Date: Wed, 24 Feb 2016 01:30:50 +0000 (UTC) From: Sepherosa Ziehau <sephe@FreeBSD.org> To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-10@freebsd.org Subject: svn commit: r295948 - in stable/10/sys/dev/hyperv: netvsc vmbus Message-ID: <201602240130.u1O1Uoko011921@repo.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: sephe Date: Wed Feb 24 01:30:50 2016 New Revision: 295948 URL: https://svnweb.freebsd.org/changeset/base/295948 Log: MFC [Hyper-V]: r294553, r294700 r294553 hyperv/vmbus: Lookup channel through id table Vmbus event handler will need to find the channel by its relative id, when software interrupt for event happens. The original lookup searches the channel list, which is not very efficient. We now create a table indexed by the channel relative id to speed up the channel lookup. Submitted by: Hongjiang Zhang <honzhan microsoft com> Reviewed by: delphij, adrain, sephe, Dexuan Cui <decui microsoft com> Approved by: adrian (mentor) Sponsored by: Microsoft OSTC Differential Revision: https://reviews.freebsd.org/D4802 ------------- r294700 hyperv/hn: Partly rework transmission path - Avoid unnecessary malloc/free on transmission path. - busdma(9)-fy transmission path. - Properly handle IFF_DRV_OACTIVE. This should fix the network stalls reported by many. - Properly setup TSO parameters. - Properly handle bpf(4) tapping. This 5 times the performance during TCP sending test, when there is one bpf(4) attached. - Allow size of chimney sending be tuned on a running system. Default value still needs more test to determine. Reviewed by: adrian, delphij Approved by: adrian (mentor) Sponsored by: Microsoft OSTC Differential Revision: https://reviews.freebsd.org/D4972 Approved by: re (marius) Sponsored by: Microsoft OSTC Modified: stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.c stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.h stable/10/sys/dev/hyperv/netvsc/hv_netvsc_drv_freebsd.c stable/10/sys/dev/hyperv/netvsc/hv_rndis.h stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.c stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.h stable/10/sys/dev/hyperv/vmbus/hv_channel_mgmt.c stable/10/sys/dev/hyperv/vmbus/hv_connection.c stable/10/sys/dev/hyperv/vmbus/hv_vmbus_priv.h Directory Properties: stable/10/ (props changed) Modified: stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.c ============================================================================== --- stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.c Wed Feb 24 01:11:51 2016 (r295947) +++ stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.c Wed Feb 24 01:30:50 2016 (r295948) @@ -1027,4 +1027,6 @@ hv_nv_on_channel_callback(void *context) if (bufferlen > NETVSC_PACKET_SIZE) free(buffer, M_NETVSC); + + hv_rf_channel_rollup(net_dev); } Modified: stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.h ============================================================================== --- stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.h Wed Feb 24 01:11:51 2016 (r295947) +++ stable/10/sys/dev/hyperv/netvsc/hv_net_vsc.h Wed Feb 24 01:30:50 2016 (r295948) @@ -38,12 +38,16 @@ #ifndef __HV_NET_VSC_H__ #define __HV_NET_VSC_H__ -#include <sys/types.h> #include <sys/param.h> #include <sys/lock.h> #include <sys/malloc.h> +#include <sys/queue.h> #include <sys/sx.h> +#include <machine/bus.h> +#include <sys/bus.h> +#include <sys/bus_dma.h> + #include <netinet/in.h> #include <netinet/tcp_lro.h> @@ -984,6 +988,9 @@ typedef struct { hv_bool_uint8_t link_state; } netvsc_device_info; +struct hn_txdesc; +SLIST_HEAD(hn_txdesc_list, hn_txdesc); + /* * Device-specific softc structure */ @@ -1002,6 +1009,18 @@ typedef struct hn_softc { struct hv_device *hn_dev_obj; netvsc_dev *net_dev; + int hn_txdesc_cnt; + struct hn_txdesc *hn_txdesc; + bus_dma_tag_t hn_tx_data_dtag; + bus_dma_tag_t hn_tx_rndis_dtag; + int hn_tx_chimney_size; + int hn_tx_chimney_max; + + struct mtx hn_txlist_spin; + struct hn_txdesc_list hn_txlist; + int hn_txdesc_avail; + int hn_txeof; + struct lro_ctrl hn_lro; int hn_lro_hiwat; @@ -1013,6 +1032,11 @@ typedef struct hn_softc { u_long hn_csum_trusted; u_long hn_lro_tried; u_long hn_small_pkts; + u_long hn_no_txdescs; + u_long hn_send_failed; + u_long hn_txdma_failed; + u_long hn_tx_collapsed; + u_long hn_tx_chimney; } hn_softc_t; Modified: stable/10/sys/dev/hyperv/netvsc/hv_netvsc_drv_freebsd.c ============================================================================== --- stable/10/sys/dev/hyperv/netvsc/hv_netvsc_drv_freebsd.c Wed Feb 24 01:11:51 2016 (r295947) +++ stable/10/sys/dev/hyperv/netvsc/hv_netvsc_drv_freebsd.c Wed Feb 24 01:30:50 2016 (r295948) @@ -129,6 +129,41 @@ __FBSDID("$FreeBSD$"); #define HV_NV_SC_PTR_OFFSET_IN_BUF 0 #define HV_NV_PACKET_OFFSET_IN_BUF 16 +/* YYY should get it from the underlying channel */ +#define HN_TX_DESC_CNT 512 + +#define HN_RNDIS_MSG_LEN \ + (sizeof(rndis_msg) + \ + RNDIS_VLAN_PPI_SIZE + \ + RNDIS_TSO_PPI_SIZE + \ + RNDIS_CSUM_PPI_SIZE) +#define HN_RNDIS_MSG_BOUNDARY PAGE_SIZE +#define HN_RNDIS_MSG_ALIGN CACHE_LINE_SIZE + +#define HN_TX_DATA_BOUNDARY PAGE_SIZE +#define HN_TX_DATA_MAXSIZE IP_MAXPACKET +#define HN_TX_DATA_SEGSIZE PAGE_SIZE +#define HN_TX_DATA_SEGCNT_MAX \ + (NETVSC_PACKET_MAXPAGE - HV_RF_NUM_TX_RESERVED_PAGE_BUFS) + +struct hn_txdesc { + SLIST_ENTRY(hn_txdesc) link; + struct mbuf *m; + struct hn_softc *sc; + int refs; + uint32_t flags; /* HN_TXD_FLAG_ */ + netvsc_packet netvsc_pkt; /* XXX to be removed */ + + bus_dmamap_t data_dmap; + + bus_addr_t rndis_msg_paddr; + rndis_msg *rndis_msg; + bus_dmamap_t rndis_msg_dmap; +}; + +#define HN_TXD_FLAG_ONLIST 0x1 +#define HN_TXD_FLAG_DMAMAP 0x2 + /* * A unified flag for all outbound check sum flags is useful, * and it helps avoiding unnecessary check sum calculation in @@ -174,6 +209,16 @@ int hv_promisc_mode = 0; /* normal mo static int hn_trust_hosttcp = 0; TUNABLE_INT("dev.hn.trust_hosttcp", &hn_trust_hosttcp); +#if __FreeBSD_version >= 1100045 +/* Limit TSO burst size */ +static int hn_tso_maxlen = 0; +TUNABLE_INT("dev.hn.tso_maxlen", &hn_tso_maxlen); +#endif + +/* Limit chimney send size */ +static int hn_tx_chimney_size = 0; +TUNABLE_INT("dev.hn.tx_chimney_size", &hn_tx_chimney_size); + /* * Forward declarations */ @@ -181,14 +226,17 @@ static void hn_stop(hn_softc_t *sc); static void hn_ifinit_locked(hn_softc_t *sc); static void hn_ifinit(void *xsc); static int hn_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data); -static int hn_start_locked(struct ifnet *ifp); +static void hn_start_locked(struct ifnet *ifp); static void hn_start(struct ifnet *ifp); static int hn_ifmedia_upd(struct ifnet *ifp); static void hn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr); #ifdef HN_LRO_HIWAT static int hn_lro_hiwat_sysctl(SYSCTL_HANDLER_ARGS); #endif +static int hn_tx_chimney_size_sysctl(SYSCTL_HANDLER_ARGS); static int hn_check_iplen(const struct mbuf *, int); +static int hn_create_tx_ring(struct hn_softc *sc); +static void hn_destroy_tx_ring(struct hn_softc *sc); static __inline void hn_set_lro_hiwat(struct hn_softc *sc, int hiwat) @@ -318,10 +366,13 @@ netvsc_attach(device_t dev) netvsc_device_info device_info; hn_softc_t *sc; int unit = device_get_unit(dev); - struct ifnet *ifp; + struct ifnet *ifp = NULL; struct sysctl_oid_list *child; struct sysctl_ctx_list *ctx; - int ret; + int error; +#if __FreeBSD_version >= 1100045 + int tso_maxlen; +#endif sc = device_get_softc(dev); if (sc == NULL) { @@ -334,6 +385,10 @@ netvsc_attach(device_t dev) sc->hn_lro_hiwat = HN_LRO_HIWAT_DEF; sc->hn_trust_hosttcp = hn_trust_hosttcp; + error = hn_create_tx_ring(sc); + if (error) + goto failed; + NV_LOCK_INIT(sc, "NetVSCLock"); sc->hn_dev_obj = device_ctx; @@ -381,12 +436,10 @@ netvsc_attach(device_t dev) else ifp->if_hwassist = CSUM_TCP | CSUM_TSO; - ret = hv_rf_on_device_add(device_ctx, &device_info); - if (ret != 0) { - if_free(ifp); + error = hv_rf_on_device_add(device_ctx, &device_info); + if (error) + goto failed; - return (ret); - } if (device_info.link_state == 0) { sc->hn_carrier = 1; } @@ -400,8 +453,30 @@ netvsc_attach(device_t dev) #endif #endif /* INET || INET6 */ +#if __FreeBSD_version >= 1100045 + tso_maxlen = hn_tso_maxlen; + if (tso_maxlen <= 0 || tso_maxlen > IP_MAXPACKET) + tso_maxlen = IP_MAXPACKET; + + ifp->if_hw_tsomaxsegcount = HN_TX_DATA_SEGCNT_MAX; + ifp->if_hw_tsomaxsegsize = PAGE_SIZE; + ifp->if_hw_tsomax = tso_maxlen - + (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN); +#endif + ether_ifattach(ifp, device_info.mac_addr); +#if __FreeBSD_version >= 1100045 + if_printf(ifp, "TSO: %u/%u/%u\n", ifp->if_hw_tsomax, + ifp->if_hw_tsomaxsegcount, ifp->if_hw_tsomaxsegsize); +#endif + + sc->hn_tx_chimney_max = sc->net_dev->send_section_size; + sc->hn_tx_chimney_size = sc->hn_tx_chimney_max; + if (hn_tx_chimney_size > 0 && + hn_tx_chimney_size < sc->hn_tx_chimney_max) + sc->hn_tx_chimney_size = hn_tx_chimney_size; + ctx = device_get_sysctl_ctx(dev); child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); @@ -429,6 +504,26 @@ netvsc_attach(device_t dev) "# of TCP segements that we trust host's csum verification"); SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "small_pkts", CTLFLAG_RW, &sc->hn_small_pkts, "# of small packets received"); + SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "no_txdescs", + CTLFLAG_RW, &sc->hn_no_txdescs, "# of times short of TX descs"); + SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "send_failed", + CTLFLAG_RW, &sc->hn_send_failed, "# of hyper-v sending failure"); + SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "txdma_failed", + CTLFLAG_RW, &sc->hn_txdma_failed, "# of TX DMA failure"); + SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "tx_collapsed", + CTLFLAG_RW, &sc->hn_tx_collapsed, "# of TX mbuf collapsed"); + SYSCTL_ADD_ULONG(ctx, child, OID_AUTO, "tx_chimney", + CTLFLAG_RW, &sc->hn_tx_chimney, "# of chimney send"); + SYSCTL_ADD_INT(ctx, child, OID_AUTO, "txdesc_cnt", + CTLFLAG_RD, &sc->hn_txdesc_cnt, 0, "# of total TX descs"); + SYSCTL_ADD_INT(ctx, child, OID_AUTO, "txdesc_avail", + CTLFLAG_RD, &sc->hn_txdesc_avail, 0, "# of available TX descs"); + SYSCTL_ADD_INT(ctx, child, OID_AUTO, "tx_chimney_max", + CTLFLAG_RD, &sc->hn_tx_chimney_max, 0, + "Chimney send packet size upper boundary"); + SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "tx_chimney_size", + CTLTYPE_INT | CTLFLAG_RW, sc, 0, hn_tx_chimney_size_sysctl, + "I", "Chimney send packet size limit"); if (unit == 0) { struct sysctl_ctx_list *dc_ctx; @@ -446,9 +541,21 @@ netvsc_attach(device_t dev) CTLFLAG_RD, &hn_trust_hosttcp, 0, "Trust tcp segement verification on host side, " "when csum info is missing (global setting)"); + SYSCTL_ADD_INT(dc_ctx, dc_child, OID_AUTO, "tx_chimney_size", + CTLFLAG_RD, &hn_tx_chimney_size, 0, + "Chimney send packet size limit"); +#if __FreeBSD_version >= 1100045 + SYSCTL_ADD_INT(dc_ctx, dc_child, OID_AUTO, "tso_maxlen", + CTLFLAG_RD, &hn_tso_maxlen, 0, "TSO burst limit"); +#endif } return (0); +failed: + hn_destroy_tx_ring(sc); + if (ifp != NULL) + if_free(ifp); + return (error); } /* @@ -480,6 +587,7 @@ netvsc_detach(device_t dev) #if defined(INET) || defined(INET6) tcp_lro_free(&sc->hn_lro); #endif + hn_destroy_tx_ring(sc); return (0); } @@ -493,6 +601,112 @@ netvsc_shutdown(device_t dev) return (0); } +static __inline int +hn_txdesc_dmamap_load(struct hn_softc *sc, struct hn_txdesc *txd, + struct mbuf **m_head, bus_dma_segment_t *segs, int *nsegs) +{ + struct mbuf *m = *m_head; + int error; + + error = bus_dmamap_load_mbuf_sg(sc->hn_tx_data_dtag, txd->data_dmap, + m, segs, nsegs, BUS_DMA_NOWAIT); + if (error == EFBIG) { + struct mbuf *m_new; + + m_new = m_collapse(m, M_NOWAIT, HN_TX_DATA_SEGCNT_MAX); + if (m_new == NULL) + return ENOBUFS; + else + *m_head = m = m_new; + sc->hn_tx_collapsed++; + + error = bus_dmamap_load_mbuf_sg(sc->hn_tx_data_dtag, + txd->data_dmap, m, segs, nsegs, BUS_DMA_NOWAIT); + } + if (!error) { + bus_dmamap_sync(sc->hn_tx_data_dtag, txd->data_dmap, + BUS_DMASYNC_PREWRITE); + txd->flags |= HN_TXD_FLAG_DMAMAP; + } + return error; +} + +static __inline void +hn_txdesc_dmamap_unload(struct hn_softc *sc, struct hn_txdesc *txd) +{ + + if (txd->flags & HN_TXD_FLAG_DMAMAP) { + bus_dmamap_sync(sc->hn_tx_data_dtag, + txd->data_dmap, BUS_DMASYNC_POSTWRITE); + bus_dmamap_unload(sc->hn_tx_data_dtag, + txd->data_dmap); + txd->flags &= ~HN_TXD_FLAG_DMAMAP; + } +} + +static __inline int +hn_txdesc_put(struct hn_softc *sc, struct hn_txdesc *txd) +{ + + KASSERT((txd->flags & HN_TXD_FLAG_ONLIST) == 0, + ("put an onlist txd %#x", txd->flags)); + + KASSERT(txd->refs > 0, ("invalid txd refs %d", txd->refs)); + if (atomic_fetchadd_int(&txd->refs, -1) != 1) + return 0; + + hn_txdesc_dmamap_unload(sc, txd); + if (txd->m != NULL) { + m_freem(txd->m); + txd->m = NULL; + } + + txd->flags |= HN_TXD_FLAG_ONLIST; + + mtx_lock_spin(&sc->hn_txlist_spin); + KASSERT(sc->hn_txdesc_avail >= 0 && + sc->hn_txdesc_avail < sc->hn_txdesc_cnt, + ("txdesc_put: invalid txd avail %d", sc->hn_txdesc_avail)); + sc->hn_txdesc_avail++; + SLIST_INSERT_HEAD(&sc->hn_txlist, txd, link); + mtx_unlock_spin(&sc->hn_txlist_spin); + + return 1; +} + +static __inline struct hn_txdesc * +hn_txdesc_get(struct hn_softc *sc) +{ + struct hn_txdesc *txd; + + mtx_lock_spin(&sc->hn_txlist_spin); + txd = SLIST_FIRST(&sc->hn_txlist); + if (txd != NULL) { + KASSERT(sc->hn_txdesc_avail > 0, + ("txdesc_get: invalid txd avail %d", sc->hn_txdesc_avail)); + sc->hn_txdesc_avail--; + SLIST_REMOVE_HEAD(&sc->hn_txlist, link); + } + mtx_unlock_spin(&sc->hn_txlist_spin); + + if (txd != NULL) { + KASSERT(txd->m == NULL && txd->refs == 0 && + (txd->flags & HN_TXD_FLAG_ONLIST), ("invalid txd")); + txd->flags &= ~HN_TXD_FLAG_ONLIST; + txd->refs = 1; + } + return txd; +} + +static __inline void +hn_txdesc_hold(struct hn_txdesc *txd) +{ + + /* 0->1 transition will never work */ + KASSERT(txd->refs > 0, ("invalid refs %d", txd->refs)); + atomic_add_int(&txd->refs, 1); +} + /* * Send completion processing * @@ -503,34 +717,46 @@ netvsc_shutdown(device_t dev) void netvsc_xmit_completion(void *context) { - netvsc_packet *packet = (netvsc_packet *)context; - struct mbuf *mb; - uint8_t *buf; + netvsc_packet *packet = context; + struct hn_txdesc *txd; + struct hn_softc *sc; + + txd = (struct hn_txdesc *)(uintptr_t) + packet->compl.send.send_completion_tid; + + sc = txd->sc; + sc->hn_txeof = 1; + hn_txdesc_put(sc, txd); +} - mb = (struct mbuf *)(uintptr_t)packet->compl.send.send_completion_tid; - buf = ((uint8_t *)packet) - HV_NV_PACKET_OFFSET_IN_BUF; +void +netvsc_channel_rollup(struct hv_device *device_ctx) +{ + struct hn_softc *sc = device_get_softc(device_ctx->device); + struct ifnet *ifp; - free(buf, M_NETVSC); + if (!sc->hn_txeof) + return; - if (mb != NULL) { - m_freem(mb); - } + sc->hn_txeof = 0; + ifp = sc->hn_ifp; + NV_LOCK(sc); + ifp->if_drv_flags &= ~IFF_DRV_OACTIVE; + hn_start_locked(ifp); + NV_UNLOCK(sc); } /* * Start a transmit of one or more packets */ -static int +static void hn_start_locked(struct ifnet *ifp) { hn_softc_t *sc = ifp->if_softc; struct hv_device *device_ctx = vmbus_get_devctx(sc->hn_dev); netvsc_dev *net_dev = sc->net_dev; - device_t dev = device_ctx->device; - uint8_t *buf; netvsc_packet *packet; struct mbuf *m_head, *m; - struct mbuf *mc_head = NULL; struct ether_vlan_header *eh; rndis_msg *rndis_mesg; rndis_packet *rndis_pkt; @@ -539,84 +765,40 @@ hn_start_locked(struct ifnet *ifp) rndis_tcp_ip_csum_info *csum_info; rndis_tcp_tso_info *tso_info; int ether_len; - int i; - int num_frags; - int len; - int retries = 0; - int ret = 0; uint32_t rndis_msg_size = 0; uint32_t trans_proto_type; uint32_t send_buf_section_idx = NVSP_1_CHIMNEY_SEND_INVALID_SECTION_INDEX; - while (!IFQ_DRV_IS_EMPTY(&sc->hn_ifp->if_snd)) { - IFQ_DRV_DEQUEUE(&sc->hn_ifp->if_snd, m_head); - if (m_head == NULL) { - break; - } - - len = 0; - num_frags = 0; - - /* Walk the mbuf list computing total length and num frags */ - for (m = m_head; m != NULL; m = m->m_next) { - if (m->m_len != 0) { - num_frags++; - len += m->m_len; - } - } + if ((ifp->if_drv_flags & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != + IFF_DRV_RUNNING) + return; - /* - * Reserve the number of pages requested. Currently, - * one page is reserved for the message in the RNDIS - * filter packet - */ - num_frags += HV_RF_NUM_TX_RESERVED_PAGE_BUFS; + while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) { + bus_dma_segment_t segs[HN_TX_DATA_SEGCNT_MAX]; + int error, nsegs, i, send_failed = 0; + struct hn_txdesc *txd; - /* If exceeds # page_buffers in netvsc_packet */ - if (num_frags > NETVSC_PACKET_MAXPAGE) { - device_printf(dev, "exceed max page buffers,%d,%d\n", - num_frags, NETVSC_PACKET_MAXPAGE); - m_freem(m_head); - if_inc_counter(ifp, IFCOUNTER_OERRORS, 1); - return (EINVAL); - } + IFQ_DRV_DEQUEUE(&ifp->if_snd, m_head); + if (m_head == NULL) + break; - /* - * Allocate a buffer with space for a netvsc packet plus a - * number of reserved areas. First comes a (currently 16 - * bytes, currently unused) reserved data area. Second is - * the netvsc_packet. Third is an area reserved for an - * rndis_filter_packet struct. Fourth (optional) is a - * rndis_per_packet_info struct. - * Changed malloc to M_NOWAIT to avoid sleep under spin lock. - * No longer reserving extra space for page buffers, as they - * are already part of the netvsc_packet. - */ - buf = malloc(HV_NV_PACKET_OFFSET_IN_BUF + - sizeof(netvsc_packet) + - sizeof(rndis_msg) + - RNDIS_VLAN_PPI_SIZE + - RNDIS_TSO_PPI_SIZE + - RNDIS_CSUM_PPI_SIZE, - M_NETVSC, M_ZERO | M_NOWAIT); - if (buf == NULL) { - device_printf(dev, "hn:malloc packet failed\n"); - m_freem(m_head); - if_inc_counter(ifp, IFCOUNTER_OERRORS, 1); - return (ENOMEM); + txd = hn_txdesc_get(sc); + if (txd == NULL) { + sc->hn_no_txdescs++; + IF_PREPEND(&ifp->if_snd, m_head); + ifp->if_drv_flags |= IFF_DRV_OACTIVE; + break; } - packet = (netvsc_packet *)(buf + HV_NV_PACKET_OFFSET_IN_BUF); - *(vm_offset_t *)buf = HV_NV_SC_PTR_OFFSET_IN_BUF; + packet = &txd->netvsc_pkt; + /* XXX not necessary */ + memset(packet, 0, sizeof(*packet)); packet->is_data_pkt = TRUE; - /* Set up the rndis header */ - packet->page_buf_count = num_frags; - /* Initialize it from the mbuf */ - packet->tot_data_buf_len = len; + packet->tot_data_buf_len = m_head->m_pkthdr.len; /* * extension points to the area reserved for the @@ -624,8 +806,9 @@ hn_start_locked(struct ifnet *ifp) * the netvsc_packet (and rppi struct, if present; * length is updated later). */ - packet->rndis_mesg = packet + 1; - rndis_mesg = (rndis_msg *)packet->rndis_mesg; + rndis_mesg = txd->rndis_msg; + /* XXX not necessary */ + memset(rndis_mesg, 0, HN_RNDIS_MSG_LEN); rndis_mesg->ndis_msg_type = REMOTE_NDIS_PACKET_MSG; rndis_pkt = &rndis_mesg->msg.packet; @@ -644,8 +827,6 @@ hn_start_locked(struct ifnet *ifp) * set up some additional fields so the Hyper-V infrastructure will stuff the VLAN tag * into the frame. */ - packet->vlan_tci = m_head->m_pkthdr.ether_vtag; - rndis_msg_size += RNDIS_VLAN_PPI_SIZE; rppi = hv_set_rppi_data(rndis_mesg, RNDIS_VLAN_PPI_SIZE, @@ -656,7 +837,7 @@ hn_start_locked(struct ifnet *ifp) rppi->per_packet_info_offset); /* FreeBSD does not support CFI or priority */ rppi_vlan_info->u1.s1.vlan_id = - packet->vlan_tci & 0xfff; + m_head->m_pkthdr.ether_vtag & 0xfff; } /* Only check the flags for outbound and ignore the ones for inbound */ @@ -758,7 +939,7 @@ pre_send: packet->tot_data_buf_len = rndis_mesg->msg_len; /* send packet with send buffer */ - if (packet->tot_data_buf_len < net_dev->send_section_size) { + if (packet->tot_data_buf_len < sc->hn_tx_chimney_size) { send_buf_section_idx = hv_nv_get_next_send_section(net_dev); if (send_buf_section_idx != @@ -783,33 +964,49 @@ pre_send: packet->send_buf_section_size = packet->tot_data_buf_len; packet->page_buf_count = 0; + sc->hn_tx_chimney++; goto do_send; } } + error = hn_txdesc_dmamap_load(sc, txd, &m_head, segs, &nsegs); + if (error) { + int freed; + + /* + * This mbuf is not linked w/ the txd yet, so free + * it now. + */ + m_freem(m_head); + freed = hn_txdesc_put(sc, txd); + KASSERT(freed != 0, + ("fail to free txd upon txdma error")); + + sc->hn_txdma_failed++; + if_inc_counter(ifp, IFCOUNTER_OERRORS, 1); + continue; + } + + packet->page_buf_count = nsegs + + HV_RF_NUM_TX_RESERVED_PAGE_BUFS; + /* send packet with page buffer */ - packet->page_buffers[0].pfn = - atop(hv_get_phys_addr(rndis_mesg)); + packet->page_buffers[0].pfn = atop(txd->rndis_msg_paddr); packet->page_buffers[0].offset = - (unsigned long)rndis_mesg & PAGE_MASK; + txd->rndis_msg_paddr & PAGE_MASK; packet->page_buffers[0].length = rndis_msg_size; /* * Fill the page buffers with mbuf info starting at index * HV_RF_NUM_TX_RESERVED_PAGE_BUFS. */ - i = HV_RF_NUM_TX_RESERVED_PAGE_BUFS; - for (m = m_head; m != NULL; m = m->m_next) { - if (m->m_len) { - vm_offset_t paddr = - vtophys(mtod(m, vm_offset_t)); - packet->page_buffers[i].pfn = - paddr >> PAGE_SHIFT; - packet->page_buffers[i].offset = - paddr & (PAGE_SIZE - 1); - packet->page_buffers[i].length = m->m_len; - i++; - } + for (i = 0; i < nsegs; ++i) { + hv_vmbus_page_buffer *pb = &packet->page_buffers[ + i + HV_RF_NUM_TX_RESERVED_PAGE_BUFS]; + + pb->pfn = atop(segs[i].ds_addr); + pb->offset = segs[i].ds_addr & PAGE_MASK; + pb->length = segs[i].ds_len; } packet->send_buf_section_idx = @@ -817,63 +1014,65 @@ pre_send: packet->send_buf_section_size = 0; do_send: + txd->m = m_head; - /* - * If bpf, copy the mbuf chain. This is less expensive than - * it appears; the mbuf clusters are not copied, only their - * reference counts are incremented. - * Needed to avoid a race condition where the completion - * callback is invoked, freeing the mbuf chain, before the - * bpf_mtap code has a chance to run. - */ - if (ifp->if_bpf) { - mc_head = m_copypacket(m_head, M_DONTWAIT); - } -retry_send: /* Set the completion routine */ packet->compl.send.on_send_completion = netvsc_xmit_completion; packet->compl.send.send_completion_context = packet; - packet->compl.send.send_completion_tid = (uint64_t)(uintptr_t)m_head; + packet->compl.send.send_completion_tid = + (uint64_t)(uintptr_t)txd; - /* Removed critical_enter(), does not appear necessary */ - ret = hv_nv_on_send(device_ctx, packet); - if (ret == 0) { - ifp->if_opackets++; - /* if bpf && mc_head, call bpf_mtap code */ - if (mc_head) { - ETHER_BPF_MTAP(ifp, mc_head); - } - } else { - retries++; - if (retries < 4) { - goto retry_send; - } +again: + /* + * Make sure that txd is not freed before ETHER_BPF_MTAP. + */ + hn_txdesc_hold(txd); + error = hv_nv_on_send(device_ctx, packet); + if (!error) { + ETHER_BPF_MTAP(ifp, m_head); + if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1); + } + hn_txdesc_put(sc, txd); - IF_PREPEND(&ifp->if_snd, m_head); - ifp->if_drv_flags |= IFF_DRV_OACTIVE; + if (__predict_false(error)) { + int freed; /* - * Null the mbuf pointer so the completion function - * does not free the mbuf chain. We just pushed the - * mbuf chain back on the if_snd queue. + * This should "really rarely" happen. + * + * XXX Too many RX to be acked or too many sideband + * commands to run? Ask netvsc_channel_rollup() + * to kick start later. */ - packet->compl.send.send_completion_tid = 0; + sc->hn_txeof = 1; + if (!send_failed) { + sc->hn_send_failed++; + send_failed = 1; + /* + * Try sending again after set hn_txeof; + * in case that we missed the last + * netvsc_channel_rollup(). + */ + goto again; + } + if_printf(ifp, "send failed\n"); /* - * Release the resources since we will not get any - * send completion + * This mbuf will be prepended, don't free it + * in hn_txdesc_put(); only unload it from the + * DMA map in hn_txdesc_put(), if it was loaded. */ - netvsc_xmit_completion(packet); - if_inc_counter(ifp, IFCOUNTER_OERRORS, 1); - } + txd->m = NULL; + freed = hn_txdesc_put(sc, txd); + KASSERT(freed != 0, + ("fail to free txd upon send error")); - /* if bpf && mc_head, free the mbuf chain copy */ - if (mc_head) { - m_freem(mc_head); + sc->hn_send_failed++; + IF_PREPEND(&ifp->if_snd, m_head); + ifp->if_drv_flags |= IFF_DRV_OACTIVE; + break; } } - - return (ret); } /* @@ -1222,6 +1421,9 @@ hn_ioctl(struct ifnet *ifp, u_long cmd, break; } + sc->hn_tx_chimney_max = sc->net_dev->send_section_size; + if (sc->hn_tx_chimney_size > sc->hn_tx_chimney_max) + sc->hn_tx_chimney_size = sc->hn_tx_chimney_max; hn_ifinit_locked(sc); NV_LOCK(sc); @@ -1479,6 +1681,25 @@ hn_lro_hiwat_sysctl(SYSCTL_HANDLER_ARGS) #endif /* HN_LRO_HIWAT */ static int +hn_tx_chimney_size_sysctl(SYSCTL_HANDLER_ARGS) +{ + struct hn_softc *sc = arg1; + int chimney_size, error; + + chimney_size = sc->hn_tx_chimney_size; + error = sysctl_handle_int(oidp, &chimney_size, 0, req); + if (error || req->newptr == NULL) + return error; + + if (chimney_size > sc->hn_tx_chimney_max || chimney_size <= 0) + return EINVAL; + + if (sc->hn_tx_chimney_size != chimney_size) + sc->hn_tx_chimney_size = chimney_size; + return 0; +} + +static int hn_check_iplen(const struct mbuf *m, int hoff) { const struct ip *ip; @@ -1553,6 +1774,150 @@ hn_check_iplen(const struct mbuf *m, int return ip->ip_p; } +static void +hn_dma_map_paddr(void *arg, bus_dma_segment_t *segs, int nseg, int error) +{ + bus_addr_t *paddr = arg; + + if (error) + return; + + KASSERT(nseg == 1, ("too many segments %d!", nseg)); + *paddr = segs->ds_addr; +} + +static int +hn_create_tx_ring(struct hn_softc *sc) +{ + bus_dma_tag_t parent_dtag; + int error, i; + + sc->hn_txdesc_cnt = HN_TX_DESC_CNT; + sc->hn_txdesc = malloc(sizeof(struct hn_txdesc) * sc->hn_txdesc_cnt, + M_NETVSC, M_WAITOK | M_ZERO); + SLIST_INIT(&sc->hn_txlist); + mtx_init(&sc->hn_txlist_spin, "hn txlist", NULL, MTX_SPIN); + + parent_dtag = bus_get_dma_tag(sc->hn_dev); + + /* DMA tag for RNDIS messages. */ + error = bus_dma_tag_create(parent_dtag, /* parent */ + HN_RNDIS_MSG_ALIGN, /* alignment */ + HN_RNDIS_MSG_BOUNDARY, /* boundary */ + BUS_SPACE_MAXADDR, /* lowaddr */ + BUS_SPACE_MAXADDR, /* highaddr */ + NULL, NULL, /* filter, filterarg */ + HN_RNDIS_MSG_LEN, /* maxsize */ + 1, /* nsegments */ + HN_RNDIS_MSG_LEN, /* maxsegsize */ + 0, /* flags */ + NULL, /* lockfunc */ + NULL, /* lockfuncarg */ + &sc->hn_tx_rndis_dtag); + if (error) { + device_printf(sc->hn_dev, "failed to create rndis dmatag\n"); + return error; + } + + /* DMA tag for data. */ + error = bus_dma_tag_create(parent_dtag, /* parent */ + 1, /* alignment */ + HN_TX_DATA_BOUNDARY, /* boundary */ + BUS_SPACE_MAXADDR, /* lowaddr */ + BUS_SPACE_MAXADDR, /* highaddr */ + NULL, NULL, /* filter, filterarg */ + HN_TX_DATA_MAXSIZE, /* maxsize */ + HN_TX_DATA_SEGCNT_MAX, /* nsegments */ + HN_TX_DATA_SEGSIZE, /* maxsegsize */ + 0, /* flags */ + NULL, /* lockfunc */ + NULL, /* lockfuncarg */ + &sc->hn_tx_data_dtag); + if (error) { + device_printf(sc->hn_dev, "failed to create data dmatag\n"); + return error; + } + + for (i = 0; i < sc->hn_txdesc_cnt; ++i) { + struct hn_txdesc *txd = &sc->hn_txdesc[i]; + + txd->sc = sc; + + /* + * Allocate and load RNDIS messages. + */ + error = bus_dmamem_alloc(sc->hn_tx_rndis_dtag, + (void **)&txd->rndis_msg, + BUS_DMA_WAITOK | BUS_DMA_COHERENT, + &txd->rndis_msg_dmap); + if (error) { + device_printf(sc->hn_dev, + "failed to allocate rndis_msg, %d\n", i); + return error; + } + + error = bus_dmamap_load(sc->hn_tx_rndis_dtag, + txd->rndis_msg_dmap, + txd->rndis_msg, HN_RNDIS_MSG_LEN, + hn_dma_map_paddr, &txd->rndis_msg_paddr, + BUS_DMA_NOWAIT); + if (error) { + device_printf(sc->hn_dev, + "failed to load rndis_msg, %d\n", i); + bus_dmamem_free(sc->hn_tx_rndis_dtag, + txd->rndis_msg, txd->rndis_msg_dmap); + return error; + } + + /* DMA map for TX data. */ + error = bus_dmamap_create(sc->hn_tx_data_dtag, 0, + &txd->data_dmap); + if (error) { + device_printf(sc->hn_dev, + "failed to allocate tx data dmamap\n"); + bus_dmamap_unload(sc->hn_tx_rndis_dtag, + txd->rndis_msg_dmap); + bus_dmamem_free(sc->hn_tx_rndis_dtag, + txd->rndis_msg, txd->rndis_msg_dmap); + return error; + } + + /* All set, put it to list */ + txd->flags |= HN_TXD_FLAG_ONLIST; + SLIST_INSERT_HEAD(&sc->hn_txlist, txd, link); + } + sc->hn_txdesc_avail = sc->hn_txdesc_cnt; + + return 0; +} + +static void +hn_destroy_tx_ring(struct hn_softc *sc) +{ + struct hn_txdesc *txd; + + while ((txd = SLIST_FIRST(&sc->hn_txlist)) != NULL) { + KASSERT(txd->m == NULL, ("still has mbuf installed")); + KASSERT((txd->flags & HN_TXD_FLAG_DMAMAP) == 0, + ("still dma mapped")); + SLIST_REMOVE_HEAD(&sc->hn_txlist, link); + + bus_dmamap_unload(sc->hn_tx_rndis_dtag, + txd->rndis_msg_dmap); + bus_dmamem_free(sc->hn_tx_rndis_dtag, + txd->rndis_msg, txd->rndis_msg_dmap); + + bus_dmamap_destroy(sc->hn_tx_data_dtag, txd->data_dmap); + } + + if (sc->hn_tx_data_dtag != NULL) + bus_dma_tag_destroy(sc->hn_tx_data_dtag); + if (sc->hn_tx_rndis_dtag != NULL) + bus_dma_tag_destroy(sc->hn_tx_rndis_dtag); + free(sc->hn_txdesc, M_NETVSC); + mtx_destroy(&sc->hn_txlist_spin); +} + static device_method_t netvsc_methods[] = { /* Device interface */ DEVMETHOD(device_probe, netvsc_probe), Modified: stable/10/sys/dev/hyperv/netvsc/hv_rndis.h ============================================================================== --- stable/10/sys/dev/hyperv/netvsc/hv_rndis.h Wed Feb 24 01:11:51 2016 (r295947) +++ stable/10/sys/dev/hyperv/netvsc/hv_rndis.h Wed Feb 24 01:30:50 2016 (r295948) @@ -1050,6 +1050,7 @@ int netvsc_recv(struct hv_device *device netvsc_packet *packet, rndis_tcp_ip_csum_info *csum_info); void netvsc_recv_rollup(struct hv_device *device_ctx); +void netvsc_channel_rollup(struct hv_device *device_ctx); void* hv_set_rppi_data(rndis_msg *rndis_mesg, uint32_t rppi_size, Modified: stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.c ============================================================================== --- stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.c Wed Feb 24 01:11:51 2016 (r295947) +++ stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.c Wed Feb 24 01:30:50 2016 (r295948) @@ -974,3 +974,21 @@ hv_rf_receive_rollup(netvsc_dev *net_dev rndis_dev = (rndis_device *)net_dev->extension; netvsc_recv_rollup(rndis_dev->net_dev->dev); } + +void +hv_rf_channel_rollup(netvsc_dev *net_dev) +{ + rndis_device *rndis_dev; + + rndis_dev = (rndis_device *)net_dev->extension; + + /* + * This could be called pretty early, so we need + * to make sure everything has been setup. + */ + if (rndis_dev == NULL || + rndis_dev->net_dev == NULL || + rndis_dev->net_dev->dev == NULL) + return; + netvsc_channel_rollup(rndis_dev->net_dev->dev); +} Modified: stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.h ============================================================================== --- stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.h Wed Feb 24 01:11:51 2016 (r295947) +++ stable/10/sys/dev/hyperv/netvsc/hv_rndis_filter.h Wed Feb 24 01:30:50 2016 (r295948) @@ -99,6 +99,7 @@ typedef struct rndis_device_ { int hv_rf_on_receive(netvsc_dev *net_dev, struct hv_device *device, netvsc_packet *pkt); void hv_rf_receive_rollup(netvsc_dev *net_dev); +void hv_rf_channel_rollup(netvsc_dev *net_dev); int hv_rf_on_device_add(struct hv_device *device, void *additl_info); *** DIFF OUTPUT TRUNCATED AT 1000 LINES ***
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201602240130.u1O1Uoko011921>