X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/d1ecb069dfe24481e4a83f44cb5217a2b06746d7..143464d58d2bd6378e74eec636961ceb0d32fb91:/bsd/netinet/tcp_output.c diff --git a/bsd/netinet/tcp_output.c b/bsd/netinet/tcp_output.c index d7f76c663..fa6e5348c 100644 --- a/bsd/netinet/tcp_output.c +++ b/bsd/netinet/tcp_output.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2000-2008 Apple Inc. All rights reserved. + * Copyright (c) 2000-2014 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * @@ -81,7 +81,11 @@ #include #include +#include #include +#include +#include +#include #include #include @@ -89,6 +93,7 @@ #include #include #include +#include #if INET6 #include #include @@ -101,10 +106,12 @@ #include #include #include +#include #if TCPDEBUG #include #endif #include +#include #if IPSEC #include @@ -114,52 +121,89 @@ #include #endif /* MAC_SOCKET */ +#include +#if MPTCP +#include +#include +#include +#endif + #define DBG_LAYER_BEG NETDBG_CODE(DBG_NETTCP, 1) #define DBG_LAYER_END NETDBG_CODE(DBG_NETTCP, 3) #define DBG_FNC_TCP_OUTPUT NETDBG_CODE(DBG_NETTCP, (4 << 8) | 1) - -#ifdef notyet -extern struct mbuf *m_copypack(); -#endif - int path_mtu_discovery = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW | CTLFLAG_LOCKED, &path_mtu_discovery, 1, "Enable Path MTU Discovery"); int ss_fltsz = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW | CTLFLAG_LOCKED, &ss_fltsz, 1, "Slow start flight size"); int ss_fltsz_local = 8; /* starts with eight segments max */ -SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW | CTLFLAG_LOCKED, &ss_fltsz_local, 1, "Slow start flight size for local networks"); -int tcp_do_newreno = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW, &tcp_do_newreno, - 0, "Enable NewReno Algorithms"); - int tcp_do_tso = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_do_tso, 0, "Enable TCP Segmentation Offload"); int tcp_ecn_outbound = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, ecn_initiate_out, CTLFLAG_RW, &tcp_ecn_outbound, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, ecn_initiate_out, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_outbound, 0, "Initiate ECN for outbound connections"); int tcp_ecn_inbound = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, ecn_negotiate_in, CTLFLAG_RW, &tcp_ecn_inbound, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, ecn_negotiate_in, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_ecn_inbound, 0, "Allow ECN negotiation for inbound connections"); int tcp_packet_chaining = 50; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, packetchain, CTLFLAG_RW, &tcp_packet_chaining, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, packetchain, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_packet_chaining, 0, "Enable TCP output packet chaining"); int tcp_output_unlocked = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, socket_unlocked_on_output, CTLFLAG_RW, &tcp_output_unlocked, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, socket_unlocked_on_output, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_output_unlocked, 0, "Unlock TCP when sending packets down to IP"); +int tcp_do_rfc3390 = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_do_rfc3390, 1, "Calculate intial slowstart cwnd depending on MSS"); + +int tcp_min_iaj_win = MIN_IAJ_WIN; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, min_iaj_win, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_min_iaj_win, 1, "Minimum recv win based on inter-packet arrival jitter"); + +int tcp_acc_iaj_react_limit = ACC_IAJ_REACT_LIMIT; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, acc_iaj_react_limit, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_acc_iaj_react_limit, 1, "Accumulated IAJ when receiver starts to react"); + +uint32_t tcp_do_autosendbuf = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, doautosndbuf, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_do_autosendbuf, 1, "Enable send socket buffer auto-tuning"); + +uint32_t tcp_autosndbuf_inc = 8 * 1024; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, autosndbufinc, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_autosndbuf_inc, 1, "Increment in send socket bufffer size"); + +uint32_t tcp_autosndbuf_max = 512 * 1024; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, autosndbufmax, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_autosndbuf_max, 1, "Maximum send socket buffer size"); + +uint32_t tcp_prioritize_acks = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, ack_prioritize, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_prioritize_acks, 1, "Prioritize pure acks"); + +uint32_t tcp_use_rtt_recvbg = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, rtt_recvbg, + CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_use_rtt_recvbg, 1, "Use RTT for bg recv algorithm"); + +uint32_t tcp_recv_throttle_minwin = 16 * 1024; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, recv_throttle_minwin, + CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_recv_throttle_minwin, 1, "Minimum recv win for throttling"); + + static int32_t packchain_newlist = 0; static int32_t packchain_looped = 0; static int32_t packchain_sent = 0; @@ -175,30 +219,17 @@ extern int fw_enable; /* firewall check for packet chaining */ extern int fw_bypass; /* firewall check: disable packet chaining if there is rules */ #endif /* IPFIREWALL */ -extern vm_size_t so_cache_zone_element_size; -#if RANDOM_IP_ID -extern int ip_use_randomid; -#endif /* RANDOM_IP_ID */ -extern u_int32_t dlil_filter_count; +extern u_int32_t dlil_filter_disable_tso_count; extern u_int32_t kipf_count; +extern int tcp_recv_bg; +extern int maxseg_unacked; static int tcp_ip_output(struct socket *, struct tcpcb *, struct mbuf *, int, - struct mbuf *, int, int, int32_t); + struct mbuf *, int, int, int32_t, boolean_t); -static __inline__ u_int16_t -get_socket_id(struct socket * s) -{ - u_int16_t val; - - if (so_cache_zone_element_size == 0) { - return (0); - } - val = (u_int16_t)(((uintptr_t)s) / so_cache_zone_element_size); - if (val == 0) { - val = 0xffff; - } - return (val); -} +extern uint32_t get_base_rtt(struct tcpcb *tp); +static struct mbuf* tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th); +static int tcp_recv_throttle(struct tcpcb *tp); /* * Tcp output routine: figure out what should be sent and send it. @@ -217,12 +248,20 @@ get_socket_id(struct socket * s) * ip_output_list:EMSGSIZE * ip_output_list:ENOBUFS * ip_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL] - * ip6_output:??? [IPV6 only] + * ip6_output_list:EINVAL + * ip6_output_list:EOPNOTSUPP + * ip6_output_list:EHOSTUNREACH + * ip6_output_list:EADDRNOTAVAIL + * ip6_output_list:ENETUNREACH + * ip6_output_list:EMSGSIZE + * ip6_output_list:ENOBUFS + * ip6_output_list:??? [ignorable: mostly IPSEC/firewall/DLIL] */ int tcp_output(struct tcpcb *tp) { - struct socket *so = tp->t_inpcb->inp_socket; + struct inpcb *inp = tp->t_inpcb; + struct socket *so = inp->inp_socket; int32_t len, recwin, sendwin, off; int flags, error; register struct mbuf *m; @@ -239,24 +278,35 @@ tcp_output(struct tcpcb *tp) int tso = 0; int sack_bytes_rxmt; struct sackhole *p; -#ifdef IPSEC +#if IPSEC unsigned ipsec_optlen = 0; -#endif - int maxburst = TCP_MAXBURST; +#endif /* IPSEC */ int last_off = 0; - int m_off; - struct mbuf *m_last = NULL; + int m_off = 0; + int idle_time = 0; + struct mbuf *m_lastm = NULL; struct mbuf *m_head = NULL; struct mbuf *packetlist = NULL; - struct mbuf *tp_inp_options = tp->t_inpcb->inp_depend4.inp4_options; + struct mbuf *tp_inp_options = inp->inp_depend4.inp4_options; #if INET6 - int isipv6 = tp->t_inpcb->inp_vflag & INP_IPV6 ; - struct ip6_pktopts *inp6_pktopts = tp->t_inpcb->inp_depend6.inp6_outputopts; + int isipv6 = inp->inp_vflag & INP_IPV6 ; #endif short packchain_listadd = 0; - u_int16_t socket_id = get_socket_id(so); int so_options = so->so_options; struct rtentry *rt; + u_int32_t basertt, svc_flags = 0, allocated_len; + u_int32_t lro_ackmore = (tp->t_lropktlen != 0) ? 1 : 0; + struct mbuf *mnext = NULL; + int sackoptlen = 0; +#if MPTCP + unsigned int *dlenp = NULL; + u_int8_t *finp = NULL; + u_int32_t *sseqp = NULL; + u_int64_t dss_val = 0; + int mptcp_acknow = 0; +#endif /* MPTCP */ + boolean_t cell = FALSE; + boolean_t wifi = FALSE; /* * Determine length of data that should be transmitted, @@ -265,28 +315,17 @@ tcp_output(struct tcpcb *tp) * to send, then transmit; otherwise, investigate further. */ idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); - if (idle && tp->t_rcvtime >= tp->t_rxtcur) { - /* - * We have been idle for "a while" and no acks are - * expected to clock out any data we send -- - * slow start to get ack "clock" running again. - * - * Set the slow-start flight size depending on whether - * this is a local network or not. - */ - if ( -#if INET6 - (isipv6 && in6_localaddr(&tp->t_inpcb->in6p_faddr)) || - (!isipv6 && -#endif - in_localaddr(tp->t_inpcb->inp_faddr) -#if INET6 - ) -#endif - ) - tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local; - else - tp->snd_cwnd = tp->t_maxseg * ss_fltsz; + + /* Since idle_time is signed integer, the following integer subtraction + * will take care of wrap around of tcp_now + */ + idle_time = tcp_now - tp->t_rcvtime; + if (idle && idle_time >= TCP_IDLETIMEOUT(tp)) { + if (CC_ALGO(tp)->after_idle != NULL) + CC_ALGO(tp)->after_idle(tp); + DTRACE_TCP5(cc, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, struct tcphdr *, NULL, + int32_t, TCP_CC_IDLE_TIMEOUT); } tp->t_flags &= ~TF_LASTIDLE; if (idle) { @@ -295,38 +334,50 @@ tcp_output(struct tcpcb *tp) idle = 0; } } +#if MPTCP + if (tp->t_mpflags & TMPF_RESET) { + tcp_check_timer_state(tp); + /* + * Once a RST has been sent for an MPTCP subflow, + * the subflow socket stays around until deleted. + * No packets such as FINs must be sent after RST. + */ + return (0); + } +#endif /* MPTCP */ + again: KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_START, 0,0,0,0,0); #if INET6 if (isipv6) { - KERNEL_DEBUG(DBG_LAYER_BEG, - ((tp->t_inpcb->inp_fport << 16) | tp->t_inpcb->inp_lport), - (((tp->t_inpcb->in6p_laddr.s6_addr16[0] & 0xffff) << 16) | - (tp->t_inpcb->in6p_faddr.s6_addr16[0] & 0xffff)), + ((inp->inp_fport << 16) | inp->inp_lport), + (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) | + (inp->in6p_faddr.s6_addr16[0] & 0xffff)), sendalot,0,0); - } - else + } else #endif { KERNEL_DEBUG(DBG_LAYER_BEG, - ((tp->t_inpcb->inp_fport << 16) | tp->t_inpcb->inp_lport), - (((tp->t_inpcb->inp_laddr.s_addr & 0xffff) << 16) | - (tp->t_inpcb->inp_faddr.s_addr & 0xffff)), + ((inp->inp_fport << 16) | inp->inp_lport), + (((inp->inp_laddr.s_addr & 0xffff) << 16) | + (inp->inp_faddr.s_addr & 0xffff)), sendalot,0,0); + } /* * If the route generation id changed, we need to check that our * local (source) IP address is still valid. If it isn't either * return error or silently do nothing (assuming the address will * come back before the TCP connection times out). */ - rt = tp->t_inpcb->inp_route.ro_rt; - if (rt != NULL && (!(rt->rt_flags & RTF_UP) || - rt->generation_id != route_generation)) { + rt = inp->inp_route.ro_rt; + if (rt != NULL && ROUTE_UNUSABLE(&tp->t_inpcb->inp_route)) { struct ifnet *ifp; - struct in_ifaddr *ia; + struct in_ifaddr *ia = NULL; + struct in6_ifaddr *ia6 = NULL; + int found_srcaddr = 0; /* disable multipages at the socket */ somultipages(so, FALSE); @@ -334,8 +385,20 @@ again: /* Disable TSO for the socket until we know more */ tp->t_flags &= ~TF_TSO; + if (isipv6) { + ia6 = ifa_foraddr6(&inp->in6p_laddr); + if (ia6 != NULL) + found_srcaddr = 1; + } else { + ia = ifa_foraddr(inp->inp_laddr.s_addr); + if (ia != NULL) + found_srcaddr = 1; + } + /* check that the source address is still valid */ - if ((ia = ifa_foraddr(tp->t_inpcb->inp_laddr.s_addr)) == NULL) { + if (found_srcaddr == 0) { + soevent(so, + (SO_FILT_HINT_LOCKED | SO_FILT_HINT_NOSRCADDR)); if (tp->t_state >= TCPS_CLOSE_WAIT) { tcp_drop(tp, EADDRNOTAVAIL); @@ -344,14 +407,16 @@ again: /* set Retransmit timer if it wasn't set * reset Persist timer and shift register as the - * adversed peer window may not be valid anymore + * advertised peer window may not be valid anymore */ - if (!tp->t_timer[TCPT_REXMT]) { - tp->t_timer[TCPT_REXMT] = tp->t_rxtcur; + if (!tp->t_timer[TCPT_REXMT]) { + tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur); if (tp->t_timer[TCPT_PERSIST]) { tp->t_timer[TCPT_PERSIST] = 0; tp->t_rxtshift = 0; + tp->t_persist_stop = 0; + tp->t_rxtstart = 0; } } @@ -364,10 +429,16 @@ again: tcp_drop(tp, EADDRNOTAVAIL); return(EADDRNOTAVAIL); } - else + else { + tcp_check_timer_state(tp); return(0); /* silently ignore, keep data in socket: address may be back */ + } } - ifafree(&ia->ia_ifa); + if (ia != NULL) + IFA_REMREF(&ia->ia_ifa); + + if (ia6 != NULL) + IFA_REMREF(&ia6->ia_ifa); /* * Address is still valid; check for multipages capability @@ -379,7 +450,7 @@ again: tcp_set_tso(tp, ifp); } if (rt->rt_flags & RTF_UP) - rt->generation_id = route_generation; + RT_GENID_SYNC(rt); /* * See if we should do MTU discovery. Don't do it if: * 1) it is disabled via the sysctl @@ -395,7 +466,11 @@ again: tp->t_flags |= TF_PMTUD; RT_UNLOCK(rt); - } + } + + if (rt != NULL) { + cell = IFNET_IS_CELLULAR(rt->rt_ifp); + wifi = (!cell && IFNET_IS_WIFI(rt->rt_ifp)); } /* @@ -403,7 +478,7 @@ again: * snd_nxt. There may be SACK information that allows us to avoid * resending already delivered data. Adjust snd_nxt accordingly. */ - if (tp->sack_enable && SEQ_LT(tp->snd_nxt, tp->snd_max)) + if (SACK_ENABLED(tp) && SEQ_LT(tp->snd_nxt, tp->snd_max)) tcp_sack_adjust(tp); sendalot = 0; off = tp->snd_nxt - tp->snd_una; @@ -427,7 +502,7 @@ again: sack_bytes_rxmt = 0; len = 0; p = NULL; - if (tp->sack_enable && IN_FASTRECOVERY(tp) && + if (SACK_ENABLED(tp) && IN_FASTRECOVERY(tp) && (p = tcp_sack_output(tp, &sack_bytes_rxmt))) { int32_t cwin; @@ -454,8 +529,9 @@ again: /* Can rexmit part of the current hole */ len = ((int32_t)min(cwin, tp->snd_recover - p->rxmit)); - } else + } else { len = ((int32_t)min(cwin, p->end - p->rxmit)); + } if (len > 0) { off = p->rxmit - tp->snd_una; /* update off only if we really transmit SACK data */ sack_rxmit = 1; @@ -463,9 +539,18 @@ again: tcpstat.tcps_sack_rexmits++; tcpstat.tcps_sack_rexmit_bytes += min(len, tp->t_maxseg); - } - else + if (nstat_collect) { + nstat_route_tx(inp->inp_route.ro_rt, 1, + min(len, tp->t_maxseg), + NSTAT_TX_FLAG_RETRANSMIT); + INP_ADD_STAT(inp, cell, wifi, txpackets, 1); + INP_ADD_STAT(inp, cell, wifi, txbytes, + min(len, tp->t_maxseg)); + tp->t_stat.txretransmitbytes += min(len, tp->t_maxseg); + } + } else { len = 0; + } } after_sack_rexmit: /* @@ -507,6 +592,8 @@ after_sack_rexmit: } else { tp->t_timer[TCPT_PERSIST] = 0; tp->t_rxtshift = 0; + tp->t_rxtstart = 0; + tp->t_persist_stop = 0; } } @@ -565,32 +652,43 @@ after_sack_rexmit: * know that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) { - flags &= ~TH_SYN; + if (tp->t_state != TCPS_SYN_RECEIVED) + flags &= ~TH_SYN; off--, len++; if (len > 0 && tp->t_state == TCPS_SYN_SENT) { - while (!(tp->t_flags & TF_SENDINPROG) && - tp->t_pktlist_head != NULL) { + while (inp->inp_sndinprog_cnt == 0 && + tp->t_pktlist_head != NULL) { packetlist = tp->t_pktlist_head; packchain_listadd = tp->t_lastchain; packchain_sent++; TCP_PKTLIST_CLEAR(tp); - tp->t_flags |= TF_SENDINPROG; error = tcp_ip_output(so, tp, packetlist, packchain_listadd, tp_inp_options, - (so_options & SO_DONTROUTE), (sack_rxmit | (sack_bytes_rxmt != 0)), 0); + (so_options & SO_DONTROUTE), + (sack_rxmit | (sack_bytes_rxmt != 0)), 0, +#if INET6 + isipv6); +#else /* INET6 */ + 0); +#endif /* !INET6 */ + - tp->t_flags &= ~TF_SENDINPROG; } - /* tcp was closed while we were in ip; resume close */ - if ((tp->t_flags & - (TF_CLOSING|TF_SENDINPROG)) == TF_CLOSING) { + /* + * tcp was closed while we were in ip, + * resume close + */ + if (inp->inp_sndinprog_cnt == 0 && + (tp->t_flags & TF_CLOSING)) { tp->t_flags &= ~TF_CLOSING; (void) tcp_close(tp); + } else { + tcp_check_timer_state(tp); } KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0); - return 0; + return(0); } } @@ -604,7 +702,12 @@ after_sack_rexmit: flags &= ~TH_FIN; } - if (len < 0) { + /* The check here used to be (len < 0). Some times len is zero when + * the congestion window is closed and we need to check if persist timer + * has to be set in that case. But don't set persist until connection + * is established. + */ + if (len <= 0 && !(flags & TH_SYN)) { /* * If FIN has been sent but not acked, * but we haven't been called to retransmit, @@ -619,20 +722,62 @@ after_sack_rexmit: if (sendwin == 0) { tp->t_timer[TCPT_REXMT] = 0; tp->t_rxtshift = 0; + tp->t_rxtstart = 0; tp->snd_nxt = tp->snd_una; if (tp->t_timer[TCPT_PERSIST] == 0) tcp_setpersist(tp); } } + /* Automatic sizing of send socket buffer. Increase the send socket buffer + * size if all of the following criteria are met + * 1. the receiver has enough buffer space for this data + * 2. send buffer is filled to 7/8th with data (so we actually + * have data to make use of it); + * 3. our send window (slow start and congestion controlled) is + * larger than sent but unacknowledged data in send buffer. + */ + basertt = get_base_rtt(tp); + if (tcp_do_autosendbuf == 1 && + !INP_WAIT_FOR_IF_FEEDBACK(inp) && !IN_FASTRECOVERY(tp) && + (so->so_snd.sb_flags & (SB_AUTOSIZE | SB_TRIM)) == SB_AUTOSIZE && + tcp_cansbgrow(&so->so_snd)) { + if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && + so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) && + sendwin >= (so->so_snd.sb_cc - + (tp->snd_nxt - tp->snd_una))) { + /* Also increase the send buffer only if the + * round-trip time is not increasing because we do + * not want to contribute to latency by filling buffers. + * We also do not want to hold onto application's + * old data for too long. Interactive applications would + * rather discard old data. + */ + if (tp->t_rttcur <= + (basertt + 25)) { + if (sbreserve(&so->so_snd, + min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc, + tcp_autosndbuf_max)) == 1) { + so->so_snd.sb_idealsize = so->so_snd.sb_hiwat; + } + } else { + so->so_snd.sb_idealsize = + max(tcp_sendspace, so->so_snd.sb_hiwat - + (2 * tcp_autosndbuf_inc)); + so->so_snd.sb_flags |= SB_TRIM; + } + } + } + /* * Truncate to the maximum segment length or enable TCP Segmentation * Offloading (if supported by hardware) and ensure that FIN is removed * if the length no longer contains the last data byte. * - * TSO may only be used if we are in a pure bulk sending state. The - * presence of TCP-MD5, SACK retransmits, SACK advertizements, ipfw rules - * and IP options prevent using TSO. With TSO the TCP header is the same + * TSO may only be used if we are in a pure bulk sending state. + * The presence of TCP-MD5, SACK retransmits, SACK advertizements, + * ipfw rules and IP options, as well as disabling hardware checksum + * offload prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per generated * segment or packet. @@ -651,14 +796,11 @@ after_sack_rexmit: #endif if (len > tp->t_maxseg) { - if ((tp->t_flags & TF_TSO) && tcp_do_tso && -#if RANDOM_IP_ID - ip_use_randomid && -#endif /* RANDOM_IP_ID */ - kipf_count == 0 && dlil_filter_count == 0 && + if ((tp->t_flags & TF_TSO) && tcp_do_tso && hwcksum_tx && + ip_use_randomid && kipf_count == 0 && dlil_filter_disable_tso_count == 0 && tp->rcv_numsacks == 0 && sack_rxmit == 0 && sack_bytes_rxmt == 0 && - tp->t_inpcb->inp_options == NULL && - tp->t_inpcb->in6p_options == NULL + inp->inp_options == NULL && + inp->in6p_options == NULL #if IPSEC && ipsec_optlen == 0 #endif @@ -674,6 +816,34 @@ after_sack_rexmit: tso = 0; } } +#if MPTCP + if (so->so_flags & SOF_MP_SUBFLOW) { + int newlen = len; + if ((tp->t_mpflags & TMPF_SND_MPPRIO) || + (tp->t_mpflags & TMPF_SND_REM_ADDR) || + (tp->t_mpflags & TMPF_SND_MPFAIL)) { + if (len > 0) { + len = 0; + } + sendalot = 1; + mptcp_acknow = 1; + } else { + mptcp_acknow = 0; + } + /* + * The contiguous bytes in the subflow socket buffer can be + * discontiguous at the MPTCP level. Since only one DSS + * option can be sent in one packet, reduce length to match + * the contiguous MPTCP level. Set sendalot to send remainder. + */ + if (len > 0) + newlen = mptcp_adj_sendlen(so, off, len); + if (newlen < len) { + len = newlen; + sendalot = 1; + } + } +#endif /* MPTCP */ if (sack_rxmit) { if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc)) flags &= ~TH_FIN; @@ -684,51 +854,79 @@ after_sack_rexmit: recwin = tcp_sbspace(tp); + + /* + * If the socket is capable of doing unordered send, + * pull the amount of data that can be sent from the + * unordered priority queues to the serial queue in + * the socket buffer. If bytes are not yet available + * in the highest priority message, we may not be able + * to send any new data. + */ + if (so->so_flags & SOF_ENABLE_MSGS) { + if ((off + len) > + so->so_msg_state->msg_serial_bytes) { + sbpull_unordered_data(so, off, len); + + /* check if len needs to be modified */ + if ((off + len) > + so->so_msg_state->msg_serial_bytes) { + len = so->so_msg_state->msg_serial_bytes - off; + if (len <= 0) { + len = 0; + tcpstat.tcps_msg_sndwaithipri++; + } + } + } + } + /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * + * - we've timed out (e.g. persist timer) + * - we need to retransmit * - We have a full segment (or more with TSO) * - This is the last buffer in a write()/send() and we are * either idle or running NODELAY - * - we've timed out (e.g. persist timer) * - we have more then 1/2 the maximum send window's worth of * data (receiver may be limited the window size) - * - we need to retransmit */ if (len) { - if (len >= tp->t_maxseg) { - tp->t_flags |= TF_MAXSEGSNT; - goto send; - } - if (!(tp->t_flags & TF_MORETOCOME) && - (idle || tp->t_flags & TF_NODELAY || tp->t_flags & TF_MAXSEGSNT) && - (tp->t_flags & TF_NOPUSH) == 0 && - len + off >= so->so_snd.sb_cc) { - tp->t_flags &= ~TF_MAXSEGSNT; - goto send; - } - if (tp->t_force) { - tp->t_flags &= ~TF_MAXSEGSNT; - goto send; - } - if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { - tp->t_flags &= ~TF_MAXSEGSNT; + if (tp->t_force) goto send; - } - if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { - tp->t_flags &= ~TF_MAXSEGSNT; + if (SEQ_LT(tp->snd_nxt, tp->snd_max)) goto send; - } if (sack_rxmit) goto send; + + /* + * Send new data on the connection only if it is + * not flow controlled + */ + if (!INP_WAIT_FOR_IF_FEEDBACK(inp) || + tp->t_state != TCPS_ESTABLISHED) { + if (len >= tp->t_maxseg) + goto send; + if (!(tp->t_flags & TF_MORETOCOME) && + (idle || tp->t_flags & TF_NODELAY || + tp->t_flags & TF_MAXSEGSNT || + ALLOW_LIMITED_TRANSMIT(tp)) && + (tp->t_flags & TF_NOPUSH) == 0 && + len + off >= so->so_snd.sb_cc) + goto send; + if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) + goto send; + } else { + tcpstat.tcps_fcholdpacket++; + } } /* * Compare available window to amount of window * known to peer (as advertised window less * next expected input). If the difference is at least two - * max size segments, or at least 50% of the maximum possible + * max size segments, or at least 25% of the maximum possible * window, then want to send a window update to peer. * Skip this if the connection is in T/TCP half-open state. */ @@ -738,21 +936,41 @@ after_sack_rexmit: * taking into account that we are limited by * TCP_MAXWIN << tp->rcv_scale. */ - int32_t adv = imin(recwin, (int)TCP_MAXWIN << tp->rcv_scale) - + int32_t adv, oldwin = 0; + adv = imin(recwin, (int)TCP_MAXWIN << tp->rcv_scale) - (tp->rcv_adv - tp->rcv_nxt); + if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) + oldwin = tp->rcv_adv - tp->rcv_nxt; + if (adv >= (int32_t) (2 * tp->t_maxseg)) { - - /* - * Update only if the resulting scaled value of the window changed, or + /* Update only if the resulting scaled value of the window changed, or * if there is a change in the sequence since the last ack. * This avoids what appears as dupe ACKS (see rdar://5640997) + * + * If streaming is detected avoid sending too many window updates. + * We will depend on the delack timer to send a window update + * when needed. */ - - if ((tp->last_ack_sent != tp->rcv_nxt) || (((recwin + adv) >> tp->rcv_scale) > recwin)) + if (!(tp->t_flags & TF_STRETCHACK) && + (tp->last_ack_sent != tp->rcv_nxt || + ((oldwin + adv) >> tp->rcv_scale) > + (oldwin >> tp->rcv_scale))) { goto send; + } + + /* Make sure that the delayed ack timer is set if we + * delayed sending a window update because of streaming + * detection. + */ + if ((tp->t_flags & TF_STRETCHACK) && + !(tp->t_flags & TF_DELACK)) { + tp->t_flags |= TF_DELACK; + tp->t_timer[TCPT_DELACK] = + OFFSET_FROM_START(tp, tcp_delack); + } } - if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) + if (4 * adv >= (int32_t) so->so_rcv.sb_hiwat) goto send; } @@ -767,23 +985,28 @@ after_sack_rexmit: goto send; if (SEQ_GT(tp->snd_up, tp->snd_una)) goto send; +#if MPTCP + if (mptcp_acknow) + goto send; +#endif /* MPTCP */ /* * If our state indicates that FIN should be sent * and we have not yet done so, then we need to send. */ - if (flags & TH_FIN && - ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una)) + if ((flags & TH_FIN) && + (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una)) goto send; /* * In SACK, it is possible for tcp_output to fail to send a segment * after the retransmission timer has been turned off. Make sure * that the retransmission timer is set. */ - if (tp->sack_enable && (tp->t_state >= TCPS_ESTABLISHED) && SEQ_GT(tp->snd_max, tp->snd_una) && - tp->t_timer[TCPT_REXMT] == 0 && + if (SACK_ENABLED(tp) && (tp->t_state >= TCPS_ESTABLISHED) && + SEQ_GT(tp->snd_max, tp->snd_una) && + tp->t_timer[TCPT_REXMT] == 0 && tp->t_timer[TCPT_PERSIST] == 0) { - tp->t_timer[TCPT_REXMT] = tp->t_rxtcur; - goto just_return; + tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur); + goto just_return; } /* * TCP window updates are not reliable, rather a polling protocol @@ -810,6 +1033,7 @@ after_sack_rexmit: if (so->so_snd.sb_cc && tp->t_timer[TCPT_REXMT] == 0 && tp->t_timer[TCPT_PERSIST] == 0) { tp->t_rxtshift = 0; + tp->t_rxtstart = 0; tcp_setpersist(tp); } just_return: @@ -817,27 +1041,44 @@ just_return: * If there is no reason to send a segment, just return. * but if there is some packets left in the packet list, send them now. */ - while (!(tp->t_flags & TF_SENDINPROG) && tp->t_pktlist_head != NULL) { + while (inp->inp_sndinprog_cnt == 0 && + tp->t_pktlist_head != NULL) { packetlist = tp->t_pktlist_head; packchain_listadd = tp->t_lastchain; packchain_sent++; TCP_PKTLIST_CLEAR(tp); - tp->t_flags |= TF_SENDINPROG; error = tcp_ip_output(so, tp, packetlist, packchain_listadd, - tp_inp_options, (so_options & SO_DONTROUTE), (sack_rxmit | (sack_bytes_rxmt != 0)), recwin); - - tp->t_flags &= ~TF_SENDINPROG; + tp_inp_options, (so_options & SO_DONTROUTE), + (sack_rxmit | (sack_bytes_rxmt != 0)), recwin, +#if INET6 + isipv6); +#else /* INET6 */ + 0); +#endif /* !INET6 */ } /* tcp was closed while we were in ip; resume close */ - if ((tp->t_flags & (TF_CLOSING|TF_SENDINPROG)) == TF_CLOSING) { + if (inp->inp_sndinprog_cnt == 0 && + (tp->t_flags & TF_CLOSING)) { tp->t_flags &= ~TF_CLOSING; (void) tcp_close(tp); + } else { + tcp_check_timer_state(tp); } KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0); return (0); send: + /* + * Set TF_MAXSEGSNT flag if the segment size is greater than + * the max segment size. + */ + if (len > 0) { + if (len >= tp->t_maxseg) + tp->t_flags |= TF_MAXSEGSNT; + else + tp->t_flags &= ~TF_MAXSEGSNT; + } /* * Before ESTABLISHED, force sending of initial options * unless TCP set not to do any options. @@ -852,7 +1093,7 @@ send: hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); else #endif - hdrlen = sizeof (struct tcpiphdr); + hdrlen = sizeof (struct tcpiphdr); if (flags & TH_SYN) { tp->snd_nxt = tp->iss; if ((tp->t_flags & TF_NOOPT) == 0) { @@ -867,15 +1108,20 @@ send: if ((tp->t_flags & TF_REQ_SCALE) && ((flags & TH_ACK) == 0 || (tp->t_flags & TF_RCVD_SCALE))) { - *((u_int32_t *)(opt + optlen)) = htonl( + *((u_int32_t *)(void *)(opt + optlen)) = htonl( TCPOPT_NOP << 24 | TCPOPT_WINDOW << 16 | TCPOLEN_WINDOW << 8 | tp->request_r_scale); optlen += 4; } +#if MPTCP + if (mptcp_enable) { + optlen = mptcp_setup_syn_opts(so, flags, opt, + optlen); + } +#endif /* MPTCP */ } - } /* @@ -948,7 +1194,7 @@ send: * the CWR flag on data packets. Pure acks don't have this set. */ if ((tp->ecn_flags & TE_SENDCWR) != 0 && len != 0 && - !SEQ_LT(tp->snd_nxt, tp->snd_max)) { + !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) { flags |= TH_CWR; tp->ecn_flags &= ~TE_SENDCWR; } @@ -969,7 +1215,7 @@ send: (flags & TH_RST) == 0 && ((flags & TH_ACK) == 0 || (tp->t_flags & TF_RCVD_TSTMP))) { - u_int32_t *lp = (u_int32_t *)(opt + optlen); + u_int32_t *lp = (u_int32_t *)(void *)(opt + optlen); /* Form timestamp option as shown in appendix A of RFC 1323. */ *lp++ = htonl(TCPOPT_TSTAMP_HDR); @@ -978,7 +1224,11 @@ send: optlen += TCPOLEN_TSTAMP_APPA; } - if (tp->sack_enable && ((tp->t_flags & TF_NOOPT) == 0)) { + /* Note the timestamp for receive buffer autosizing */ + if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) + tp->rfbuf_ts = tcp_now; + + if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) { /* * Tack on the SACK permitted option *last*. * And do padding of options after tacking this on. @@ -1002,7 +1252,27 @@ send: *bp++ = TCPOLEN_SACK_PERMITTED; optlen += TCPOLEN_SACK_PERMITTED; } + } +#if MPTCP + if (so->so_flags & SOF_MP_SUBFLOW) { + /* + * Its important to piggyback acks with data as ack only packets + * may get lost and data packets that don't send Data ACKs + * still advance the subflow level ACK and therefore make it + * hard for the remote end to recover in low cwnd situations. + */ + if (len != 0) + tp->t_mpflags |= (TMPF_SEND_DSN | + TMPF_MPTCP_ACKNOW); + else + tp->t_mpflags |= TMPF_MPTCP_ACKNOW; + optlen = mptcp_setup_opts(tp, off, &opt[0], optlen, flags, + len, &dlenp, &finp, &dss_val, &sseqp); + tp->t_mpflags &= ~TMPF_SEND_DSN; + } +#endif /* MPTCP */ + if (SACK_ENABLED(tp) && ((tp->t_flags & TF_NOOPT) == 0)) { /* * Send SACKs if necessary. This should be the last * option processed. Only as many SACKs are sent as @@ -1019,7 +1289,7 @@ send: if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0 && MAX_TCPOPTLEN - optlen - 2 >= TCPOLEN_SACK) { - int nsack, sackoptlen, padlen; + int nsack, padlen; u_char *bp = (u_char *)opt + optlen; u_int32_t *lp; @@ -1040,7 +1310,7 @@ send: tcpstat.tcps_sack_send_blocks++; *bp++ = TCPOPT_SACK; *bp++ = sackoptlen; - lp = (u_int32_t *)bp; + lp = (u_int32_t *)(void *)bp; for (i = 0; i < nsack; i++) { struct sackblk sack = tp->sackblks[i]; *lp++ = htonl(sack.start); @@ -1066,7 +1336,7 @@ send: #if INET6 if (isipv6) - ipoptlen = ip6_optlen(tp->t_inpcb); + ipoptlen = ip6_optlen(inp); else #endif { @@ -1114,17 +1384,45 @@ send: sendalot = 1; } } +#if MPTCP + /* Adjust the length in the DSS option, if it is lesser than len */ + if (dlenp) { + /* + * To test this path without SACK, artificially + * decrement len with something like + * if (len > 10) + len -= 10; + */ + if (ntohs(*dlenp) > len) { + *dlenp = htons(len); + /* Unset the FIN flag, if len was adjusted */ + if (finp) { + *finp &= ~MDSS_F; + } + sendalot = 1; + } + } +#endif /* MPTCP */ -/*#ifdef DIAGNOSTIC*/ -#if INET6 if (max_linkhdr + hdrlen > MCLBYTES) panic("tcphdr too big"); -#else - if (max_linkhdr + hdrlen > MHLEN) - panic("tcphdr too big"); -#endif -/*#endif*/ + /* Check if there is enough data in the send socket + * buffer to start measuring bw + */ + if ((tp->t_flagsext & TF_MEASURESNDBW) != 0 && + (tp->t_bwmeas != NULL) && + (tp->t_flagsext & TF_BWMEAS_INPROGRESS) == 0 && + (so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)) >= + tp->t_bwmeas->bw_minsize) { + tp->t_bwmeas->bw_size = min((so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)), + tp->t_bwmeas->bw_maxsize); + tp->t_flagsext |= TF_BWMEAS_INPROGRESS; + tp->t_bwmeas->bw_start = tp->snd_max; + tp->t_bwmeas->bw_ts = tcp_now; + } + + VERIFY(inp->inp_flowhash != 0); /* * Grab a header mbuf, attaching a copy of data to * be transmitted, and initialize the header from @@ -1136,22 +1434,28 @@ send: else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { tcpstat.tcps_sndrexmitpack++; tcpstat.tcps_sndrexmitbyte += len; + if (nstat_collect) { + nstat_route_tx(inp->inp_route.ro_rt, 1, + len, NSTAT_TX_FLAG_RETRANSMIT); + INP_ADD_STAT(inp, cell, wifi, txpackets, 1); + INP_ADD_STAT(inp, cell, wifi, txbytes, len); + tp->t_stat.txretransmitbytes += len; + } } else { tcpstat.tcps_sndpack++; tcpstat.tcps_sndbyte += len; + + if (nstat_collect) { + INP_ADD_STAT(inp, cell, wifi, txpackets, 1); + INP_ADD_STAT(inp, cell, wifi, txbytes, len); + } } -#ifdef notyet - if ((m = m_copypack(so->so_snd.sb_mb, off, - (int)len, max_linkhdr + hdrlen)) == 0) { - error = ENOBUFS; - goto out; +#if MPTCP + if (tp->t_mpflags & TMPF_MPTCP_TRUE) { + tcpstat.tcps_mp_sndpacks++; + tcpstat.tcps_mp_sndbytes += len; } - /* - * m_copypack left space for our hdr; use it. - */ - m->m_len += hdrlen; - m->m_data -= hdrlen; -#else +#endif /* MPTCP */ /* * try to use the new interface that allocates all * the necessary mbuf hdrs under 1 mbuf lock and @@ -1165,11 +1469,13 @@ send: * data area (no cluster attached) */ m = NULL; -#if INET6 + + /* minimum length we are going to allocate */ + allocated_len = MHLEN; if (MHLEN < hdrlen + max_linkhdr) { - MGETHDR(m, M_DONTWAIT, MT_HEADER); /* MAC-OK */ + MGETHDR(m, M_DONTWAIT, MT_HEADER); if (m == NULL) { - error = ENOBUFS; + error = ENOBUFS; goto out; } MCLGET(m, M_DONTWAIT); @@ -1180,13 +1486,14 @@ send: } m->m_data += max_linkhdr; m->m_len = hdrlen; + allocated_len = MCLBYTES; } -#endif - if (len <= MHLEN - hdrlen - max_linkhdr) { + if (len <= allocated_len - hdrlen - max_linkhdr) { if (m == NULL) { - MGETHDR(m, M_DONTWAIT, MT_HEADER); /* MAC-OK */ + VERIFY(allocated_len <= MHLEN); + MGETHDR(m, M_DONTWAIT, MT_HEADER); if (m == NULL) { - error = ENOBUFS; + error = ENOBUFS; goto out; } m->m_data += max_linkhdr; @@ -1202,40 +1509,71 @@ send: mtod(m, caddr_t) + hdrlen); m->m_len += len; } else { - if (m != NULL) { - m->m_next = m_copy(so->so_snd.sb_mb, off, (int) len); - if (m->m_next == 0) { - (void) m_free(m); + uint32_t copymode; + /* + * Retain packet header metadata at the socket + * buffer if this is is an MPTCP subflow, + * otherwise move it. + */ + copymode = M_COPYM_MOVE_HDR; +#if MPTCP + if ((tp->t_mpflags & TMPF_MPTCP_TRUE) || + (tp->t_mpflags & TMPF_TCP_FALLBACK)) { + copymode = M_COPYM_NOOP_HDR; + } +#endif /* MPTCP */ + if (m != NULL) { + m->m_next = m_copym_mode(so->so_snd.sb_mb, off, + (int) len, M_DONTWAIT, copymode); + if (m->m_next == NULL) { + (void) m_free(m); error = ENOBUFS; goto out; } } else { - /* - * determine whether the mbuf pointer and offset passed back by the 'last' call - * to m_copym_with_hdrs are still valid... if the head of the socket chain has - * changed (due to an incoming ACK for instance), or the offset into the chain we - * just computed is different from the one last returned by m_copym_with_hdrs (perhaps - * we're re-transmitting a packet sent earlier), than we can't pass the mbuf pointer and - * offset into it as valid hints for m_copym_with_hdrs to use (if valid, these hints allow - * m_copym_with_hdrs to avoid rescanning from the beginning of the socket buffer mbuf list. - * setting the mbuf pointer to NULL is sufficient to disable the hint mechanism. + /* + * determine whether the mbuf pointer and + * offset passed back by the 'last' call to + * m_copym_with_hdrs are still valid... if the + * head of the socket chain has changed (due + * to an incoming ACK for instance), or the + * offset into the chain we just computed is + * different from the one last returned by + * m_copym_with_hdrs (perhaps we're re- + * transmitting a packet sent earlier), then + * we can't pass the mbuf pointer and offset + * into it as valid hints for m_copym_with_hdrs + * to use (if valid, these hints allow + * m_copym_with_hdrs to avoid rescanning from + * the beginning of the socket buffer mbuf list. + * + * Setting the mbuf pointer to NULL is + * sufficient to disable the hint mechanism. */ - if (m_head != so->so_snd.sb_mb || sack_rxmit || last_off != off) - m_last = NULL; + if (m_head != so->so_snd.sb_mb || sack_rxmit || + last_off != off) + m_lastm = NULL; last_off = off + len; m_head = so->so_snd.sb_mb; - - /* makes sure we still have data left to be sent at this point */ + + /* + * make sure we still have data left + * to be sent at this point + */ if (m_head == NULL) { error = 0; /* should we return an error? */ goto out; } - + /* - * m_copym_with_hdrs will always return the last mbuf pointer and the offset into it that - * it acted on to fullfill the current request, whether a valid 'hint' was passed in or not + * m_copym_with_hdrs will always return the + * last mbuf pointer and the offset into it that + * it acted on to fullfill the current request, + * whether a valid 'hint' was passed in or not. */ - if ((m = m_copym_with_hdrs(so->so_snd.sb_mb, off, len, M_DONTWAIT, &m_last, &m_off)) == NULL) { + if ((m = m_copym_with_hdrs(so->so_snd.sb_mb, + off, len, M_DONTWAIT, &m_lastm, &m_off, + copymode)) == NULL) { error = ENOBUFS; goto out; } @@ -1243,7 +1581,6 @@ send: m->m_len = hdrlen; } } -#endif /* * If we're sending everything we've got, set PUSH. * (This will keep happy those implementations which only @@ -1267,34 +1604,45 @@ send: error = ENOBUFS; goto out; } -#if INET6 - if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && - MHLEN >= hdrlen) { - MH_ALIGN(m, hdrlen); - } else -#endif + if (MHLEN < (hdrlen + max_linkhdr)) { + MCLGET(m, M_DONTWAIT); + if ((m->m_flags & M_EXT) == 0) { + m_freem(m); + error = ENOBUFS; + goto out; + } + } m->m_data += max_linkhdr; m->m_len = hdrlen; } m->m_pkthdr.rcvif = 0; +#if MPTCP + /* Before opt is copied to the mbuf, set the csum field */ + mptcp_output_csum(tp, m, len, hdrlen, dss_val, sseqp); +#endif /* MPTCP */ #if CONFIG_MACF_NET - mac_mbuf_label_associate_inpcb(tp->t_inpcb, m); + mac_mbuf_label_associate_inpcb(inp, m); #endif #if INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); - th = (struct tcphdr *)(ip6 + 1); + th = (struct tcphdr *)(void *)(ip6 + 1); tcp_fillheaders(tp, ip6, th); + if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len && + !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) { + ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); + } + svc_flags |= PKT_SCF_IPV6; } else #endif /* INET6 */ { ip = mtod(m, struct ip *); ipov = (struct ipovly *)ip; - th = (struct tcphdr *)(ip + 1); + th = (struct tcphdr *)(void *)(ip + 1); /* this picks up the pseudo header (w/o the length) */ tcp_fillheaders(tp, ip, th); if ((tp->ecn_flags & TE_SENDIPECT) != 0 && len && - !SEQ_LT(tp->snd_nxt, tp->snd_max)) { + !SEQ_LT(tp->snd_nxt, tp->snd_max) && !sack_rxmit) { ip->ip_tos = IPTOS_ECN_ECT0; } } @@ -1304,7 +1652,7 @@ send: * window for use in delaying messages about window sizes. * If resending a FIN, be sure not to use a new sequence number. */ - if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && + if (flags & TH_FIN && (tp->t_flags & TF_SENTFIN) && tp->snd_nxt == tp->snd_max) tp->snd_nxt--; /* @@ -1349,13 +1697,27 @@ send: if (tp->t_flags & TF_SLOWLINK && slowlink_wsize > 0) { if (recwin > (int32_t)slowlink_wsize) recwin = slowlink_wsize; - th->th_win = htons((u_short) (recwin>>tp->rcv_scale)); } - else { - if (recwin > (int32_t)(TCP_MAXWIN << tp->rcv_scale)) - recwin = (int32_t)(TCP_MAXWIN << tp->rcv_scale); - th->th_win = htons((u_short) (recwin>>tp->rcv_scale)); + +#if TRAFFIC_MGT + if (tcp_recv_bg == 1 || IS_TCP_RECV_BG(so)) { + if (tcp_recv_throttle(tp)) { + uint32_t min_iaj_win = + tcp_min_iaj_win * tp->t_maxseg; + if (tp->iaj_rwintop == 0 || + SEQ_LT(tp->iaj_rwintop, tp->rcv_adv)) + tp->iaj_rwintop = tp->rcv_adv; + if (SEQ_LT(tp->iaj_rwintop, + tp->rcv_nxt + min_iaj_win)) + tp->iaj_rwintop = tp->rcv_nxt + min_iaj_win; + recwin = min(tp->iaj_rwintop - tp->rcv_nxt, recwin); + } } +#endif /* TRAFFIC_MGT */ + + if (recwin > (int32_t)(TCP_MAXWIN << tp->rcv_scale)) + recwin = (int32_t)(TCP_MAXWIN << tp->rcv_scale); + th->th_win = htons((u_short) (recwin>>tp->rcv_scale)); /* * Adjust the RXWIN0SENT flag - indicate that we have advertised @@ -1365,14 +1727,14 @@ send: * to read more data then can be buffered prior to transmitting on * the connection. */ - if (recwin == 0) + if (th->th_win == 0) tp->t_flags |= TF_RXWIN0SENT; else tp->t_flags &= ~TF_RXWIN0SENT; if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); th->th_flags |= TH_URG; - } else + } else { /* * If no urgent pointer to send, then we pull * the urgent pointer to the left edge of the send window @@ -1380,6 +1742,7 @@ send: * number wraparound. */ tp->snd_up = tp->snd_una; /* drag it along */ + } /* * Put TCP length in extended header, and then @@ -1387,13 +1750,17 @@ send: */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #if INET6 - if (isipv6) + if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ - th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr), - sizeof(struct tcphdr) + optlen + len); + m->m_pkthdr.csum_flags = CSUM_TCPIPV6; + m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); + if (len + optlen) + th->th_sum = in_addword(th->th_sum, + htons((u_short)(optlen + len))); + } else #endif /* INET6 */ { @@ -1407,20 +1774,19 @@ send: /* * Enable TSO and specify the size of the segments. * The TCP pseudo header checksum is always provided. - * XXX: Fixme: This is currently not the case for IPv6. */ if (tso) { #if INET6 - if (isipv6) - m->m_pkthdr.csum_flags = CSUM_TSO_IPV6; + if (isipv6) + m->m_pkthdr.csum_flags |= CSUM_TSO_IPV6; else #endif /* INET6 */ - m->m_pkthdr.csum_flags = CSUM_TSO_IPV4; + m->m_pkthdr.csum_flags |= CSUM_TSO_IPV4; m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen; - } - else + } else { m->m_pkthdr.tso_segsz = 0; + } /* * In transmit state, time the transmission and arrange for @@ -1435,7 +1801,8 @@ send: if (flags & (TH_SYN|TH_FIN)) { if (flags & TH_SYN) tp->snd_nxt++; - if (flags & TH_FIN) { + if ((flags & TH_FIN) && + !(tp->t_flags & TF_SENTFIN)) { tp->snd_nxt++; tp->t_flags |= TF_SENTFIN; } @@ -1450,7 +1817,7 @@ send: * not currently timing anything. */ if (tp->t_rtttime == 0) { - tp->t_rtttime = 1; + tp->t_rtttime = tcp_now; tp->t_rtseq = startseq; tcpstat.tcps_segstimed++; } @@ -1467,12 +1834,15 @@ send: timer: if (tp->t_timer[TCPT_REXMT] == 0 && ((sack_rxmit && tp->snd_nxt != tp->snd_max) || - tp->snd_nxt != tp->snd_una)) { + tp->snd_nxt != tp->snd_una || + (flags & TH_FIN))) { if (tp->t_timer[TCPT_PERSIST]) { tp->t_timer[TCPT_PERSIST] = 0; tp->t_rxtshift = 0; + tp->t_rxtstart = 0; + tp->t_persist_stop = 0; } - tp->t_timer[TCPT_REXMT] = tp->t_rxtcur; + tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur); } } else { /* @@ -1482,7 +1852,8 @@ timer: int xlen = len; if (flags & TH_SYN) ++xlen; - if (flags & TH_FIN) { + if ((flags & TH_FIN) && + !(tp->t_flags & TF_SENTFIN)) { ++xlen; tp->t_flags |= TF_SENTFIN; } @@ -1504,11 +1875,11 @@ timer: * to handle ttl and tos; we could keep them in * the template, but need a way to checksum without them. */ +#if INET6 /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ -#if INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the @@ -1516,61 +1887,26 @@ timer: * Also, desired default hop limit might be changed via * Neighbor Discovery. */ - ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, - tp->t_inpcb->in6p_route.ro_rt ? - tp->t_inpcb->in6p_route.ro_rt->rt_ifp - : NULL); + ip6->ip6_hlim = in6_selecthlim(inp, inp->in6p_route.ro_rt ? + inp->in6p_route.ro_rt->rt_ifp : NULL); /* TODO: IPv6 IP6TOS_ECT bit on */ -#if IPSEC - if (ipsec_bypass == 0 && ipsec_setsocket(m, so) != 0) { - m_freem(m); - error = ENOBUFS; - goto out; - } -#endif /*IPSEC*/ - m->m_pkthdr.socket_id = socket_id; -#if PKT_PRIORITY - if (soisbackground(so)) - m_prio_background(m); -#endif /* PKT_PRIORITY */ - error = ip6_output(m, - inp6_pktopts, - &tp->t_inpcb->in6p_route, - (so_options & SO_DONTROUTE), NULL, NULL, 0); + KERNEL_DEBUG(DBG_LAYER_BEG, + ((inp->inp_fport << 16) | inp->inp_lport), + (((inp->in6p_laddr.s6_addr16[0] & 0xffff) << 16) | + (inp->in6p_faddr.s6_addr16[0] & 0xffff)), + sendalot,0,0); } else #endif /* INET6 */ - { - ip->ip_len = m->m_pkthdr.len; -#if INET6 - if (isipv6) - ip->ip_ttl = in6_selecthlim(tp->t_inpcb, - tp->t_inpcb->in6p_route.ro_rt ? - tp->t_inpcb->in6p_route.ro_rt->rt_ifp - : NULL); - else -#endif /* INET6 */ - ip->ip_ttl = tp->t_inpcb->inp_ip_ttl; /* XXX */ - ip->ip_tos |= (tp->t_inpcb->inp_ip_tos & ~IPTOS_ECN_MASK); /* XXX */ - - -#if INET6 - if (isipv6) { - KERNEL_DEBUG(DBG_LAYER_BEG, - ((tp->t_inpcb->inp_fport << 16) | tp->t_inpcb->inp_lport), - (((tp->t_inpcb->in6p_laddr.s6_addr16[0] & 0xffff) << 16) | - (tp->t_inpcb->in6p_faddr.s6_addr16[0] & 0xffff)), - 0,0,0); - } - else -#endif { - KERNEL_DEBUG(DBG_LAYER_BEG, - ((tp->t_inpcb->inp_fport << 16) | tp->t_inpcb->inp_lport), - (((tp->t_inpcb->inp_laddr.s_addr & 0xffff) << 16) | - (tp->t_inpcb->inp_faddr.s_addr & 0xffff)), - 0,0,0); - } + ip->ip_len = m->m_pkthdr.len; + ip->ip_ttl = inp->inp_ip_ttl; /* XXX */ + ip->ip_tos |= (inp->inp_ip_tos & ~IPTOS_ECN_MASK);/* XXX */ + KERNEL_DEBUG(DBG_LAYER_BEG, + ((inp->inp_fport << 16) | inp->inp_lport), + (((inp->inp_laddr.s_addr & 0xffff) << 16) | + (inp->inp_faddr.s_addr & 0xffff)), 0,0,0); + } /* * See if we should do MTU discovery. @@ -1581,9 +1917,11 @@ timer: * 4) the MTU is not locked (if it is, then discovery has been * disabled for that route) */ - - if (path_mtu_discovery && (tp->t_flags & TF_PMTUD)) - ip->ip_off |= IP_DF; +#if INET6 + if (!isipv6) +#endif /* INET6 */ + if (path_mtu_discovery && (tp->t_flags & TF_PMTUD)) + ip->ip_off |= IP_DF; #if IPSEC if (ipsec_bypass == 0) @@ -1594,10 +1932,57 @@ timer: * The socket is kept locked while sending out packets in ip_output, even if packet chaining is not active. */ lost = 0; - m->m_pkthdr.socket_id = socket_id; + + /* + * Embed the flow hash in pkt hdr and mark the packet as + * capable of flow controlling + */ + m->m_pkthdr.pkt_flowsrc = FLOWSRC_INPCB; + m->m_pkthdr.pkt_flowid = inp->inp_flowhash; + m->m_pkthdr.pkt_flags |= PKTF_FLOW_ID | PKTF_FLOW_LOCALSRC; +#if MPTCP + /* Disable flow advisory when using MPTCP. */ + if (!(tp->t_mpflags & TMPF_MPTCP_TRUE)) +#endif /* MPTCP */ + m->m_pkthdr.pkt_flags |= PKTF_FLOW_ADV; + m->m_pkthdr.pkt_proto = IPPROTO_TCP; + m->m_nextpkt = NULL; + + if (inp->inp_last_outifp != NULL && + !(inp->inp_last_outifp->if_flags & IFF_LOOPBACK)) { + /* Hint to prioritize this packet if + * 1. if the packet has no data + * 2. the interface supports transmit-start model and did + * not disable ACK prioritization. + * 3. Only ACK flag is set. + * 4. there is no outstanding data on this connection. + */ + if (tcp_prioritize_acks != 0 && len == 0 && + (inp->inp_last_outifp->if_eflags & + (IFEF_TXSTART | IFEF_NOACKPRI)) == IFEF_TXSTART && + th->th_flags == TH_ACK && tp->snd_una == tp->snd_max && + tp->t_timer[TCPT_REXMT] == 0) { + svc_flags |= PKT_SCF_TCP_ACK; + } + set_packet_service_class(m, so, MBUF_SC_UNSPEC, svc_flags); + } + tp->t_pktlist_sentlen += len; tp->t_lastchain++; + +#if INET6 + if (isipv6) { + DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp, + struct ip6 *, ip6, struct tcpcb *, tp, struct tcphdr *, + th); + } else +#endif /* INET6 */ + { + DTRACE_TCP5(send, struct mbuf *, m, struct inpcb *, inp, + struct ip *, ip, struct tcpcb *, tp, struct tcphdr *, th); + } + if (tp->t_pktlist_head != NULL) { tp->t_pktlist_tail->m_nextpkt = m; tp->t_pktlist_tail = m; @@ -1606,25 +1991,59 @@ timer: tp->t_pktlist_head = tp->t_pktlist_tail = m; } + if ((lro_ackmore) && (!sackoptlen) && (!tp->t_timer[TCPT_PERSIST]) && + ((th->th_flags & TH_ACK) == TH_ACK) && (!len) && + (tp->t_state == TCPS_ESTABLISHED)) { + /* For a pure ACK, see if you need to send more of them */ + mnext = tcp_send_lroacks(tp, m, th); + if (mnext) { + tp->t_pktlist_tail->m_nextpkt = mnext; + if (mnext->m_nextpkt == NULL) { + tp->t_pktlist_tail = mnext; + tp->t_lastchain++; + } else { + struct mbuf *tail, *next; + next = mnext->m_nextpkt; + tail = next->m_nextpkt; + while (tail) { + next = tail; + tail = tail->m_nextpkt; + tp->t_lastchain++; + } + tp->t_pktlist_tail = next; + } + } + } + if (sendalot == 0 || (tp->t_state != TCPS_ESTABLISHED) || (tp->snd_cwnd <= (tp->snd_wnd / 8)) || (tp->t_flags & (TH_PUSH | TF_ACKNOW)) || tp->t_force != 0 || tp->t_lastchain >= tcp_packet_chaining) { error = 0; - while (!(tp->t_flags & TF_SENDINPROG) && - tp->t_pktlist_head != NULL) { + + /* + * Reset the stack memory of offset as the socket + * may get unlocked + */ + m_lastm = NULL; + while (inp->inp_sndinprog_cnt == 0 && + tp->t_pktlist_head != NULL) { packetlist = tp->t_pktlist_head; packchain_listadd = tp->t_lastchain; packchain_sent++; lost = tp->t_pktlist_sentlen; TCP_PKTLIST_CLEAR(tp); - tp->t_flags |= TF_SENDINPROG; error = tcp_ip_output(so, tp, packetlist, packchain_listadd, tp_inp_options, - (so_options & SO_DONTROUTE), (sack_rxmit | (sack_bytes_rxmt != 0)), recwin); + (so_options & SO_DONTROUTE), + (sack_rxmit | (sack_bytes_rxmt != 0)), recwin, +#if INET6 + isipv6); +#else /* INET6 */ + 0); +#endif /* !INET6 */ - tp->t_flags &= ~TF_SENDINPROG; if (error) { /* * Take into account the rest of unsent @@ -1639,20 +2058,19 @@ timer: } } /* tcp was closed while we were in ip; resume close */ - if ((tp->t_flags & (TF_CLOSING|TF_SENDINPROG)) == TF_CLOSING) { + if (inp->inp_sndinprog_cnt == 0 && + (tp->t_flags & TF_CLOSING)) { tp->t_flags &= ~TF_CLOSING; (void) tcp_close(tp); return (0); } - } - else { + } else { error = 0; packchain_looped++; tcpstat.tcps_sndtotal++; goto again; } - } if (error) { /* * Assume that the packets were lost, so back out the @@ -1668,10 +2086,21 @@ timer: */ if ((flags & TH_SYN) == 0) { if (sack_rxmit) { - p->rxmit -= lost; + if (SEQ_GT((p->rxmit - lost), + tp->snd_una)) { + p->rxmit -= lost; + } else { + lost = p->rxmit - tp->snd_una; + p->rxmit = tp->snd_una; + } tp->sackhint.sack_bytes_rexmit -= lost; - } else - tp->snd_nxt -= lost; + } else { + if (SEQ_GT((tp->snd_nxt - lost), + tp->snd_una)) + tp->snd_nxt -= lost; + else + tp->snd_nxt = tp->snd_una; + } } } out: @@ -1681,13 +2110,19 @@ out: if (error == ENOBUFS) { if (!tp->t_timer[TCPT_REXMT] && - !tp->t_timer[TCPT_PERSIST]) - tp->t_timer[TCPT_REXMT] = tp->t_rxtcur; + !tp->t_timer[TCPT_PERSIST]) + tp->t_timer[TCPT_REXMT] = + OFFSET_FROM_START(tp, tp->t_rxtcur); tp->snd_cwnd = tp->t_maxseg; tp->t_bytes_acked = 0; + tcp_check_timer_state(tp); KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0); + + DTRACE_TCP5(cc, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, struct tcphdr *, NULL, + int32_t, TCP_CC_OUTPUT_ERROR); return (0); } if (error == EMSGSIZE) { @@ -1706,70 +2141,102 @@ out: if (tso) tp->t_flags &= ~TF_TSO; - tcp_mtudisc(tp->t_inpcb, 0); + tcp_mtudisc(inp, 0); + tcp_check_timer_state(tp); + KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0); return 0; } - if ((error == EHOSTUNREACH || error == ENETDOWN) - && TCPS_HAVERCVDSYN(tp->t_state)) { + /* + * Unless this is due to interface restriction policy, + * treat EHOSTUNREACH/ENETDOWN as a soft error. + */ + if ((error == EHOSTUNREACH || error == ENETDOWN) && + TCPS_HAVERCVDSYN(tp->t_state) && + !((inp->inp_flags & INP_NO_IFT_CELLULAR) && + inp->inp_last_outifp != NULL && + IFNET_IS_CELLULAR(inp->inp_last_outifp))) { tp->t_softerror = error; - KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0); + tcp_check_timer_state(tp); + KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, + 0, 0, 0, 0, 0); return (0); } + tcp_check_timer_state(tp); KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END, 0,0,0,0,0); return (error); } tcpstat.tcps_sndtotal++; - /* - * Data sent (as far as we can tell). - * If this advertises a larger window than any other segment, - * then remember the size of the advertised window. - * Make sure ACK/DELACK conditions are cleared before - * we unlock the socket. - * NOTE: for now, this is done in tcp_ip_output for IPv4 - */ -#if INET6 - if (isipv6) { - if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) - tp->rcv_adv = tp->rcv_nxt + recwin; - tp->last_ack_sent = tp->rcv_nxt; - tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); - } -#endif - KERNEL_DEBUG(DBG_FNC_TCP_OUTPUT | DBG_FUNC_END,0,0,0,0,0); - if (sendalot && (!tcp_do_newreno || --maxburst)) + if (sendalot) goto again; + + tcp_check_timer_state(tp); return (0); } static int tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt, - int cnt, struct mbuf *opt, int flags, int sack_in_progress, int recwin) + int cnt, struct mbuf *opt, int flags, int sack_in_progress, int recwin, + boolean_t isipv6) { int error = 0; boolean_t chain; boolean_t unlocked = FALSE; + boolean_t ifdenied = FALSE; struct inpcb *inp = tp->t_inpcb; - struct ip_out_args ipoa; + struct ip_out_args ipoa = + { IFSCOPE_NONE, { 0 }, IPOAF_SELECT_SRCIF|IPOAF_BOUND_SRCADDR, 0 }; struct route ro; -#if PKT_PRIORITY - boolean_t bg = FALSE; -#endif /* PKT_PRIORITY */ + struct ifnet *outif = NULL; +#if INET6 + struct ip6_out_args ip6oa = + { IFSCOPE_NONE, { 0 }, IP6OAF_SELECT_SRCIF|IP6OAF_BOUND_SRCADDR, 0 }; + struct route_in6 ro6; + struct flowadv *adv = + (isipv6 ? &ip6oa.ip6oa_flowadv : &ipoa.ipoa_flowadv); +#else /* INET6 */ + struct flowadv *adv = &ipoa.ipoa_flowadv; +#endif /* !INET6 */ /* If socket was bound to an ifindex, tell ip_output about it */ - ipoa.ipoa_ifscope = (inp->inp_flags & INP_BOUND_IF) ? - inp->inp_boundif : IFSCOPE_NONE; - flags |= IP_OUTARGS; + if (inp->inp_flags & INP_BOUND_IF) { +#if INET6 + if (isipv6) { + ip6oa.ip6oa_boundif = inp->inp_boundifp->if_index; + ip6oa.ip6oa_flags |= IP6OAF_BOUND_IF; + } else +#endif /* INET6 */ + { + ipoa.ipoa_boundif = inp->inp_boundifp->if_index; + ipoa.ipoa_flags |= IPOAF_BOUND_IF; + } + } - /* Copy the cached route and take an extra reference */ - inp_route_copyout(inp, &ro); + if (inp->inp_flags & INP_NO_IFT_CELLULAR) { +#if INET6 + if (isipv6) + ip6oa.ip6oa_flags |= IP6OAF_NO_CELLULAR; + else +#endif /* INET6 */ + ipoa.ipoa_flags |= IPOAF_NO_CELLULAR; + } +#if INET6 + if (isipv6) + flags |= IPV6_OUTARGS; + else +#endif /* INET6 */ + flags |= IP_OUTARGS; -#if PKT_PRIORITY - bg = soisbackground(so); -#endif /* PKT_PRIORITY */ + /* Copy the cached route and take an extra reference */ +#if INET6 + if (isipv6) + in6p_route_copyout(inp, &ro6); + else +#endif /* INET6 */ + inp_route_copyout(inp, &ro); /* * Data sent (as far as we can tell). @@ -1782,14 +2249,25 @@ tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt, tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); + tp->t_timer[TCPT_DELACK] = 0; + tp->t_unacksegs = 0; + + /* Increment the count of outstanding send operations */ + inp->inp_sndinprog_cnt++; /* * If allowed, unlock TCP socket while in IP * but only if the connection is established and - * if we're not sending from an upcall. + * in a normal mode where reentrancy on the tcpcb won't be + * an issue: + * - there is no SACK episode + * - we're not in Fast Recovery mode + * - if we're not sending from an upcall. */ - if (tcp_output_unlocked && ((so->so_flags & SOF_UPCALLINUSE) == 0) && - (tp->t_state == TCPS_ESTABLISHED) && (sack_in_progress == 0)) { + if (tcp_output_unlocked && !so->so_upcallusecount && + (tp->t_state == TCPS_ESTABLISHED) && (sack_in_progress == 0) && + ((tp->t_flags & TF_FASTRECOVERY) == 0)) { + unlocked = TRUE; socket_unlock(so, 0); } @@ -1829,11 +2307,19 @@ tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt, */ cnt = 0; } -#if PKT_PRIORITY - if (bg) - m_prio_background(pkt); -#endif /* PKT_PRIORITY */ - error = ip_output_list(pkt, cnt, opt, &ro, flags, 0, &ipoa); +#if INET6 + if (isipv6) { + error = ip6_output_list(pkt, cnt, + inp->in6p_outputopts, &ro6, flags, NULL, NULL, + &ip6oa); + ifdenied = (ip6oa.ip6oa_retflags & IP6OARF_IFDENIED); + } else { +#endif /* INET6 */ + error = ip_output_list(pkt, cnt, opt, &ro, flags, NULL, + &ipoa); + ifdenied = (ipoa.ipoa_retflags & IPOARF_IFDENIED); + } + if (chain || error) { /* * If we sent down a chain then we are done since @@ -1850,9 +2336,70 @@ tcp_ip_output(struct socket *so, struct tcpcb *tp, struct mbuf *pkt, if (unlocked) socket_lock(so, 0); - /* Synchronize cached PCB route */ - inp_route_copyin(inp, &ro); + /* + * Enter flow controlled state if the connection is established + * and is not in recovery. + * + * A connection will enter suspended state even if it is in + * recovery. + */ + if (((adv->code == FADV_FLOW_CONTROLLED && !IN_FASTRECOVERY(tp)) || + adv->code == FADV_SUSPENDED) && + !(tp->t_flags & TF_CLOSING) && + tp->t_state == TCPS_ESTABLISHED) { + int rc; + rc = inp_set_fc_state(inp, adv->code); + + if (rc == 1) + DTRACE_TCP5(cc, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, struct tcphdr *, NULL, + int32_t, ((adv->code == FADV_FLOW_CONTROLLED) ? + TCP_CC_FLOW_CONTROL : TCP_CC_SUSPEND)); + } + + /* + * When an interface queue gets suspended, some of the + * packets are dropped. Return ENOBUFS, to update the + * pcb state. + */ + if (adv->code == FADV_SUSPENDED) + error = ENOBUFS; + + VERIFY(inp->inp_sndinprog_cnt > 0); + if ( --inp->inp_sndinprog_cnt == 0) + inp->inp_flags &= ~(INP_FC_FEEDBACK); + +#if INET6 + if (isipv6) { + if (ro6.ro_rt != NULL && (outif = ro6.ro_rt->rt_ifp) != + inp->in6p_last_outifp) + inp->in6p_last_outifp = outif; + } else +#endif /* INET6 */ + if (ro.ro_rt != NULL && (outif = ro.ro_rt->rt_ifp) != + inp->inp_last_outifp) + inp->inp_last_outifp = outif; + + if (error != 0 && ifdenied && (inp->inp_flags & INP_NO_IFT_CELLULAR)) + soevent(inp->inp_socket, + (SO_FILT_HINT_LOCKED|SO_FILT_HINT_IFDENIED)); + + /* Synchronize cached PCB route & options */ +#if INET6 + if (isipv6) + in6p_route_copyin(inp, &ro6); + else +#endif /* INET6 */ + inp_route_copyin(inp, &ro); + if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift == 0 && + tp->t_inpcb->inp_route.ro_rt != NULL) { + /* If we found the route and there is an rtt on it + * reset the retransmit timer + */ + tcp_getrt_rtt(tp, tp->t_inpcb->in6p_route.ro_rt); + tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur); + } return (error); } @@ -1862,14 +2409,187 @@ tcp_setpersist(tp) { int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1; - if (tp->t_timer[TCPT_REXMT]) - panic("tcp_setpersist: retransmit pending"); + /* If a PERSIST_TIMER option was set we will limit the + * time the persist timer will be active for that connection + * in order to avoid DOS by using zero window probes. + * see rdar://5805356 + */ + + if ((tp->t_persist_timeout != 0) && + (tp->t_timer[TCPT_PERSIST] == 0) && + (tp->t_persist_stop == 0)) { + tp->t_persist_stop = tcp_now + tp->t_persist_timeout; + } + /* * Start/restart persistance timer. */ TCPT_RANGESET(tp->t_timer[TCPT_PERSIST], t * tcp_backoff[tp->t_rxtshift], - TCPTV_PERSMIN, TCPTV_PERSMAX); + TCPTV_PERSMIN, TCPTV_PERSMAX, 0); + tp->t_timer[TCPT_PERSIST] = OFFSET_FROM_START(tp, tp->t_timer[TCPT_PERSIST]); + if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; } + +/* + * Send as many acks as data coalesced. Every other packet when stretch + * ACK is not enabled. Every 8 packets, if stretch ACK is enabled. + */ +static struct mbuf* +tcp_send_lroacks(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th) +{ + struct mbuf *mnext = NULL, *ack_chain = NULL, *tail = NULL; + int count = 0; + tcp_seq org_ack = ntohl(th->th_ack); + tcp_seq prev_ack = 0; + int tack_offset = 28; /* XXX IPv6 and IP options not supported */ + int twin_offset = 34; /* XXX IPv6 and IP options not supported */ + int ack_size = (tp->t_flags & TF_STRETCHACK) ? + (maxseg_unacked * tp->t_maxseg) : (tp->t_maxseg << 1); + int segs_acked = (tp->t_flags & TF_STRETCHACK) ? maxseg_unacked : 2; + struct mbuf *prev_ack_pkt = NULL; + struct socket *so = tp->t_inpcb->inp_socket; + unsigned short winsz = ntohs(th->th_win); + unsigned int scaled_win = winsz<rcv_scale; + tcp_seq win_rtedge = org_ack + scaled_win; + + count = tp->t_lropktlen/tp->t_maxseg; + + prev_ack = (org_ack - tp->t_lropktlen) + ack_size; + if (prev_ack < org_ack) { + ack_chain = m_dup(m, M_DONTWAIT); + if (ack_chain) { + th->th_ack = htonl(prev_ack); + /* Keep adv window constant for duplicated ACK packets */ + scaled_win = win_rtedge - prev_ack; + if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale)) + scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale); + th->th_win = htons(scaled_win>>tp->rcv_scale); + if (lrodebug == 5) { + printf("%s: win = %d winsz = %d sc = %d" + " lro_len %d %d\n", + __func__, scaled_win>>tp->rcv_scale, winsz, + tp->rcv_scale, tp->t_lropktlen, count); + } + tail = ack_chain; + count -= segs_acked; /* accounts for prev_ack packet */ + count = (count <= segs_acked) ? 0 : count - segs_acked; + tcpstat.tcps_sndacks++; + so_tc_update_stats(m, so, m_get_service_class(m)); + } else { + return NULL; + } + } + else { + tp->t_lropktlen = 0; + return NULL; + } + + prev_ack_pkt = ack_chain; + + while (count > 0) { + if ((prev_ack + ack_size) < org_ack) { + prev_ack += ack_size; + } else { + /* + * The last ACK sent must have the ACK number that TCP + * thinks is the last sent ACK number. + */ + prev_ack = org_ack; + } + mnext = m_dup(prev_ack_pkt, M_DONTWAIT); + if (mnext) { + /* Keep adv window constant for duplicated ACK packets */ + scaled_win = win_rtedge - prev_ack; + if (scaled_win > (int32_t)(TCP_MAXWIN << tp->rcv_scale)) + scaled_win = (int32_t)(TCP_MAXWIN << tp->rcv_scale); + winsz = htons(scaled_win>>tp->rcv_scale); + if (lrodebug == 5) { + printf("%s: winsz = %d ack %x count %d\n", + __func__, scaled_win>>tp->rcv_scale, + prev_ack, count); + } + bcopy(&winsz, mtod(prev_ack_pkt, caddr_t) + twin_offset, 2); + HTONL(prev_ack); + bcopy(&prev_ack, mtod(prev_ack_pkt, caddr_t) + tack_offset, 4); + NTOHL(prev_ack); + tail->m_nextpkt = mnext; + tail = mnext; + count -= segs_acked; + tcpstat.tcps_sndacks++; + so_tc_update_stats(m, so, m_get_service_class(m)); + } else { + if (lrodebug == 5) { + printf("%s: failed to alloc mbuf.\n", __func__); + } + break; + } + prev_ack_pkt = mnext; + } + tp->t_lropktlen = 0; + return ack_chain; +} + +static int +tcp_recv_throttle (struct tcpcb *tp) +{ + uint32_t base_rtt, newsize; + int32_t qdelay; + struct sockbuf *sbrcv = &tp->t_inpcb->inp_socket->so_rcv; + + if (tcp_use_rtt_recvbg == 1 && + TSTMP_SUPPORTED(tp)) { + /* + * Timestamps are supported on this connection. Use + * RTT to look for an increase in latency. + */ + + /* + * If the connection is already being throttled, leave it + * in that state until rtt comes closer to base rtt + */ + if (tp->t_flagsext & TF_RECV_THROTTLE) + return (1); + + base_rtt = get_base_rtt(tp); + + if (base_rtt != 0 && tp->t_rttcur != 0) { + qdelay = tp->t_rttcur - base_rtt; + /* + * if latency increased on a background flow, + * return 1 to start throttling. + */ + if (qdelay > target_qdelay) { + tp->t_flagsext |= TF_RECV_THROTTLE; + + /* + * Reduce the recv socket buffer size to + * minimize latecy. + */ + if (sbrcv->sb_idealsize > + tcp_recv_throttle_minwin) { + newsize = sbrcv->sb_idealsize >> 1; + /* Set a minimum of 16 K */ + newsize = + max(newsize, + tcp_recv_throttle_minwin); + sbrcv->sb_idealsize = newsize; + } + return (1); + } else { + return (0); + } + } + } + + /* + * Timestamps are not supported or there is no good RTT + * measurement. Use IPDV in this case. + */ + if (tp->acc_iaj > tcp_acc_iaj_react_limit) + return (1); + + return (0); +}