X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/b7266188b87f3620ec3f9f717e57194a7dd989fe..a1c7dba18ef36983396c282fe85292db066e39db:/bsd/netinet/tcp_input.c diff --git a/bsd/netinet/tcp_input.c b/bsd/netinet/tcp_input.c index a293cc24a..ab5242853 100644 --- a/bsd/netinet/tcp_input.c +++ b/bsd/netinet/tcp_input.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@ * @@ -78,7 +78,8 @@ #include #include #include - +#include +#include #include /* before tcp_seq.h, for tcp_random18() */ #include @@ -86,6 +87,8 @@ #include #include #include +#include +#include #include #include @@ -95,6 +98,7 @@ #include /* for ICMP_BANDLIM */ #include #include +#include #if INET6 #include #include @@ -107,6 +111,9 @@ #include #include #include +#include +#include +#include #if INET6 #include #endif @@ -130,138 +137,208 @@ struct tcphdr tcp_savetcp; #endif /* CONFIG_MACF_NET || CONFIG_MACF_SOCKET */ #include - -#ifndef __APPLE__ -MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry"); -#endif +#include +#if MPTCP +#include +#include +#include +#endif /* MPTCP */ #define DBG_LAYER_BEG NETDBG_CODE(DBG_NETTCP, 0) #define DBG_LAYER_END NETDBG_CODE(DBG_NETTCP, 2) #define DBG_FNC_TCP_INPUT NETDBG_CODE(DBG_NETTCP, (3 << 8)) #define DBG_FNC_TCP_NEWCONN NETDBG_CODE(DBG_NETTCP, (7 << 8)) -static int tcprexmtthresh = 2; tcp_cc tcp_ccgen; -#if IPSEC -extern int ipsec_bypass; -#endif - struct tcpstat tcpstat; static int log_in_vain = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW | CTLFLAG_LOCKED, &log_in_vain, 0, "Log all incoming TCP connections"); static int blackhole = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_RW | CTLFLAG_LOCKED, &blackhole, 0, "Do not send RST when dropping refused connections"); int tcp_delack_enabled = 3; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_delack_enabled, 0, "Delay ACK to try and piggyback it onto a data packet"); int tcp_lq_overflow = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_lq_overflow, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_lq_overflow, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_lq_overflow, 0, "Listen Queue Overflow"); +int tcp_recv_bg = 0; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbg, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_recv_bg, 0, + "Receive background"); + #if TCP_DROP_SYNFIN static int drop_synfin = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_RW | CTLFLAG_LOCKED, &drop_synfin, 0, "Drop TCP packets with SYN+FIN set"); #endif SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW|CTLFLAG_LOCKED, 0, "TCP Segment Reassembly Queue"); -__private_extern__ int tcp_reass_maxseg = 0; -SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RW, - &tcp_reass_maxseg, 0, - "Global maximum number of TCP Segments in Reassembly Queue"); - -__private_extern__ int tcp_reass_qsize = 0; -SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD, - &tcp_reass_qsize, 0, - "Global number of TCP Segments currently in Reassembly Queue"); - static int tcp_reass_overflows = 0; -SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD, +SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD | CTLFLAG_LOCKED, &tcp_reass_overflows, 0, "Global number of TCP Segment Reassembly Queue Overflows"); __private_extern__ int slowlink_wsize = 8192; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowlink_wsize, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowlink_wsize, CTLFLAG_RW | CTLFLAG_LOCKED, &slowlink_wsize, 0, "Maximum advertised window size for slowlink"); -static int maxseg_unacked = 8; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, maxseg_unacked, CTLFLAG_RW, +int maxseg_unacked = 8; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, maxseg_unacked, CTLFLAG_RW | CTLFLAG_LOCKED, &maxseg_unacked, 0, "Maximum number of outstanding segments left unacked"); -static int tcp_do_rfc3465 = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465, CTLFLAG_RW, +int tcp_do_rfc3465 = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_do_rfc3465, 0, ""); -static int tcp_do_rfc3465_lim2 = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465_lim2, CTLFLAG_RW, +int tcp_do_rfc3465_lim2 = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465_lim2, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_do_rfc3465_lim2, 0, "Appropriate bytes counting w/ L=2*SMSS"); +int rtt_samples_per_slot = 20; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, rtt_samples_per_slot, CTLFLAG_RW | CTLFLAG_LOCKED, + &rtt_samples_per_slot, 0, "Number of RTT samples stored for rtt history"); + +int tcp_allowed_iaj = ALLOWED_IAJ; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, recv_allowed_iaj, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_allowed_iaj, 0, "Allowed inter-packet arrival jiter"); + +int tcp_acc_iaj_high_thresh = ACC_IAJ_HIGH_THRESH; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, acc_iaj_high_thresh, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_acc_iaj_high_thresh, 0, "Used in calculating maximum accumulated IAJ"); + +u_int32_t tcp_do_autorcvbuf = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, doautorcvbuf, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_do_autorcvbuf, 0, "Enable automatic socket buffer tuning"); + +u_int32_t tcp_autorcvbuf_inc_shift = 3; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, autorcvbufincshift, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_autorcvbuf_inc_shift, 0, "Shift for increment in receive socket buffer size"); + +u_int32_t tcp_autorcvbuf_max = 512 * 1024; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, autorcvbufmax, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_autorcvbuf_max, 0, "Maximum receive socket buffer size"); + +int sw_lro = 0; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_LOCKED, + &sw_lro, 0, "Used to coalesce TCP packets"); + +int lrodebug = 0; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, lrodbg, CTLFLAG_RW | CTLFLAG_LOCKED, + &lrodebug, 0, "Used to debug SW LRO"); + +int lro_start = 4; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, lro_startcnt, CTLFLAG_RW | CTLFLAG_LOCKED, + &lro_start, 0, "Segments for starting LRO computed as power of 2"); + +extern int tcp_do_autosendbuf; + +int limited_txmt = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, limited_transmit, CTLFLAG_RW | CTLFLAG_LOCKED, + &limited_txmt, 0, "Enable limited transmit"); + +int early_rexmt = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, early_rexmt, CTLFLAG_RW | CTLFLAG_LOCKED, + &early_rexmt, 0, "Enable Early Retransmit"); + +int sack_ackadv = 1; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack_ackadv, CTLFLAG_RW | CTLFLAG_LOCKED, + &sack_ackadv, 0, "Use SACK with cumulative ack advancement as a dupack"); + #if CONFIG_IFEF_NOWINDOWSCALE int tcp_obey_ifef_nowindowscale = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, obey_ifef_nowindowscale, CTLFLAG_RW, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, obey_ifef_nowindowscale, CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_obey_ifef_nowindowscale, 0, ""); #endif extern int tcp_TCPTV_MIN; +extern int tcp_acc_iaj_high; +extern int tcp_acc_iaj_react_limit; +extern struct zone *tcp_reass_zone; + +int tcprexmtthresh = 3; u_int32_t tcp_now; +struct timeval tcp_uptime; /* uptime when tcp_now was last updated */ +lck_spin_t *tcp_uptime_lock; /* Used to sychronize updates to tcp_now */ struct inpcbhead tcb; #define tcb6 tcb /* for KAME src sync over BSD*'s */ struct inpcbinfo tcbinfo; -static void tcp_dooptions(struct tcpcb *, - u_char *, int, struct tcphdr *, struct tcpopt *, unsigned int); +static void tcp_dooptions(struct tcpcb *, u_char *, int, struct tcphdr *, + struct tcpopt *, unsigned int); static void tcp_pulloutofband(struct socket *, struct tcphdr *, struct mbuf *, int); -static int tcp_reass(struct tcpcb *, struct tcphdr *, int *, - struct mbuf *); -static void tcp_xmit_timer(struct tcpcb *, int); +static int tcp_reass(struct tcpcb *, struct tcphdr *, int *, struct mbuf *, + struct ifnet *); +static void tcp_xmit_timer(struct tcpcb *, int, u_int32_t, tcp_seq); static inline unsigned int tcp_maxmtu(struct rtentry *); -#if INET6 -static inline unsigned int tcp_maxmtu6(struct rtentry *); -#endif +static inline int tcp_stretch_ack_enable(struct tcpcb *tp); +static inline void tcp_adaptive_rwtimo_check(struct tcpcb *, int); + +#if TRAFFIC_MGT +static inline void update_iaj_state(struct tcpcb *tp, uint32_t tlen, + int reset_size); +void compute_iaj(struct tcpcb *tp, int nlropkts, int lro_delay_factor); +static void compute_iaj_meat(struct tcpcb *tp, uint32_t cur_iaj); +#endif /* TRAFFIC_MGT */ -/* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */ #if INET6 -#define ND6_HINT(tp) \ -do { \ - if ((tp) && (tp)->t_inpcb && \ - ((tp)->t_inpcb->inp_vflag & INP_IPV6) != 0 && \ - (tp)->t_inpcb->in6p_route.ro_rt) \ - nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \ -} while (0) -#else -#define ND6_HINT(tp) +static inline unsigned int tcp_maxmtu6(struct rtentry *); #endif -extern u_int32_t *delack_bitmask; +static void tcp_sbrcv_grow(struct tcpcb *tp, struct sockbuf *sb, + struct tcpopt *to, u_int32_t tlen); + +void tcp_sbrcv_trim(struct tcpcb *tp, struct sockbuf *sb); +static void tcp_sbsnd_trim(struct sockbuf *sbsnd); +static inline void tcp_sbrcv_tstmp_check(struct tcpcb *tp); +static inline void tcp_sbrcv_reserve(struct tcpcb *tp, struct sockbuf *sb, + u_int32_t newsize, u_int32_t idealsize); +static void tcp_bad_rexmt_restore_state(struct tcpcb *tp, struct tcphdr *th); +static void tcp_compute_rtt(struct tcpcb *tp, struct tcpopt *to, + struct tcphdr *th); +static void tcp_early_rexmt_check(struct tcpcb *tp, struct tcphdr *th); +static void tcp_bad_rexmt_check(struct tcpcb *tp, struct tcphdr *th, + struct tcpopt *to); +/* + * Constants used for resizing receive socket buffer + * when timestamps are not supported + */ +#define TCPTV_RCVNOTS_QUANTUM 100 +#define TCP_RCVNOTS_BYTELEVEL 204800 -extern void add_to_time_wait(struct tcpcb *); -extern void postevent(struct socket *, struct sockbuf *, int); +/* + * Constants used for limiting early retransmits + * to 10 per minute. + */ +#define TCP_EARLY_REXMT_WIN (60 * TCP_RETRANSHZ) /* 60 seconds */ +#define TCP_EARLY_REXMT_LIMIT 10 extern void ipfwsyslog( int level, const char *format,...); -extern int ChkAddressOK( __uint32_t dstaddr, __uint32_t srcaddr ); extern int fw_verbose; -__private_extern__ int tcp_sockthreshold; -__private_extern__ int tcp_win_scale; #if IPFIREWALL +extern void ipfw_stealth_stats_incr_tcp(void); + #define log_in_vain_log( a ) { \ if ( (log_in_vain == 3 ) && (fw_verbose == 2)) { /* Apple logging, log to ipfw.log */ \ ipfwsyslog a ; \ + } else if ( (log_in_vain == 4 ) && (fw_verbose == 2)) { \ + ipfw_stealth_stats_incr_tcp(); \ } \ else log a ; \ } @@ -269,52 +346,230 @@ __private_extern__ int tcp_win_scale; #define log_in_vain_log( a ) { log a; } #endif +int tcp_rcvunackwin = TCPTV_UNACKWIN; +int tcp_maxrcvidle = TCPTV_MAXRCVIDLE; +int tcp_rcvsspktcnt = TCP_RCV_SS_PKTCOUNT; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, rcvsspktcnt, CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_rcvsspktcnt, 0, "packets to be seen before receiver stretches acks"); -/* - * Indicate whether this ack should be delayed. - * We can delay the ack if: - * - delayed acks are enabled (set to 1) and - * - our last ack wasn't a 0-sized window. We never want to delay - * the ack that opens up a 0-sized window. - * - delayed acks are enabled (set to 2, "more compatible") and - * - our last ack wasn't a 0-sized window. - * - if the peer hasn't sent us a TH_PUSH data packet (this solves 3649245) - * - the peer hasn't sent us a TH_PUSH data packet, if he did, take this as a clue that we - * need to ACK with no delay. This helps higher level protocols who won't send - * us more data even if the window is open because their last "segment" hasn't been ACKed - * - delayed acks are enabled (set to 3, "streaming detection") and - * - if we receive more than "maxseg_unacked" full packets per second on this socket - * - if we don't have more than "maxseg_unacked" delayed so far - * - if those criteria aren't met, acts like "2". Allowing faster acking while browsing for example. - * - */ -#define DELAY_ACK(tp) \ - (((tcp_delack_enabled == 1) && ((tp->t_flags & TF_RXWIN0SENT) == 0)) || \ - (((tcp_delack_enabled == 2) && (tp->t_flags & TF_RXWIN0SENT) == 0) && \ - ((thflags & TH_PUSH) == 0) && ((tp->t_flags & TF_DELACK) == 0)) || \ - (((tcp_delack_enabled == 3) && (tp->t_flags & TF_RXWIN0SENT) == 0) && \ - (tp->t_rcvtime == 0) && ((thflags & TH_PUSH) == 0) && \ - (((tp->t_unacksegs == 0)) || \ - ((tp->rcv_byps > (maxseg_unacked * tp->t_maxseg)) && (tp->t_unacksegs < maxseg_unacked))))) +#define DELAY_ACK(tp, th) \ + (CC_ALGO(tp)->delay_ack != NULL && CC_ALGO(tp)->delay_ack(tp, th)) static int tcp_dropdropablreq(struct socket *head); static void tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th); +static void update_base_rtt(struct tcpcb *tp, uint32_t rtt); +void tcp_set_background_cc(struct socket *so); +void tcp_set_foreground_cc(struct socket *so); +static void tcp_set_new_cc(struct socket *so, uint16_t cc_index); +static void tcp_bwmeas_check(struct tcpcb *tp); + +#if TRAFFIC_MGT +void +reset_acc_iaj(struct tcpcb *tp) +{ + tp->acc_iaj = 0; + tp->iaj_rwintop = 0; + CLEAR_IAJ_STATE(tp); +} + +static inline void +update_iaj_state(struct tcpcb *tp, uint32_t size, int rst_size) +{ + if (rst_size > 0) + tp->iaj_size = 0; + if (tp->iaj_size == 0 || size >= tp->iaj_size) { + tp->iaj_size = size; + tp->iaj_rcv_ts = tcp_now; + tp->iaj_small_pkt = 0; + } +} + +/* For every 32 bit unsigned integer(v), this function will find the + * largest integer n such that (n*n <= v). This takes at most 16 iterations + * irrespective of the value of v and does not involve multiplications. + */ +static inline int +isqrt(unsigned int val) { + unsigned int sqrt_cache[11] = {0, 1, 4, 9, 16, 25, 36, 49, 64, 81, 100}; + unsigned int temp, g=0, b=0x8000, bshft=15; + if ( val <= 100) { + for (g = 0; g <= 10; ++g) { + if (sqrt_cache[g] > val) { + g--; + break; + } else if (sqrt_cache[g] == val) { + break; + } + } + } else { + do { + temp = (((g << 1) + b) << (bshft--)); + if (val >= temp) { + g += b; + val -= temp; + } + b >>= 1; + } while ( b > 0 && val > 0); + } + return(g); +} + +/* +* With LRO, roughly estimate the inter arrival time between +* each sub coalesced packet as an average. Count the delay +* cur_iaj to be the delay between the last packet received +* and the first packet of the LRO stream. Due to round off errors +* cur_iaj may be the same as lro_delay_factor. Averaging has +* round off errors too. lro_delay_factor may be close to 0 +* in steady state leading to lower values fed to compute_iaj_meat. +*/ +void +compute_iaj(struct tcpcb *tp, int nlropkts, int lro_delay_factor) +{ + uint32_t cur_iaj = tcp_now - tp->iaj_rcv_ts; + uint32_t timediff = 0; + + if (cur_iaj >= lro_delay_factor) { + cur_iaj = cur_iaj - lro_delay_factor; + } + + compute_iaj_meat(tp, cur_iaj); + + if (nlropkts <= 1) + return; + + nlropkts--; + + timediff = lro_delay_factor/nlropkts; + + while (nlropkts > 0) + { + compute_iaj_meat(tp, timediff); + nlropkts--; + } +} + +static +void compute_iaj_meat(struct tcpcb *tp, uint32_t cur_iaj) +{ + /* When accumulated IAJ reaches MAX_ACC_IAJ in milliseconds, + * throttle the receive window to a minimum of MIN_IAJ_WIN packets + */ +#define MAX_ACC_IAJ (tcp_acc_iaj_high_thresh + tcp_acc_iaj_react_limit) +#define IAJ_DIV_SHIFT 4 +#define IAJ_ROUNDUP_CONST (1 << (IAJ_DIV_SHIFT - 1)) + + uint32_t allowed_iaj, acc_iaj = 0; + + uint32_t mean, temp; + int32_t cur_iaj_dev; + + cur_iaj_dev = (cur_iaj - tp->avg_iaj); + + /* Allow a jitter of "allowed_iaj" milliseconds. Some connections + * may have a constant jitter more than that. We detect this by + * using standard deviation. + */ + allowed_iaj = tp->avg_iaj + tp->std_dev_iaj; + if (allowed_iaj < tcp_allowed_iaj) + allowed_iaj = tcp_allowed_iaj; + + /* Initially when the connection starts, the senders congestion + * window is small. During this period we avoid throttling a + * connection because we do not have a good starting point for + * allowed_iaj. IAJ_IGNORE_PKTCNT is used to quietly gloss over + * the first few packets. + */ + if (tp->iaj_pktcnt > IAJ_IGNORE_PKTCNT) { + if ( cur_iaj <= allowed_iaj ) { + if (tp->acc_iaj >= 2) + acc_iaj = tp->acc_iaj - 2; + else + acc_iaj = 0; + + } else { + acc_iaj = tp->acc_iaj + (cur_iaj - allowed_iaj); + } + + if (acc_iaj > MAX_ACC_IAJ) + acc_iaj = MAX_ACC_IAJ; + tp->acc_iaj = acc_iaj; + } + + /* Compute weighted average where the history has a weight of + * 15 out of 16 and the current value has a weight of 1 out of 16. + * This will make the short-term measurements have more weight. + * + * The addition of 8 will help to round-up the value + * instead of round-down + */ + tp->avg_iaj = (((tp->avg_iaj << IAJ_DIV_SHIFT) - tp->avg_iaj) + + cur_iaj + IAJ_ROUNDUP_CONST) >> IAJ_DIV_SHIFT; + + /* Compute Root-mean-square of deviation where mean is a weighted + * average as described above. + */ + temp = tp->std_dev_iaj * tp->std_dev_iaj; + mean = (((temp << IAJ_DIV_SHIFT) - temp) + + (cur_iaj_dev * cur_iaj_dev) + + IAJ_ROUNDUP_CONST) >> IAJ_DIV_SHIFT; + + tp->std_dev_iaj = isqrt(mean); + + DTRACE_TCP3(iaj, struct tcpcb *, tp, uint32_t, cur_iaj, + uint32_t, allowed_iaj); + + return; +} +#endif /* TRAFFIC_MGT */ + +/* Check if enough amount of data has been acknowledged since + * bw measurement was started + */ +static void +tcp_bwmeas_check(struct tcpcb *tp) +{ + int32_t bw_meas_bytes; + uint32_t bw, bytes, elapsed_time; + bw_meas_bytes = tp->snd_una - tp->t_bwmeas->bw_start; + if ((tp->t_flagsext & TF_BWMEAS_INPROGRESS) != 0 && + bw_meas_bytes >= (int32_t)(tp->t_bwmeas->bw_size)) { + bytes = bw_meas_bytes; + elapsed_time = tcp_now - tp->t_bwmeas->bw_ts; + if (elapsed_time > 0) { + bw = bytes / elapsed_time; + if ( bw > 0) { + if (tp->t_bwmeas->bw_sndbw > 0) { + tp->t_bwmeas->bw_sndbw = + (((tp->t_bwmeas->bw_sndbw << 3) - tp->t_bwmeas->bw_sndbw) + bw) >> 3; + } else { + tp->t_bwmeas->bw_sndbw = bw; + } + } + } + tp->t_flagsext &= ~(TF_BWMEAS_INPROGRESS); + } +} static int -tcp_reass(tp, th, tlenp, m) - register struct tcpcb *tp; - register struct tcphdr *th; - int *tlenp; - struct mbuf *m; +tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m, + struct ifnet *ifp) { struct tseg_qent *q; struct tseg_qent *p = NULL; struct tseg_qent *nq; struct tseg_qent *te = NULL; - struct socket *so = tp->t_inpcb->inp_socket; - int flags; + struct inpcb *inp = tp->t_inpcb; + struct socket *so = inp->inp_socket; + int flags = 0; int dowakeup = 0; + struct mbuf *oodata = NULL; + int copy_oodata = 0; + u_int16_t qlimit; + boolean_t cell = IFNET_IS_CELLULAR(ifp); + boolean_t wifi = (!cell && IFNET_IS_WIFI(ifp)); + boolean_t wired = (!wifi && IFNET_IS_WIRED(ifp)); /* * Call with th==0 after become established to @@ -322,6 +577,18 @@ tcp_reass(tp, th, tlenp, m) */ if (th == NULL) goto present; + + /* + * If the reassembly queue already has entries or if we are going + * to add a new one, then the connection has reached a loss state. + * Reset the stretch-ack algorithm at this point. + */ + tcp_reset_stretch_ack(tp); + +#if TRAFFIC_MGT + if (tp->acc_iaj > 0) + reset_acc_iaj(tp); +#endif /* TRAFFIC_MGT */ /* * Limit the number of segments in the reassembly queue to prevent @@ -330,8 +597,10 @@ tcp_reass(tp, th, tlenp, m) * queue. Always keep one global queue entry spare to be able to * process the missing segment. */ + qlimit = min(max(100, so->so_rcv.sb_hiwat >> 10), + tcp_autorcvbuf_max >> 10); if (th->th_seq != tp->rcv_nxt && - tcp_reass_qsize + 1 >= tcp_reass_maxseg) { + (tp->t_reassqlen + 1) >= qlimit) { tcp_reass_overflows++; tcpstat.tcps_rcvmemdrop++; m_freem(m); @@ -340,14 +609,13 @@ tcp_reass(tp, th, tlenp, m) } /* Allocate a new queue entry. If we can't, just drop the pkt. XXX */ - MALLOC(te, struct tseg_qent *, sizeof (struct tseg_qent), M_TSEGQ, - M_NOWAIT); + te = (struct tseg_qent *) zalloc(tcp_reass_zone); if (te == NULL) { tcpstat.tcps_rcvmemdrop++; m_freem(m); return (0); } - tcp_reass_qsize++; + tp->t_reassqlen++; /* * Find a segment which begins after this one does. @@ -371,16 +639,27 @@ tcp_reass(tp, th, tlenp, m) if (i >= *tlenp) { tcpstat.tcps_rcvduppack++; tcpstat.tcps_rcvdupbyte += *tlenp; + if (nstat_collect) { + nstat_route_rx(inp->inp_route.ro_rt, + 1, *tlenp, + NSTAT_RX_FLAG_DUPLICATE); + INP_ADD_STAT(inp, cell, wifi, wired, + rxpackets, 1); + INP_ADD_STAT(inp, cell, wifi, wired, + rxbytes, *tlenp); + tp->t_stat.rxduplicatebytes += *tlenp; + } m_freem(m); - FREE(te, M_TSEGQ); - tcp_reass_qsize--; + zfree(tcp_reass_zone, te); + te = NULL; + tp->t_reassqlen--; /* * Try to present any queued data * at the left window edge to the user. * This is needed after the 3-WHS * completes. */ - goto present; /* ??? */ + goto present; } m_adj(m, i); *tlenp -= i; @@ -389,6 +668,13 @@ tcp_reass(tp, th, tlenp, m) } tcpstat.tcps_rcvoopack++; tcpstat.tcps_rcvoobyte += *tlenp; + if (nstat_collect) { + nstat_route_rx(inp->inp_route.ro_rt, 1, *tlenp, + NSTAT_RX_FLAG_OUT_OF_ORDER); + INP_ADD_STAT(inp, cell, wifi, wired, rxpackets, 1); + INP_ADD_STAT(inp, cell, wifi, wired, rxbytes, *tlenp); + tp->t_stat.rxoutoforderbytes += *tlenp; + } /* * While we overlap succeeding segments trim them or, @@ -408,8 +694,8 @@ tcp_reass(tp, th, tlenp, m) nq = LIST_NEXT(q, tqe_q); LIST_REMOVE(q, tqe_q); m_freem(q->tqe_m); - FREE(q, M_TSEGQ); - tcp_reass_qsize--; + zfree(tcp_reass_zone, q); + tp->t_reassqlen--; q = nq; } @@ -424,6 +710,15 @@ tcp_reass(tp, th, tlenp, m) LIST_INSERT_AFTER(p, te, tqe_q); } + /* + * New out-of-order data exists, and is pointed to by + * queue entry te. Set copy_oodata to 1 so out-of-order data + * can be copied off to sockbuf after in-order data + * is copied off. + */ + if (!(so->so_state & SS_CANTRCVMORE)) + copy_oodata = 1; + present: /* * Present data to user, advancing rcv_nxt through @@ -432,47 +727,104 @@ present: if (!TCPS_HAVEESTABLISHED(tp->t_state)) return (0); q = LIST_FIRST(&tp->t_segq); - if (!q || q->tqe_th->th_seq != tp->rcv_nxt) + if (!q || q->tqe_th->th_seq != tp->rcv_nxt) { + /* Stop using LRO once out of order packets arrive */ + if (tp->t_flagsext & TF_LRO_OFFLOADED) { + tcp_lro_remove_state(inp->inp_laddr, inp->inp_faddr, + th->th_dport, th->th_sport); + tp->t_flagsext &= ~TF_LRO_OFFLOADED; + } + + /* + * continue processing if out-of-order data + * can be delivered + */ + if (q && (so->so_flags & SOF_ENABLE_MSGS)) + goto msg_unordered_delivery; + return (0); + } + + /* lost packet was recovered, so ooo data can be returned */ + tcpstat.tcps_recovered_pkts++; + do { tp->rcv_nxt += q->tqe_len; flags = q->tqe_th->th_flags & TH_FIN; nq = LIST_NEXT(q, tqe_q); LIST_REMOVE(q, tqe_q); - if (so->so_state & SS_CANTRCVMORE) + if (so->so_state & SS_CANTRCVMORE) { m_freem(q->tqe_m); - else { - if (sbappendstream(&so->so_rcv, q->tqe_m)) + } else { + so_recv_data_stat(so, q->tqe_m, 0); /* XXXX */ + if (so->so_flags & SOF_ENABLE_MSGS) { + /* + * Append the inorder data as a message to the + * receive socket buffer. Also check to see if + * the data we are about to deliver is the same + * data that we wanted to pass up to the user + * out of order. If so, reset copy_oodata -- + * the received data filled a gap, and + * is now in order! + */ + if (q == te) + copy_oodata = 0; + } + if (sbappendstream_rcvdemux(so, q->tqe_m, + q->tqe_th->th_seq - (tp->irs + 1), 0)) dowakeup = 1; + if (tp->t_flagsext & TF_LRO_OFFLOADED) { + tcp_update_lro_seq(tp->rcv_nxt, + inp->inp_laddr, inp->inp_faddr, + th->th_dport, th->th_sport); + } } - FREE(q, M_TSEGQ); - tcp_reass_qsize--; + zfree(tcp_reass_zone, q); + tp->t_reassqlen--; q = nq; } while (q && q->tqe_th->th_seq == tp->rcv_nxt); - ND6_HINT(tp); #if INET6 - if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) { + if ((inp->inp_vflag & INP_IPV6) != 0) { 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)), 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)), + ((inp->inp_fport << 16) | inp->inp_lport), + (((inp->inp_laddr.s_addr & 0xffff) << 16) | + (inp->inp_faddr.s_addr & 0xffff)), 0,0,0); } + +msg_unordered_delivery: + /* Deliver out-of-order data as a message */ + if (te && (so->so_flags & SOF_ENABLE_MSGS) && copy_oodata && te->tqe_len) { + /* + * make a copy of the mbuf to be delivered up to + * the user, and add it to the sockbuf + */ + oodata = m_copym(te->tqe_m, 0, M_COPYALL, M_DONTWAIT); + if (oodata != NULL) { + if (sbappendmsgstream_rcv(&so->so_rcv, oodata, + te->tqe_th->th_seq - (tp->irs + 1), 1)) { + dowakeup = 1; + tcpstat.tcps_msg_unopkts++; + } else { + tcpstat.tcps_msg_unoappendfail++; + } + } + } + if (dowakeup) sorwakeup(so); /* done with socket lock held */ return (flags); - } /* @@ -482,56 +834,653 @@ static void tcp_reduce_congestion_window( struct tcpcb *tp) { - u_int win; - - win = min(tp->snd_wnd, tp->snd_cwnd) / - 2 / tp->t_maxseg; - if (win < 2) - win = 2; - tp->snd_ssthresh = win * tp->t_maxseg; + /* + * If the current tcp cc module has + * defined a hook for tasks to run + * before entering FR, call it + */ + if (CC_ALGO(tp)->pre_fr != NULL) + CC_ALGO(tp)->pre_fr(tp); ENTER_FASTRECOVERY(tp); tp->snd_recover = tp->snd_max; tp->t_timer[TCPT_REXMT] = 0; + tp->t_timer[TCPT_PTO] = 0; + tp->t_rtttime = 0; + tp->snd_cwnd = tp->snd_ssthresh + + tp->t_maxseg * tcprexmtthresh; +} + +/* + * The application wants to get an event if there + * is a stall during read. Set the initial keepalive + * timeout to be equal to twice RTO. + */ +static inline void +tcp_adaptive_rwtimo_check(struct tcpcb *tp, int tlen) +{ + if (tp->t_adaptive_rtimo > 0 && tlen > 0 && + tp->t_state == TCPS_ESTABLISHED) { + tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, + (TCP_REXMTVAL(tp) << 1)); + tp->t_flagsext |= TF_DETECT_READSTALL; + tp->t_rtimo_probes = 0; + } +} + +inline void +tcp_keepalive_reset(struct tcpcb *tp) +{ + tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, + TCP_CONN_KEEPIDLE(tp)); + tp->t_flagsext &= ~(TF_DETECT_READSTALL); + tp->t_rtimo_probes = 0; +} + +/* + * TCP input routine, follows pages 65-76 of the + * protocol specification dated September, 1981 very closely. + */ +#if INET6 +int +tcp6_input(struct mbuf **mp, int *offp, int proto) +{ +#pragma unused(proto) + register struct mbuf *m = *mp; + uint32_t ia6_flags; + struct ifnet *ifp = m->m_pkthdr.rcvif; + + IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), return IPPROTO_DONE); + + /* Expect 32-bit aligned data pointer on strict-align platforms */ + MBUF_STRICT_DATA_ALIGNMENT_CHECK_32(m); + + /* + * draft-itojun-ipv6-tcp-to-anycast + * better place to put this in? + */ + if (ip6_getdstifaddr_info(m, NULL, &ia6_flags) == 0) { + if (ia6_flags & IN6_IFF_ANYCAST) { + struct ip6_hdr *ip6; + + ip6 = mtod(m, struct ip6_hdr *); + icmp6_error(m, ICMP6_DST_UNREACH, + ICMP6_DST_UNREACH_ADDR, + (caddr_t)&ip6->ip6_dst - (caddr_t)ip6); + + IF_TCP_STATINC(ifp, icmp6unreach); + + return (IPPROTO_DONE); + } + } + + tcp_input(m, *offp); + return (IPPROTO_DONE); +} +#endif + +/* Depending on the usage of mbuf space in the system, this function + * will return true or false. This is used to determine if a socket + * buffer can take more memory from the system for auto-tuning or not. + */ +u_int8_t +tcp_cansbgrow(struct sockbuf *sb) +{ + /* Calculate the host level space limit in terms of MSIZE buffers. + * We can use a maximum of half of the available mbuf space for + * socket buffers. + */ + u_int32_t mblim = ((nmbclusters >> 1) << (MCLSHIFT - MSIZESHIFT)); + + /* Calculate per sb limit in terms of bytes. We optimize this limit + * for upto 16 socket buffers. + */ + + u_int32_t sbspacelim = ((nmbclusters >> 4) << MCLSHIFT); + + if ((total_sbmb_cnt < mblim) && + (sb->sb_hiwat < sbspacelim)) { + return(1); + } else { + OSIncrementAtomic64(&sbmb_limreached); + } + return(0); +} + +static void +tcp_sbrcv_reserve(struct tcpcb *tp, struct sockbuf *sbrcv, + u_int32_t newsize, u_int32_t idealsize) +{ + + /* newsize should not exceed max */ + newsize = min(newsize, tcp_autorcvbuf_max); + + /* The receive window scale negotiated at the + * beginning of the connection will also set a + * limit on the socket buffer size + */ + newsize = min(newsize, TCP_MAXWIN << tp->rcv_scale); + + /* Set new socket buffer size */ + if (newsize > sbrcv->sb_hiwat && + (sbreserve(sbrcv, newsize) == 1)) { + sbrcv->sb_idealsize = min(max(sbrcv->sb_idealsize, + (idealsize != 0) ? idealsize : newsize), + tcp_autorcvbuf_max); + + /* Again check the limit set by the advertised + * window scale + */ + sbrcv->sb_idealsize = min(sbrcv->sb_idealsize, + TCP_MAXWIN << tp->rcv_scale); + } +} + +/* + * This function is used to grow a receive socket buffer. It + * will take into account system-level memory usage and the + * bandwidth available on the link to make a decision. + */ +static void +tcp_sbrcv_grow(struct tcpcb *tp, struct sockbuf *sbrcv, + struct tcpopt *to, u_int32_t pktlen) { + + /* + * Do not grow the receive socket buffer if + * - auto resizing is disabled, globally or on this socket + * - the high water mark already reached the maximum + * - the stream is in background and receive side is being + * throttled + * - if there are segments in reassembly queue indicating loss, + * do not need to increase recv window during recovery as more + * data is not going to be sent. A duplicate ack sent during + * recovery should not change the receive window + */ + if (tcp_do_autorcvbuf == 0 || + (sbrcv->sb_flags & SB_AUTOSIZE) == 0 || + tcp_cansbgrow(sbrcv) == 0 || + sbrcv->sb_hiwat >= tcp_autorcvbuf_max || + (tp->t_flagsext & TF_RECV_THROTTLE) || + !LIST_EMPTY(&tp->t_segq)) { + /* Can not resize the socket buffer, just return */ + goto out; + } + + if (TSTMP_GT(tcp_now, + tp->rfbuf_ts + TCPTV_RCVBUFIDLE)) { + /* If there has been an idle period in the + * connection, just restart the measurement + */ + goto out; + } + + if (!TSTMP_SUPPORTED(tp)) { + /* + * Timestamp option is not supported on this connection. + * If the connection reached a state to indicate that + * the receive socket buffer needs to grow, increase + * the high water mark. + */ + if (TSTMP_GEQ(tcp_now, + tp->rfbuf_ts + TCPTV_RCVNOTS_QUANTUM)) { + if (tp->rfbuf_cnt >= TCP_RCVNOTS_BYTELEVEL) { + tcp_sbrcv_reserve(tp, sbrcv, + tcp_autorcvbuf_max, 0); + } + goto out; + } else { + tp->rfbuf_cnt += pktlen; + return; + } + } else if (to->to_tsecr != 0) { + /* + * If the timestamp shows that one RTT has + * completed, we can stop counting the + * bytes. Here we consider increasing + * the socket buffer if the bandwidth measured in + * last rtt, is more than half of sb_hiwat, this will + * help to scale the buffer according to the bandwidth + * on the link. + */ + if (TSTMP_GEQ(to->to_tsecr, tp->rfbuf_ts)) { + if (tp->rfbuf_cnt > (sbrcv->sb_hiwat - + (sbrcv->sb_hiwat >> 1))) { + int32_t rcvbuf_inc, min_incr; + /* + * Increment the receive window by a + * multiple of maximum sized segments. + * This will prevent a connection from + * sending smaller segments on wire if it + * is limited by the receive window. + * + * Set the ideal size based on current + * bandwidth measurements. We set the + * ideal size on receive socket buffer to + * be twice the bandwidth delay product. + */ + rcvbuf_inc = (tp->rfbuf_cnt << 1) + - sbrcv->sb_hiwat; + + /* + * Make the increment equal to 8 segments + * at least + */ + min_incr = tp->t_maxseg << tcp_autorcvbuf_inc_shift; + if (rcvbuf_inc < min_incr) + rcvbuf_inc = min_incr; + + rcvbuf_inc = + (rcvbuf_inc / tp->t_maxseg) * tp->t_maxseg; + tcp_sbrcv_reserve(tp, sbrcv, + sbrcv->sb_hiwat + rcvbuf_inc, + (tp->rfbuf_cnt * 2)); + } + goto out; + } else { + tp->rfbuf_cnt += pktlen; + return; + } + } +out: + /* Restart the measurement */ + tp->rfbuf_ts = 0; + tp->rfbuf_cnt = 0; + return; +} + +/* This function will trim the excess space added to the socket buffer + * to help a slow-reading app. The ideal-size of a socket buffer depends + * on the link bandwidth or it is set by an application and we aim to + * reach that size. + */ +void +tcp_sbrcv_trim(struct tcpcb *tp, struct sockbuf *sbrcv) { + if (tcp_do_autorcvbuf == 1 && sbrcv->sb_idealsize > 0 && + sbrcv->sb_hiwat > sbrcv->sb_idealsize) { + int32_t trim; + /* compute the difference between ideal and current sizes */ + u_int32_t diff = sbrcv->sb_hiwat - sbrcv->sb_idealsize; + + /* Compute the maximum advertised window for + * this connection. + */ + u_int32_t advwin = tp->rcv_adv - tp->rcv_nxt; + + /* How much can we trim the receive socket buffer? + * 1. it can not be trimmed beyond the max rcv win advertised + * 2. if possible, leave 1/16 of bandwidth*delay to + * avoid closing the win completely + */ + u_int32_t leave = max(advwin, (sbrcv->sb_idealsize >> 4)); + + /* Sometimes leave can be zero, in that case leave at least + * a few segments worth of space. + */ + if (leave == 0) + leave = tp->t_maxseg << tcp_autorcvbuf_inc_shift; + + trim = sbrcv->sb_hiwat - (sbrcv->sb_cc + leave); + trim = imin(trim, (int32_t)diff); + + if (trim > 0) + sbreserve(sbrcv, (sbrcv->sb_hiwat - trim)); + } +} + +/* We may need to trim the send socket buffer size for two reasons: + * 1. if the rtt seen on the connection is climbing up, we do not + * want to fill the buffers any more. + * 2. if the congestion win on the socket backed off, there is no need + * to hold more mbufs for that connection than what the cwnd will allow. + */ +void +tcp_sbsnd_trim(struct sockbuf *sbsnd) { + if (tcp_do_autosendbuf == 1 && + ((sbsnd->sb_flags & (SB_AUTOSIZE | SB_TRIM)) == + (SB_AUTOSIZE | SB_TRIM)) && + (sbsnd->sb_idealsize > 0) && + (sbsnd->sb_hiwat > sbsnd->sb_idealsize)) { + u_int32_t trim = 0; + if (sbsnd->sb_cc <= sbsnd->sb_idealsize) { + trim = sbsnd->sb_hiwat - sbsnd->sb_idealsize; + } else { + trim = sbsnd->sb_hiwat - sbsnd->sb_cc; + } + sbreserve(sbsnd, (sbsnd->sb_hiwat - trim)); + } + if (sbsnd->sb_hiwat <= sbsnd->sb_idealsize) + sbsnd->sb_flags &= ~(SB_TRIM); +} + +/* + * If timestamp option was not negotiated on this connection + * and this connection is on the receiving side of a stream + * then we can not measure the delay on the link accurately. + * Instead of enabling automatic receive socket buffer + * resizing, just give more space to the receive socket buffer. + */ +static inline void +tcp_sbrcv_tstmp_check(struct tcpcb *tp) { + struct socket *so = tp->t_inpcb->inp_socket; + u_int32_t newsize = 2 * tcp_recvspace; + struct sockbuf *sbrcv = &so->so_rcv; + + if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP)) != + (TF_REQ_TSTMP | TF_RCVD_TSTMP) && + (sbrcv->sb_flags & SB_AUTOSIZE) != 0) { + tcp_sbrcv_reserve(tp, sbrcv, newsize, 0); + } +} + +/* A receiver will evaluate the flow of packets on a connection + * to see if it can reduce ack traffic. The receiver will start + * stretching acks if all of the following conditions are met: + * 1. tcp_delack_enabled is set to 3 + * 2. If the bytes received in the last 100ms is greater than a threshold + * defined by maxseg_unacked + * 3. If the connection has not been idle for tcp_maxrcvidle period. + * 4. If the connection has seen enough packets to let the slow-start + * finish after connection establishment or after some packet loss. + * + * The receiver will stop stretching acks if there is congestion/reordering + * as indicated by packets on reassembly queue or an ECN. If the delayed-ack + * timer fires while stretching acks, it means that the packet flow has gone + * below the threshold defined by maxseg_unacked and the receiver will stop + * stretching acks. The receiver gets no indication when slow-start is completed + * or when the connection reaches an idle state. That is why we use + * tcp_rcvsspktcnt to cover slow-start and tcp_maxrcvidle to identify idle + * state. + */ +static inline int +tcp_stretch_ack_enable(struct tcpcb *tp) +{ + if (!(tp->t_flagsext & (TF_NOSTRETCHACK|TF_DISABLE_STRETCHACK)) && + tp->rcv_by_unackwin >= (maxseg_unacked * tp->t_maxseg) && + TSTMP_GT(tp->rcv_unackwin + tcp_maxrcvidle, tcp_now) && + (!(tp->t_flagsext & TF_RCVUNACK_WAITSS) || + (tp->rcv_waitforss >= tcp_rcvsspktcnt))) { + return(1); + } + + return(0); +} + +/* + * Reset the state related to stretch-ack algorithm. This will make + * the receiver generate an ack every other packet. The receiver + * will start re-evaluating the rate at which packets come to decide + * if it can benefit by lowering the ack traffic. + */ +void +tcp_reset_stretch_ack(struct tcpcb *tp) +{ + tp->t_flags &= ~(TF_STRETCHACK); + tp->rcv_by_unackwin = 0; + tp->rcv_unackwin = tcp_now + tcp_rcvunackwin; + + /* + * When there is packet loss or packet re-ordering or CWR due to + * ECN, the sender's congestion window is reduced. In these states, + * generate an ack for every other packet for some time to allow + * the sender's congestion window to grow. + */ + tp->t_flagsext |= TF_RCVUNACK_WAITSS; + tp->rcv_waitforss = 0; +} + +/* + * The last packet was a retransmission, check if this ack + * indicates that the retransmission was spurious. + * + * If the connection supports timestamps, we could use it to + * detect if the last retransmit was not needed. Otherwise, + * we check if the ACK arrived within RTT/2 window, then it + * was a mistake to do the retransmit in the first place. + * + * This function will return 1 if it is a spurious retransmit, + * 0 otherwise. + */ +int +tcp_detect_bad_rexmt(struct tcpcb *tp, struct tcphdr *th, + struct tcpopt *to, u_int32_t rxtime) +{ + int32_t tdiff, bad_rexmt_win; + bad_rexmt_win = (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); + + /* If the ack has ECN CE bit, then cwnd has to be adjusted */ + if ((tp->ecn_flags & (TE_ECN_ON)) == (TE_ECN_ON) + && (th->th_flags & TH_ECE)) + return (0); + if (TSTMP_SUPPORTED(tp)) { + if (rxtime > 0 && (to->to_flags & TOF_TS) + && to->to_tsecr != 0 + && TSTMP_LT(to->to_tsecr, rxtime)) + return (1); + } else { + if ((tp->t_rxtshift == 1 + || (tp->t_flagsext & TF_SENT_TLPROBE)) + && rxtime > 0) { + tdiff = (int32_t)(tcp_now - rxtime); + if (tdiff < bad_rexmt_win) + return(1); + } + } + return(0); +} + + +/* + * Restore congestion window state if a spurious timeout + * was detected. + */ +static void +tcp_bad_rexmt_restore_state(struct tcpcb *tp, struct tcphdr *th) +{ + if (TSTMP_SUPPORTED(tp)) { + u_int32_t fsize, acked; + fsize = tp->snd_max - th->th_ack; + acked = BYTES_ACKED(th, tp); + + /* + * Implement bad retransmit recovery as + * described in RFC 4015. + */ + tp->snd_ssthresh = tp->snd_ssthresh_prev; + + /* Initialize cwnd to the initial window */ + if (CC_ALGO(tp)->cwnd_init != NULL) + CC_ALGO(tp)->cwnd_init(tp); + + tp->snd_cwnd = fsize + min(acked, tp->snd_cwnd); + + } else { + tp->snd_cwnd = tp->snd_cwnd_prev; + tp->snd_ssthresh = tp->snd_ssthresh_prev; + if (tp->t_flags & TF_WASFRECOVERY) + ENTER_FASTRECOVERY(tp); + } + tp->snd_cwnd = max(tp->snd_cwnd, TCP_CC_CWND_INIT_BYTES); + tp->snd_recover = tp->snd_recover_prev; + tp->snd_nxt = tp->snd_max; + tp->t_rxtshift = 0; + tp->t_rxtstart = 0; + + /* Fix send socket buffer to reflect the change in cwnd */ + tcp_bad_rexmt_fix_sndbuf(tp); + + /* + * This RTT might reflect the extra delay induced + * by the network. Skip using this sample for RTO + * calculation and mark the connection so we can + * recompute RTT when the next eligible sample is + * found. + */ + tp->t_flagsext |= TF_RECOMPUTE_RTT; + tp->t_badrexmt_time = tcp_now; tp->t_rtttime = 0; - tp->ecn_flags |= TE_SENDCWR; - tp->snd_cwnd = tp->snd_ssthresh + - tp->t_maxseg * tcprexmtthresh; } - /* - * TCP input routine, follows pages 65-76 of the - * protocol specification dated September, 1981 very closely. + * If the previous packet was sent in retransmission timer, and it was + * not needed, then restore the congestion window to the state before that + * transmission. + * + * If the last packet was sent in tail loss probe timeout, check if that + * recovered the last packet. If so, that will indicate a real loss and + * the congestion window needs to be lowered. */ -#if INET6 -int -tcp6_input(mp, offp) - struct mbuf **mp; - int *offp; +static void +tcp_bad_rexmt_check(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to) { - register struct mbuf *m = *mp; - struct in6_ifaddr *ia6; + if (tp->t_rxtshift > 0 && + tcp_detect_bad_rexmt(tp, th, to, tp->t_rxtstart)) { + ++tcpstat.tcps_sndrexmitbad; + tcp_bad_rexmt_restore_state(tp, th); + tcp_ccdbg_trace(tp, th, TCP_CC_BAD_REXMT_RECOVERY); + } else if ((tp->t_flagsext & TF_SENT_TLPROBE) + && tp->t_tlphighrxt > 0 + && SEQ_GEQ(th->th_ack, tp->t_tlphighrxt) + && !tcp_detect_bad_rexmt(tp, th, to, tp->t_tlpstart)) { + /* + * The tail loss probe recovered the last packet and + * we need to adjust the congestion window to take + * this loss into account. + */ + ++tcpstat.tcps_tlp_recoverlastpkt; + if (!IN_FASTRECOVERY(tp)) { + tcp_reduce_congestion_window(tp); + EXIT_FASTRECOVERY(tp); + } + tcp_ccdbg_trace(tp, th, TCP_CC_TLP_RECOVER_LASTPACKET); + } - IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), return IPPROTO_DONE); + tp->t_flagsext &= ~(TF_SENT_TLPROBE); + tp->t_tlphighrxt = 0; + tp->t_tlpstart = 0; /* - * draft-itojun-ipv6-tcp-to-anycast - * better place to put this in? + * check if the latest ack was for a segment sent during PMTU + * blackhole detection. If the timestamp on the ack is before + * PMTU blackhole detection, then revert the size of the max + * segment to previous size. */ - ia6 = ip6_getdstifaddr(m); - if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) { - struct ip6_hdr *ip6; + if (tp->t_rxtshift > 0 && (tp->t_flags & TF_BLACKHOLE) && + tp->t_pmtud_start_ts > 0 && TSTMP_SUPPORTED(tp)) { + if ((to->to_flags & TOF_TS) && to->to_tsecr != 0 + && TSTMP_LT(to->to_tsecr, tp->t_pmtud_start_ts)) { + tcp_pmtud_revert_segment_size(tp); + } + } + if (tp->t_pmtud_start_ts > 0) + tp->t_pmtud_start_ts = 0; +} - ip6 = mtod(m, struct ip6_hdr *); - icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR, - (caddr_t)&ip6->ip6_dst - (caddr_t)ip6); - return IPPROTO_DONE; +/* + * Check if early retransmit can be attempted according to RFC 5827. + * + * If packet reordering is detected on a connection, fast recovery will + * be delayed until it is clear that the packet was lost and not reordered. + * But reordering detection is done only when SACK is enabled. + * + * On connections that do not support SACK, there is a limit on the number + * of early retransmits that can be done per minute. This limit is needed + * to make sure that too many packets are not retransmitted when there is + * packet reordering. + */ +static void +tcp_early_rexmt_check (struct tcpcb *tp, struct tcphdr *th) +{ + u_int32_t obytes, snd_off; + int32_t snd_len; + struct socket *so = tp->t_inpcb->inp_socket; + + if (early_rexmt && (SACK_ENABLED(tp) || + tp->t_early_rexmt_count < TCP_EARLY_REXMT_LIMIT) && + SEQ_GT(tp->snd_max, tp->snd_una) && + (tp->t_dupacks == 1 || + (SACK_ENABLED(tp) && + !TAILQ_EMPTY(&tp->snd_holes)))) { + /* + * If there are only a few outstanding + * segments on the connection, we might need + * to lower the retransmit threshold. This + * will allow us to do Early Retransmit as + * described in RFC 5827. + */ + if (SACK_ENABLED(tp) && + !TAILQ_EMPTY(&tp->snd_holes)) { + obytes = (tp->snd_max - tp->snd_fack) + + tp->sackhint.sack_bytes_rexmit; + } else { + obytes = (tp->snd_max - tp->snd_una); + } + + /* + * In order to lower retransmit threshold the + * following two conditions must be met. + * 1. the amount of outstanding data is less + * than 4*SMSS bytes + * 2. there is no unsent data ready for + * transmission or the advertised window + * will limit sending new segments. + */ + snd_off = tp->snd_max - tp->snd_una; + snd_len = min(so->so_snd.sb_cc, tp->snd_wnd) - snd_off; + if (obytes < (tp->t_maxseg << 2) && + snd_len <= 0) { + u_int32_t osegs; + + osegs = obytes / tp->t_maxseg; + if ((osegs * tp->t_maxseg) < obytes) + osegs++; + + /* + * Since the connection might have already + * received some dupacks, we add them to + * to the outstanding segments count to get + * the correct retransmit threshold. + * + * By checking for early retransmit after + * receiving some duplicate acks when SACK + * is supported, the connection will + * enter fast recovery even if multiple + * segments are lost in the same window. + */ + osegs += tp->t_dupacks; + if (osegs < 4) { + tp->t_rexmtthresh = + ((osegs - 1) > 1) ? (osegs - 1) : 1; + tp->t_rexmtthresh = + min(tp->t_rexmtthresh, tcprexmtthresh); + tp->t_rexmtthresh = + max(tp->t_rexmtthresh, tp->t_dupacks); + + if (tp->t_early_rexmt_count == 0) + tp->t_early_rexmt_win = tcp_now; + + if (tp->t_flagsext & TF_SENT_TLPROBE) { + tcpstat.tcps_tlp_recovery++; + tcp_ccdbg_trace(tp, th, + TCP_CC_TLP_RECOVERY); + } else { + tcpstat.tcps_early_rexmt++; + tp->t_early_rexmt_count++; + tcp_ccdbg_trace(tp, th, + TCP_CC_EARLY_RETRANSMIT); + } + } + } } - tcp_input(m, *offp); - return IPPROTO_DONE; + /* + * If we ever sent a TLP probe, the acknowledgement will trigger + * early retransmit because the value of snd_fack will be close + * to snd_max. This will take care of adjustments to the + * congestion window. So we can reset TF_SENT_PROBE flag. + */ + tp->t_flagsext &= ~(TF_SENT_TLPROBE); + tp->t_tlphighrxt = 0; + tp->t_tlpstart = 0; } -#endif void tcp_input(m, off0) @@ -540,11 +1489,10 @@ tcp_input(m, off0) { register struct tcphdr *th; register struct ip *ip = NULL; - register struct ipovly *ipov; register struct inpcb *inp; u_char *optp = NULL; int optlen = 0; - int len, tlen, off; + int tlen, off; int drop_hdrlen; register struct tcpcb *tp = 0; register int thflags; @@ -555,9 +1503,8 @@ tcp_input(m, off0) struct in6_addr laddr6; #endif int dropsocket = 0; - int iss = 0; - int nosock = 0; - u_int32_t tiwin; + int iss = 0, nosock = 0; + u_int32_t tiwin, sack_bytes_acked = 0; struct tcpopt to; /* options in this segment */ struct sockaddr_in *next_hop = NULL; #if TCPDEBUG @@ -566,18 +1513,25 @@ tcp_input(m, off0) struct m_tag *fwd_tag; u_char ip_ecn = IPTOS_ECN_NOTECT; unsigned int ifscope; + uint8_t isconnected, isdisconnected; + struct ifnet *ifp = m->m_pkthdr.rcvif; + int pktf_sw_lro_pkt = (m->m_pkthdr.pkt_flags & PKTF_SW_LRO_PKT) ? 1 : 0; + int nlropkts = (pktf_sw_lro_pkt == 1) ? m->m_pkthdr.lro_npkts : 1; + int turnoff_lro = 0, win; +#if MPTCP + struct mptcb *mp_tp = NULL; + uint16_t mptcp_csum = 0; +#endif /* MPTCP */ + boolean_t cell = IFNET_IS_CELLULAR(ifp); + boolean_t wifi = (!cell && IFNET_IS_WIFI(ifp)); + boolean_t wired = (!wifi && IFNET_IS_WIRED(ifp)); + struct tcp_respond_args tra; + +#define TCP_INC_VAR(stat, npkts) do { \ + stat += npkts; \ +} while (0) - /* - * Record the interface where this segment arrived on; this does not - * affect normal data output (for non-detached TCP) as it provides a - * hint about which route and interface to use for sending in the - * absence of a PCB, when scoped routing (and thus source interface - * selection) are enabled. - */ - if ((m->m_flags & M_PKTHDR) && m->m_pkthdr.rcvif != NULL) - ifscope = m->m_pkthdr.rcvif->if_index; - else - ifscope = IFSCOPE_NONE; + TCP_INC_VAR(tcpstat.tcps_rcvtotal, nlropkts); /* Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain. */ if (!SLIST_EMPTY(&m->m_pkthdr.tags)) { @@ -587,19 +1541,20 @@ tcp_input(m, off0) fwd_tag = NULL; } if (fwd_tag != NULL) { - struct ip_fwd_tag *ipfwd_tag = (struct ip_fwd_tag *)(fwd_tag+1); - + struct ip_fwd_tag *ipfwd_tag = + (struct ip_fwd_tag *)(fwd_tag+1); + next_hop = ipfwd_tag->next_hop; m_tag_delete(m, fwd_tag); } - + #if INET6 struct ip6_hdr *ip6 = NULL; int isipv6; #endif /* INET6 */ int rstreason; /* For badport_bandlim accounting purposes */ struct proc *proc0=current_proc(); - + KERNEL_DEBUG(DBG_FNC_TCP_INPUT | DBG_FUNC_START,0,0,0,0,0); #if INET6 @@ -607,20 +1562,21 @@ tcp_input(m, off0) #endif bzero((char *)&to, sizeof(to)); - tcpstat.tcps_rcvtotal++; - - - #if INET6 if (isipv6) { + /* + * Expect 32-bit aligned data pointer on + * strict-align platforms + */ + MBUF_STRICT_DATA_ALIGNMENT_CHECK_32(m); + /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */ ip6 = mtod(m, struct ip6_hdr *); tlen = sizeof(*ip6) + ntohs(ip6->ip6_plen) - off0; - if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) { - tcpstat.tcps_rcvbadsum++; + th = (struct tcphdr *)(void *)((caddr_t)ip6 + off0); + + if (tcp_input_checksum(AF_INET6, m, th, off0, tlen)) goto dropnosock; - } - th = (struct tcphdr *)((caddr_t)ip6 + off0); KERNEL_DEBUG(DBG_LAYER_BEG, ((th->th_dport << 16) | th->th_sport), (((ip6->ip6_src.s6_addr16[0]) << 16) | (ip6->ip6_dst.s6_addr16[0])), @@ -635,8 +1591,14 @@ tcp_input(m, off0) */ if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) { /* XXX stat */ + IF_TCP_STATINC(ifp, unspecv6); goto dropnosock; } + DTRACE_TCP5(receive, struct mbuf *, m, struct inpcb *, NULL, + struct ip6_hdr *, ip6, struct tcpcb *, NULL, + struct tcphdr *, th); + + ip_ecn = (ntohl(ip6->ip6_flow) >> 20) & IPTOS_ECN_MASK; } else #endif /* INET6 */ { @@ -647,9 +1609,6 @@ tcp_input(m, off0) if (off0 > sizeof (struct ip)) { ip_stripoptions(m, (struct mbuf *)0); off0 = sizeof(struct ip); - if (m->m_pkthdr.csum_flags & CSUM_TCP_SUM16) - m->m_pkthdr.csum_flags = 0; /* invalidate hwcksuming */ - } if (m->m_len < sizeof (struct tcpiphdr)) { if ((m = m_pullup(m, sizeof (struct tcpiphdr))) == 0) { @@ -657,80 +1616,30 @@ tcp_input(m, off0) return; } } - ip = mtod(m, struct ip *); - ipov = (struct ipovly *)ip; - th = (struct tcphdr *)((caddr_t)ip + off0); - tlen = ip->ip_len; - - KERNEL_DEBUG(DBG_LAYER_BEG, ((th->th_dport << 16) | th->th_sport), - (((ip->ip_src.s_addr & 0xffff) << 16) | (ip->ip_dst.s_addr & 0xffff)), - th->th_seq, th->th_ack, th->th_win); - - if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { - if (m->m_pkthdr.csum_flags & CSUM_TCP_SUM16) { - u_short pseudo; - char b[9]; - *(uint32_t*)&b[0] = *(uint32_t*)&ipov->ih_x1[0]; - *(uint32_t*)&b[4] = *(uint32_t*)&ipov->ih_x1[4]; - *(uint8_t*)&b[8] = *(uint8_t*)&ipov->ih_x1[8]; - - bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); - ipov->ih_len = (u_short)tlen; - -#if BYTE_ORDER != BIG_ENDIAN - HTONS(ipov->ih_len); -#endif - - pseudo = in_cksum(m, sizeof (struct ip)); - - *(uint32_t*)&ipov->ih_x1[0] = *(uint32_t*)&b[0]; - *(uint32_t*)&ipov->ih_x1[4] = *(uint32_t*)&b[4]; - *(uint8_t*)&ipov->ih_x1[8] = *(uint8_t*)&b[8]; - - th->th_sum = in_addword(pseudo, (m->m_pkthdr.csum_data & 0xFFFF)); - } else { - if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) - th->th_sum = m->m_pkthdr.csum_data; - else - th->th_sum = in_pseudo(ip->ip_src.s_addr, - ip->ip_dst.s_addr, htonl(m->m_pkthdr.csum_data + - ip->ip_len + IPPROTO_TCP)); - } - th->th_sum ^= 0xffff; - } else { - char b[9]; - /* - * Checksum extended TCP header and data. - */ - *(uint32_t*)&b[0] = *(uint32_t*)&ipov->ih_x1[0]; - *(uint32_t*)&b[4] = *(uint32_t*)&ipov->ih_x1[4]; - *(uint8_t*)&b[8] = *(uint8_t*)&ipov->ih_x1[8]; - - len = sizeof (struct ip) + tlen; - bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); - ipov->ih_len = (u_short)tlen; -#if BYTE_ORDER != BIG_ENDIAN - HTONS(ipov->ih_len); -#endif + /* Expect 32-bit aligned data pointer on strict-align platforms */ + MBUF_STRICT_DATA_ALIGNMENT_CHECK_32(m); - th->th_sum = in_cksum(m, len); - - *(uint32_t*)&ipov->ih_x1[0] = *(uint32_t*)&b[0]; - *(uint32_t*)&ipov->ih_x1[4] = *(uint32_t*)&b[4]; - *(uint8_t*)&ipov->ih_x1[8] = *(uint8_t*)&b[8]; + ip = mtod(m, struct ip *); + th = (struct tcphdr *)(void *)((caddr_t)ip + off0); + tlen = ip->ip_len; - tcp_in_cksum_stats(len); - } - if (th->th_sum) { - tcpstat.tcps_rcvbadsum++; + if (tcp_input_checksum(AF_INET, m, th, off0, tlen)) goto dropnosock; - } + #if INET6 /* Re-initialization for later version check */ ip->ip_v = IPVERSION; #endif ip_ecn = (ip->ip_tos & IPTOS_ECN_MASK); + + DTRACE_TCP5(receive, struct mbuf *, m, struct inpcb *, NULL, + struct ip *, ip, struct tcpcb *, NULL, struct tcphdr *, th); + + KERNEL_DEBUG(DBG_LAYER_BEG, ((th->th_dport << 16) | th->th_sport), + (((ip->ip_src.s_addr & 0xffff) << 16) | (ip->ip_dst.s_addr & 0xffff)), + th->th_seq, th->th_ack, th->th_win); + } /* @@ -740,6 +1649,7 @@ tcp_input(m, off0) off = th->th_off << 2; if (off < sizeof (struct tcphdr) || off > tlen) { tcpstat.tcps_rcvbadoff++; + IF_TCP_STATINC(ifp, badformat); goto dropnosock; } tlen -= off; /* tlen is used instead of ti->ti_len */ @@ -748,7 +1658,7 @@ tcp_input(m, off0) if (isipv6) { IP6_EXTHDR_CHECK(m, off0, off, return); ip6 = mtod(m, struct ip6_hdr *); - th = (struct tcphdr *)((caddr_t)ip6 + off0); + th = (struct tcphdr *)(void *)((caddr_t)ip6 + off0); } else #endif /* INET6 */ { @@ -758,8 +1668,7 @@ tcp_input(m, off0) return; } ip = mtod(m, struct ip *); - ipov = (struct ipovly *)ip; - th = (struct tcphdr *)((caddr_t)ip + off0); + th = (struct tcphdr *)(void *)((caddr_t)ip + off0); } } optlen = off - sizeof (struct tcphdr); @@ -774,11 +1683,11 @@ tcp_input(m, off0) if ((optlen == TCPOLEN_TSTAMP_APPA || (optlen > TCPOLEN_TSTAMP_APPA && optp[TCPOLEN_TSTAMP_APPA] == TCPOPT_EOL)) && - *(u_int32_t *)optp == htonl(TCPOPT_TSTAMP_HDR) && + *(u_int32_t *)(void *)optp == htonl(TCPOPT_TSTAMP_HDR) && (th->th_flags & TH_SYN) == 0) { to.to_flags |= TOF_TS; - to.to_tsval = ntohl(*(u_int32_t *)(optp + 4)); - to.to_tsecr = ntohl(*(u_int32_t *)(optp + 8)); + to.to_tsval = ntohl(*(u_int32_t *)(void *)(optp + 4)); + to.to_tsecr = ntohl(*(u_int32_t *)(void *)(optp + 8)); optp = NULL; /* we've parsed the options */ } } @@ -792,19 +1701,10 @@ tcp_input(m, off0) * * This is a violation of the TCP specification. */ - if (drop_synfin && (thflags & (TH_SYN|TH_FIN)) == (TH_SYN|TH_FIN)) + if (drop_synfin && (thflags & (TH_SYN|TH_FIN)) == (TH_SYN|TH_FIN)) { + IF_TCP_STATINC(ifp, synfin); goto dropnosock; -#endif - - /* - * Convert TCP protocol specific fields to host format. - */ - -#if BYTE_ORDER != BIG_ENDIAN - NTOHL(th->th_seq); - NTOHL(th->th_ack); - NTOHS(th->th_win); - NTOHS(th->th_urp); + } #endif /* @@ -816,11 +1716,43 @@ tcp_input(m, off0) * parameters to be unchanged. */ drop_hdrlen = off0 + off; + + /* Since this is an entry point for input processing of tcp packets, we + * can update the tcp clock here. + */ + calculate_tcp_clock(); + + /* + * Record the interface where this segment arrived on; this does not + * affect normal data output (for non-detached TCP) as it provides a + * hint about which route and interface to use for sending in the + * absence of a PCB, when scoped routing (and thus source interface + * selection) are enabled. + */ + if ((m->m_pkthdr.pkt_flags & PKTF_LOOP) || m->m_pkthdr.rcvif == NULL) + ifscope = IFSCOPE_NONE; + else + ifscope = m->m_pkthdr.rcvif->if_index; + + /* + * Convert TCP protocol specific fields to host format. + */ + +#if BYTE_ORDER != BIG_ENDIAN + NTOHL(th->th_seq); + NTOHL(th->th_ack); + NTOHS(th->th_win); + NTOHS(th->th_urp); +#endif /* * Locate pcb for segment. */ findpcb: + + isconnected = FALSE; + isdisconnected = FALSE; + #if IPFIREWALL_FORWARD if (next_hop != NULL #if INET6 @@ -870,27 +1802,25 @@ findpcb: * the segment arrived on. */ if (inp != NULL && (inp->inp_flags & INP_BOUND_IF)) - ifscope = inp->inp_boundif; -#if IPSEC - if (ipsec_bypass == 0) { + ifscope = inp->inp_boundifp->if_index; +#if NECP + if (inp != NULL && ( #if INET6 - if (isipv6) { - if (inp != NULL && ipsec6_in_reject_so(m, inp->inp_socket)) { - IPSEC_STAT_INCREMENT(ipsec6stat.in_polvio); - if (in_pcb_checkstate(inp, WNT_RELEASE, 0) == WNT_STOPUSING) - inp = NULL; // pretend we didn't find it - goto dropnosock; - } - } else -#endif /* INET6 */ - if (inp != NULL && ipsec4_in_reject_so(m, inp->inp_socket)) { - IPSEC_STAT_INCREMENT(ipsecstat.in_polvio); - if (in_pcb_checkstate(inp, WNT_RELEASE, 0) == WNT_STOPUSING) - inp = NULL; // pretend we didn't find it - goto dropnosock; - } + isipv6 ? !necp_socket_is_allowed_to_send_recv_v6(inp, + th->th_dport, th->th_sport, &ip6->ip6_dst, + &ip6->ip6_src, ifp, NULL) : +#endif + !necp_socket_is_allowed_to_send_recv_v4(inp, th->th_dport, + th->th_sport, &ip->ip_dst, &ip->ip_src, + ifp, NULL))) { + if (in_pcb_checkstate(inp, WNT_RELEASE, 0) + == WNT_STOPUSING) { + inp = NULL; /* pretend we didn't find it */ + } + IF_TCP_STATINC(ifp, badformatipsec); + goto dropnosock; } -#endif /*IPSEC*/ +#endif /* NECP */ /* * If the state is CLOSED (i.e., TCB does not exist) then @@ -932,7 +1862,8 @@ findpcb: ntohs(th->th_sport), thflags); break; case 3: - if ((thflags & TH_SYN) && + case 4: + if ((thflags & TH_SYN) && !(thflags & TH_ACK) && !(m->m_flags & (M_BCAST | M_MCAST)) && #if INET6 ((isipv6 && !IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &ip6->ip6_src)) || @@ -953,6 +1884,7 @@ findpcb: } if (blackhole) { if (m->m_pkthdr.rcvif && m->m_pkthdr.rcvif->if_type != IFT_LOOP) + switch (blackhole) { case 1: if (thflags & TH_SYN) @@ -965,6 +1897,7 @@ findpcb: } } rstreason = BANDLIM_RST_CLOSEDPORT; + IF_TCP_STATINC(ifp, noconnnolist); goto dropwithresetnosock; } so = inp->inp_socket; @@ -981,7 +1914,7 @@ findpcb: goto dropnosock; } - tcp_lock(so, 1, (void *)2); + tcp_lock(so, 1, 0); if (in_pcb_checkstate(inp, WNT_RELEASE, 1) == WNT_STOPUSING) { tcp_unlock(so, 1, (void *)2); inp = NULL; // pretend we didn't find it @@ -991,6 +1924,7 @@ findpcb: tp = intotcpcb(inp); if (tp == 0) { rstreason = BANDLIM_RST_CLOSEDPORT; + IF_TCP_STATINC(ifp, noconnlist); goto dropwithreset; } if (tp->t_state == TCPS_CLOSED) @@ -1007,8 +1941,9 @@ findpcb: goto drop; #endif - /* Radar 7377561: Avoid processing packets while closing a listen socket */ - if (tp->t_state == TCPS_LISTEN && (so->so_options & SO_ACCEPTCONN) == 0) + /* Avoid processing packets while closing a listen socket */ + if (tp->t_state == TCPS_LISTEN && + (so->so_options & SO_ACCEPTCONN) == 0) goto drop; if (so->so_options & (SO_DEBUG|SO_ACCEPTCONN)) { @@ -1033,11 +1968,20 @@ findpcb: #if INET6 struct inpcb *oinp = sotoinpcb(so); #endif /* INET6 */ - unsigned int head_ifscope; + struct ifnet *head_ifscope; + unsigned int head_nocell, head_recvanyif, + head_noexpensive, head_awdl_unrestricted; /* Get listener's bound-to-interface, if any */ head_ifscope = (inp->inp_flags & INP_BOUND_IF) ? - inp->inp_boundif : IFSCOPE_NONE; + inp->inp_boundifp : NULL; + /* Get listener's no-cellular information, if any */ + head_nocell = INP_NO_CELLULAR(inp); + /* Get listener's recv-any-interface, if any */ + head_recvanyif = (inp->inp_flags & INP_RECV_ANYIF); + /* Get listener's no-expensive information, if any */ + head_noexpensive = INP_NO_EXPENSIVE(inp); + head_awdl_unrestricted = INP_AWDL_UNRESTRICTED(inp); /* * If the state is LISTEN then ignore segment if it contains an RST. @@ -1046,6 +1990,8 @@ findpcb: * If it is from this socket, drop it, it must be forged. */ if ((thflags & (TH_RST|TH_ACK|TH_SYN)) != TH_SYN) { + IF_TCP_STATINC(ifp, listbadsyn); + if (thflags & TH_RST) { goto drop; } @@ -1109,32 +2055,20 @@ findpcb: * for the exchange. * * If we do not forbid deprecated addresses, we accept - * the SYN packet. RFC2462 does not suggest dropping - * SYN in this case. - * If we decipher RFC2462 5.5.4, it says like this: - * 1. use of deprecated addr with existing - * communication is okay - "SHOULD continue to be - * used" - * 2. use of it with new communication: - * (2a) "SHOULD NOT be used if alternate address - * with sufficient scope is available" - * (2b) nothing mentioned otherwise. - * Here we fall into (2b) case as we have no choice in - * our source address selection - we must obey the peer. - * - * The wording in RFC2462 is confusing, and there are - * multiple description text for deprecated address - * handling - worse, they are not exactly the same. - * I believe 5.5.4 is the best one, so we follow 5.5.4. + * the SYN packet. RFC 4862 forbids dropping SYN in + * this case. */ if (isipv6 && !ip6_use_deprecated) { - struct in6_ifaddr *ia6; - - if ((ia6 = ip6_getdstifaddr(m)) && - (ia6->ia6_flags & IN6_IFF_DEPRECATED)) { - tp = NULL; - rstreason = BANDLIM_RST_OPENPORT; - goto dropwithreset; + uint32_t ia6_flags; + + if (ip6_getdstifaddr_info(m, NULL, + &ia6_flags) == 0) { + if (ia6_flags & IN6_IFF_DEPRECATED) { + tp = NULL; + rstreason = BANDLIM_RST_OPENPORT; + IF_TCP_STATINC(ifp, deprecate6); + goto dropwithreset; + } } } #endif @@ -1199,18 +2133,32 @@ findpcb: dropsocket++; /* * Inherit INP_BOUND_IF from listener; testing if - * head_ifscope is non-zero is sufficient, since it + * head_ifscope is non-NULL is sufficient, since it * can only be set to a non-zero value earlier if * the listener has such a flag set. */ -#if INET6 - if (head_ifscope != IFSCOPE_NONE && !isipv6) { -#else - if (head_ifscope != IFSCOPE_NONE) { -#endif /* INET6 */ + if (head_ifscope != NULL) { inp->inp_flags |= INP_BOUND_IF; - inp->inp_boundif = head_ifscope; + inp->inp_boundifp = head_ifscope; + } else { + inp->inp_flags &= ~INP_BOUND_IF; } + /* + * Inherit restrictions from listener. + */ + if (head_nocell) + inp_set_nocellular(inp); + if (head_noexpensive) + inp_set_noexpensive(inp); + if (head_awdl_unrestricted) + inp_set_awdl_unrestricted(inp); + /* + * Inherit {IN,IN6}_RECV_ANYIF from listener. + */ + if (head_recvanyif) + inp->inp_flags |= INP_RECV_ANYIF; + else + inp->inp_flags &= ~INP_RECV_ANYIF; #if INET6 if (isipv6) inp->in6p_laddr = ip6->ip6_dst; @@ -1233,7 +2181,7 @@ findpcb: inp->in6p_laddr = in6addr_any; else #endif /* INET6 */ - inp->inp_laddr.s_addr = INADDR_ANY; + inp->inp_laddr.s_addr = INADDR_ANY; inp->inp_lport = 0; tcp_lock(oso, 0, 0); /* release ref on parent */ tcp_unlock(oso, 1, 0); @@ -1262,7 +2210,7 @@ findpcb: M_NOWAIT); } else #endif /* INET6 */ - inp->inp_options = ip_srcroute(); + inp->inp_options = ip_srcroute(); tcp_lock(oso, 0, 0); #if IPSEC /* copy old policy into new socket's */ @@ -1277,165 +2225,144 @@ findpcb: } #endif /* inherit states from the listener */ + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_LISTEN); tp->t_state = TCPS_LISTEN; tp->t_flags |= tp0->t_flags & (TF_NOPUSH|TF_NOOPT|TF_NODELAY); + tp->t_flagsext |= (tp0->t_flagsext & (TF_RXTFINDROP|TF_NOTIMEWAIT)); tp->t_keepinit = tp0->t_keepinit; + tp->t_keepcnt = tp0->t_keepcnt; + tp->t_keepintvl = tp0->t_keepintvl; + tp->t_adaptive_wtimo = tp0->t_adaptive_wtimo; + tp->t_adaptive_rtimo = tp0->t_adaptive_rtimo; tp->t_inpcb->inp_ip_ttl = tp0->t_inpcb->inp_ip_ttl; + if ((so->so_flags & SOF_NOTSENT_LOWAT) != 0) + tp->t_notsent_lowat = tp0->t_notsent_lowat; /* now drop the reference on the listener */ tcp_unlock(oso, 1, 0); - /* Compute proper scaling value from buffer space */ - if (inp->inp_pcbinfo->ipi_count < tcp_sockthreshold) { - tp->request_r_scale = max(tcp_win_scale, tp->request_r_scale); - so->so_rcv.sb_hiwat = imin(TCP_MAXWIN << tp->request_r_scale, (sb_max / (MSIZE+MCLBYTES)) * MCLBYTES); - } - else { - while (tp->request_r_scale < TCP_MAX_WINSHIFT && - TCP_MAXWIN << tp->request_r_scale < - so->so_rcv.sb_hiwat) - tp->request_r_scale++; - } + tcp_set_max_rwinscale(tp, so); KERNEL_DEBUG(DBG_FNC_TCP_NEWCONN | DBG_FUNC_END,0,0,0,0,0); } } -#if 1 - lck_mtx_assert(((struct inpcb *)so->so_pcb)->inpcb_mtx, LCK_MTX_ASSERT_OWNED); -#endif - /* - * Radar 3529618 - * This is the second part of the MSS DoS prevention code (after - * minmss on the sending side) and it deals with too many too small - * tcp packets in a too short timeframe (1 second). - * - * For every full second we count the number of received packets - * and bytes. If we get a lot of packets per second for this connection - * (tcp_minmssoverload) we take a closer look at it and compute the - * average packet size for the past second. If that is less than - * tcp_minmss we get too many packets with very small payload which - * is not good and burdens our system (and every packet generates - * a wakeup to the process connected to our socket). We can reasonable - * expect this to be small packet DoS attack to exhaust our CPU - * cycles. - * - * Care has to be taken for the minimum packet overload value. This - * value defines the minimum number of packets per second before we - * start to worry. This must not be too low to avoid killing for - * example interactive connections with many small packets like - * telnet or SSH. - * - * Setting either tcp_minmssoverload or tcp_minmss to "0" disables - * this check. - * - * Account for packet if payload packet, skip over ACK, etc. - * - * The packet per second count is done all the time and is also used - * by "DELAY_ACK" to detect streaming situations. - * - */ + lck_mtx_assert(&((struct inpcb *)so->so_pcb)->inpcb_mtx, + LCK_MTX_ASSERT_OWNED); + if (tp->t_state == TCPS_ESTABLISHED && tlen > 0) { - if (tp->rcv_reset > tcp_now) { - tp->rcv_pps++; - tp->rcv_byps += tlen + off; - if (tp->rcv_byps > tp->rcv_maxbyps) - tp->rcv_maxbyps = tp->rcv_byps; - /* - * Setting either tcp_minmssoverload or tcp_minmss to "0" disables - * the check. + /* + * Evaluate the rate of arrival of packets to see if the + * receiver can reduce the ack traffic. The algorithm to + * stretch acks will be enabled if the connection meets + * certain criteria defined in tcp_stretch_ack_enable function. */ - if (tcp_minmss && tcp_minmssoverload && tp->rcv_pps > tcp_minmssoverload) { - if ((tp->rcv_byps / tp->rcv_pps) < tcp_minmss) { - char ipstrbuf[MAX_IPv6_STR_LEN]; - printf("too many small tcp packets from " - "%s:%u, av. %ubyte/packet, " - "dropping connection\n", -#if INET6 - isipv6 ? - inet_ntop(AF_INET6, &inp->in6p_faddr, ipstrbuf, - sizeof(ipstrbuf)) : -#endif - inet_ntop(AF_INET, &inp->inp_faddr, ipstrbuf, - sizeof(ipstrbuf)), - inp->inp_fport, - tp->rcv_byps / tp->rcv_pps); - tp = tcp_drop(tp, ECONNRESET); -/* tcpstat.tcps_minmssdrops++; */ - goto drop; - } - } + if ((tp->t_flagsext & TF_RCVUNACK_WAITSS) != 0) { + TCP_INC_VAR(tp->rcv_waitforss, nlropkts); + } + if (tcp_stretch_ack_enable(tp)) { + tp->t_flags |= TF_STRETCHACK; + tp->t_flagsext &= ~(TF_RCVUNACK_WAITSS); + tp->rcv_waitforss = 0; } else { - tp->rcv_reset = tcp_now + TCP_RETRANSHZ; - tp->rcv_pps = 1; - tp->rcv_byps = tlen + off; + tp->t_flags &= ~(TF_STRETCHACK); + } + if (TSTMP_GT(tp->rcv_unackwin, tcp_now)) { + tp->rcv_by_unackwin += (tlen + off); + } else { + tp->rcv_unackwin = tcp_now + tcp_rcvunackwin; + tp->rcv_by_unackwin = tlen + off; } } - -#if TRAFFIC_MGT - if (so->so_traffic_mgt_flags & TRAFFIC_MGT_SO_BACKGROUND) { - tcpstat.tcps_bg_rcvtotal++; - - /* Take snapshots of pkts recv; - * tcpcb should have been initialized to 0 when allocated, - * so if 0 then this is the first time we're doing this - */ - if (!tp->tot_recv_snapshot) { - tp->tot_recv_snapshot = tcpstat.tcps_rcvtotal; - } - if (!tp->bg_recv_snapshot) { - tp->bg_recv_snapshot = tcpstat.tcps_bg_rcvtotal; - } - } -#endif /* TRAFFIC_MGT */ + /* + * Keep track of how many bytes were received in the LRO packet + */ + if ((pktf_sw_lro_pkt) && (nlropkts > 2)) { + tp->t_lropktlen += tlen; + } /* - Explicit Congestion Notification - Flag that we need to send ECT if - + The IP Congestion experienced flag was set. - + Socket is in established state - + We negotiated ECN in the TCP setup - + This isn't a pure ack (tlen > 0) - + The data is in the valid window - - TE_SENDECE will be cleared when we receive a packet with TH_CWR set. + * Explicit Congestion Notification - Flag that we need to send ECT if + * + The IP Congestion experienced flag was set. + * + Socket is in established state + * + We negotiated ECN in the TCP setup + * + This isn't a pure ack (tlen > 0) + * + The data is in the valid window + * + * TE_SENDECE will be cleared when we receive a packet with TH_CWR set. */ if (ip_ecn == IPTOS_ECN_CE && tp->t_state == TCPS_ESTABLISHED && - (tp->ecn_flags & (TE_SETUPSENT | TE_SETUPRECEIVED)) == - (TE_SETUPSENT | TE_SETUPRECEIVED) && tlen > 0 && + ((tp->ecn_flags & (TE_ECN_ON)) == (TE_ECN_ON)) && tlen > 0 && SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { tp->ecn_flags |= TE_SENDECE; } /* - Clear TE_SENDECE if TH_CWR is set. This is harmless, so we don't - bother doing extensive checks for state and whatnot. + * Clear TE_SENDECE if TH_CWR is set. This is harmless, so we don't + * bother doing extensive checks for state and whatnot. */ if ((thflags & TH_CWR) == TH_CWR) { tp->ecn_flags &= ~TE_SENDECE; } - + + /* + * If we received an explicit notification of congestion in + * ip tos ecn bits or by the CWR bit in TCP header flags, reset + * the ack-strteching state. We need to handle ECN notification if + * an ECN setup SYN was sent even once. + */ + if (tp->t_state == TCPS_ESTABLISHED + && (tp->ecn_flags & TE_SETUPSENT) + && (ip_ecn == IPTOS_ECN_CE || (thflags & TH_CWR))) + tcp_reset_stretch_ack(tp); + + /* + * Try to determine if we are receiving a packet after a long time. + * Use our own approximation of idletime to roughly measure remote + * end's idle time. Since slowstart is used after an idle period + * we want to avoid doing LRO if the remote end is not up to date + * on initial window support and starts with 1 or 2 packets as its IW. + */ + if (sw_lro && (tp->t_flagsext & TF_LRO_OFFLOADED) && + ((tcp_now - tp->t_rcvtime) >= (TCP_IDLETIMEOUT(tp)))) { + turnoff_lro = 1; + } + + /* Update rcvtime as a new segment was received on the connection */ + tp->t_rcvtime = tcp_now; + /* * Segment received on connection. * Reset idle time and keep-alive timer. */ - tp->t_rcvtime = 0; if (TCPS_HAVEESTABLISHED(tp->t_state)) - tp->t_timer[TCPT_KEEP] = TCP_KEEPIDLE(tp); + tcp_keepalive_reset(tp); /* * Process options if not in LISTEN state, * else do it below (after getting remote address). */ - if (tp->t_state != TCPS_LISTEN && optp) + if (tp->t_state != TCPS_LISTEN && optp) { tcp_dooptions(tp, optp, optlen, th, &to, ifscope); - - if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { - if (to.to_flags & TOF_SCALE) { - tp->t_flags |= TF_RCVD_SCALE; - tp->requested_s_scale = to.to_requested_s_scale; - tp->snd_wnd = th->th_win << tp->snd_scale; - tiwin = tp->snd_wnd; +#if MPTCP + mptcp_csum = mptcp_input_csum(tp, m, drop_hdrlen); + if (mptcp_csum) { + tp->t_mpflags |= TMPF_SND_MPFAIL; + tp->t_mpflags &= ~TMPF_EMBED_DSN; + mptcp_notify_mpfail(so); + m_freem(m); + tcpstat.tcps_mp_badcsum++; + tcp_check_timer_state(tp); + tcp_unlock(so, 1, 0); + KERNEL_DEBUG(DBG_FNC_TCP_INPUT | + DBG_FUNC_END,0,0,0,0,0); + return; } + mptcp_insert_rmap(tp, m); +#endif /* MPTCP */ + } + if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { if (to.to_flags & TOF_TS) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; @@ -1443,14 +2370,80 @@ findpcb: } if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss, ifscope); - if (tp->sack_enable) { + if (SACK_ENABLED(tp)) { if (!(to.to_flags & TOF_SACK)) - tp->sack_enable = 0; + tp->t_flagsext &= ~(TF_SACK_ENABLE); else tp->t_flags |= TF_SACK_PERMIT; } } +#if TRAFFIC_MGT + /* Compute inter-packet arrival jitter. According to RFC 3550, inter-packet + * arrival jitter is defined as the difference in packet spacing at the + * receiver compared to the sender for a pair of packets. When two packets + * of maximum segment size come one after the other with consecutive + * sequence numbers, we consider them as packets sent together at the + * sender and use them as a pair to compute inter-packet arrival jitter. + * This metric indicates the delay induced by the network components due + * to queuing in edge/access routers. + */ + if (tp->t_state == TCPS_ESTABLISHED && + (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK|TH_ECE|TH_PUSH)) == TH_ACK && + ((tp->t_flags & (TF_NEEDSYN|TF_NEEDFIN)) == 0) && + ((to.to_flags & TOF_TS) == 0 || + TSTMP_GEQ(to.to_tsval, tp->ts_recent)) && + th->th_seq == tp->rcv_nxt && + LIST_EMPTY(&tp->t_segq)) { + int seg_size = tlen; + if (tp->iaj_pktcnt <= IAJ_IGNORE_PKTCNT) { + TCP_INC_VAR(tp->iaj_pktcnt, nlropkts); + } + + if (m->m_pkthdr.pkt_flags & PKTF_SW_LRO_PKT) { + seg_size = m->m_pkthdr.lro_pktlen; + } + if ( tp->iaj_size == 0 || seg_size > tp->iaj_size || + (seg_size == tp->iaj_size && tp->iaj_rcv_ts == 0)) { + /* State related to inter-arrival jitter is uninitialized + * or we are trying to find a good first packet to start + * computing the metric + */ + update_iaj_state(tp, seg_size, 0); + } else { + if (seg_size == tp->iaj_size) { + /* Compute inter-arrival jitter taking this packet + * as the second packet + */ + if (pktf_sw_lro_pkt) + compute_iaj(tp, nlropkts, + m->m_pkthdr.lro_elapsed); + else + compute_iaj(tp, 1, 0); + } + if (seg_size < tp->iaj_size) { + /* There is a smaller packet in the stream. + * Some times the maximum size supported on a path can + * change if there is a new link with smaller MTU. + * The receiver will not know about this change. + * If there are too many packets smaller than iaj_size, + * we try to learn the iaj_size again. + */ + TCP_INC_VAR(tp->iaj_small_pkt, nlropkts); + if (tp->iaj_small_pkt > RESET_IAJ_SIZE_THRESH) { + update_iaj_state(tp, seg_size, 1); + } else { + CLEAR_IAJ_STATE(tp); + } + } else { + update_iaj_state(tp, seg_size, 0); + } + } + } else { + CLEAR_IAJ_STATE(tp); + } +#endif /* TRAFFIC_MGT */ + /* * Header prediction: check for the two common cases * of a uni-directional data xfer. If the packet has @@ -1498,78 +2491,56 @@ findpcb: if (SEQ_GT(th->th_ack, tp->snd_una) && SEQ_LEQ(th->th_ack, tp->snd_max) && tp->snd_cwnd >= tp->snd_ssthresh && - ((!tcp_do_newreno && !tp->sack_enable && - tp->t_dupacks < tcprexmtthresh) || - ((tcp_do_newreno || tp->sack_enable) && - !IN_FASTRECOVERY(tp) && to.to_nsacks == 0 && - TAILQ_EMPTY(&tp->snd_holes)))) { + (!IN_FASTRECOVERY(tp) && + ((!(SACK_ENABLED(tp)) && + tp->t_dupacks < tp->t_rexmtthresh) || + (SACK_ENABLED(tp) && to.to_nsacks == 0 && + TAILQ_EMPTY(&tp->snd_holes))))) { /* * this is a pure ack for outstanding data. */ ++tcpstat.tcps_predack; - /* - * "bad retransmit" recovery - */ - if (tp->t_rxtshift == 1 && - tcp_now < tp->t_badrxtwin) { - ++tcpstat.tcps_sndrexmitbad; - tp->snd_cwnd = tp->snd_cwnd_prev; - tp->snd_ssthresh = - tp->snd_ssthresh_prev; - tp->snd_recover = tp->snd_recover_prev; - if (tp->t_flags & TF_WASFRECOVERY) - ENTER_FASTRECOVERY(tp); - tp->snd_nxt = tp->snd_max; - tp->t_badrxtwin = 0; - } - /* - * Recalculate the transmit timer / rtt. - * - * Some boxes send broken timestamp replies - * during the SYN+ACK phase, ignore - * timestamps of 0 or we could calculate a - * huge RTT and blow up the retransmit timer. - */ - if (((to.to_flags & TOF_TS) != 0) && (to.to_tsecr != 0)) { /* Makes sure we already have a TS */ - if (!tp->t_rttlow || - tp->t_rttlow > tcp_now - to.to_tsecr) - tp->t_rttlow = tcp_now - to.to_tsecr; - tcp_xmit_timer(tp, - tcp_now - to.to_tsecr); - } else if (tp->t_rtttime && - SEQ_GT(th->th_ack, tp->t_rtseq)) { - if (!tp->t_rttlow || - tp->t_rttlow > tcp_now - tp->t_rtttime) - tp->t_rttlow = tcp_now - tp->t_rtttime; - tcp_xmit_timer(tp, tp->t_rtttime); - } - acked = th->th_ack - tp->snd_una; + + tcp_bad_rexmt_check(tp, th, &to), + + /* Recalculate the RTT */ + tcp_compute_rtt(tp, &to, th); + + acked = BYTES_ACKED(th, tp); tcpstat.tcps_rcvackpack++; tcpstat.tcps_rcvackbyte += acked; - /* - * Grow the congestion window, if the - * connection is cwnd bound. - */ - if (tp->snd_cwnd < tp->snd_wnd) { - tp->t_bytes_acked += acked; - if (tp->t_bytes_acked > tp->snd_cwnd) { - tp->t_bytes_acked -= tp->snd_cwnd; - tp->snd_cwnd += tp->t_maxseg; - } - } + + /* Handle an ack that is in sequence during congestion + * avoidance phase. The calculations in this function + * assume that snd_una is not updated yet. + */ + if (CC_ALGO(tp)->congestion_avd != NULL) + CC_ALGO(tp)->congestion_avd(tp, th); + tcp_ccdbg_trace(tp, th, TCP_CC_INSEQ_ACK_RCVD); sbdrop(&so->so_snd, acked); + if (so->so_flags & SOF_ENABLE_MSGS) { + VERIFY(acked <= so->so_msg_state->msg_serial_bytes); + so->so_msg_state->msg_serial_bytes -= acked; + } + tcp_sbsnd_trim(&so->so_snd); + if (SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; tp->snd_una = th->th_ack; + /* * pull snd_wl2 up to prevent seq wrap relative * to th_ack. */ tp->snd_wl2 = th->th_ack; - tp->t_dupacks = 0; + + if (tp->t_dupacks > 0) { + tp->t_dupacks = 0; + tp->t_rexmtthresh = tcprexmtthresh; + } + m_freem(m); - ND6_HINT(tp); /* some progress has been done */ /* * If all outstanding data are acked, stop @@ -1580,16 +2551,24 @@ findpcb: * are ready to send, let tcp_output * decide between more output or persist. */ - if (tp->snd_una == tp->snd_max) + if (tp->snd_una == tp->snd_max) { tp->t_timer[TCPT_REXMT] = 0; - else if (tp->t_timer[TCPT_PERSIST] == 0) - tp->t_timer[TCPT_REXMT] = tp->t_rxtcur; + tp->t_timer[TCPT_PTO] = 0; + } else if (tp->t_timer[TCPT_PERSIST] == 0) { + tp->t_timer[TCPT_REXMT] = + OFFSET_FROM_START(tp, + tp->t_rxtcur); + } + if ((tp->t_flagsext & TF_MEASURESNDBW) != 0 && + tp->t_bwmeas != NULL) + tcp_bwmeas_check(tp); sowwakeup(so); /* has to be done with socket lock held */ if ((so->so_snd.sb_cc) || (tp->t_flags & TF_ACKNOW)) { - tp->t_unacksegs = 0; (void) tcp_output(tp); } + + tcp_check_timer_state(tp); tcp_unlock(so, 1, 0); KERNEL_DEBUG(DBG_FNC_TCP_INPUT | DBG_FUNC_END,0,0,0,0,0); return; @@ -1602,8 +2581,33 @@ findpcb: * with nothing on the reassembly queue and * we have enough buffer space to take it. */ + + /* + * If this is a connection in steady state, start + * coalescing packets belonging to this flow. + */ + if (turnoff_lro) { + tcp_lro_remove_state(tp->t_inpcb->inp_laddr, + tp->t_inpcb->inp_faddr, + tp->t_inpcb->inp_lport, + tp->t_inpcb->inp_fport); + tp->t_flagsext &= ~TF_LRO_OFFLOADED; + tp->t_idleat = tp->rcv_nxt; + } else if (sw_lro && !pktf_sw_lro_pkt && !isipv6 && + (so->so_flags & SOF_USELRO) && + !IFNET_IS_CELLULAR(m->m_pkthdr.rcvif) && + (m->m_pkthdr.rcvif->if_type != IFT_LOOP) && + ((th->th_seq - tp->irs) > + (tp->t_maxseg << lro_start)) && + ((tp->t_idleat == 0) || ((th->th_seq - + tp->t_idleat) > (tp->t_maxseg << lro_start)))) { + tp->t_flagsext |= TF_LRO_OFFLOADED; + tcp_start_coalescing(ip, th, tlen); + tp->t_idleat = 0; + } + /* Clean receiver SACK report if present */ - if (tp->sack_enable && tp->rcv_numsacks) + if (SACK_ENABLED(tp) && tp->rcv_numsacks) tcp_clean_sackreport(tp); ++tcpstat.tcps_preddat; tp->rcv_nxt += tlen; @@ -1617,15 +2621,46 @@ findpcb: * rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; - tcpstat.tcps_rcvpack++; + TCP_INC_VAR(tcpstat.tcps_rcvpack, nlropkts); tcpstat.tcps_rcvbyte += tlen; - ND6_HINT(tp); /* some progress has been done */ + if (nstat_collect) { + if (m->m_pkthdr.pkt_flags & PKTF_SW_LRO_PKT) { + INP_ADD_STAT(inp, cell, wifi, wired, + rxpackets, m->m_pkthdr.lro_npkts); + } else { + INP_ADD_STAT(inp, cell, wifi, wired, + rxpackets, 1); + } + INP_ADD_STAT(inp, cell, wifi, wired,rxbytes, + tlen); + } + + /* + * Calculate the RTT on the receiver only if the + * connection is in streaming mode and the last + * packet was not an end-of-write + */ + if ((tp->t_flags & TF_STRETCHACK) && + !(tp->t_flagsext & TF_STREAMEOW)) + tcp_compute_rtt(tp, &to, th); + + tcp_sbrcv_grow(tp, &so->so_rcv, &to, tlen); + /* * Add data to socket buffer. */ + so_recv_data_stat(so, m, 0); m_adj(m, drop_hdrlen); /* delayed header drop */ - if (sbappendstream(&so->so_rcv, m)) + + /* + * If message delivery (SOF_ENABLE_MSGS) is enabled on + * this socket, deliver the packet received as an + * in-order message with sequence number attached to it. + */ + if (sbappendstream_rcvdemux(so, m, + th->th_seq - (tp->irs + 1), 0)) { sorwakeup(so); + } #if INET6 if (isipv6) { KERNEL_DEBUG(DBG_LAYER_END, ((th->th_dport << 16) | th->th_sport), @@ -1639,14 +2674,20 @@ findpcb: (((ip->ip_src.s_addr & 0xffff) << 16) | (ip->ip_dst.s_addr & 0xffff)), th->th_seq, th->th_ack, th->th_win); } - if (DELAY_ACK(tp)) { - tp->t_flags |= TF_DELACK; - tp->t_unacksegs++; + TCP_INC_VAR(tp->t_unacksegs, nlropkts); + if (DELAY_ACK(tp, th)) { + if ((tp->t_flags & TF_DELACK) == 0) { + tp->t_flags |= TF_DELACK; + tp->t_timer[TCPT_DELACK] = OFFSET_FROM_START(tp, tcp_delack); + } } else { - tp->t_unacksegs = 0; tp->t_flags |= TF_ACKNOW; tcp_output(tp); } + + tcp_adaptive_rwtimo_check(tp, tlen); + + tcp_check_timer_state(tp); tcp_unlock(so, 1, 0); KERNEL_DEBUG(DBG_FNC_TCP_INPUT | DBG_FUNC_END,0,0,0,0,0); return; @@ -1659,13 +2700,9 @@ findpcb: * Receive window is amount of space in rcv queue, * but not less than advertised window. */ -#if 1 - lck_mtx_assert(((struct inpcb *)so->so_pcb)->inpcb_mtx, LCK_MTX_ASSERT_OWNED); -#endif - { int win; - + lck_mtx_assert(&((struct inpcb *)so->so_pcb)->inpcb_mtx, + LCK_MTX_ASSERT_OWNED); win = tcp_sbspace(tp); - if (win < 0) win = 0; else { /* clip rcv window to 4K for modems */ @@ -1673,7 +2710,21 @@ findpcb: win = min(win, slowlink_wsize); } tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); +#if MPTCP + /* + * Ensure that the subflow receive window isn't greater + * than the connection level receive window. + */ + if ((tp->t_mpflags & TMPF_MPTCP_TRUE) && + (mp_tp = tptomptp(tp))) { + MPT_LOCK(mp_tp); + if (tp->rcv_wnd > mp_tp->mpt_rcvwnd) { + tp->rcv_wnd = mp_tp->mpt_rcvwnd; + tcpstat.tcps_mp_reducedwin++; + } + MPT_UNLOCK(mp_tp); } +#endif /* MPTCP */ switch (tp->t_state) { @@ -1692,9 +2743,8 @@ findpcb: register struct sockaddr_in6 *sin6; #endif -#if 1 - lck_mtx_assert(((struct inpcb *)so->so_pcb)->inpcb_mtx, LCK_MTX_ASSERT_OWNED); -#endif + lck_mtx_assert(&((struct inpcb *)so->so_pcb)->inpcb_mtx, + LCK_MTX_ASSERT_OWNED); #if INET6 if (isipv6) { MALLOC(sin6, struct sockaddr_in6 *, sizeof *sin6, @@ -1719,9 +2769,7 @@ findpcb: } else #endif { -#if 0 - lck_mtx_assert(((struct inpcb *)so->so_pcb)->inpcb_mtx, LCK_MTX_ASSERT_OWNED); -#endif + lck_mtx_assert(&((struct inpcb *)so->so_pcb)->inpcb_mtx, LCK_MTX_ASSERT_OWNED); MALLOC(sin, struct sockaddr_in *, sizeof *sin, M_SONAME, M_NOWAIT); if (sin == NULL) @@ -1734,7 +2782,8 @@ findpcb: laddr = inp->inp_laddr; if (inp->inp_laddr.s_addr == INADDR_ANY) inp->inp_laddr = ip->ip_dst; - if (in_pcbconnect(inp, (struct sockaddr *)sin, proc0)) { + if (in_pcbconnect(inp, (struct sockaddr *)sin, proc0, + IFSCOPE_NONE, NULL)) { inp->inp_laddr = laddr; FREE(sin, M_SONAME); goto drop; @@ -1744,9 +2793,9 @@ findpcb: tcp_dooptions(tp, optp, optlen, th, &to, ifscope); - if (tp->sack_enable) { + if (SACK_ENABLED(tp)) { if (!(to.to_flags & TOF_SACK)) - tp->sack_enable = 0; + tp->t_flagsext &= ~(TF_SACK_ENABLE); else tp->t_flags |= TF_SACK_PERMIT; } @@ -1768,14 +2817,33 @@ findpcb: tp->snd_wnd = tiwin; /* initial send-window */ tp->t_flags |= TF_ACKNOW; tp->t_unacksegs = 0; + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_SYN_RECEIVED); tp->t_state = TCPS_SYN_RECEIVED; - tp->t_timer[TCPT_KEEP] = tp->t_keepinit ? tp->t_keepinit : tcp_keepinit; + tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, + TCP_CONN_KEEPINIT(tp)); dropsocket = 0; /* committed to socket */ + + if (inp->inp_flowhash == 0) + inp->inp_flowhash = inp_calc_flowhash(inp); +#if INET6 + /* update flowinfo - RFC 6437 */ + if (inp->inp_flow == 0 && + inp->in6p_flags & IN6P_AUTOFLOWLABEL) { + inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; + inp->inp_flow |= + (htonl(inp->inp_flowhash) & IPV6_FLOWLABEL_MASK); + } +#endif /* INET6 */ + + /* reset the incomp processing flag */ + so->so_flags &= ~(SOF_INCOMP_INPROGRESS); tcpstat.tcps_accepts++; if ((thflags & (TH_ECE | TH_CWR)) == (TH_ECE | TH_CWR)) { /* ECN-setup SYN */ tp->ecn_flags |= (TE_SETUPRECEIVED | TE_SENDIPECT); } + #if CONFIG_IFEF_NOWINDOWSCALE if (tcp_obey_ifef_nowindowscale && m->m_pkthdr.rcvif != NULL && (m->m_pkthdr.rcvif->if_eflags & IFEF_NOWINDOWSCALE)) { @@ -1795,8 +2863,19 @@ findpcb: (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { rstreason = BANDLIM_RST_OPENPORT; + IF_TCP_STATINC(ifp, ooopacket); goto dropwithreset; } + + /* + * In SYN_RECEIVED state, if we recv some SYNS with + * window scale and others without, window scaling should + * be disabled. Otherwise the window advertised will be + * lower if we assume scaling and the other end does not. + */ + if ((thflags & TH_SYN) && + !(to.to_flags & TOF_SCALE)) + tp->t_flags &= ~TF_RCVD_SCALE; break; /* @@ -1816,10 +2895,22 @@ findpcb: (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { rstreason = BANDLIM_UNLIMITED; + IF_TCP_STATINC(ifp, ooopacket); goto dropwithreset; } if (thflags & TH_RST) { if ((thflags & TH_ACK) != 0) { +#if MPTCP + if ((so->so_flags & SOF_MPTCP_FASTJOIN) && + SEQ_GT(th->th_ack, tp->iss+1)) { + so->so_flags &= ~SOF_MPTCP_FASTJOIN; + /* ignore the RST and retransmit SYN */ + goto drop; + } +#endif /* MPTCP */ + soevent(so, + (SO_FILT_HINT_LOCKED | + SO_FILT_HINT_CONNRESET)); tp = tcp_drop(tp, ECONNREFUSED); postevent(so, 0, EV_RESET); } @@ -1837,6 +2928,7 @@ findpcb: if ((thflags & (TH_ECE | TH_CWR)) == (TH_ECE)) { /* ECN-setup SYN-ACK */ tp->ecn_flags |= TE_SETUPRECEIVED; + tcpstat.tcps_ecn_setup++; } else { /* non-ECN-setup SYN-ACK */ @@ -1854,19 +2946,21 @@ findpcb: tp->snd_scale = tp->requested_s_scale; tp->rcv_scale = tp->request_r_scale; } - tp->rcv_adv += tp->rcv_wnd; + tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); tp->snd_una++; /* SYN is acked */ /* * If there's data, delay ACK; if there's also a FIN * ACKNOW will be turned on later. */ - if (DELAY_ACK(tp) && tlen != 0) { - tp->t_flags |= TF_DELACK; - tp->t_unacksegs++; + TCP_INC_VAR(tp->t_unacksegs, nlropkts); + if (DELAY_ACK(tp, th) && tlen != 0 ) { + if ((tp->t_flags & TF_DELACK) == 0) { + tp->t_flags |= TF_DELACK; + tp->t_timer[TCPT_DELACK] = OFFSET_FROM_START(tp, tcp_delack); + } } else { tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; } /* * Received in SYN_SENT[*] state. @@ -1874,30 +2968,52 @@ findpcb: * SYN_SENT --> ESTABLISHED * SYN_SENT* --> FIN_WAIT_1 */ - tp->t_starttime = 0; + tp->t_starttime = tcp_now; + tcp_sbrcv_tstmp_check(tp); if (tp->t_flags & TF_NEEDFIN) { + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_FIN_WAIT_1); tp->t_state = TCPS_FIN_WAIT_1; tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_ESTABLISHED); tp->t_state = TCPS_ESTABLISHED; - tp->t_timer[TCPT_KEEP] = TCP_KEEPIDLE(tp); + tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, + TCP_CONN_KEEPIDLE(tp)); + if (nstat_collect) + nstat_route_connect_success(tp->t_inpcb->inp_route.ro_rt); } - /* soisconnected may lead to socket_unlock in case of upcalls, - * make sure this is done when everything is setup. +#if MPTCP + /* + * Do not send the connect notification for additional + * subflows until ACK for 3-way handshake arrives. */ - soisconnected(so); + if ((!(tp->t_mpflags & TMPF_MPTCP_TRUE)) && + (tp->t_mpflags & TMPF_SENT_JOIN)) { + isconnected = FALSE; + /* Start data xmit if fastjoin */ + if (mptcp_fastjoin && (so->so_flags & SOF_MPTCP_FASTJOIN)) { + soevent(so, (SO_FILT_HINT_LOCKED | + SO_FILT_HINT_MPFASTJ)); + } + } else +#endif /* MPTCP */ + isconnected = TRUE; } else { - /* - * Received initial SYN in SYN-SENT[*] state => simul- - * taneous open. If segment contains CC option and there is - * a cached CC, apply TAO test; if it succeeds, connection is - * half-synchronized. Otherwise, do 3-way handshake: - * SYN-SENT -> SYN-RECEIVED - * SYN-SENT* -> SYN-RECEIVED* - */ + /* + * Received initial SYN in SYN-SENT[*] state => simul- + * taneous open. If segment contains CC option and there is + * a cached CC, apply TAO test; if it succeeds, connection is + * half-synchronized. Otherwise, do 3-way handshake: + * SYN-SENT -> SYN-RECEIVED + * SYN-SENT* -> SYN-RECEIVED* + */ tp->t_flags |= TF_ACKNOW; tp->t_timer[TCPT_REXMT] = 0; + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_SYN_RECEIVED); tp->t_state = TCPS_SYN_RECEIVED; } @@ -2023,10 +3139,12 @@ trimthenstep6: if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || (tp->rcv_wnd == 0 && - ((tp->last_ack_sent == th->th_seq) || ((tp->last_ack_sent -1) == th->th_seq)))) { + ((tp->last_ack_sent == th->th_seq) || + ((tp->last_ack_sent -1) == th->th_seq)))) { switch (tp->t_state) { case TCPS_SYN_RECEIVED: + IF_TCP_STATINC(ifp, rstinsynrcv); so->so_error = ECONNREFUSED; goto close; @@ -2044,7 +3162,10 @@ trimthenstep6: so->so_error = ECONNRESET; close: postevent(so, 0, EV_RESET); - tp->t_state = TCPS_CLOSED; + soevent(so, + (SO_FILT_HINT_LOCKED | + SO_FILT_HINT_CONNRESET)); + tcpstat.tcps_drops++; tp = tcp_close(tp); break; @@ -2061,10 +3182,6 @@ trimthenstep6: goto drop; } -#if 0 - lck_mtx_assert(((struct inpcb *)so->so_pcb)->inpcb_mtx, LCK_MTX_ASSERT_OWNED); -#endif - /* * RFC 1323 PAWS: If we have a timestamp reply on this segment * and it's less than ts_recent, drop it. @@ -2090,6 +3207,15 @@ trimthenstep6: tcpstat.tcps_rcvduppack++; tcpstat.tcps_rcvdupbyte += tlen; tcpstat.tcps_pawsdrop++; + if (nstat_collect) { + nstat_route_rx(tp->t_inpcb->inp_route.ro_rt, + 1, tlen, NSTAT_RX_FLAG_DUPLICATE); + INP_ADD_STAT(inp, cell, wifi, wired, + rxpackets, 1); + INP_ADD_STAT(inp, cell, wifi, wired, + rxbytes, tlen); + tp->t_stat.rxduplicatebytes += tlen; + } if (tlen) goto dropafterack; goto drop; @@ -2105,6 +3231,7 @@ trimthenstep6: */ if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) { rstreason = BANDLIM_RST_OPENPORT; + IF_TCP_STATINC(ifp, dospacket); goto dropwithreset; } @@ -2136,14 +3263,25 @@ trimthenstep6: * But keep on processing for RST or ACK. */ tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; + if (todrop == 1) { + /* This could be a keepalive */ + soevent(so, SO_FILT_HINT_LOCKED | + SO_FILT_HINT_KEEPALIVE); + } todrop = tlen; tcpstat.tcps_rcvduppack++; - tcpstat.tcps_rcvdupbyte += todrop; + tcpstat.tcps_rcvdupbyte += todrop; } else { tcpstat.tcps_rcvpartduppack++; tcpstat.tcps_rcvpartdupbyte += todrop; } + if (nstat_collect) { + nstat_route_rx(tp->t_inpcb->inp_route.ro_rt, 1, + todrop, NSTAT_RX_FLAG_DUPLICATE); + INP_ADD_STAT(inp, cell, wifi, wired, rxpackets, 1); + INP_ADD_STAT(inp, cell, wifi, wired, rxbytes, todrop); + tp->t_stat.rxduplicatebytes += todrop; + } drop_hdrlen += todrop; /* drop from the top afterwards */ th->th_seq += todrop; tlen -= todrop; @@ -2156,14 +3294,19 @@ trimthenstep6: } /* - * If new data are received on a connection after the - * user processes are gone, then RST the other end. + * If new data are received on a connection after the user processes + * are gone, then RST the other end. Note that an MPTCP subflow socket + * would have SS_NOFDREF set by default, so check to make sure that + * we test for SOF_MP_SUBFLOW socket flag (which would be cleared when + * the socket is closed.) */ - if ((so->so_state & SS_NOFDREF) && + if (!(so->so_flags & SOF_MP_SUBFLOW) && + (so->so_state & SS_NOFDREF) && tp->t_state > TCPS_CLOSE_WAIT && tlen) { tp = tcp_close(tp); tcpstat.tcps_rcvafterclose++; rstreason = BANDLIM_UNLIMITED; + IF_TCP_STATINC(ifp, cleanup); goto dropwithreset; } @@ -2199,7 +3342,6 @@ trimthenstep6: */ if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; tcpstat.tcps_rcvwinprobe++; } else goto dropafterack; @@ -2244,6 +3386,7 @@ trimthenstep6: tp = tcp_drop(tp, ECONNRESET); rstreason = BANDLIM_UNLIMITED; postevent(so, 0, EV_RESET); + IF_TCP_STATINC(ifp, synwindow); goto dropwithreset; } @@ -2265,6 +3408,7 @@ trimthenstep6: /* * Ack processing. */ + switch (tp->t_state) { /* @@ -2281,19 +3425,29 @@ trimthenstep6: (TF_RCVD_SCALE|TF_REQ_SCALE)) { tp->snd_scale = tp->requested_s_scale; tp->rcv_scale = tp->request_r_scale; + tp->snd_wnd = th->th_win << tp->snd_scale; + tiwin = tp->snd_wnd; } /* * Make transitions: * SYN-RECEIVED -> ESTABLISHED * SYN-RECEIVED* -> FIN-WAIT-1 */ - tp->t_starttime = 0; + tp->t_starttime = tcp_now; + tcp_sbrcv_tstmp_check(tp); if (tp->t_flags & TF_NEEDFIN) { + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_FIN_WAIT_1); tp->t_state = TCPS_FIN_WAIT_1; tp->t_flags &= ~TF_NEEDFIN; } else { + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_ESTABLISHED); tp->t_state = TCPS_ESTABLISHED; - tp->t_timer[TCPT_KEEP] = TCP_KEEPIDLE(tp); + tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, + TCP_CONN_KEEPIDLE(tp)); + if (nstat_collect) + nstat_route_connect_success(tp->t_inpcb->inp_route.ro_rt); } /* * If segment contains data or ACK, will call tcp_reass() @@ -2301,15 +3455,21 @@ trimthenstep6: */ if (tlen == 0 && (thflags & TH_FIN) == 0) (void) tcp_reass(tp, (struct tcphdr *)0, &tlen, - (struct mbuf *)0); + NULL, ifp); tp->snd_wl1 = th->th_seq - 1; /* FALLTHROUGH */ - - /* soisconnected may lead to socket_unlock in case of upcalls, - * make sure this is done when everything is setup. +#if MPTCP + /* + * Do not send the connect notification for additional subflows + * until ACK for 3-way handshake arrives. */ - soisconnected(so); + if ((!(tp->t_mpflags & TMPF_MPTCP_TRUE)) && + (tp->t_mpflags & TMPF_SENT_JOIN)) { + isconnected = FALSE; + } else +#endif /* MPTCP */ + isconnected = TRUE; /* * In ESTABLISHED state: drop duplicate ACKs; ACK out of range @@ -2330,19 +3490,109 @@ trimthenstep6: tcpstat.tcps_rcvacktoomuch++; goto dropafterack; } - if (tp->sack_enable && + if (SACK_ENABLED(tp) && (to.to_nsacks > 0 || !TAILQ_EMPTY(&tp->snd_holes))) - tcp_sack_doack(tp, &to, th->th_ack); + tcp_sack_doack(tp, &to, th, &sack_bytes_acked); + +#if MPTCP + if ((tp->t_mpuna) && (SEQ_GEQ(th->th_ack, tp->t_mpuna))) { + if (tp->t_mpflags & TMPF_PREESTABLISHED) { + /* MP TCP establishment succeeded */ + tp->t_mpuna = 0; + if (tp->t_mpflags & TMPF_JOINED_FLOW) { + if (tp->t_mpflags & TMPF_SENT_JOIN) { + tp->t_mpflags &= + ~TMPF_PREESTABLISHED; + tp->t_mpflags |= + TMPF_MPTCP_TRUE; + so->so_flags |= SOF_MPTCP_TRUE; + if (mptcp_dbg >= MP_ERR_DEBUG) + printf("MPTCP SUCCESS" + " %s \n",__func__); + tp->t_timer[TCPT_JACK_RXMT] = 0; + tp->t_mprxtshift = 0; + isconnected = TRUE; + } else { + isconnected = FALSE; + } + } else { + isconnected = TRUE; + tp->t_mpflags &= ~TMPF_SENT_KEYS; + } + } + } +#endif /* MPTCP */ + /* + * If we have outstanding data (other than + * a window probe), this is a completely + * duplicate ack (ie, window info didn't + * change) and the ack is the biggest we've seen. + */ if (SEQ_LEQ(th->th_ack, tp->snd_una)) { if (tlen == 0 && tiwin == tp->snd_wnd) { + /* + * If both ends send FIN at the same time, + * then the ack will be a duplicate ack + * but we have to process the FIN. Check + * for this condition and process the FIN + * instead of the dupack + */ + if ((thflags & TH_FIN) && + (tp->t_flags & TF_SENTFIN) && + !TCPS_HAVERCVDFIN(tp->t_state) && + (th->th_ack + 1) == tp->snd_max) { + break; + } +process_dupack: +#if MPTCP + /* + * MPTCP options that are ignored must + * not be treated as duplicate ACKs. + */ + if (to.to_flags & TOF_MPTCP) { + goto drop; + } + + if ((isconnected) && (tp->t_mpflags & TMPF_JOINED_FLOW)) { + if (mptcp_dbg >= MP_ERR_DEBUG) + printf("%s: bypass ack recovery\n",__func__); + break; + } +#endif /* MPTCP */ + /* + * If a duplicate acknowledgement was seen + * after ECN, it indicates packet loss in + * addition to ECN. Reset INRECOVERY flag + * so that we can process partial acks + * correctly + */ + if (tp->ecn_flags & TE_INRECOVERY) + tp->ecn_flags &= ~TE_INRECOVERY; + tcpstat.tcps_rcvdupack++; + ++tp->t_dupacks; + + /* + * Check if we need to reset the limit on + * early retransmit + */ + if (tp->t_early_rexmt_count > 0 && + TSTMP_GEQ(tcp_now, + (tp->t_early_rexmt_win + + TCP_EARLY_REXMT_WIN))) + tp->t_early_rexmt_count = 0; + + /* + * Is early retransmit needed? We check for + * this when the connection is waiting for + * duplicate acks to enter fast recovery. + */ + if (!IN_FASTRECOVERY(tp)) + tcp_early_rexmt_check(tp, th); + /* - * If we have outstanding data (other than - * a window probe), this is a completely - * duplicate ack (ie, window info didn't - * change), the ack is the biggest we've - * seen and we've seen exactly our rexmt - * threshhold of them, assume a packet + * If we've seen exactly rexmt threshold + * of duplicate acks, assume a packet * has been dropped and retransmit it. * Kludge snd_nxt & the congestion * window so we send only this one @@ -2361,12 +3611,32 @@ trimthenstep6: * network. */ if (tp->t_timer[TCPT_REXMT] == 0 || - th->th_ack != tp->snd_una) + (th->th_ack != tp->snd_una + && sack_bytes_acked == 0)) { tp->t_dupacks = 0; - else if (++tp->t_dupacks > tcprexmtthresh || - ((tcp_do_newreno || tp->sack_enable) && - IN_FASTRECOVERY(tp))) { - if (tp->sack_enable && IN_FASTRECOVERY(tp)) { + tp->t_rexmtthresh = tcprexmtthresh; + } else if (tp->t_dupacks > tp->t_rexmtthresh || + IN_FASTRECOVERY(tp)) { + + /* + * If this connection was seeing packet + * reordering, then recovery might be + * delayed to disambiguate between + * reordering and loss + */ + if (SACK_ENABLED(tp) && !IN_FASTRECOVERY(tp) && + (tp->t_flagsext & + (TF_PKTS_REORDERED|TF_DELAY_RECOVERY)) == + (TF_PKTS_REORDERED|TF_DELAY_RECOVERY)) { + /* + * Since the SACK information is already + * updated, this ACK will be dropped + */ + break; + } + + if (SACK_ENABLED(tp) + && IN_FASTRECOVERY(tp)) { int awnd; /* @@ -2384,12 +3654,13 @@ trimthenstep6: } } else tp->snd_cwnd += tp->t_maxseg; - tp->t_unacksegs = 0; + + tcp_ccdbg_trace(tp, th, TCP_CC_IN_FASTRECOVERY); + (void) tcp_output(tp); goto drop; - } else if (tp->t_dupacks == tcprexmtthresh) { + } else if (tp->t_dupacks == tp->t_rexmtthresh) { tcp_seq onxt = tp->snd_nxt; - u_int win; /* * If we're doing sack, check to @@ -2398,116 +3669,168 @@ trimthenstep6: * check to see if we're in newreno * recovery. */ - if (tp->sack_enable) { + if (SACK_ENABLED(tp)) { if (IN_FASTRECOVERY(tp)) { tp->t_dupacks = 0; break; + } else if (tp->t_flagsext & TF_DELAY_RECOVERY) { + break; } - } else if (tcp_do_newreno) { + } else { if (SEQ_LEQ(th->th_ack, tp->snd_recover)) { tp->t_dupacks = 0; break; } } - win = min(tp->snd_wnd, tp->snd_cwnd) / - 2 / tp->t_maxseg; - if (win < 2) - win = 2; - tp->snd_ssthresh = win * tp->t_maxseg; - ENTER_FASTRECOVERY(tp); + tp->snd_recover = tp->snd_max; - tp->t_timer[TCPT_REXMT] = 0; + tp->t_timer[TCPT_PTO] = 0; tp->t_rtttime = 0; - tp->ecn_flags |= TE_SENDCWR; - if (tp->sack_enable) { + + /* + * If the connection has seen pkt + * reordering, delay recovery until + * it is clear that the packet + * was lost. + */ + if (SACK_ENABLED(tp) && + (tp->t_flagsext & + (TF_PKTS_REORDERED|TF_DELAY_RECOVERY)) + == TF_PKTS_REORDERED && + !IN_FASTRECOVERY(tp) && + tp->t_reorderwin > 0 && + tp->t_state == TCPS_ESTABLISHED) { + tp->t_timer[TCPT_DELAYFR] = + OFFSET_FROM_START(tp, + tp->t_reorderwin); + tp->t_flagsext |= TF_DELAY_RECOVERY; + tcpstat.tcps_delay_recovery++; + tcp_ccdbg_trace(tp, th, + TCP_CC_DELAY_FASTRECOVERY); + break; + } + + /* + * If the current tcp cc module has + * defined a hook for tasks to run + * before entering FR, call it + */ + if (CC_ALGO(tp)->pre_fr != NULL) + CC_ALGO(tp)->pre_fr(tp); + ENTER_FASTRECOVERY(tp); + tp->t_timer[TCPT_REXMT] = 0; + if ((tp->ecn_flags & TE_ECN_ON) + == TE_ECN_ON) + tp->ecn_flags |= TE_SENDCWR; + + if (SACK_ENABLED(tp)) { tcpstat.tcps_sack_recovery_episode++; tp->sack_newdata = tp->snd_nxt; tp->snd_cwnd = tp->t_maxseg; - tp->t_unacksegs = 0; + + /* + * Enable probe timeout to detect + * a tail loss in the recovery + * window. + */ + tp->t_timer[TCPT_PTO] = + OFFSET_FROM_START(tp, + max(10, (tp->t_srtt >> TCP_RTT_SHIFT))); + + tcp_ccdbg_trace(tp, th, + TCP_CC_ENTER_FASTRECOVERY); + (void) tcp_output(tp); goto drop; } tp->snd_nxt = th->th_ack; tp->snd_cwnd = tp->t_maxseg; - tp->t_unacksegs = 0; (void) tcp_output(tp); tp->snd_cwnd = tp->snd_ssthresh + tp->t_maxseg * tp->t_dupacks; if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; + tcp_ccdbg_trace(tp, th, + TCP_CC_ENTER_FASTRECOVERY); goto drop; + } else if (limited_txmt && + ALLOW_LIMITED_TRANSMIT(tp) && + (!(SACK_ENABLED(tp)) || sack_bytes_acked > 0) && + (so->so_snd.sb_cc - (tp->snd_max - tp->snd_una)) > 0) { + u_int32_t incr = (tp->t_maxseg * tp->t_dupacks); + + /* Use Limited Transmit algorithm on the first two + * duplicate acks when there is new data to transmit + */ + tp->snd_cwnd += incr; + tcpstat.tcps_limited_txt++; + (void) tcp_output(tp); + + tcp_ccdbg_trace(tp, th, TCP_CC_LIMITED_TRANSMIT); + + /* Reset snd_cwnd back to normal */ + tp->snd_cwnd -= incr; } - } else + } else { tp->t_dupacks = 0; + tp->t_rexmtthresh = tcprexmtthresh; + } break; } /* * If the congestion window was inflated to account * for the other side's cached packets, retract it. */ - if (!IN_FASTRECOVERY(tp)) { - /* - * We were not in fast recovery. Reset the duplicate ack - * counter. - */ - tp->t_dupacks = 0; - } - /* - * If the congestion window was inflated to account - * for the other side's cached packets, retract it. - */ - else { - if (tcp_do_newreno || tp->sack_enable) { - if (SEQ_LT(th->th_ack, tp->snd_recover)) { - if (tp->sack_enable) - tcp_sack_partialack(tp, th); - else - tcp_newreno_partial_ack(tp, th); - } - else { - if (tcp_do_newreno) { - int32_t ss = tp->snd_max - th->th_ack; - - /* - * Complete ack. Inflate the congestion window to - * ssthresh and exit fast recovery. - * - * Window inflation should have left us with approx. - * snd_ssthresh outstanding data. But in case we - * would be inclined to send a burst, better to do - * it via the slow start mechanism. - */ - if (ss < tp->snd_ssthresh) - tp->snd_cwnd = ss + tp->t_maxseg; - else - tp->snd_cwnd = tp->snd_ssthresh; - } - else { - /* - * Clamp the congestion window to the crossover point - * and exit fast recovery. - */ - if (tp->snd_cwnd > tp->snd_ssthresh) - tp->snd_cwnd = tp->snd_ssthresh; - } - - EXIT_FASTRECOVERY(tp); - tp->t_dupacks = 0; - tp->t_bytes_acked = 0; - } - } - else { + if (IN_FASTRECOVERY(tp)) { + if (SEQ_LT(th->th_ack, tp->snd_recover)) { /* - * Clamp the congestion window to the crossover point - * and exit fast recovery in non-newreno and non-SACK case. + * If we received an ECE and entered + * recovery, the subsequent ACKs should + * not be treated as partial acks. */ - if (tp->snd_cwnd > tp->snd_ssthresh) - tp->snd_cwnd = tp->snd_ssthresh; + if (tp->ecn_flags & TE_INRECOVERY) + goto process_ACK; + + if (SACK_ENABLED(tp)) + tcp_sack_partialack(tp, th); + else + tcp_newreno_partial_ack(tp, th); + tcp_ccdbg_trace(tp, th, TCP_CC_PARTIAL_ACK); + } else { EXIT_FASTRECOVERY(tp); - tp->t_dupacks = 0; - tp->t_bytes_acked = 0; + if (CC_ALGO(tp)->post_fr != NULL) + CC_ALGO(tp)->post_fr(tp, th); + + tcp_ccdbg_trace(tp, th, + TCP_CC_EXIT_FASTRECOVERY); + } + } else if ((tp->t_flagsext & + (TF_PKTS_REORDERED|TF_DELAY_RECOVERY)) + == (TF_PKTS_REORDERED|TF_DELAY_RECOVERY)) { + /* + * If the ack acknowledges upto snd_recover or if + * it acknowledges all the snd holes, exit + * recovery and cancel the timer. Otherwise, + * this is a partial ack. Wait for recovery timer + * to enter recovery. The snd_holes have already + * been updated. + */ + if (SEQ_GEQ(th->th_ack, tp->snd_recover) || + TAILQ_EMPTY(&tp->snd_holes)) { + tp->t_timer[TCPT_DELAYFR] = 0; + tp->t_flagsext &= ~TF_DELAY_RECOVERY; + EXIT_FASTRECOVERY(tp); + tcp_ccdbg_trace(tp, th, + TCP_CC_EXIT_FASTRECOVERY); } + } else { + /* + * We were not in fast recovery. Reset the + * duplicate ack counter. + */ + tp->t_dupacks = 0; + tp->t_rexmtthresh = tcprexmtthresh; } @@ -2534,52 +3857,22 @@ trimthenstep6: } process_ACK: - acked = th->th_ack - tp->snd_una; + acked = BYTES_ACKED(th, tp); tcpstat.tcps_rcvackpack++; tcpstat.tcps_rcvackbyte += acked; /* - * If we just performed our first retransmit, and the ACK - * arrives within our recovery window, then it was a mistake - * to do the retransmit in the first place. Recover our - * original cwnd and ssthresh, and proceed to transmit where - * we left off. - */ - if (tp->t_rxtshift == 1 && tcp_now < tp->t_badrxtwin) { - ++tcpstat.tcps_sndrexmitbad; - tp->snd_cwnd = tp->snd_cwnd_prev; - tp->snd_ssthresh = tp->snd_ssthresh_prev; - tp->snd_recover = tp->snd_recover_prev; - if (tp->t_flags & TF_WASFRECOVERY) - ENTER_FASTRECOVERY(tp); - tp->snd_nxt = tp->snd_max; - tp->t_badrxtwin = 0; /* XXX probably not required */ - } - - /* - * If we have a timestamp reply, update smoothed - * round trip time. If no timestamp is present but - * transmit timer is running and timed sequence - * number was acked, update smoothed round trip time. - * Since we now have an rtt measurement, cancel the - * timer backoff (cf., Phil Karn's retransmit alg.). - * Recompute the initial retransmit timer. - * Also makes sure we have a valid time stamp in hand + * If the last packet was a retransmit, make sure + * it was not spurious. * - * Some boxes send broken timestamp replies - * during the SYN+ACK phase, ignore - * timestamps of 0 or we could calculate a - * huge RTT and blow up the retransmit timer. + * This will also take care of congestion window + * adjustment if a last packet was recovered due to a + * tail loss probe. */ - if (((to.to_flags & TOF_TS) != 0) && (to.to_tsecr != 0)) { - if (!tp->t_rttlow || tp->t_rttlow > tcp_now - to.to_tsecr) - tp->t_rttlow = tcp_now - to.to_tsecr; - tcp_xmit_timer(tp, tcp_now - to.to_tsecr); - } else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) { - if (!tp->t_rttlow || tp->t_rttlow > tcp_now - tp->t_rtttime) - tp->t_rttlow = tcp_now - tp->t_rtttime; - tcp_xmit_timer(tp, tp->t_rtttime); - } + tcp_bad_rexmt_check(tp, th, &to); + + /* Recalculate the RTT */ + tcp_compute_rtt(tp, &to, th); /* * If all outstanding data is acked, stop retransmit @@ -2589,117 +3882,85 @@ process_ACK: */ if (th->th_ack == tp->snd_max) { tp->t_timer[TCPT_REXMT] = 0; + tp->t_timer[TCPT_PTO] = 0; needoutput = 1; } else if (tp->t_timer[TCPT_PERSIST] == 0) - tp->t_timer[TCPT_REXMT] = tp->t_rxtcur; + tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, + tp->t_rxtcur); /* - * If no data (only SYN) was ACK'd, - * skip rest of ACK processing. + * If no data (only SYN) was ACK'd, skip rest of ACK + * processing. */ if (acked == 0) goto step6; - /* - * When new data is acked, open the congestion window. - */ + if ((thflags & TH_ECE) != 0 && - (tp->ecn_flags & TE_SETUPSENT) != 0) { + ((tp->ecn_flags & TE_ECN_ON) == TE_ECN_ON)) { /* - * Reduce the congestion window if we haven't done so. + * Reduce the congestion window if we haven't + * done so. */ - if (!(tp->sack_enable && IN_FASTRECOVERY(tp)) && - !(tcp_do_newreno && SEQ_LEQ(th->th_ack, tp->snd_recover))) { + if (!IN_FASTRECOVERY(tp)) { tcp_reduce_congestion_window(tp); + tp->ecn_flags |= (TE_INRECOVERY|TE_SENDCWR); + tcp_ccdbg_trace(tp, th, TCP_CC_ECN_RCVD); } - } else if ((!tcp_do_newreno && !tp->sack_enable) || - !IN_FASTRECOVERY(tp)) { - /* - * RFC 3465 - Appropriate Byte Counting. - * - * If the window is currently less than ssthresh, - * open the window by the number of bytes ACKed by - * the last ACK, however clamp the window increase - * to an upper limit "L". - * - * In congestion avoidance phase, open the window by - * one segment each time "bytes_acked" grows to be - * greater than or equal to the congestion window. - */ - - register u_int cw = tp->snd_cwnd; - register u_int incr = tp->t_maxseg; - - if (tcp_do_rfc3465) { - - if (cw >= tp->snd_ssthresh) { - tp->t_bytes_acked += acked; - if (tp->t_bytes_acked >= cw) { - /* Time to increase the window. */ - tp->t_bytes_acked -= cw; - } else { - /* No need to increase yet. */ - incr = 0; - } - } else { - /* - * If the user explicitly enables RFC3465 - * use 2*SMSS for the "L" param. Otherwise - * use the more conservative 1*SMSS. - * - * (See RFC 3465 2.3 Choosing the Limit) - */ - u_int abc_lim; - - abc_lim = (tcp_do_rfc3465_lim2 && - tp->snd_nxt == tp->snd_max) ? incr * 2 : incr; - - incr = lmin(acked, abc_lim); - } - } - else { - /* - * If the window gives us less than ssthresh packets - * in flight, open exponentially (segsz per packet). - * Otherwise open linearly: segsz per window - * (segsz^2 / cwnd per packet). - */ - - if (cw >= tp->snd_ssthresh) { - incr = max((incr * incr / cw), 1); - } - } - + } - tp->snd_cwnd = min(cw+incr, TCP_MAXWIN<snd_scale); + /* + * When new data is acked, open the congestion window. + * The specifics of how this is achieved are up to the + * congestion control algorithm in use for this connection. + * + * The calculations in this function assume that snd_una is + * not updated yet. + */ + if (!IN_FASTRECOVERY(tp)) { + if (CC_ALGO(tp)->ack_rcvd != NULL) + CC_ALGO(tp)->ack_rcvd(tp, th); + tcp_ccdbg_trace(tp, th, TCP_CC_ACK_RCVD); } if (acked > so->so_snd.sb_cc) { tp->snd_wnd -= so->so_snd.sb_cc; sbdrop(&so->so_snd, (int)so->so_snd.sb_cc); + if (so->so_flags & SOF_ENABLE_MSGS) { + so->so_msg_state->msg_serial_bytes -= + (int)so->so_snd.sb_cc; + } ourfinisacked = 1; } else { sbdrop(&so->so_snd, acked); + if (so->so_flags & SOF_ENABLE_MSGS) { + so->so_msg_state->msg_serial_bytes -= + acked; + } + tcp_sbsnd_trim(&so->so_snd); tp->snd_wnd -= acked; ourfinisacked = 0; } /* detect una wraparound */ - if ((tcp_do_newreno || tp->sack_enable) && - !IN_FASTRECOVERY(tp) && + if ( !IN_FASTRECOVERY(tp) && SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; - if ((tcp_do_newreno || tp->sack_enable) && - IN_FASTRECOVERY(tp) && + + if (IN_FASTRECOVERY(tp) && SEQ_GEQ(th->th_ack, tp->snd_recover)) EXIT_FASTRECOVERY(tp); + tp->snd_una = th->th_ack; - if (tp->sack_enable) { + if (SACK_ENABLED(tp)) { if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; } if (SEQ_LT(tp->snd_nxt, tp->snd_una)) tp->snd_nxt = tp->snd_una; - + if ((tp->t_flagsext & TF_MEASURESNDBW) != 0 && + tp->t_bwmeas != NULL) + tcp_bwmeas_check(tp); + /* * sowwakeup must happen after snd_una, et al. are updated so that * the sequence numbers are in sync with so_snd @@ -2718,17 +3979,24 @@ process_ACK: /* * If we can't receive any more * data, then closing user can proceed. - * Starting the timer is contrary to the + * Starting the TCPT_2MSL timer is contrary to the * specification, but if we don't get a FIN * we'll hang forever. */ if (so->so_state & SS_CANTRCVMORE) { - tp->t_timer[TCPT_2MSL] = tcp_maxidle; - add_to_time_wait(tp); - soisdisconnected(so); + tp->t_timer[TCPT_2MSL] = OFFSET_FROM_START(tp, + TCP_CONN_MAXIDLE(tp)); + isconnected = FALSE; + isdisconnected = TRUE; } + DTRACE_TCP4(state__change, void, NULL, + struct inpcb *, inp, + struct tcpcb *, tp, + int32_t, TCPS_FIN_WAIT_2); tp->t_state = TCPS_FIN_WAIT_2; - /* fall through and make sure we also recognize data ACKed with the FIN */ + /* fall through and make sure we also recognize + * data ACKed with the FIN + */ } tp->t_flags |= TF_ACKNOW; break; @@ -2741,17 +4009,19 @@ process_ACK: */ case TCPS_CLOSING: if (ourfinisacked) { + DTRACE_TCP4(state__change, void, NULL, + struct inpcb *, inp, + struct tcpcb *, tp, + int32_t, TCPS_TIME_WAIT); tp->t_state = TCPS_TIME_WAIT; tcp_canceltimers(tp); - /* Shorten TIME_WAIT [RFC-1644, p.28] */ - if (tp->cc_recv != 0 && - tp->t_starttime < (u_int32_t)tcp_msl) - tp->t_timer[TCPT_2MSL] = - tp->t_rxtcur * TCPTV_TWTRUNC; - else - tp->t_timer[TCPT_2MSL] = 2 * tcp_msl; - add_to_time_wait(tp); - soisdisconnected(so); + if (tp->t_flagsext & TF_NOTIMEWAIT) { + tp->t_flags |= TF_CLOSING; + } else { + add_to_time_wait(tp, 2 * tcp_msl); + } + isconnected = FALSE; + isdisconnected = TRUE; } tp->t_flags |= TF_ACKNOW; break; @@ -2775,10 +4045,29 @@ process_ACK: * it and restart the finack timer. */ case TCPS_TIME_WAIT: - tp->t_timer[TCPT_2MSL] = 2 * tcp_msl; - add_to_time_wait(tp); + add_to_time_wait(tp, 2 * tcp_msl); goto dropafterack; } + + /* + * If there is a SACK option on the ACK and we + * haven't seen any duplicate acks before, count + * it as a duplicate ack even if the cumulative + * ack is advanced. If the receiver delayed an + * ack and detected loss afterwards, then the ack + * will advance cumulative ack and will also have + * a SACK option. So counting it as one duplicate + * ack is ok. + */ + if (sack_ackadv == 1 && + tp->t_state == TCPS_ESTABLISHED && + SACK_ENABLED(tp) && sack_bytes_acked > 0 && + to.to_nsacks > 0 && tp->t_dupacks == 0 && + SEQ_LEQ(th->th_ack, tp->snd_una) && tlen == 0 && + !(tp->t_flagsext & TF_PKTS_REORDERED)) { + tcpstat.tcps_sack_ackadv++; + goto process_dupack; + } } step6: @@ -2856,7 +4145,7 @@ step6: ) tcp_pulloutofband(so, th, m, drop_hdrlen); /* hdr drop is delayed */ - } else + } else { /* * If no out of band data is expected, * pull receive urgent pointer along @@ -2864,8 +4153,27 @@ step6: */ if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) tp->rcv_up = tp->rcv_nxt; -dodata: /* XXX */ + } +dodata: + + /* Set socket's connect or disconnect state correcly before doing data. + * The following might unlock the socket if there is an upcall or a socket + * filter. + */ + if (isconnected) { + soisconnected(so); + } else if (isdisconnected) { + soisdisconnected(so); + } + /* Let's check the state of pcb just to make sure that it did not get closed + * when we unlocked above + */ + if (inp->inp_state == INPCB_STATE_DEAD) { + /* Just drop the packet that we are processing and return */ + goto drop; + } + /* * Process the segment text, merging it into the TCP sequencing queue, * and arranging for acknowledgment of receipt if necessary. @@ -2894,30 +4202,59 @@ dodata: /* XXX */ if (th->th_seq == tp->rcv_nxt && LIST_EMPTY(&tp->t_segq) && TCPS_HAVEESTABLISHED(tp->t_state)) { - if (DELAY_ACK(tp) && ((tp->t_flags & TF_ACKNOW) == 0)) { - tp->t_flags |= TF_DELACK; - tp->t_unacksegs++; + TCP_INC_VAR(tp->t_unacksegs, nlropkts); + /* + * Calculate the RTT on the receiver only if the + * connection is in streaming mode and the last + * packet was not an end-of-write + */ + if ((tp->t_flags & TF_STRETCHACK) && + !(tp->t_flagsext & TF_STREAMEOW)) + tcp_compute_rtt(tp, &to, th); + + if (DELAY_ACK(tp, th) && + ((tp->t_flags & TF_ACKNOW) == 0) ) { + if ((tp->t_flags & TF_DELACK) == 0) { + tp->t_flags |= TF_DELACK; + tp->t_timer[TCPT_DELACK] = + OFFSET_FROM_START(tp, tcp_delack); + } } else { - tp->t_unacksegs = 0; tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt += tlen; thflags = th->th_flags & TH_FIN; - tcpstat.tcps_rcvpack++; + TCP_INC_VAR(tcpstat.tcps_rcvpack, nlropkts); tcpstat.tcps_rcvbyte += tlen; - ND6_HINT(tp); - if (sbappendstream(&so->so_rcv, m)) + if (nstat_collect) { + if (m->m_pkthdr.pkt_flags & PKTF_SW_LRO_PKT) { + INP_ADD_STAT(inp, cell, wifi, wired, + rxpackets, m->m_pkthdr.lro_npkts); + } else { + INP_ADD_STAT(inp, cell, wifi, wired, + rxpackets, 1); + } + INP_ADD_STAT(inp, cell, wifi, wired, + rxbytes, tlen); + } + tcp_sbrcv_grow(tp, &so->so_rcv, &to, tlen); + so_recv_data_stat(so, m, drop_hdrlen); + + if (sbappendstream_rcvdemux(so, m, + th->th_seq - (tp->irs + 1), 0)) { sorwakeup(so); + } } else { - thflags = tcp_reass(tp, th, &tlen, m); + thflags = tcp_reass(tp, th, &tlen, m, ifp); tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; } - if (tlen > 0 && tp->sack_enable) + if (tlen > 0 && SACK_ENABLED(tp)) tcp_update_sack_list(tp, save_start, save_end); + tcp_adaptive_rwtimo_check(tp, tlen); + if (tp->t_flags & TF_DELACK) { #if INET6 @@ -2935,14 +4272,6 @@ dodata: /* XXX */ } } - /* - * Note the amount of data that peer has sent into - * our window, in order to estimate the sender's - * buffer size. - */ - len = (u_int)(so->so_rcv.sb_hiwat - (tp->rcv_adv - tp->rcv_nxt)); - if (len > so->so_rcv.sb_maxused) - so->so_rcv.sb_maxused = len; } else { m_freem(m); thflags &= ~TH_FIN; @@ -2957,21 +4286,21 @@ dodata: /* XXX */ socantrcvmore(so); postevent(so, 0, EV_FIN); /* - * If connection is half-synchronized - * (ie NEEDSYN flag on) then delay ACK, * If connection is half-synchronized * (ie NEEDSYN flag on) then delay ACK, * so it may be piggybacked when SYN is sent. * Otherwise, since we received a FIN then no * more input can be expected, send ACK now. */ - if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) { - tp->t_flags |= TF_DELACK; - tp->t_unacksegs++; + TCP_INC_VAR(tp->t_unacksegs, nlropkts); + if (DELAY_ACK(tp, th) && (tp->t_flags & TF_NEEDSYN)) { + if ((tp->t_flags & TF_DELACK) == 0) { + tp->t_flags |= TF_DELACK; + tp->t_timer[TCPT_DELACK] = OFFSET_FROM_START(tp, tcp_delack); + } } else { tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; } tp->rcv_nxt++; } @@ -2982,8 +4311,10 @@ dodata: /* XXX */ * enter the CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: - tp->t_starttime = 0; + tp->t_starttime = tcp_now; case TCPS_ESTABLISHED: + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_CLOSE_WAIT); tp->t_state = TCPS_CLOSE_WAIT; break; @@ -2992,6 +4323,8 @@ dodata: /* XXX */ * enter the CLOSING state. */ case TCPS_FIN_WAIT_1: + DTRACE_TCP4(state__change, void, NULL, struct inpcb *, inp, + struct tcpcb *, tp, int32_t, TCPS_CLOSING); tp->t_state = TCPS_CLOSING; break; @@ -3001,21 +4334,18 @@ dodata: /* XXX */ * standard timers. */ case TCPS_FIN_WAIT_2: + DTRACE_TCP4(state__change, void, NULL, + struct inpcb *, inp, + struct tcpcb *, tp, + int32_t, TCPS_TIME_WAIT); tp->t_state = TCPS_TIME_WAIT; tcp_canceltimers(tp); - /* Shorten TIME_WAIT [RFC-1644, p.28] */ - if (tp->cc_recv != 0 && - tp->t_starttime < (u_int32_t)tcp_msl) { - tp->t_timer[TCPT_2MSL] = - tp->t_rxtcur * TCPTV_TWTRUNC; - /* For transaction client, force ACK now. */ - tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; + tp->t_flags |= TF_ACKNOW; + if (tp->t_flagsext & TF_NOTIMEWAIT) { + tp->t_flags |= TF_CLOSING; + } else { + add_to_time_wait(tp, 2 * tcp_msl); } - else - tp->t_timer[TCPT_2MSL] = 2 * tcp_msl; - - add_to_time_wait(tp); soisdisconnected(so); break; @@ -3023,8 +4353,7 @@ dodata: /* XXX */ * In TIME_WAIT state restart the 2 MSL time_wait timer. */ case TCPS_TIME_WAIT: - tp->t_timer[TCPT_2MSL] = 2 * tcp_msl; - add_to_time_wait(tp); + add_to_time_wait(tp, 2 * tcp_msl); break; } } @@ -3038,9 +4367,12 @@ dodata: /* XXX */ * Return any desired output. */ if (needoutput || (tp->t_flags & TF_ACKNOW)) { - tp->t_unacksegs = 0; (void) tcp_output(tp); } + + tcp_check_timer_state(tp); + + tcp_unlock(so, 1, 0); KERNEL_DEBUG(DBG_FNC_TCP_INPUT | DBG_FUNC_END,0,0,0,0,0); return; @@ -3065,6 +4397,7 @@ dropafterack: (SEQ_GT(tp->snd_una, th->th_ack) || SEQ_GT(th->th_ack, tp->snd_max)) ) { rstreason = BANDLIM_RST_OPENPORT; + IF_TCP_STATINC(ifp, dospacket); goto dropwithreset; } #if TCPDEBUG @@ -3074,8 +4407,9 @@ dropafterack: #endif m_freem(m); tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; (void) tcp_output(tp); + + /* Don't need to check timer state as we should have done it during tcp_output */ tcp_unlock(so, 1, 0); KERNEL_DEBUG(DBG_FNC_TCP_INPUT | DBG_FUNC_END,0,0,0,0,0); return; @@ -3116,25 +4450,27 @@ dropwithreset: tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif + bzero(&tra, sizeof(tra)); + tra.ifscope = ifscope; + tra.awdl_unrestricted = 1; if (thflags & TH_ACK) /* mtod() below is safe as long as hdr dropping is delayed */ tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack, - TH_RST, ifscope); + TH_RST, &tra); else { if (thflags & TH_SYN) tlen++; /* mtod() below is safe as long as hdr dropping is delayed */ tcp_respond(tp, mtod(m, void *), th, m, th->th_seq+tlen, - (tcp_seq)0, TH_RST|TH_ACK, ifscope); + (tcp_seq)0, TH_RST|TH_ACK, &tra); } /* destroy temporarily created socket */ if (dropsocket) { (void) soabort(so); tcp_unlock(so, 1, 0); + } else if ((inp != NULL) && (nosock == 0)) { + tcp_unlock(so, 1, 0); } - else - if ((inp != NULL) && (nosock == 0)) - tcp_unlock(so, 1, 0); KERNEL_DEBUG(DBG_FNC_TCP_INPUT | DBG_FUNC_END,0,0,0,0,0); return; dropnosock: @@ -3154,9 +4490,9 @@ drop: (void) soabort(so); tcp_unlock(so, 1, 0); } - else - if (nosock == 0) - tcp_unlock(so, 1, 0); + else if (nosock == 0) { + tcp_unlock(so, 1, 0); + } KERNEL_DEBUG(DBG_FNC_TCP_INPUT | DBG_FUNC_END,0,0,0,0,0); return; } @@ -3200,11 +4536,7 @@ tcp_dooptions(tp, cp, cnt, th, to, input_ifscope) if (!(th->th_flags & TH_SYN)) continue; bcopy((char *) cp + 2, (char *) &mss, sizeof(mss)); - -#if BYTE_ORDER != BIG_ENDIAN NTOHS(mss); -#endif - break; case TCPOPT_WINDOW: @@ -3212,6 +4544,7 @@ tcp_dooptions(tp, cp, cnt, th, to, input_ifscope) continue; if (!(th->th_flags & TH_SYN)) continue; + to->to_flags |= TOF_SCALE; tp->t_flags |= TF_RCVD_SCALE; tp->requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT); break; @@ -3222,18 +4555,10 @@ tcp_dooptions(tp, cp, cnt, th, to, input_ifscope) to->to_flags |= TOF_TS; bcopy((char *)cp + 2, (char *)&to->to_tsval, sizeof(to->to_tsval)); - -#if BYTE_ORDER != BIG_ENDIAN NTOHL(to->to_tsval); -#endif - bcopy((char *)cp + 6, (char *)&to->to_tsecr, sizeof(to->to_tsecr)); - -#if BYTE_ORDER != BIG_ENDIAN NTOHL(to->to_tsecr); -#endif - /* * A timestamp received in a SYN makes * it ok to send timestamp requests and replies. @@ -3259,6 +4584,12 @@ tcp_dooptions(tp, cp, cnt, th, to, input_ifscope) tcpstat.tcps_sack_rcv_blocks++; break; + +#if MPTCP + case TCPOPT_MULTIPATH: + tcp_do_mptcp_options(tp, cp, th, to, optlen); + break; +#endif /* MPTCP */ } } if (th->th_flags & TH_SYN) @@ -3301,28 +4632,138 @@ tcp_pulloutofband(so, th, m, off) panic("tcp_pulloutofband"); } +uint32_t +get_base_rtt(struct tcpcb *tp) +{ + uint32_t base_rtt = 0, i; + for (i = 0; i < N_RTT_BASE; ++i) { + if (tp->rtt_hist[i] != 0 && + (base_rtt == 0 || tp->rtt_hist[i] < base_rtt)) + base_rtt = tp->rtt_hist[i]; + } + return base_rtt; +} + +/* Each value of RTT base represents the minimum RTT seen in a minute. + * We keep upto N_RTT_BASE minutes worth of history. + */ +void +update_base_rtt(struct tcpcb *tp, uint32_t rtt) +{ + int32_t i, qdelay; + u_int32_t base_rtt; + + if (++tp->rtt_count >= rtt_samples_per_slot) { +#if TRAFFIC_MGT + /* + * If the recv side is being throttled, check if the + * current RTT is closer to the base RTT seen in + * first (recent) two slots. If so, unthrottle the stream. + */ + if (tp->t_flagsext & TF_RECV_THROTTLE) { + base_rtt = min(tp->rtt_hist[0], tp->rtt_hist[1]); + qdelay = tp->t_rttcur - base_rtt; + if (qdelay < target_qdelay) + tp->t_flagsext &= ~(TF_RECV_THROTTLE); + } +#endif /* TRAFFIC_MGT */ + + for (i = (N_RTT_BASE-1); i > 0; --i) { + tp->rtt_hist[i] = tp->rtt_hist[i-1]; + } + tp->rtt_hist[0] = rtt; + tp->rtt_count = 0; + } else { + tp->rtt_hist[0] = min(tp->rtt_hist[0], rtt); + } +} + +/* + * If we have a timestamp reply, update smoothed RTT. If no timestamp is + * present but transmit timer is running and timed sequence number was + * acked, update smoothed RTT. + * + * If timestamps are supported, a receiver can update RTT even if + * there is no outstanding data. + * + * Some boxes send broken timestamp replies during the SYN+ACK phase, + * ignore timestamps of 0or we could calculate a huge RTT and blow up + * the retransmit timer. + */ +static void +tcp_compute_rtt(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) +{ + VERIFY(to != NULL && th != NULL); + if (((to->to_flags & TOF_TS) != 0) && + (to->to_tsecr != 0) && + TSTMP_GEQ(tcp_now, to->to_tsecr)) { + tcp_xmit_timer(tp, tcp_now - to->to_tsecr, + to->to_tsecr, th->th_ack); + } else if (tp->t_rtttime != 0 && SEQ_GT(th->th_ack, tp->t_rtseq)) { + tcp_xmit_timer(tp, tcp_now - tp->t_rtttime, 0, + th->th_ack); + } +} + /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void -tcp_xmit_timer(tp, rtt) - register struct tcpcb *tp; - int rtt; +tcp_xmit_timer(register struct tcpcb *tp, int rtt, + u_int32_t tsecr, tcp_seq th_ack) { register int delta; + if (tp->t_flagsext & TF_RECOMPUTE_RTT) { + if (SEQ_GT(th_ack, tp->snd_una) && + SEQ_LEQ(th_ack, tp->snd_max) && + (tsecr == 0 || + TSTMP_GEQ(tsecr, tp->t_badrexmt_time))) { + /* + * We received a new ACk after a + * spurious timeout. Adapt retransmission + * timer as described in rfc 4015. + */ + tp->t_flagsext &= ~(TF_RECOMPUTE_RTT); + tp->t_badrexmt_time = 0; + tp->t_srtt = max(tp->t_srtt_prev, rtt); + tp->t_srtt = tp->t_srtt << TCP_RTT_SHIFT; + tp->t_rttvar = max(tp->t_rttvar_prev, (rtt >> 1)); + tp->t_rttvar = tp->t_rttvar << TCP_RTTVAR_SHIFT; + + if (tp->t_rttbest > (tp->t_srtt + tp->t_rttvar)) + tp->t_rttbest = tp->t_srtt + tp->t_rttvar; + + goto compute_rto; + } else { + return; + } + } + tcpstat.tcps_rttupdated++; tp->t_rttupdated++; + + if (rtt > 0) { + tp->t_rttcur = rtt; + update_base_rtt(tp, rtt); + } + if (tp->t_srtt != 0) { /* * srtt is stored as fixed point with 5 bits after the - * binary point (i.e., scaled by 8). The following magic + * binary point (i.e., scaled by 32). The following magic * is equivalent to the smoothing algorithm in rfc793 with * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed - * point). Adjust rtt to origin 0. + * point). + * + * Freebsd adjusts rtt to origin 0 by subtracting 1 + * from the provided rtt value. This was required because + * of the way t_rtttime was initiailised to 1 before. + * Since we changed t_rtttime to be based on + * tcp_now, this extra adjustment is not needed. */ - delta = ((rtt - 1) << TCP_DELTA_SHIFT) + delta = (rtt << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); if ((tp->t_srtt += delta) <= 0) @@ -3343,8 +4784,9 @@ tcp_xmit_timer(tp, rtt) delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); if ((tp->t_rttvar += delta) <= 0) tp->t_rttvar = 1; - if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) - tp->t_rttbest = tp->t_srtt + tp->t_rttvar; + if (tp->t_rttbest == 0 || + tp->t_rttbest > (tp->t_srtt + tp->t_rttvar)) + tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. @@ -3353,10 +4795,14 @@ tcp_xmit_timer(tp, rtt) */ tp->t_srtt = rtt << TCP_RTT_SHIFT; tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); - tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } + +compute_rto: + nstat_route_rtt(tp->t_inpcb->inp_route.ro_rt, tp->t_srtt, + tp->t_rttvar); tp->t_rtttime = 0; tp->t_rxtshift = 0; + tp->t_rxtstart = 0; /* * the retransmit should happen at rtt + 4 * rttvar. @@ -3370,7 +4816,8 @@ tcp_xmit_timer(tp, rtt) * the minimum feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), - max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX); + max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX, + TCP_ADD_REXMTSLOP(tp)); /* * We received an ack for a packet that wasn't retransmitted; @@ -3401,13 +4848,20 @@ static inline unsigned int tcp_maxmtu6(struct rtentry *rt) { unsigned int maxmtu; + struct nd_ifinfo *ndi; RT_LOCK_ASSERT_HELD(rt); lck_rw_lock_shared(nd_if_rwlock); + if ((ndi = ND_IFINFO(rt->rt_ifp)) != NULL && !ndi->initialized) + ndi = NULL; + if (ndi != NULL) + lck_mtx_lock(&ndi->lock); if (rt->rt_rmx.rmx_mtu == 0) maxmtu = IN6_LINKMTU(rt->rt_ifp); else maxmtu = MIN(rt->rt_rmx.rmx_mtu, IN6_LINKMTU(rt->rt_ifp)); + if (ndi != NULL) + lck_mtx_unlock(&ndi->lock); lck_rw_done(nd_if_rwlock); return (maxmtu); @@ -3425,9 +4879,9 @@ tcp_maxmtu6(struct rtentry *rt) * size (usually 512 or the default IP max size, but no more than the mtu * of the interface), as we can't discover anything about intervening * gateways or networks. We also initialize the congestion/slow start - * window to be a single segment if the destination isn't local. - * While looking at the routing entry, we also initialize other path-dependent - * parameters from pre-set or cached values in the routing entry. + * window. While looking at the routing entry, we also initialize + * other path-dependent parameters from pre-set or cached values + * in the routing entry. * * Also take into account the space needed for options that we * send regularly. Make maxseg shorter by that amount to assure @@ -3471,22 +4925,15 @@ tcp_mss(tp, offer, input_ifscope) #if INET6 if (isipv6) { - rt = tcp_rtlookup6(inp); - if (rt != NULL && - (IN6_IS_ADDR_LOOPBACK(&inp->in6p_faddr) || - IN6_IS_ADDR_LINKLOCAL(&inp->in6p_faddr) || - rt->rt_gateway->sa_family == AF_LINK)) - isnetlocal = TRUE; + rt = tcp_rtlookup6(inp, input_ifscope); } else #endif /* INET6 */ { rt = tcp_rtlookup(inp, input_ifscope); - if (rt != NULL && - (rt->rt_gateway->sa_family == AF_LINK || - rt->rt_ifp->if_flags & IFF_LOOPBACK)) - isnetlocal = TRUE; } + isnetlocal = (tp->t_flags & TF_LOCAL); + if (rt == NULL) { tp->t_maxopd = tp->t_maxseg = #if INET6 @@ -3498,9 +4945,10 @@ tcp_mss(tp, offer, input_ifscope) ifp = rt->rt_ifp; /* * Slower link window correction: - * If a value is specificied for slowlink_wsize use it for PPP links - * believed to be on a serial modem (speed <128Kbps). Excludes 9600bps as - * it is the default value adversized by pseudo-devices over ppp. + * If a value is specificied for slowlink_wsize use it for + * PPP links believed to be on a serial modem (speed <128Kbps). + * Excludes 9600bps as it is the default value adversized + * by pseudo-devices over ppp. */ if (ifp->if_type == IFT_PPP && slowlink_wsize > 0 && ifp->if_baudrate > 9600 && ifp->if_baudrate <= 128000) { @@ -3546,32 +4994,11 @@ tcp_mss(tp, offer, input_ifscope) * or rttvar. Convert from the route-table units * to scaled multiples of the slow timeout timer. */ - if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) { - /* - * XXX the lock bit for RTT indicates that the value - * is also a minimum value; this is subject to time. - */ - if (rt->rt_rmx.rmx_locks & RTV_RTT) - tp->t_rttmin = rtt / (RTM_RTTUNIT / TCP_RETRANSHZ); - else - tp->t_rttmin = isnetlocal ? tcp_TCPTV_MIN : TCP_RETRANSHZ; - tp->t_srtt = rtt / (RTM_RTTUNIT / (TCP_RETRANSHZ * TCP_RTT_SCALE)); - tcpstat.tcps_usedrtt++; - if (rt->rt_rmx.rmx_rttvar) { - tp->t_rttvar = rt->rt_rmx.rmx_rttvar / - (RTM_RTTUNIT / (TCP_RETRANSHZ * TCP_RTTVAR_SCALE)); - tcpstat.tcps_usedrttvar++; - } else { - /* default variation is +- 1 rtt */ - tp->t_rttvar = - tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE; - } - TCPT_RANGESET(tp->t_rxtcur, - ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1, - tp->t_rttmin, TCPTV_REXMTMAX); + if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt) != 0) { + tcp_getrt_rtt(tp, rt); + } else { + tp->t_rttmin = isnetlocal ? tcp_TCPTV_MIN : TCPTV_REXMTMIN; } - else - tp->t_rttmin = isnetlocal ? tcp_TCPTV_MIN : TCP_RETRANSHZ; #if INET6 mss = (isipv6 ? tcp_maxmtu6(rt) : tcp_maxmtu(rt)); @@ -3609,8 +5036,12 @@ tcp_mss(tp, offer, input_ifscope) (origoffer == -1 || (tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP)) mss -= TCPOLEN_TSTAMP_APPA; - tp->t_maxseg = mss; +#if MPTCP + mss -= mptcp_adj_mss(tp, FALSE); +#endif /* MPTCP */ + tp->t_maxseg = mss; + /* * Calculate corrected value for sb_max; ensure to upgrade the * numerator for large sb_max values else it will overflow. @@ -3651,21 +5082,14 @@ tcp_mss(tp, offer, input_ifscope) (void)sbreserve(&so->so_rcv, bufsize); } - /* - * Set the slow-start flight size depending on whether this - * is a local network or not. - */ - if (isnetlocal) - tp->snd_cwnd = mss * ss_fltsz_local; - else - tp->snd_cwnd = mss * ss_fltsz; + set_tcp_stream_priority(so); if (rt->rt_rmx.rmx_ssthresh) { /* * There's some sort of gateway or interface * buffer limit on the path. Use this to set - * the slow start threshhold, but set the - * threshold to no less than 2*mss. + * slow-start threshold, but set the threshold to + * no less than 2*mss. */ tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh); tcpstat.tcps_usedssthresh++; @@ -3673,6 +5097,15 @@ tcp_mss(tp, offer, input_ifscope) tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; } + /* + * Set the slow-start flight size depending on whether this + * is a local network or not. + */ + if (CC_ALGO(tp)->cwnd_init != NULL) + CC_ALGO(tp)->cwnd_init(tp); + + tcp_ccdbg_trace(tp, NULL, TCP_CC_CWND_INIT); + /* Route locked during lookup above */ RT_UNLOCK(rt); } @@ -3701,7 +5134,7 @@ tcp_mssopt(tp) #if INET6 if (isipv6) - rt = tcp_rtlookup6(tp->t_inpcb); + rt = tcp_rtlookup6(tp->t_inpcb, IFSCOPE_NONE); else #endif /* INET6 */ rt = tcp_rtlookup(tp->t_inpcb, IFSCOPE_NONE); @@ -3736,7 +5169,7 @@ tcp_mssopt(tp) /* * On a partial ack arrives, force the retransmission of the * next unacknowledged segment. Do not clear tp->t_dupacks. - * By setting snd_nxt to ti_ack, this forces retransmission timer to + * By setting snd_nxt to th_ack, this forces retransmission timer to * be started again. */ static void @@ -3747,6 +5180,7 @@ tcp_newreno_partial_ack(tp, th) tcp_seq onxt = tp->snd_nxt; u_int32_t ocwnd = tp->snd_cwnd; tp->t_timer[TCPT_REXMT] = 0; + tp->t_timer[TCPT_PTO] = 0; tp->t_rtttime = 0; tp->snd_nxt = th->th_ack; /* @@ -3754,9 +5188,8 @@ tcp_newreno_partial_ack(tp, th) * (tp->snd_una has not yet been updated when this function * is called) */ - tp->snd_cwnd = tp->t_maxseg + (th->th_ack - tp->snd_una); + tp->snd_cwnd = tp->t_maxseg + BYTES_ACKED(th, tp); tp->t_flags |= TF_ACKNOW; - tp->t_unacksegs = 0; (void) tcp_output(tp); tp->snd_cwnd = ocwnd; if (SEQ_GT(onxt, tp->snd_nxt)) @@ -3765,8 +5198,8 @@ tcp_newreno_partial_ack(tp, th) * Partial window deflation. Relies on fact that tp->snd_una * not updated yet. */ - if (tp->snd_cwnd > th->th_ack - tp->snd_una) - tp->snd_cwnd -= th->th_ack - tp->snd_una; + if (tp->snd_cwnd > BYTES_ACKED(th, tp)) + tp->snd_cwnd -= BYTES_ACKED(th, tp); else tp->snd_cwnd = 0; tp->snd_cwnd += tp->t_maxseg; @@ -3787,29 +5220,61 @@ tcp_dropdropablreq(struct socket *head) { struct socket *so, *sonext; unsigned int i, j, qlen; - static int rnd; - static struct timeval old_runtime; + static u_int32_t rnd = 0; + static u_int64_t old_runtime; static unsigned int cur_cnt, old_cnt; - struct timeval tv; + u_int64_t now_sec; struct inpcb *inp = NULL; struct tcpcb *tp; if ((head->so_options & SO_ACCEPTCONN) == 0) - return 0; + return (0); + + if (TAILQ_EMPTY(&head->so_incomp)) + return (0); + + /* + * Check if there is any socket in the incomp queue + * that is closed because of a reset from the peer and is + * waiting to be garbage collected. If so, pick that as + * the victim + */ + TAILQ_FOREACH_SAFE(so, &head->so_incomp, so_list, sonext) { + inp = sotoinpcb(so); + tp = intotcpcb(inp); + if (tp != NULL && tp->t_state == TCPS_CLOSED && + so->so_head != NULL && + (so->so_state & (SS_INCOMP|SS_CANTSENDMORE|SS_CANTRCVMORE)) == + (SS_INCOMP|SS_CANTSENDMORE|SS_CANTRCVMORE)) { + /* + * The listen socket is already locked but we + * can lock this socket here without lock ordering + * issues because it is in the incomp queue and + * is not visible to others. + */ + if (lck_mtx_try_lock(&inp->inpcb_mtx)) { + so->so_usecount++; + goto found_victim; + } else { + continue; + } + } + } so = TAILQ_FIRST(&head->so_incomp); - if (!so) - return 0; - microtime(&tv); - if ((i = (tv.tv_sec - old_runtime.tv_sec)) != 0) { - old_runtime = tv; + now_sec = net_uptime(); + if ((i = (now_sec - old_runtime)) != 0) { + old_runtime = now_sec; old_cnt = cur_cnt / i; cur_cnt = 0; } qlen = head->so_incqlen; + if (rnd == 0) + rnd = RandomULong(); + if (++cur_cnt > qlen || old_cnt > qlen) { rnd = (314159 * rnd + 66329) & 0xffff; j = ((qlen + 1) * rnd) >> 16; @@ -3823,23 +5288,34 @@ tcp_dropdropablreq(struct socket *head) sonext = TAILQ_NEXT(so, so_list); - if (in_pcb_checkstate(inp, WNT_ACQUIRE, 0) != WNT_STOPUSING) { - /* Avoid the issue of a socket being accepted by one input thread - * and being dropped by another input thread. - * If we can't get a hold on this mutex, then grab the next socket in line. + if (in_pcb_checkstate(inp, WNT_ACQUIRE, 0) + != WNT_STOPUSING) { + /* + * Avoid the issue of a socket being accepted + * by one input thread and being dropped by + * another input thread. If we can't get a hold + * on this mutex, then grab the next socket in + * line. */ - if (lck_mtx_try_lock(inp->inpcb_mtx)) { + if (lck_mtx_try_lock(&inp->inpcb_mtx)) { so->so_usecount++; - if ((so->so_usecount == 2) && so->so_state & SS_INCOMP) + if ((so->so_usecount == 2) && + (so->so_state & SS_INCOMP) && + !(so->so_flags & SOF_INCOMP_INPROGRESS)) { break; - else {/* don't use if beeing accepted or used in any other way */ + } else { + /* + * don't use if being accepted or + * used in any other way + */ in_pcb_checkstate(inp, WNT_RELEASE, 1); tcp_unlock(so, 1, 0); } - } - else { - /* do not try to lock the inp in in_pcb_checkstate - * because the lock is already held in some other thread. + } else { + /* + * do not try to lock the inp in + * in_pcb_checkstate because the lock + * is already held in some other thread. * Only drop the inp_wntcnt reference. */ in_pcb_checkstate(inp, WNT_RELEASE, 1); @@ -3848,49 +5324,181 @@ tcp_dropdropablreq(struct socket *head) so = sonext; } - if (!so) - return 0; - - TAILQ_REMOVE(&head->so_incomp, so, so_list); - tcp_unlock(head, 0, 0); + if (so == NULL) { + return (0); + } /* Makes sure socket is still in the right state to be discarded */ if (in_pcb_checkstate(inp, WNT_RELEASE, 1) == WNT_STOPUSING) { tcp_unlock(so, 1, 0); - tcp_lock(head, 0, 0); - return 0; + return (0); } +found_victim: if (so->so_usecount != 2 || !(so->so_state & SS_INCOMP)) { - /* do not discard: that socket is beeing accepted */ + /* do not discard: that socket is being accepted */ tcp_unlock(so, 1, 0); - tcp_lock(head, 0, 0); - return 0; + return (0); } - so->so_head = NULL; + TAILQ_REMOVE(&head->so_incomp, so, so_list); + tcp_unlock(head, 0, 0); - /* - * We do not want to lose track of the PCB right away in case we receive - * more segments from the peer - */ + lck_mtx_assert(&inp->inpcb_mtx, LCK_MTX_ASSERT_OWNED); tp = sototcpcb(so); so->so_flags |= SOF_OVERFLOW; - tp->t_state = TCPS_TIME_WAIT; - (void) tcp_close(tp); - tp->t_unacksegs = 0; + so->so_head = NULL; + + tcp_close(tp); + if (inp->inp_wantcnt > 0 && inp->inp_wantcnt != WNT_STOPUSING) { + /* + * Some one has a wantcnt on this pcb. Since WNT_ACQUIRE + * doesn't require a lock, it could have happened while + * we are holding the lock. This pcb will have to + * be garbage collected later. + * Release the reference held for so_incomp queue + */ + so->so_usecount--; + tcp_unlock(so, 1, 0); + } else { + /* + * Unlock this socket and leave the reference on. + * We need to acquire the pcbinfo lock in order to + * fully dispose it off + */ + tcp_unlock(so, 0, 0); + + lck_rw_lock_exclusive(tcbinfo.ipi_lock); + + tcp_lock(so, 0, 0); + /* Release the reference held for so_incomp queue */ + so->so_usecount--; + + if (so->so_usecount != 1 || + (inp->inp_wantcnt > 0 && + inp->inp_wantcnt != WNT_STOPUSING)) { + /* + * There is an extra wantcount or usecount + * that must have been added when the socket + * was unlocked. This socket will have to be + * garbage collected later + */ + tcp_unlock(so, 1, 0); + } else { + + /* Drop the reference held for this function */ + so->so_usecount--; + + in_pcbdispose(inp); + } + lck_rw_done(tcbinfo.ipi_lock); + } tcpstat.tcps_drops++; - tcp_canceltimers(tp); - add_to_time_wait(tp); - - tcp_unlock(so, 1, 0); + tcp_lock(head, 0, 0); head->so_incqlen--; head->so_qlen--; - return 1; + return(1); +} + +/* Set background congestion control on a socket */ +void +tcp_set_background_cc(struct socket *so) +{ + tcp_set_new_cc(so, TCP_CC_ALGO_BACKGROUND_INDEX); +} + +/* Set foreground congestion control on a socket */ +void +tcp_set_foreground_cc(struct socket *so) +{ + if (tcp_use_newreno) + tcp_set_new_cc(so, TCP_CC_ALGO_NEWRENO_INDEX); + else + tcp_set_new_cc(so, TCP_CC_ALGO_CUBIC_INDEX); +} + +static void +tcp_set_new_cc(struct socket *so, uint16_t cc_index) +{ + struct inpcb *inp = sotoinpcb(so); + struct tcpcb *tp = intotcpcb(inp); + u_char old_cc_index = 0; + if (tp->tcp_cc_index != cc_index) { + + old_cc_index = tp->tcp_cc_index; + + if (CC_ALGO(tp)->cleanup != NULL) + CC_ALGO(tp)->cleanup(tp); + tp->tcp_cc_index = cc_index; + + tcp_cc_allocate_state(tp); + + if (CC_ALGO(tp)->switch_to != NULL) + CC_ALGO(tp)->switch_to(tp, old_cc_index); + + tcp_ccdbg_trace(tp, NULL, TCP_CC_CHANGE_ALGO); + } +} + +void +tcp_set_recv_bg(struct socket *so) +{ + if (!IS_TCP_RECV_BG(so)) + so->so_traffic_mgt_flags |= TRAFFIC_MGT_TCP_RECVBG; + + /* Unset Large Receive Offload on background sockets */ + so_set_lro(so, SO_TC_BK); +} + +void +tcp_clear_recv_bg(struct socket *so) +{ + if (IS_TCP_RECV_BG(so)) + so->so_traffic_mgt_flags &= ~(TRAFFIC_MGT_TCP_RECVBG); + + /* + * Set/unset use of Large Receive Offload depending on + * the traffic class + */ + so_set_lro(so, so->so_traffic_class); } +void +inp_fc_unthrottle_tcp(struct inpcb *inp) +{ + struct tcpcb *tp = inp->inp_ppcb; + /* + * Back off the slow-start threshold and enter + * congestion avoidance phase + */ + if (CC_ALGO(tp)->pre_fr != NULL) + CC_ALGO(tp)->pre_fr(tp); + + tp->snd_cwnd = tp->snd_ssthresh; + + /* + * Restart counting for ABC as we changed the + * congestion window just now. + */ + tp->t_bytes_acked = 0; + + /* Reset retransmit shift as we know that the reason + * for delay in sending a packet is due to flow + * control on the outgoing interface. There is no need + * to backoff retransmit timer. + */ + tp->t_rxtshift = 0; + + /* + * Start the output stream again. Since we are + * not retransmitting data, do not reset the + * retransmit timer or rtt calculation. + */ + tcp_output(tp); +} + static int tcp_getstat SYSCTL_HANDLER_ARGS { @@ -3898,6 +5506,41 @@ tcp_getstat SYSCTL_HANDLER_ARGS int error; + proc_t caller = PROC_NULL; + proc_t caller_parent = PROC_NULL; + char command_name[MAXCOMLEN + 1] = ""; + char parent_name[MAXCOMLEN + 1] = ""; + + if ((caller = proc_self()) != PROC_NULL) { + /* get process name */ + strlcpy(command_name, caller->p_comm, sizeof(command_name)); + + /* get parent process name if possible */ + if ((caller_parent = proc_find(caller->p_ppid)) != PROC_NULL) { + strlcpy(parent_name, caller_parent->p_comm, + sizeof(parent_name)); + proc_rele(caller_parent); + } + + if ((escape_str(command_name, strlen(command_name), + sizeof(command_name)) == 0) && + (escape_str(parent_name, strlen(parent_name), + sizeof(parent_name)) == 0)) { + kern_asl_msg(LOG_DEBUG, "messagetracer", + 5, + "com.apple.message.domain", + "com.apple.kernel.tcpstat", /* 1 */ + "com.apple.message.signature", + "tcpstat", /* 2 */ + "com.apple.message.signature2", command_name, /* 3 */ + "com.apple.message.signature3", parent_name, /* 4 */ + "com.apple.message.summarize", "YES", /* 5 */ + NULL); + } + } + if (caller != PROC_NULL) + proc_rele(caller); + if (req->oldptr == 0) { req->oldlen= (size_t)sizeof(struct tcpstat); } @@ -3908,7 +5551,155 @@ tcp_getstat SYSCTL_HANDLER_ARGS } -SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RD, 0, 0, +/* + * Checksum extended TCP header and data. + */ +int +tcp_input_checksum(int af, struct mbuf *m, struct tcphdr *th, int off, int tlen) +{ + struct ifnet *ifp = m->m_pkthdr.rcvif; + + switch (af) { + case AF_INET: { + struct ip *ip = mtod(m, struct ip *); + struct ipovly *ipov = (struct ipovly *)ip; + + if (m->m_pkthdr.pkt_flags & PKTF_SW_LRO_DID_CSUM) + return (0); + + if ((hwcksum_rx || (ifp->if_flags & IFF_LOOPBACK) || + (m->m_pkthdr.pkt_flags & PKTF_LOOP)) && + (m->m_pkthdr.csum_flags & CSUM_DATA_VALID)) { + if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) { + th->th_sum = m->m_pkthdr.csum_rx_val; + } else { + uint16_t sum = m->m_pkthdr.csum_rx_val; + uint16_t start = m->m_pkthdr.csum_rx_start; + + /* + * Perform 1's complement adjustment of octets + * that got included/excluded in the hardware- + * calculated checksum value. Ignore cases + * where the value includes or excludes the IP + * header span, as the sum for those octets + * would already be 0xffff and thus no-op. + */ + if ((m->m_pkthdr.csum_flags & CSUM_PARTIAL) && + start != 0 && (off - start) != off) { +#if BYTE_ORDER != BIG_ENDIAN + if (start < off) { + HTONS(ip->ip_len); + HTONS(ip->ip_off); + } +#endif + /* callee folds in sum */ + sum = m_adj_sum16(m, start, off, sum); +#if BYTE_ORDER != BIG_ENDIAN + if (start < off) { + NTOHS(ip->ip_off); + NTOHS(ip->ip_len); + } +#endif + } + + /* callee folds in sum */ + th->th_sum = in_pseudo(ip->ip_src.s_addr, + ip->ip_dst.s_addr, + sum + htonl(tlen + IPPROTO_TCP)); + } + th->th_sum ^= 0xffff; + } else { + uint16_t ip_sum; + int len; + char b[9]; + + bcopy(ipov->ih_x1, b, sizeof (ipov->ih_x1)); + bzero(ipov->ih_x1, sizeof (ipov->ih_x1)); + ip_sum = ipov->ih_len; + ipov->ih_len = (u_short)tlen; +#if BYTE_ORDER != BIG_ENDIAN + HTONS(ipov->ih_len); +#endif + len = sizeof (struct ip) + tlen; + th->th_sum = in_cksum(m, len); + bcopy(b, ipov->ih_x1, sizeof (ipov->ih_x1)); + ipov->ih_len = ip_sum; + + tcp_in_cksum_stats(len); + } + break; + } +#if INET6 + case AF_INET6: { + struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *); + + if (m->m_pkthdr.pkt_flags & PKTF_SW_LRO_DID_CSUM) + return (0); + + if ((hwcksum_rx || (ifp->if_flags & IFF_LOOPBACK) || + (m->m_pkthdr.pkt_flags & PKTF_LOOP)) && + (m->m_pkthdr.csum_flags & CSUM_DATA_VALID)) { + if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) { + th->th_sum = m->m_pkthdr.csum_rx_val; + } else { + uint16_t sum = m->m_pkthdr.csum_rx_val; + uint16_t start = m->m_pkthdr.csum_rx_start; + + /* + * Perform 1's complement adjustment of octets + * that got included/excluded in the hardware- + * calculated checksum value. + */ + if ((m->m_pkthdr.csum_flags & CSUM_PARTIAL) && + start != off) { + uint16_t s, d; + + if (IN6_IS_SCOPE_EMBED(&ip6->ip6_src)) { + s = ip6->ip6_src.s6_addr16[1]; + ip6->ip6_src.s6_addr16[1] = 0 ; + } + if (IN6_IS_SCOPE_EMBED(&ip6->ip6_dst)) { + d = ip6->ip6_dst.s6_addr16[1]; + ip6->ip6_dst.s6_addr16[1] = 0; + } + + /* callee folds in sum */ + sum = m_adj_sum16(m, start, off, sum); + + if (IN6_IS_SCOPE_EMBED(&ip6->ip6_src)) + ip6->ip6_src.s6_addr16[1] = s; + if (IN6_IS_SCOPE_EMBED(&ip6->ip6_dst)) + ip6->ip6_dst.s6_addr16[1] = d; + } + + th->th_sum = in6_pseudo( + &ip6->ip6_src, &ip6->ip6_dst, + sum + htonl(tlen + IPPROTO_TCP)); + } + th->th_sum ^= 0xffff; + } else { + tcp_in6_cksum_stats(tlen); + th->th_sum = in6_cksum(m, IPPROTO_TCP, off, tlen); + } + break; + } +#endif /* INET6 */ + default: + VERIFY(0); + /* NOTREACHED */ + } + + if (th->th_sum != 0) { + tcpstat.tcps_rcvbadsum++; + IF_TCP_STATINC(ifp, badformat); + return (-1); + } + + return (0); +} + +SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, + CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_LOCKED, 0, 0, tcp_getstat, "S,tcpstat", "TCP statistics (struct tcpstat, netinet/tcp_var.h)"); static int @@ -3936,5 +5727,5 @@ sysctl_rexmtthresh SYSCTL_HANDLER_ARGS return (0); } -SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmt_thresh, CTLTYPE_INT|CTLFLAG_RW, +SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmt_thresh, CTLTYPE_INT|CTLFLAG_RW | CTLFLAG_LOCKED, &tcprexmtthresh, 0, &sysctl_rexmtthresh, "I", "Duplicate ACK Threshold for Fast Retransmit");