rxq_entries + sc->txq_entries + 128); /* Initialise TX completion list */ evq->txqs = &evq->txq; /* Allocate DMA space. */ if ((rc = sfxge_dma_alloc(sc, EFX_EVQ_SIZE(evq->entries), esmp)) != 0) return (rc); /* Allocate buffer table entries. */ sfxge_sram_buf_tbl_alloc(sc, EFX_EVQ_NBUFS(evq->entries), &evq->buf_base_id); SFXGE_EVQ_LOCK_INIT(evq, device_get_nameunit(sc->dev), index); evq->init_state = SFXGE_EVQ_INITIALIZED; #if EFSYS_OPT_QSTATS rc = sfxge_evq_stat_init(evq); if (rc != 0) goto fail_evq_stat_init; #endif return (0); #if EFSYS_OPT_QSTATS fail_evq_stat_init: evq->init_state = SFXGE_EVQ_UNINITIALIZED; SFXGE_EVQ_LOCK_DESTROY(evq); sfxge_dma_free(esmp); sc->evq[index] = NULL; free(evq, M_SFXGE); return (rc); #endif } void sfxge_ev_fini(struct sfxge_softc *sc) { struct sfxge_intr *intr; int index; intr = &sc->intr; KASSERT(intr->state == SFXGE_INTR_INITIALIZED, ("intr->state != SFXGE_INTR_INITIALIZED")); sc->ev_moderation = 0; /* Tear down the event queue(s). */ index = sc->evq_count; while (--index >= 0) sfxge_ev_qfini(sc, index); sc->evq_count = 0; } int sfxge_ev_init(struct sfxge_softc *sc) { struct sysctl_ctx_list *sysctl_ctx = device_get_sysctl_ctx(sc->dev); struct sysctl_oid *sysctl_tree = device_get_sysctl_tree(sc->dev); struct sfxge_intr *intr; int index; int rc; intr = &sc->intr; sc->evq_count = intr->n_alloc; KASSERT(intr->state == SFXGE_INTR_INITIALIZED, ("intr->state != SFXGE_INTR_INITIALIZED")); /* Set default interrupt moderation; add a sysctl to * read and change it. */ sc->ev_moderation = SFXGE_MODERATION; SYSCTL_ADD_PROC(sysctl_ctx, SYSCTL_CHILDREN(sysctl_tree), OID_AUTO, "int_mod", CTLTYPE_UINT|CTLFLAG_RW, sc, 0, sfxge_int_mod_handler, "IU", "sfxge interrupt moderation (us)"); #if EFSYS_OPT_QSTATS sc->evqs_stats_node = SYSCTL_ADD_NODE( device_get_sysctl_ctx(sc->dev), SYSCTL_CHILDREN(sc->stats_node), OID_AUTO, "evq", CTLFLAG_RD, NULL, "Event queues stats"); if (sc->evqs_stats_node == NULL) { rc = ENOMEM; goto fail_evqs_stats_node; } #endif /* * Initialize the event queue(s) - one per interrupt. */ for (index = 0; index < sc->evq_count; index++) { if ((rc = sfxge_ev_qinit(sc, index)) != 0) goto fail; } #if EFSYS_OPT_QSTATS sfxge_ev_stat_init(sc); #endif return (0); fail: while (--index >= 0) sfxge_ev_qfini(sc, index); #if EFSYS_OPT_QSTATS fail_evqs_stats_node: #endif sc->evq_count = 0; return (rc); } Index: stable/11/sys/dev/sfxge/sfxge_tx.c =================================================================== --- stable/11/sys/dev/sfxge/sfxge_tx.c (revision 342453) +++ stable/11/sys/dev/sfxge/sfxge_tx.c (revision 342454) @@ -1,2029 +1,2029 @@ /*- * Copyright (c) 2010-2016 Solarflare Communications Inc. * All rights reserved. * * This software was developed in part by Philip Paeps under contract for * Solarflare Communications, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * The views and conclusions contained in the software and documentation are * those of the authors and should not be interpreted as representing official * policies, either expressed or implied, of the FreeBSD Project. */ /* Theory of operation: * * Tx queues allocation and mapping * * One Tx queue with enabled checksum offload is allocated per Rx channel * (event queue). Also 2 Tx queues (one without checksum offload and one * with IP checksum offload only) are allocated and bound to event queue 0. * sfxge_txq_type is used as Tx queue label. * * So, event queue plus label mapping to Tx queue index is: * if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES) * else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1 * See sfxge_get_txq_by_label() sfxge_ev.c */ #include __FBSDID("$FreeBSD$"); #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef RSS #include #endif #include "common/efx.h" #include "sfxge.h" #include "sfxge_tx.h" #define SFXGE_PARAM_TX_DPL_GET_MAX SFXGE_PARAM(tx_dpl_get_max) static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT; TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max); SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN, &sfxge_tx_dpl_get_max, 0, "Maximum number of any packets in deferred packet get-list"); #define SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \ SFXGE_PARAM(tx_dpl_get_non_tcp_max) static int sfxge_tx_dpl_get_non_tcp_max = SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT; TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max); SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN, &sfxge_tx_dpl_get_non_tcp_max, 0, "Maximum number of non-TCP packets in deferred packet get-list"); #define SFXGE_PARAM_TX_DPL_PUT_MAX SFXGE_PARAM(tx_dpl_put_max) static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT; TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max); SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN, &sfxge_tx_dpl_put_max, 0, "Maximum number of any packets in deferred packet put-list"); #define SFXGE_PARAM_TSO_FW_ASSISTED SFXGE_PARAM(tso_fw_assisted) static int sfxge_tso_fw_assisted = (SFXGE_FATSOV1 | SFXGE_FATSOV2); TUNABLE_INT(SFXGE_PARAM_TSO_FW_ASSISTED, &sfxge_tso_fw_assisted); SYSCTL_INT(_hw_sfxge, OID_AUTO, tso_fw_assisted, CTLFLAG_RDTUN, &sfxge_tso_fw_assisted, 0, "Bitmask of FW-assisted TSO allowed to use if supported by NIC firmware"); static const struct { const char *name; size_t offset; } sfxge_tx_stats[] = { #define SFXGE_TX_STAT(name, member) \ { #name, offsetof(struct sfxge_txq, member) } SFXGE_TX_STAT(tso_bursts, tso_bursts), SFXGE_TX_STAT(tso_packets, tso_packets), SFXGE_TX_STAT(tso_long_headers, tso_long_headers), SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many), SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc), SFXGE_TX_STAT(tx_collapses, collapses), SFXGE_TX_STAT(tx_drops, drops), SFXGE_TX_STAT(tx_get_overflow, get_overflow), SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow), SFXGE_TX_STAT(tx_put_overflow, put_overflow), SFXGE_TX_STAT(tx_netdown_drops, netdown_drops), }; /* Forward declarations. */ static void sfxge_tx_qdpl_service(struct sfxge_txq *txq); static void sfxge_tx_qlist_post(struct sfxge_txq *txq); static void sfxge_tx_qunblock(struct sfxge_txq *txq); static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, const bus_dma_segment_t *dma_seg, int n_dma_seg, int vlan_tagged); static int sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf) { uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ? mbuf->m_pkthdr.ether_vtag : 0); if (this_tag == txq->hw_vlan_tci) return (0); efx_tx_qdesc_vlantci_create(txq->common, bswap16(this_tag), &txq->pend_desc[0]); txq->n_pend_desc = 1; txq->hw_vlan_tci = this_tag; return (1); } static inline void sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp) { KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0")); if (__predict_false(*pstmp == &txq->stmp[txq->ptr_mask])) *pstmp = &txq->stmp[0]; else (*pstmp)++; } void sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq) { unsigned int completed; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); completed = txq->completed; while (completed != txq->pending) { struct sfxge_tx_mapping *stmp; unsigned int id; id = completed++ & txq->ptr_mask; stmp = &txq->stmp[id]; if (stmp->flags & TX_BUF_UNMAP) { bus_dmamap_unload(txq->packet_dma_tag, stmp->map); if (stmp->flags & TX_BUF_MBUF) { struct mbuf *m = stmp->u.mbuf; do m = m_free(m); while (m != NULL); } else { free(stmp->u.heap_buf, M_SFXGE); } stmp->flags = 0; } } txq->completed = completed; /* Check whether we need to unblock the queue. */ mb(); if (txq->blocked) { unsigned int level; level = txq->added - txq->completed; if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) sfxge_tx_qunblock(txq); } } static unsigned int sfxge_is_mbuf_non_tcp(struct mbuf *mbuf) { /* Absence of TCP checksum flags does not mean that it is non-TCP * but it should be true if user wants to achieve high throughput. */ return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP))); } /* * Reorder the put list and append it to the get list. */ static void sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq) { struct sfxge_tx_dpl *stdp; struct mbuf *mbuf, *get_next, **get_tailp; volatile uintptr_t *putp; uintptr_t put; unsigned int count; unsigned int non_tcp_count; SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); stdp = &txq->dpl; /* Acquire the put list. */ putp = &stdp->std_put; put = atomic_readandclear_ptr(putp); mbuf = (void *)put; if (mbuf == NULL) return; /* Reverse the put list. */ get_tailp = &mbuf->m_nextpkt; get_next = NULL; count = 0; non_tcp_count = 0; do { struct mbuf *put_next; non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf); put_next = mbuf->m_nextpkt; mbuf->m_nextpkt = get_next; get_next = mbuf; mbuf = put_next; count++; } while (mbuf != NULL); if (count > stdp->std_put_hiwat) stdp->std_put_hiwat = count; /* Append the reversed put list to the get list. */ KASSERT(*get_tailp == NULL, ("*get_tailp != NULL")); *stdp->std_getp = get_next; stdp->std_getp = get_tailp; stdp->std_get_count += count; stdp->std_get_non_tcp_count += non_tcp_count; } static void sfxge_tx_qreap(struct sfxge_txq *txq) { SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); txq->reaped = txq->completed; } static void sfxge_tx_qlist_post(struct sfxge_txq *txq) { unsigned int old_added; unsigned int block_level; unsigned int level; int rc; SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0")); KASSERT(txq->n_pend_desc <= txq->max_pkt_desc, ("txq->n_pend_desc too large")); KASSERT(!txq->blocked, ("txq->blocked")); old_added = txq->added; /* Post the fragment list. */ rc = efx_tx_qdesc_post(txq->common, txq->pend_desc, txq->n_pend_desc, txq->reaped, &txq->added); KASSERT(rc == 0, ("efx_tx_qdesc_post() failed")); /* If efx_tx_qdesc_post() had to refragment, our information about * buffers to free may be associated with the wrong * descriptors. */ KASSERT(txq->added - old_added == txq->n_pend_desc, ("efx_tx_qdesc_post() refragmented descriptors")); level = txq->added - txq->reaped; KASSERT(level <= txq->entries, ("overfilled TX queue")); /* Clear the fragment list. */ txq->n_pend_desc = 0; /* * Set the block level to ensure there is space to generate a * large number of descriptors for TSO. */ block_level = EFX_TXQ_LIMIT(txq->entries) - txq->max_pkt_desc; /* Have we reached the block level? */ if (level < block_level) return; /* Reap, and check again */ sfxge_tx_qreap(txq); level = txq->added - txq->reaped; if (level < block_level) return; txq->blocked = 1; /* * Avoid a race with completion interrupt handling that could leave * the queue blocked. */ mb(); sfxge_tx_qreap(txq); level = txq->added - txq->reaped; if (level < block_level) { mb(); txq->blocked = 0; } } static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf) { bus_dmamap_t *used_map; bus_dmamap_t map; bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG]; unsigned int id; struct sfxge_tx_mapping *stmp; efx_desc_t *desc; int n_dma_seg; int rc; int i; int eop; uint16_t hw_vlan_tci_prev; int vlan_tagged; KASSERT(!txq->blocked, ("txq->blocked")); #if SFXGE_TX_PARSE_EARLY /* * If software TSO is used, we still need to copy packet header, * even if we have already parsed it early before enqueue. */ if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) && (txq->tso_fw_assisted == 0)) prefetch_read_many(mbuf->m_data); #else /* * Prefetch packet header since we need to parse it and extract * IP ID, TCP sequence number and flags. */ if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) prefetch_read_many(mbuf->m_data); #endif if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) { rc = EINTR; goto reject; } /* Load the packet for DMA. */ id = txq->added & txq->ptr_mask; stmp = &txq->stmp[id]; rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map, mbuf, dma_seg, &n_dma_seg, 0); if (rc == EFBIG) { /* Try again. */ struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT, SFXGE_TX_MAPPING_MAX_SEG); if (new_mbuf == NULL) goto reject; ++txq->collapses; mbuf = new_mbuf; rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map, mbuf, dma_seg, &n_dma_seg, 0); } if (rc != 0) goto reject; /* Make the packet visible to the hardware. */ bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE); used_map = &stmp->map; hw_vlan_tci_prev = txq->hw_vlan_tci; vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf); if (vlan_tagged) { sfxge_next_stmp(txq, &stmp); } if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged); if (rc < 0) goto reject_mapped; stmp = &txq->stmp[(rc - 1) & txq->ptr_mask]; } else { /* Add the mapping to the fragment list, and set flags * for the buffer. */ i = 0; for (;;) { desc = &txq->pend_desc[i + vlan_tagged]; eop = (i == n_dma_seg - 1); efx_tx_qdesc_dma_create(txq->common, dma_seg[i].ds_addr, dma_seg[i].ds_len, eop, desc); if (eop) break; i++; sfxge_next_stmp(txq, &stmp); } txq->n_pend_desc = n_dma_seg + vlan_tagged; } /* * If the mapping required more than one descriptor * then we need to associate the DMA map with the last * descriptor, not the first. */ if (used_map != &stmp->map) { map = stmp->map; stmp->map = *used_map; *used_map = map; } stmp->u.mbuf = mbuf; stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF; /* Post the fragment list. */ sfxge_tx_qlist_post(txq); return (0); reject_mapped: txq->hw_vlan_tci = hw_vlan_tci_prev; bus_dmamap_unload(txq->packet_dma_tag, *used_map); reject: /* Drop the packet on the floor. */ m_freem(mbuf); ++txq->drops; return (rc); } /* * Drain the deferred packet list into the transmit queue. */ static void sfxge_tx_qdpl_drain(struct sfxge_txq *txq) { struct sfxge_softc *sc; struct sfxge_tx_dpl *stdp; struct mbuf *mbuf, *next; unsigned int count; unsigned int non_tcp_count; unsigned int pushed; int rc; SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); sc = txq->sc; stdp = &txq->dpl; pushed = txq->added; if (__predict_true(txq->init_state == SFXGE_TXQ_STARTED)) { prefetch_read_many(sc->enp); prefetch_read_many(txq->common); } mbuf = stdp->std_get; count = stdp->std_get_count; non_tcp_count = stdp->std_get_non_tcp_count; if (count > stdp->std_get_hiwat) stdp->std_get_hiwat = count; while (count != 0) { KASSERT(mbuf != NULL, ("mbuf == NULL")); next = mbuf->m_nextpkt; mbuf->m_nextpkt = NULL; ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */ if (next != NULL) prefetch_read_many(next); rc = sfxge_tx_queue_mbuf(txq, mbuf); --count; non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf); mbuf = next; if (rc != 0) continue; if (txq->blocked) break; /* Push the fragments to the hardware in batches. */ if (txq->added - pushed >= SFXGE_TX_BATCH) { efx_tx_qpush(txq->common, txq->added, pushed); pushed = txq->added; } } if (count == 0) { KASSERT(mbuf == NULL, ("mbuf != NULL")); KASSERT(non_tcp_count == 0, ("inconsistent TCP/non-TCP detection")); stdp->std_get = NULL; stdp->std_get_count = 0; stdp->std_get_non_tcp_count = 0; stdp->std_getp = &stdp->std_get; } else { stdp->std_get = mbuf; stdp->std_get_count = count; stdp->std_get_non_tcp_count = non_tcp_count; } if (txq->added != pushed) efx_tx_qpush(txq->common, txq->added, pushed); KASSERT(txq->blocked || stdp->std_get_count == 0, ("queue unblocked but count is non-zero")); } #define SFXGE_TX_QDPL_PENDING(_txq) ((_txq)->dpl.std_put != 0) /* * Service the deferred packet list. * * NOTE: drops the txq mutex! */ static void sfxge_tx_qdpl_service(struct sfxge_txq *txq) { SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); do { if (SFXGE_TX_QDPL_PENDING(txq)) sfxge_tx_qdpl_swizzle(txq); if (!txq->blocked) sfxge_tx_qdpl_drain(txq); SFXGE_TXQ_UNLOCK(txq); } while (SFXGE_TX_QDPL_PENDING(txq) && SFXGE_TXQ_TRYLOCK(txq)); } /* * Put a packet on the deferred packet get-list. */ static int sfxge_tx_qdpl_put_locked(struct sfxge_txq *txq, struct mbuf *mbuf) { struct sfxge_tx_dpl *stdp; stdp = &txq->dpl; KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); SFXGE_TXQ_LOCK_ASSERT_OWNED(txq); if (stdp->std_get_count >= stdp->std_get_max) { txq->get_overflow++; return (ENOBUFS); } if (sfxge_is_mbuf_non_tcp(mbuf)) { if (stdp->std_get_non_tcp_count >= stdp->std_get_non_tcp_max) { txq->get_non_tcp_overflow++; return (ENOBUFS); } stdp->std_get_non_tcp_count++; } *(stdp->std_getp) = mbuf; stdp->std_getp = &mbuf->m_nextpkt; stdp->std_get_count++; return (0); } /* * Put a packet on the deferred packet put-list. * * We overload the csum_data field in the mbuf to keep track of this length * because there is no cheap alternative to avoid races. */ static int sfxge_tx_qdpl_put_unlocked(struct sfxge_txq *txq, struct mbuf *mbuf) { struct sfxge_tx_dpl *stdp; volatile uintptr_t *putp; uintptr_t old; uintptr_t new; unsigned int put_count; KASSERT(mbuf->m_nextpkt == NULL, ("mbuf->m_nextpkt != NULL")); SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); stdp = &txq->dpl; putp = &stdp->std_put; new = (uintptr_t)mbuf; do { old = *putp; if (old != 0) { struct mbuf *mp = (struct mbuf *)old; put_count = mp->m_pkthdr.csum_data; } else put_count = 0; if (put_count >= stdp->std_put_max) { atomic_add_long(&txq->put_overflow, 1); return (ENOBUFS); } mbuf->m_pkthdr.csum_data = put_count + 1; mbuf->m_nextpkt = (void *)old; } while (atomic_cmpset_ptr(putp, old, new) == 0); return (0); } /* * Called from if_transmit - will try to grab the txq lock and enqueue to the * put list if it succeeds, otherwise try to push onto the defer list if space. */ static int sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m) { int rc; if (!SFXGE_LINK_UP(txq->sc)) { atomic_add_long(&txq->netdown_drops, 1); return (ENETDOWN); } /* * Try to grab the txq lock. If we are able to get the lock, * the packet will be appended to the "get list" of the deferred * packet list. Otherwise, it will be pushed on the "put list". */ if (SFXGE_TXQ_TRYLOCK(txq)) { /* First swizzle put-list to get-list to keep order */ sfxge_tx_qdpl_swizzle(txq); rc = sfxge_tx_qdpl_put_locked(txq, m); /* Try to service the list. */ sfxge_tx_qdpl_service(txq); /* Lock has been dropped. */ } else { rc = sfxge_tx_qdpl_put_unlocked(txq, m); /* * Try to grab the lock again. * * If we are able to get the lock, we need to process * the deferred packet list. If we are not able to get * the lock, another thread is processing the list. */ if ((rc == 0) && SFXGE_TXQ_TRYLOCK(txq)) { sfxge_tx_qdpl_service(txq); /* Lock has been dropped. */ } } SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(txq); return (rc); } static void sfxge_tx_qdpl_flush(struct sfxge_txq *txq) { struct sfxge_tx_dpl *stdp = &txq->dpl; struct mbuf *mbuf, *next; SFXGE_TXQ_LOCK(txq); sfxge_tx_qdpl_swizzle(txq); for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) { next = mbuf->m_nextpkt; m_freem(mbuf); } stdp->std_get = NULL; stdp->std_get_count = 0; stdp->std_get_non_tcp_count = 0; stdp->std_getp = &stdp->std_get; SFXGE_TXQ_UNLOCK(txq); } void sfxge_if_qflush(struct ifnet *ifp) { struct sfxge_softc *sc; unsigned int i; sc = ifp->if_softc; for (i = 0; i < sc->txq_count; i++) sfxge_tx_qdpl_flush(sc->txq[i]); } #if SFXGE_TX_PARSE_EARLY /* There is little space for user data in mbuf pkthdr, so we * use l*hlen fields which are not used by the driver otherwise * to store header offsets. * The fields are 8-bit, but it's ok, no header may be longer than 255 bytes. */ #define TSO_MBUF_PROTO(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[0]) /* We abuse l5hlen here because PH_loc can hold only 64 bits of data */ #define TSO_MBUF_FLAGS(_mbuf) ((_mbuf)->m_pkthdr.l5hlen) #define TSO_MBUF_PACKETID(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1]) #define TSO_MBUF_SEQNUM(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.thirtytwo[1]) static void sfxge_parse_tx_packet(struct mbuf *mbuf) { struct ether_header *eh = mtod(mbuf, struct ether_header *); const struct tcphdr *th; struct tcphdr th_copy; /* Find network protocol and header */ TSO_MBUF_PROTO(mbuf) = eh->ether_type; if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_VLAN)) { struct ether_vlan_header *veh = mtod(mbuf, struct ether_vlan_header *); TSO_MBUF_PROTO(mbuf) = veh->evl_proto; mbuf->m_pkthdr.l2hlen = sizeof(*veh); } else { mbuf->m_pkthdr.l2hlen = sizeof(*eh); } /* Find TCP header */ if (TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IP)) { const struct ip *iph = (const struct ip *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen); KASSERT(iph->ip_p == IPPROTO_TCP, ("TSO required on non-TCP packet")); mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + 4 * iph->ip_hl; TSO_MBUF_PACKETID(mbuf) = iph->ip_id; } else { KASSERT(TSO_MBUF_PROTO(mbuf) == htons(ETHERTYPE_IPV6), ("TSO required on non-IP packet")); KASSERT(((const struct ip6_hdr *)mtodo(mbuf, mbuf->m_pkthdr.l2hlen))->ip6_nxt == IPPROTO_TCP, ("TSO required on non-TCP packet")); mbuf->m_pkthdr.l3hlen = mbuf->m_pkthdr.l2hlen + sizeof(struct ip6_hdr); TSO_MBUF_PACKETID(mbuf) = 0; } KASSERT(mbuf->m_len >= mbuf->m_pkthdr.l3hlen, ("network header is fragmented in mbuf")); /* We need TCP header including flags (window is the next) */ if (mbuf->m_len < mbuf->m_pkthdr.l3hlen + offsetof(struct tcphdr, th_win)) { m_copydata(mbuf, mbuf->m_pkthdr.l3hlen, sizeof(th_copy), (caddr_t)&th_copy); th = &th_copy; } else { th = (const struct tcphdr *)mtodo(mbuf, mbuf->m_pkthdr.l3hlen); } mbuf->m_pkthdr.l4hlen = mbuf->m_pkthdr.l3hlen + 4 * th->th_off; TSO_MBUF_SEQNUM(mbuf) = ntohl(th->th_seq); /* These flags must not be duplicated */ /* * RST should not be duplicated as well, but FreeBSD kernel * generates TSO packets with RST flag. So, do not assert * its absence. */ KASSERT(!(th->th_flags & (TH_URG | TH_SYN)), ("incompatible TCP flag 0x%x on TSO packet", th->th_flags & (TH_URG | TH_SYN))); TSO_MBUF_FLAGS(mbuf) = th->th_flags; } #endif /* * TX start -- called by the stack. */ int sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) { struct sfxge_softc *sc; struct sfxge_txq *txq; int rc; sc = (struct sfxge_softc *)ifp->if_softc; /* * Transmit may be called when interface is up from the kernel * point of view, but not yet up (in progress) from the driver * point of view. I.e. link aggregation bring up. * Transmit may be called when interface is up from the driver * point of view, but already down from the kernel point of * view. I.e. Rx when interface shutdown is in progress. */ KASSERT((ifp->if_flags & IFF_UP) || (sc->if_flags & IFF_UP), ("interface not up")); /* Pick the desired transmit queue. */ if (m->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) { int index = 0; #ifdef RSS uint32_t bucket_id; /* * Select a TX queue which matches the corresponding * RX queue for the hash in order to assign both * TX and RX parts of the flow to the same CPU */ if (rss_m2bucket(m, &bucket_id) == 0) index = bucket_id % (sc->txq_count - (SFXGE_TXQ_NTYPES - 1)); #else /* check if flowid is set */ if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { uint32_t hash = m->m_pkthdr.flowid; uint32_t idx = hash % nitems(sc->rx_indir_table); index = sc->rx_indir_table[idx]; } #endif #if SFXGE_TX_PARSE_EARLY if (m->m_pkthdr.csum_flags & CSUM_TSO) sfxge_parse_tx_packet(m); #endif txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index]; } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) { txq = sc->txq[SFXGE_TXQ_IP_CKSUM]; } else { txq = sc->txq[SFXGE_TXQ_NON_CKSUM]; } rc = sfxge_tx_packet_add(txq, m); if (rc != 0) m_freem(m); return (rc); } /* * Software "TSO". Not quite as good as doing it in hardware, but * still faster than segmenting in the stack. */ struct sfxge_tso_state { /* Output position */ unsigned out_len; /* Remaining length in current segment */ unsigned seqnum; /* Current sequence number */ unsigned packet_space; /* Remaining space in current packet */ unsigned segs_space; /* Remaining number of DMA segments for the packet (FATSOv2 only) */ /* Input position */ uint64_t dma_addr; /* DMA address of current position */ unsigned in_len; /* Remaining length in current mbuf */ const struct mbuf *mbuf; /* Input mbuf (head of chain) */ u_short protocol; /* Network protocol (after VLAN decap) */ ssize_t nh_off; /* Offset of network header */ ssize_t tcph_off; /* Offset of TCP header */ unsigned header_len; /* Number of bytes of header */ unsigned seg_size; /* TCP segment size */ int fw_assisted; /* Use FW-assisted TSO */ u_short packet_id; /* IPv4 packet ID from the original packet */ uint8_t tcp_flags; /* TCP flags */ efx_desc_t header_desc; /* Precomputed header descriptor for * FW-assisted TSO */ }; #if !SFXGE_TX_PARSE_EARLY static const struct ip *tso_iph(const struct sfxge_tso_state *tso) { KASSERT(tso->protocol == htons(ETHERTYPE_IP), ("tso_iph() in non-IPv4 state")); return (const struct ip *)(tso->mbuf->m_data + tso->nh_off); } static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso) { KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), ("tso_ip6h() in non-IPv6 state")); return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off); } static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso) { return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off); } #endif /* Size of preallocated TSO header buffers. Larger blocks must be * allocated from the heap. */ #define TSOH_STD_SIZE 128 /* At most half the descriptors in the queue at any time will refer to * a TSO header buffer, since they must always be followed by a * payload descriptor referring to an mbuf. */ #define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u) #define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE) #define TSOH_PAGE_COUNT(_txq_entries) \ howmany(TSOH_COUNT(_txq_entries), TSOH_PER_PAGE) static int tso_init(struct sfxge_txq *txq) { struct sfxge_softc *sc = txq->sc; unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries); int i, rc; /* Allocate TSO header buffers */ txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]), M_SFXGE, M_WAITOK); for (i = 0; i < tsoh_page_count; i++) { rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]); if (rc != 0) goto fail; } return (0); fail: while (i-- > 0) sfxge_dma_free(&txq->tsoh_buffer[i]); free(txq->tsoh_buffer, M_SFXGE); txq->tsoh_buffer = NULL; return (rc); } static void tso_fini(struct sfxge_txq *txq) { int i; if (txq->tsoh_buffer != NULL) { for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++) sfxge_dma_free(&txq->tsoh_buffer[i]); free(txq->tsoh_buffer, M_SFXGE); } } static void tso_start(struct sfxge_txq *txq, struct sfxge_tso_state *tso, const bus_dma_segment_t *hdr_dma_seg, struct mbuf *mbuf) { const efx_nic_cfg_t *encp = efx_nic_cfg_get(txq->sc->enp); #if !SFXGE_TX_PARSE_EARLY struct ether_header *eh = mtod(mbuf, struct ether_header *); const struct tcphdr *th; struct tcphdr th_copy; #endif tso->fw_assisted = txq->tso_fw_assisted; tso->mbuf = mbuf; /* Find network protocol and header */ #if !SFXGE_TX_PARSE_EARLY tso->protocol = eh->ether_type; if (tso->protocol == htons(ETHERTYPE_VLAN)) { struct ether_vlan_header *veh = mtod(mbuf, struct ether_vlan_header *); tso->protocol = veh->evl_proto; tso->nh_off = sizeof(*veh); } else { tso->nh_off = sizeof(*eh); } #else tso->protocol = TSO_MBUF_PROTO(mbuf); tso->nh_off = mbuf->m_pkthdr.l2hlen; tso->tcph_off = mbuf->m_pkthdr.l3hlen; tso->packet_id = ntohs(TSO_MBUF_PACKETID(mbuf)); #endif #if !SFXGE_TX_PARSE_EARLY /* Find TCP header */ if (tso->protocol == htons(ETHERTYPE_IP)) { KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP, ("TSO required on non-TCP packet")); tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl; tso->packet_id = ntohs(tso_iph(tso)->ip_id); } else { KASSERT(tso->protocol == htons(ETHERTYPE_IPV6), ("TSO required on non-IP packet")); KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP, ("TSO required on non-TCP packet")); tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr); tso->packet_id = 0; } #endif if (tso->fw_assisted && __predict_false(tso->tcph_off > encp->enc_tx_tso_tcp_header_offset_limit)) { tso->fw_assisted = 0; } #if !SFXGE_TX_PARSE_EARLY KASSERT(mbuf->m_len >= tso->tcph_off, ("network header is fragmented in mbuf")); /* We need TCP header including flags (window is the next) */ if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) { m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy), (caddr_t)&th_copy); th = &th_copy; } else { th = tso_tcph(tso); } tso->header_len = tso->tcph_off + 4 * th->th_off; #else tso->header_len = mbuf->m_pkthdr.l4hlen; #endif tso->seg_size = mbuf->m_pkthdr.tso_segsz; #if !SFXGE_TX_PARSE_EARLY tso->seqnum = ntohl(th->th_seq); /* These flags must not be duplicated */ /* * RST should not be duplicated as well, but FreeBSD kernel * generates TSO packets with RST flag. So, do not assert * its absence. */ KASSERT(!(th->th_flags & (TH_URG | TH_SYN)), ("incompatible TCP flag 0x%x on TSO packet", th->th_flags & (TH_URG | TH_SYN))); tso->tcp_flags = th->th_flags; #else tso->seqnum = TSO_MBUF_SEQNUM(mbuf); tso->tcp_flags = TSO_MBUF_FLAGS(mbuf); #endif tso->out_len = mbuf->m_pkthdr.len - tso->header_len; if (tso->fw_assisted) { if (hdr_dma_seg->ds_len >= tso->header_len) efx_tx_qdesc_dma_create(txq->common, hdr_dma_seg->ds_addr, tso->header_len, B_FALSE, &tso->header_desc); else tso->fw_assisted = 0; } } /* * tso_fill_packet_with_fragment - form descriptors for the current fragment * * Form descriptors for the current fragment, until we reach the end * of fragment or end-of-packet. Return 0 on success, 1 if not enough * space. */ static void tso_fill_packet_with_fragment(struct sfxge_txq *txq, struct sfxge_tso_state *tso) { efx_desc_t *desc; int n; uint64_t dma_addr = tso->dma_addr; boolean_t eop; if (tso->in_len == 0 || tso->packet_space == 0) return; KASSERT(tso->in_len > 0, ("TSO input length went negative")); KASSERT(tso->packet_space > 0, ("TSO packet space went negative")); if (tso->fw_assisted & SFXGE_FATSOV2) { n = tso->in_len; tso->out_len -= n; tso->seqnum += n; tso->in_len = 0; if (n < tso->packet_space) { tso->packet_space -= n; tso->segs_space--; } else { tso->packet_space = tso->seg_size - (n - tso->packet_space) % tso->seg_size; tso->segs_space = EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1 - (tso->packet_space != tso->seg_size); } } else { n = min(tso->in_len, tso->packet_space); tso->packet_space -= n; tso->out_len -= n; tso->dma_addr += n; tso->in_len -= n; } /* * It is OK to use binary OR below to avoid extra branching * since all conditions may always be checked. */ eop = (tso->out_len == 0) | (tso->packet_space == 0) | (tso->segs_space == 0); desc = &txq->pend_desc[txq->n_pend_desc++]; efx_tx_qdesc_dma_create(txq->common, dma_addr, n, eop, desc); } /* Callback from bus_dmamap_load() for long TSO headers. */ static void tso_map_long_header(void *dma_addr_ret, bus_dma_segment_t *segs, int nseg, int error) { *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) && __predict_true(nseg == 1)) ? segs->ds_addr : 0); } /* * tso_start_new_packet - generate a new header and prepare for the new packet * * Generate a new header and prepare for the new packet. Return 0 on * success, or an error code if failed to alloc header. */ static int tso_start_new_packet(struct sfxge_txq *txq, struct sfxge_tso_state *tso, unsigned int *idp) { unsigned int id = *idp; struct tcphdr *tsoh_th; unsigned ip_length; caddr_t header; uint64_t dma_addr; bus_dmamap_t map; efx_desc_t *desc; int rc; if (tso->fw_assisted) { if (tso->fw_assisted & SFXGE_FATSOV2) { /* Add 2 FATSOv2 option descriptors */ desc = &txq->pend_desc[txq->n_pend_desc]; efx_tx_qdesc_tso2_create(txq->common, tso->packet_id, tso->seqnum, tso->seg_size, desc, EFX_TX_FATSOV2_OPT_NDESCS); desc += EFX_TX_FATSOV2_OPT_NDESCS; txq->n_pend_desc += EFX_TX_FATSOV2_OPT_NDESCS; KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); id = (id + EFX_TX_FATSOV2_OPT_NDESCS) & txq->ptr_mask; tso->segs_space = EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1; } else { uint8_t tcp_flags = tso->tcp_flags; if (tso->out_len > tso->seg_size) tcp_flags &= ~(TH_FIN | TH_PUSH); /* Add FATSOv1 option descriptor */ desc = &txq->pend_desc[txq->n_pend_desc++]; efx_tx_qdesc_tso_create(txq->common, tso->packet_id, tso->seqnum, tcp_flags, desc++); KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); id = (id + 1) & txq->ptr_mask; tso->seqnum += tso->seg_size; tso->segs_space = UINT_MAX; } /* Header DMA descriptor */ *desc = tso->header_desc; txq->n_pend_desc++; KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); id = (id + 1) & txq->ptr_mask; } else { /* Allocate a DMA-mapped header buffer. */ if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) { unsigned int page_index = (id / 2) / TSOH_PER_PAGE; unsigned int buf_index = (id / 2) % TSOH_PER_PAGE; header = (txq->tsoh_buffer[page_index].esm_base + buf_index * TSOH_STD_SIZE); dma_addr = (txq->tsoh_buffer[page_index].esm_addr + buf_index * TSOH_STD_SIZE); map = txq->tsoh_buffer[page_index].esm_map; KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); } else { struct sfxge_tx_mapping *stmp = &txq->stmp[id]; /* We cannot use bus_dmamem_alloc() as that may sleep */ header = malloc(tso->header_len, M_SFXGE, M_NOWAIT); if (__predict_false(!header)) return (ENOMEM); rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map, header, tso->header_len, tso_map_long_header, &dma_addr, BUS_DMA_NOWAIT); if (__predict_false(dma_addr == 0)) { if (rc == 0) { /* Succeeded but got >1 segment */ bus_dmamap_unload(txq->packet_dma_tag, stmp->map); rc = EINVAL; } free(header, M_SFXGE); return (rc); } map = stmp->map; txq->tso_long_headers++; stmp->u.heap_buf = header; stmp->flags = TX_BUF_UNMAP; } tsoh_th = (struct tcphdr *)(header + tso->tcph_off); /* Copy and update the headers. */ m_copydata(tso->mbuf, 0, tso->header_len, header); tsoh_th->th_seq = htonl(tso->seqnum); tso->seqnum += tso->seg_size; if (tso->out_len > tso->seg_size) { /* This packet will not finish the TSO burst. */ ip_length = tso->header_len - tso->nh_off + tso->seg_size; tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH); } else { /* This packet will be the last in the TSO burst. */ ip_length = tso->header_len - tso->nh_off + tso->out_len; } if (tso->protocol == htons(ETHERTYPE_IP)) { struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off); tsoh_iph->ip_len = htons(ip_length); /* XXX We should increment ip_id, but FreeBSD doesn't * currently allocate extra IDs for multiple segments. */ } else { struct ip6_hdr *tsoh_iph = (struct ip6_hdr *)(header + tso->nh_off); tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph)); } /* Make the header visible to the hardware. */ bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE); /* Form a descriptor for this header. */ desc = &txq->pend_desc[txq->n_pend_desc++]; efx_tx_qdesc_dma_create(txq->common, dma_addr, tso->header_len, 0, desc); id = (id + 1) & txq->ptr_mask; tso->segs_space = UINT_MAX; } tso->packet_space = tso->seg_size; txq->tso_packets++; *idp = id; return (0); } static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, const bus_dma_segment_t *dma_seg, int n_dma_seg, int vlan_tagged) { struct sfxge_tso_state tso; unsigned int id; unsigned skipped = 0; tso_start(txq, &tso, dma_seg, mbuf); while (dma_seg->ds_len + skipped <= tso.header_len) { skipped += dma_seg->ds_len; --n_dma_seg; KASSERT(n_dma_seg, ("no payload found in TSO packet")); ++dma_seg; } tso.in_len = dma_seg->ds_len - (tso.header_len - skipped); tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped); id = (txq->added + vlan_tagged) & txq->ptr_mask; if (__predict_false(tso_start_new_packet(txq, &tso, &id))) return (-1); while (1) { tso_fill_packet_with_fragment(txq, &tso); /* Exactly one DMA descriptor is added */ KASSERT(txq->stmp[id].flags == 0, ("stmp flags are not 0")); id = (id + 1) & txq->ptr_mask; /* Move onto the next fragment? */ if (tso.in_len == 0) { --n_dma_seg; if (n_dma_seg == 0) break; ++dma_seg; tso.in_len = dma_seg->ds_len; tso.dma_addr = dma_seg->ds_addr; } /* End of packet? */ if ((tso.packet_space == 0) | (tso.segs_space == 0)) { unsigned int n_fatso_opt_desc = (tso.fw_assisted & SFXGE_FATSOV2) ? EFX_TX_FATSOV2_OPT_NDESCS : (tso.fw_assisted & SFXGE_FATSOV1) ? 1 : 0; /* If the queue is now full due to tiny MSS, * or we can't create another header, discard * the remainder of the input mbuf but do not * roll back the work we have done. */ if (txq->n_pend_desc + n_fatso_opt_desc + 1 /* header */ + n_dma_seg > txq->max_pkt_desc) { txq->tso_pdrop_too_many++; break; } if (__predict_false(tso_start_new_packet(txq, &tso, &id))) { txq->tso_pdrop_no_rsrc++; break; } } } txq->tso_bursts++; return (id); } static void sfxge_tx_qunblock(struct sfxge_txq *txq) { struct sfxge_softc *sc; struct sfxge_evq *evq; sc = txq->sc; evq = sc->evq[txq->evq_index]; SFXGE_EVQ_LOCK_ASSERT_OWNED(evq); if (__predict_false(txq->init_state != SFXGE_TXQ_STARTED)) return; SFXGE_TXQ_LOCK(txq); if (txq->blocked) { unsigned int level; level = txq->added - txq->completed; if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) { /* reaped must be in sync with blocked */ sfxge_tx_qreap(txq); txq->blocked = 0; } } sfxge_tx_qdpl_service(txq); /* note: lock has been dropped */ } void sfxge_tx_qflush_done(struct sfxge_txq *txq) { txq->flush_state = SFXGE_FLUSH_DONE; } static void sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index) { struct sfxge_txq *txq; struct sfxge_evq *evq; unsigned int count; SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); txq = sc->txq[index]; evq = sc->evq[txq->evq_index]; SFXGE_EVQ_LOCK(evq); SFXGE_TXQ_LOCK(txq); KASSERT(txq->init_state == SFXGE_TXQ_STARTED, ("txq->init_state != SFXGE_TXQ_STARTED")); txq->init_state = SFXGE_TXQ_INITIALIZED; if (txq->flush_state != SFXGE_FLUSH_DONE) { txq->flush_state = SFXGE_FLUSH_PENDING; SFXGE_EVQ_UNLOCK(evq); SFXGE_TXQ_UNLOCK(txq); /* Flush the transmit queue. */ if (efx_tx_qflush(txq->common) != 0) { log(LOG_ERR, "%s: Flushing Tx queue %u failed\n", device_get_nameunit(sc->dev), index); txq->flush_state = SFXGE_FLUSH_DONE; } else { count = 0; do { /* Spin for 100ms. */ DELAY(100000); if (txq->flush_state != SFXGE_FLUSH_PENDING) break; } while (++count < 20); } SFXGE_EVQ_LOCK(evq); SFXGE_TXQ_LOCK(txq); KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED, ("txq->flush_state == SFXGE_FLUSH_FAILED")); if (txq->flush_state != SFXGE_FLUSH_DONE) { /* Flush timeout */ log(LOG_ERR, "%s: Cannot flush Tx queue %u\n", device_get_nameunit(sc->dev), index); txq->flush_state = SFXGE_FLUSH_DONE; } } txq->blocked = 0; txq->pending = txq->added; sfxge_tx_qcomplete(txq, evq); KASSERT(txq->completed == txq->added, ("txq->completed != txq->added")); sfxge_tx_qreap(txq); KASSERT(txq->reaped == txq->completed, ("txq->reaped != txq->completed")); txq->added = 0; txq->pending = 0; txq->completed = 0; txq->reaped = 0; /* Destroy the common code transmit queue. */ efx_tx_qdestroy(txq->common); txq->common = NULL; efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, EFX_TXQ_NBUFS(sc->txq_entries)); SFXGE_EVQ_UNLOCK(evq); SFXGE_TXQ_UNLOCK(txq); } /* * Estimate maximum number of Tx descriptors required for TSO packet. * With minimum MSS and maximum mbuf length we might need more (even * than a ring-ful of descriptors), but this should not happen in * practice except due to deliberate attack. In that case we will * truncate the output at a packet boundary. */ static unsigned int sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, enum sfxge_txq_type type, unsigned int tso_fw_assisted) { /* One descriptor for every input fragment */ unsigned int max_descs = SFXGE_TX_MAPPING_MAX_SEG; unsigned int sw_tso_max_descs; unsigned int fa_tso_v1_max_descs = 0; unsigned int fa_tso_v2_max_descs = 0; /* VLAN tagging Tx option descriptor may be required */ if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled) max_descs++; if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) { /* * Plus header and payload descriptor for each output segment. * Minus one since header fragment is already counted. * Even if FATSO is used, we should be ready to fallback * to do it in the driver. */ sw_tso_max_descs = SFXGE_TSO_MAX_SEGS * 2 - 1; /* FW assisted TSOv1 requires one more descriptor per segment * in comparison to SW TSO */ if (tso_fw_assisted & SFXGE_FATSOV1) fa_tso_v1_max_descs = sw_tso_max_descs + SFXGE_TSO_MAX_SEGS; /* FW assisted TSOv2 requires 3 (2 FATSO plus header) extra * descriptors per superframe limited by number of DMA fetches * per packet. The first packet header is already counted. */ if (tso_fw_assisted & SFXGE_FATSOV2) { fa_tso_v2_max_descs = howmany(SFXGE_TX_MAPPING_MAX_SEG, EFX_TX_FATSOV2_DMA_SEGS_PER_PKT_MAX - 1) * (EFX_TX_FATSOV2_OPT_NDESCS + 1) - 1; } max_descs += MAX(sw_tso_max_descs, MAX(fa_tso_v1_max_descs, fa_tso_v2_max_descs)); } return (max_descs); } static int sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index) { struct sfxge_txq *txq; efsys_mem_t *esmp; uint16_t flags; unsigned int tso_fw_assisted; struct sfxge_evq *evq; unsigned int desc_index; int rc; SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc); txq = sc->txq[index]; esmp = &txq->mem; evq = sc->evq[txq->evq_index]; KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, ("txq->init_state != SFXGE_TXQ_INITIALIZED")); KASSERT(evq->init_state == SFXGE_EVQ_STARTED, ("evq->init_state != SFXGE_EVQ_STARTED")); /* Program the buffer table. */ if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp, EFX_TXQ_NBUFS(sc->txq_entries))) != 0) return (rc); /* Determine the kind of queue we are creating. */ tso_fw_assisted = 0; switch (txq->type) { case SFXGE_TXQ_NON_CKSUM: flags = 0; break; case SFXGE_TXQ_IP_CKSUM: flags = EFX_TXQ_CKSUM_IPV4; break; case SFXGE_TXQ_IP_TCP_UDP_CKSUM: flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP; tso_fw_assisted = sc->tso_fw_assisted; if (tso_fw_assisted & SFXGE_FATSOV2) flags |= EFX_TXQ_FATSOV2; break; default: KASSERT(0, ("Impossible TX queue")); flags = 0; break; } /* Create the common code transmit queue. */ if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, sc->txq_entries, txq->buf_base_id, flags, evq->common, &txq->common, &desc_index)) != 0) { /* Retry if no FATSOv2 resources, otherwise fail */ if ((rc != ENOSPC) || (~flags & EFX_TXQ_FATSOV2)) goto fail; /* Looks like all FATSOv2 contexts are used */ flags &= ~EFX_TXQ_FATSOV2; tso_fw_assisted &= ~SFXGE_FATSOV2; if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, sc->txq_entries, txq->buf_base_id, flags, evq->common, &txq->common, &desc_index)) != 0) goto fail; } /* Initialise queue descriptor indexes */ txq->added = txq->pending = txq->completed = txq->reaped = desc_index; SFXGE_TXQ_LOCK(txq); /* Enable the transmit queue. */ efx_tx_qenable(txq->common); txq->init_state = SFXGE_TXQ_STARTED; txq->flush_state = SFXGE_FLUSH_REQUIRED; txq->tso_fw_assisted = tso_fw_assisted; txq->max_pkt_desc = sfxge_tx_max_pkt_desc(sc, txq->type, tso_fw_assisted); txq->hw_vlan_tci = 0; SFXGE_TXQ_UNLOCK(txq); return (0); fail: efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, EFX_TXQ_NBUFS(sc->txq_entries)); return (rc); } void sfxge_tx_stop(struct sfxge_softc *sc) { int index; index = sc->txq_count; while (--index >= 0) sfxge_tx_qstop(sc, index); /* Tear down the transmit module */ efx_tx_fini(sc->enp); } int sfxge_tx_start(struct sfxge_softc *sc) { int index; int rc; /* Initialize the common code transmit module. */ if ((rc = efx_tx_init(sc->enp)) != 0) return (rc); for (index = 0; index < sc->txq_count; index++) { if ((rc = sfxge_tx_qstart(sc, index)) != 0) goto fail; } return (0); fail: while (--index >= 0) sfxge_tx_qstop(sc, index); efx_tx_fini(sc->enp); return (rc); } static int sfxge_txq_stat_init(struct sfxge_txq *txq, struct sysctl_oid *txq_node) { struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(txq->sc->dev); struct sysctl_oid *stat_node; unsigned int id; stat_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, "stats", CTLFLAG_RD, NULL, "Tx queue statistics"); if (stat_node == NULL) return (ENOMEM); for (id = 0; id < nitems(sfxge_tx_stats); id++) { SYSCTL_ADD_ULONG( ctx, SYSCTL_CHILDREN(stat_node), OID_AUTO, sfxge_tx_stats[id].name, CTLFLAG_RD | CTLFLAG_STATS, (unsigned long *)((caddr_t)txq + sfxge_tx_stats[id].offset), ""); } return (0); } /** * Destroy a transmit queue. */ static void sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index) { struct sfxge_txq *txq; unsigned int nmaps; txq = sc->txq[index]; KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED, ("txq->init_state != SFXGE_TXQ_INITIALIZED")); if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM) tso_fini(txq); /* Free the context arrays. */ free(txq->pend_desc, M_SFXGE); nmaps = sc->txq_entries; while (nmaps-- != 0) bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); free(txq->stmp, M_SFXGE); /* Release DMA memory mapping. */ sfxge_dma_free(&txq->mem); sc->txq[index] = NULL; SFXGE_TXQ_LOCK_DESTROY(txq); free(txq, M_SFXGE); } static int sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index, enum sfxge_txq_type type, unsigned int evq_index) { const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp); char name[16]; struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); struct sysctl_oid *txq_node; struct sfxge_txq *txq; struct sfxge_evq *evq; struct sfxge_tx_dpl *stdp; struct sysctl_oid *dpl_node; efsys_mem_t *esmp; unsigned int nmaps; int rc; txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK); txq->sc = sc; txq->entries = sc->txq_entries; txq->ptr_mask = txq->entries - 1; sc->txq[txq_index] = txq; esmp = &txq->mem; evq = sc->evq[evq_index]; /* Allocate and zero DMA space for the descriptor ring. */ if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0) return (rc); /* Allocate buffer table entries. */ sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries), &txq->buf_base_id); /* Create a DMA tag for packet mappings. */ if (bus_dma_tag_create(sc->parent_dma_tag, 1, encp->enc_tx_dma_desc_boundary, MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL, NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, encp->enc_tx_dma_desc_size_max, 0, NULL, NULL, &txq->packet_dma_tag) != 0) { device_printf(sc->dev, "Couldn't allocate txq DMA tag\n"); rc = ENOMEM; goto fail; } /* Allocate pending descriptor array for batching writes. */ txq->pend_desc = malloc(sizeof(efx_desc_t) * sc->txq_entries, M_SFXGE, M_ZERO | M_WAITOK); /* Allocate and initialise mbuf DMA mapping array. */ txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries, M_SFXGE, M_ZERO | M_WAITOK); for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) { rc = bus_dmamap_create(txq->packet_dma_tag, 0, &txq->stmp[nmaps].map); if (rc != 0) goto fail2; } snprintf(name, sizeof(name), "%u", txq_index); txq_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(sc->txqs_node), OID_AUTO, name, CTLFLAG_RD, NULL, ""); if (txq_node == NULL) { rc = ENOMEM; goto fail_txq_node; } if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM && (rc = tso_init(txq)) != 0) goto fail3; /* Initialize the deferred packet list. */ stdp = &txq->dpl; stdp->std_put_max = sfxge_tx_dpl_put_max; stdp->std_get_max = sfxge_tx_dpl_get_max; stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max; stdp->std_getp = &stdp->std_get; SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index); dpl_node = SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(txq_node), OID_AUTO, "dpl", CTLFLAG_RD, NULL, "Deferred packet list statistics"); if (dpl_node == NULL) { rc = ENOMEM; goto fail_dpl_node; } SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, "get_count", CTLFLAG_RD | CTLFLAG_STATS, &stdp->std_get_count, 0, ""); SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, "get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS, &stdp->std_get_non_tcp_count, 0, ""); SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, "get_hiwat", CTLFLAG_RD | CTLFLAG_STATS, &stdp->std_get_hiwat, 0, ""); SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(dpl_node), OID_AUTO, "put_hiwat", CTLFLAG_RD | CTLFLAG_STATS, &stdp->std_put_hiwat, 0, ""); rc = sfxge_txq_stat_init(txq, txq_node); if (rc != 0) goto fail_txq_stat_init; txq->type = type; txq->evq_index = evq_index; txq->init_state = SFXGE_TXQ_INITIALIZED; return (0); fail_txq_stat_init: fail_dpl_node: fail3: fail_txq_node: free(txq->pend_desc, M_SFXGE); fail2: while (nmaps-- != 0) bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map); free(txq->stmp, M_SFXGE); bus_dma_tag_destroy(txq->packet_dma_tag); fail: sfxge_dma_free(esmp); return (rc); } static int sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS) { struct sfxge_softc *sc = arg1; unsigned int id = arg2; unsigned long sum; unsigned int index; /* Sum across all TX queues */ sum = 0; for (index = 0; index < sc->txq_count; index++) sum += *(unsigned long *)((caddr_t)sc->txq[index] + sfxge_tx_stats[id].offset); return (SYSCTL_OUT(req, &sum, sizeof(sum))); } static void sfxge_tx_stat_init(struct sfxge_softc *sc) { struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev); struct sysctl_oid_list *stat_list; unsigned int id; stat_list = SYSCTL_CHILDREN(sc->stats_node); for (id = 0; id < nitems(sfxge_tx_stats); id++) { SYSCTL_ADD_PROC( ctx, stat_list, OID_AUTO, sfxge_tx_stats[id].name, CTLTYPE_ULONG|CTLFLAG_RD, sc, id, sfxge_tx_stat_handler, "LU", ""); } } uint64_t sfxge_tx_get_drops(struct sfxge_softc *sc) { unsigned int index; uint64_t drops = 0; struct sfxge_txq *txq; /* Sum across all TX queues */ for (index = 0; index < sc->txq_count; index++) { txq = sc->txq[index]; /* * In theory, txq->put_overflow and txq->netdown_drops * should use atomic operation and other should be * obtained under txq lock, but it is just statistics. */ drops += txq->drops + txq->get_overflow + txq->get_non_tcp_overflow + txq->put_overflow + txq->netdown_drops + txq->tso_pdrop_too_many + txq->tso_pdrop_no_rsrc; } return (drops); } void sfxge_tx_fini(struct sfxge_softc *sc) { int index; index = sc->txq_count; while (--index >= 0) sfxge_tx_qfini(sc, index); sc->txq_count = 0; } int sfxge_tx_init(struct sfxge_softc *sc) { const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp); struct sfxge_intr *intr; int index; int rc; intr = &sc->intr; KASSERT(intr->state == SFXGE_INTR_INITIALIZED, ("intr->state != SFXGE_INTR_INITIALIZED")); if (sfxge_tx_dpl_get_max <= 0) { log(LOG_ERR, "%s=%d must be greater than 0", SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max); rc = EINVAL; goto fail_tx_dpl_get_max; } if (sfxge_tx_dpl_get_non_tcp_max <= 0) { log(LOG_ERR, "%s=%d must be greater than 0", SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, sfxge_tx_dpl_get_non_tcp_max); rc = EINVAL; goto fail_tx_dpl_get_non_tcp_max; } if (sfxge_tx_dpl_put_max < 0) { log(LOG_ERR, "%s=%d must be greater or equal to 0", SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max); rc = EINVAL; goto fail_tx_dpl_put_max; } - sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc; + sc->txq_count = SFXGE_EVQ0_N_TXQ(sc) - 1 + sc->intr.n_alloc; sc->tso_fw_assisted = sfxge_tso_fw_assisted; if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO) || (!encp->enc_fw_assisted_tso_enabled)) sc->tso_fw_assisted &= ~SFXGE_FATSOV1; if ((~encp->enc_features & EFX_FEATURE_FW_ASSISTED_TSO_V2) || (!encp->enc_fw_assisted_tso_v2_enabled)) sc->tso_fw_assisted &= ~SFXGE_FATSOV2; sc->txqs_node = SYSCTL_ADD_NODE( device_get_sysctl_ctx(sc->dev), SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)), OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues"); if (sc->txqs_node == NULL) { rc = ENOMEM; goto fail_txq_node; } /* Initialize the transmit queues */ if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, SFXGE_TXQ_NON_CKSUM, 0)) != 0) goto fail; if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, SFXGE_TXQ_IP_CKSUM, 0)) != 0) goto fail2; for (index = 0; - index < sc->txq_count - SFXGE_TXQ_NTYPES + 1; + index < sc->txq_count - SFXGE_EVQ0_N_TXQ(sc) + 1; index++) { - if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index, + if ((rc = sfxge_tx_qinit(sc, SFXGE_EVQ0_N_TXQ(sc) - 1 + index, SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0) goto fail3; } sfxge_tx_stat_init(sc); return (0); fail3: while (--index >= 0) sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index); sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM); fail2: sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM); fail: fail_txq_node: sc->txq_count = 0; fail_tx_dpl_put_max: fail_tx_dpl_get_non_tcp_max: fail_tx_dpl_get_max: return (rc); } Index: stable/11/sys/dev/sfxge/sfxge_tx.h =================================================================== --- stable/11/sys/dev/sfxge/sfxge_tx.h (revision 342453) +++ stable/11/sys/dev/sfxge/sfxge_tx.h (revision 342454) @@ -1,243 +1,245 @@ /*- * Copyright (c) 2010-2016 Solarflare Communications Inc. * All rights reserved. * * This software was developed in part by Philip Paeps under contract for * Solarflare Communications, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * The views and conclusions contained in the software and documentation are * those of the authors and should not be interpreted as representing official * policies, either expressed or implied, of the FreeBSD Project. * * $FreeBSD$ */ #ifndef _SFXGE_TX_H #define _SFXGE_TX_H #include #include #include /* If defined, parse TX packets directly in if_transmit * for better cache locality and reduced time under TX lock */ #define SFXGE_TX_PARSE_EARLY 1 /* Maximum size of TSO packet */ #define SFXGE_TSO_MAX_SIZE (65535) /* * Maximum number of segments to be created for a TSO packet. * Allow for a reasonable minimum MSS of 512. */ #define SFXGE_TSO_MAX_SEGS howmany(SFXGE_TSO_MAX_SIZE, 512) /* Maximum number of DMA segments needed to map an mbuf chain. With * TSO, the mbuf length may be just over 64K, divided into 2K mbuf * clusters taking into account that the first may be not 2K cluster * boundary aligned. * Packet header may be split into two segments because of, for example, * VLAN header insertion. * The chain could be longer than this initially, but can be shortened * with m_collapse(). */ #define SFXGE_TX_MAPPING_MAX_SEG \ (2 + howmany(SFXGE_TSO_MAX_SIZE, MCLBYTES) + 1) /* * Buffer mapping flags. * * Buffers and DMA mappings must be freed when the last descriptor * referring to them is completed. Set the TX_BUF_UNMAP and * TX_BUF_MBUF flags on the last descriptor generated for an mbuf * chain. Set only the TX_BUF_UNMAP flag on a descriptor referring to * a heap buffer. */ enum sfxge_tx_buf_flags { TX_BUF_UNMAP = 1, TX_BUF_MBUF = 2, }; /* * Buffer mapping information for descriptors in flight. */ struct sfxge_tx_mapping { union { struct mbuf *mbuf; caddr_t heap_buf; } u; bus_dmamap_t map; enum sfxge_tx_buf_flags flags; }; #define SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT (64 * 1024) #define SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT 1024 #define SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT 1024 /* * Deferred packet list. */ struct sfxge_tx_dpl { unsigned int std_get_max; /* Maximum number of packets * in get list */ unsigned int std_get_non_tcp_max; /* Maximum number * of non-TCP packets * in get list */ unsigned int std_put_max; /* Maximum number of packets * in put list */ uintptr_t std_put; /* Head of put list. */ struct mbuf *std_get; /* Head of get list. */ struct mbuf **std_getp; /* Tail of get list. */ unsigned int std_get_count; /* Packets in get list. */ unsigned int std_get_non_tcp_count; /* Non-TCP packets * in get list */ unsigned int std_get_hiwat; /* Packets in get list * high watermark */ unsigned int std_put_hiwat; /* Packets in put list * high watermark */ }; #define SFXGE_TX_BUFFER_SIZE 0x400 #define SFXGE_TX_HEADER_SIZE 0x100 #define SFXGE_TX_COPY_THRESHOLD 0x200 enum sfxge_txq_state { SFXGE_TXQ_UNINITIALIZED = 0, SFXGE_TXQ_INITIALIZED, SFXGE_TXQ_STARTED }; enum sfxge_txq_type { SFXGE_TXQ_NON_CKSUM = 0, SFXGE_TXQ_IP_CKSUM, SFXGE_TXQ_IP_TCP_UDP_CKSUM, SFXGE_TXQ_NTYPES }; +#define SFXGE_EVQ0_N_TXQ(_sc) SFXGE_TXQ_NTYPES + #define SFXGE_TXQ_UNBLOCK_LEVEL(_entries) (EFX_TXQ_LIMIT(_entries) / 4) #define SFXGE_TX_BATCH 64 #define SFXGE_TXQ_LOCK_INIT(_txq, _ifname, _txq_index) \ do { \ struct sfxge_txq *__txq = (_txq); \ \ snprintf((__txq)->lock_name, \ sizeof((__txq)->lock_name), \ "%s:txq%u", (_ifname), (_txq_index)); \ mtx_init(&(__txq)->lock, (__txq)->lock_name, \ NULL, MTX_DEF); \ } while (B_FALSE) #define SFXGE_TXQ_LOCK_DESTROY(_txq) \ mtx_destroy(&(_txq)->lock) #define SFXGE_TXQ_LOCK(_txq) \ mtx_lock(&(_txq)->lock) #define SFXGE_TXQ_TRYLOCK(_txq) \ mtx_trylock(&(_txq)->lock) #define SFXGE_TXQ_UNLOCK(_txq) \ mtx_unlock(&(_txq)->lock) #define SFXGE_TXQ_LOCK_ASSERT_OWNED(_txq) \ mtx_assert(&(_txq)->lock, MA_OWNED) #define SFXGE_TXQ_LOCK_ASSERT_NOTOWNED(_txq) \ mtx_assert(&(_txq)->lock, MA_NOTOWNED) struct sfxge_txq { /* The following fields should be written very rarely */ struct sfxge_softc *sc; enum sfxge_txq_state init_state; enum sfxge_flush_state flush_state; unsigned int tso_fw_assisted; enum sfxge_txq_type type; unsigned int evq_index; efsys_mem_t mem; unsigned int buf_base_id; unsigned int entries; unsigned int ptr_mask; unsigned int max_pkt_desc; struct sfxge_tx_mapping *stmp; /* Packets in flight. */ bus_dma_tag_t packet_dma_tag; efx_desc_t *pend_desc; efx_txq_t *common; efsys_mem_t *tsoh_buffer; char lock_name[SFXGE_LOCK_NAME_MAX]; /* This field changes more often and is read regularly on both * the initiation and completion paths */ int blocked __aligned(CACHE_LINE_SIZE); /* The following fields change more often, and are used mostly * on the initiation path */ struct mtx lock __aligned(CACHE_LINE_SIZE); struct sfxge_tx_dpl dpl; /* Deferred packet list. */ unsigned int n_pend_desc; unsigned int added; unsigned int reaped; /* The last VLAN TCI seen on the queue if FW-assisted tagging is used */ uint16_t hw_vlan_tci; /* Statistics */ unsigned long tso_bursts; unsigned long tso_packets; unsigned long tso_long_headers; unsigned long collapses; unsigned long drops; unsigned long get_overflow; unsigned long get_non_tcp_overflow; unsigned long put_overflow; unsigned long netdown_drops; unsigned long tso_pdrop_too_many; unsigned long tso_pdrop_no_rsrc; /* The following fields change more often, and are used mostly * on the completion path */ unsigned int pending __aligned(CACHE_LINE_SIZE); unsigned int completed; struct sfxge_txq *next; }; struct sfxge_evq; extern uint64_t sfxge_tx_get_drops(struct sfxge_softc *sc); extern int sfxge_tx_init(struct sfxge_softc *sc); extern void sfxge_tx_fini(struct sfxge_softc *sc); extern int sfxge_tx_start(struct sfxge_softc *sc); extern void sfxge_tx_stop(struct sfxge_softc *sc); extern void sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq); extern void sfxge_tx_qflush_done(struct sfxge_txq *txq); extern void sfxge_if_qflush(struct ifnet *ifp); extern int sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m); #endif Index: stable/11 =================================================================== --- stable/11 (revision 342453) +++ stable/11 (revision 342454) Property changes on: stable/11 ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head:r341784