X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/39236c6e673c41db228275375ab7fdb0f837b292..743345f9a4b36f7e2f9ba37691e70c50baecb56e:/bsd/netinet/tcp_timer.c diff --git a/bsd/netinet/tcp_timer.c b/bsd/netinet/tcp_timer.c index bc2942705..e50bab301 100644 --- a/bsd/netinet/tcp_timer.c +++ b/bsd/netinet/tcp_timer.c @@ -1,8 +1,8 @@ /* - * Copyright (c) 2000-2013 Apple Inc. All rights reserved. + * Copyright (c) 2000-2016 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ - * + * * This file contains Original Code and/or Modifications of Original Code * as defined in and that are subject to the Apple Public Source License * Version 2.0 (the 'License'). You may not use this file except in @@ -11,10 +11,10 @@ * unlawful or unlicensed copies of an Apple operating system, or to * circumvent, violate, or enable the circumvention or violation of, any * terms of an Apple operating system software license agreement. - * + * * Please obtain a copy of the License at * http://www.opensource.apple.com/apsl/ and read it before using this file. - * + * * The Original Code and all software distributed under the License are * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, @@ -22,7 +22,7 @@ * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. * Please see the License for the specific language governing rights and * limitations under the License. - * + * * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ /* @@ -79,6 +79,7 @@ #include #include +#include #include #include @@ -88,6 +89,7 @@ #endif #include #include +#include #include #include #include @@ -104,24 +106,14 @@ #include #include -extern void postevent(struct socket *, struct sockbuf *, - int); -#define DBG_FNC_TCP_FAST NETDBG_CODE(DBG_NETTCP, (5 << 8)) -#define DBG_FNC_TCP_SLOW NETDBG_CODE(DBG_NETTCP, (5 << 8) | 1) - -#define TIMERENTRY_TO_TP(te) ((struct tcpcb *)((uintptr_t)te - offsetof(struct tcpcb, tentry.le.le_next))) - -#define VERIFY_NEXT_LINK(elm,field) do { \ - if (LIST_NEXT((elm),field) != NULL && \ - LIST_NEXT((elm),field)->field.le_prev != \ - &((elm)->field.le_next)) \ - panic("Bad link elm %p next->prev != elm", (elm)); \ -} while(0) +/* Max number of times a stretch ack can be delayed on a connection */ +#define TCP_STRETCHACK_DELAY_THRESHOLD 5 -#define VERIFY_PREV_LINK(elm,field) do { \ - if (*(elm)->field.le_prev != (elm)) \ - panic("Bad link elm %p prev->next != elm", (elm)); \ -} while(0) +/* + * If the host processor has been sleeping for too long, this is the threshold + * used to avoid sending stale retransmissions. + */ +#define TCP_SLEEP_TOO_LONG (10 * 60 * 1000) /* 10 minutes in ms */ /* tcp timer list */ struct tcptimerlist tcp_timer_list; @@ -129,10 +121,6 @@ struct tcptimerlist tcp_timer_list; /* List of pcbs in timewait state, protected by tcbinfo's ipi_lock */ struct tcptailq tcp_tw_tailq; -static int background_io_trigger = 5; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, background_io_trigger, CTLFLAG_RW | CTLFLAG_LOCKED, - &background_io_trigger, 0, "Background IO Trigger Setting"); - static int sysctl_msec_to_ticks SYSCTL_HANDLER_ARGS { @@ -155,45 +143,59 @@ sysctl_msec_to_ticks SYSCTL_HANDLER_ARGS } int tcp_keepinit; -SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, +SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit, + CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_keepinit, 0, sysctl_msec_to_ticks, "I", ""); int tcp_keepidle; -SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, +SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle, + CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_keepidle, 0, sysctl_msec_to_ticks, "I", ""); int tcp_keepintvl; -SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, +SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl, + CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I", ""); int tcp_keepcnt; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, - &tcp_keepcnt, 0, "number of times to repeat keepalive"); +SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, + CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_keepcnt, 0, "number of times to repeat keepalive"); int tcp_msl; -SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, +SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, + CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_msl, 0, sysctl_msec_to_ticks, "I", "Maximum segment lifetime"); -/* - * Avoid DoS via TCP Robustness in Persist Condition (see http://www.ietf.org/id/draft-ananth-tcpm-persist-02.txt) - * by allowing a system wide maximum persistence timeout value when in Zero Window Probe mode. - * Expressed in milliseconds to be consistent without timeout related values, the TCP socket option is in seconds. +/* + * Avoid DoS via TCP Robustness in Persist Condition + * (see http://www.ietf.org/id/draft-ananth-tcpm-persist-02.txt) + * by allowing a system wide maximum persistence timeout value when in + * Zero Window Probe mode. + * + * Expressed in milliseconds to be consistent without timeout related + * values, the TCP socket option is in seconds. */ u_int32_t tcp_max_persist_timeout = 0; -SYSCTL_PROC(_net_inet_tcp, OID_AUTO, max_persist_timeout, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, - &tcp_max_persist_timeout, 0, sysctl_msec_to_ticks, "I", "Maximum persistence timout for ZWP"); +SYSCTL_PROC(_net_inet_tcp, OID_AUTO, max_persist_timeout, + CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, + &tcp_max_persist_timeout, 0, sysctl_msec_to_ticks, "I", + "Maximum persistence timeout for ZWP"); static int always_keepalive = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_RW | CTLFLAG_LOCKED, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, + CTLFLAG_RW | CTLFLAG_LOCKED, &always_keepalive , 0, "Assume SO_KEEPALIVE on all TCP connections"); -/* This parameter determines how long the timer list will stay in fast mode even - * though all connections are idle. In fast mode, the timer will fire more frequently - * anticipating new data. +/* + * This parameter determines how long the timer list will stay in fast or + * quick mode even though all connections are idle. In this state, the + * timer will run more frequently anticipating new data. */ -int timer_fastmode_idlemax = TCP_FASTMODE_IDLEGEN_MAX; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, timer_fastmode_idlemax, CTLFLAG_RW | CTLFLAG_LOCKED, - &timer_fastmode_idlemax, 0, "Maximum idle generations in fast mode"); +int timer_fastmode_idlemax = TCP_FASTMODE_IDLERUN_MAX; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, timer_fastmode_idlemax, + CTLFLAG_RW | CTLFLAG_LOCKED, + &timer_fastmode_idlemax, 0, "Maximum idle generations in fast mode"); /* * See tcp_syn_backoff[] for interval values between SYN retransmits; @@ -202,43 +204,48 @@ SYSCTL_INT(_net_inet_tcp, OID_AUTO, timer_fastmode_idlemax, CTLFLAG_RW | CTLFLAG * SYN retransmits. Setting it to 0 disables the dropping off of those * two options. */ -static int tcp_broken_peer_syn_rxmit_thres = 7; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, broken_peer_syn_rxmit_thres, CTLFLAG_RW | CTLFLAG_LOCKED, - &tcp_broken_peer_syn_rxmit_thres, 0, "Number of retransmitted SYNs before " - "TCP disables rfc1323 and rfc1644 during the rest of attempts"); - -/* A higher threshold on local connections for disabling RFC 1323 options */ -static int tcp_broken_peer_syn_rxmit_thres_local = 10; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, broken_peer_syn_rexmit_thres_local, - CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_broken_peer_syn_rxmit_thres_local, 0, - "Number of retransmitted SYNs before disabling RFC 1323 options on local connections"); +static int tcp_broken_peer_syn_rxmit_thres = 10; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, broken_peer_syn_rexmit_thres, + CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_broken_peer_syn_rxmit_thres, 0, + "Number of retransmitted SYNs before disabling RFC 1323 " + "options on local connections"); static int tcp_timer_advanced = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_timer_advanced, CTLFLAG_RD | CTLFLAG_LOCKED, - &tcp_timer_advanced, 0, "Number of times one of the timers was advanced"); +SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_timer_advanced, + CTLFLAG_RD | CTLFLAG_LOCKED, &tcp_timer_advanced, 0, + "Number of times one of the timers was advanced"); static int tcp_resched_timerlist = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_resched_timerlist, CTLFLAG_RD | CTLFLAG_LOCKED, - &tcp_resched_timerlist, 0, +SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcp_resched_timerlist, + CTLFLAG_RD | CTLFLAG_LOCKED, &tcp_resched_timerlist, 0, "Number of times timer list was rescheduled as part of processing a packet"); int tcp_pmtud_black_hole_detect = 1 ; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection, CTLFLAG_RW | CTLFLAG_LOCKED, - &tcp_pmtud_black_hole_detect, 0, "Path MTU Discovery Black Hole Detection"); +SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection, + CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_pmtud_black_hole_detect, 0, + "Path MTU Discovery Black Hole Detection"); int tcp_pmtud_black_hole_mss = 1200 ; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss, CTLFLAG_RW | CTLFLAG_LOCKED, - &tcp_pmtud_black_hole_mss, 0, "Path MTU Discovery Black Hole Detection lowered MSS"); +SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss, + CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_pmtud_black_hole_mss, 0, + "Path MTU Discovery Black Hole Detection lowered MSS"); + +static u_int32_t tcp_mss_rec_medium = 1200; +static u_int32_t tcp_mss_rec_low = 512; + +#define TCP_REPORT_STATS_INTERVAL 43200 /* 12 hours, in seconds */ +int tcp_report_stats_interval = TCP_REPORT_STATS_INTERVAL; /* performed garbage collection of "used" sockets */ static boolean_t tcp_gc_done = FALSE; - /* max idle probes */ +/* max idle probes */ int tcp_maxpersistidle; -/* TCP delack timer is set to 100 ms. Since the processing of timer list in fast - * mode will happen no faster than 100 ms, the delayed ack timer will fire some where - * between 100 and 200 ms. +/* + * TCP delack timer is set to 100 ms. Since the processing of timer list + * in fast mode will happen no faster than 100 ms, the delayed ack timer + * will fire some where between 100 and 200 ms. */ int tcp_delack = TCP_RETRANSHZ / 10; @@ -249,45 +256,161 @@ int tcp_delack = TCP_RETRANSHZ / 10; int tcp_jack_rxmt = TCP_RETRANSHZ / 2; #endif /* MPTCP */ -/* The frequency of running through the TCP timer list in - * fast and slow mode can be configured. - */ -SYSCTL_UINT(_net_inet_tcp, OID_AUTO, timer_fastquantum, CTLFLAG_RW | CTLFLAG_LOCKED, - &tcp_timer_list.fast_quantum, TCP_FASTTIMER_QUANTUM, - "Frequency of running timer list in fast mode"); - -SYSCTL_UINT(_net_inet_tcp, OID_AUTO, timer_slowquantum, CTLFLAG_RW | CTLFLAG_LOCKED, - &tcp_timer_list.slow_quantum, TCP_SLOWTIMER_QUANTUM, - "Frequency of running timer list in slow mode"); +static boolean_t tcp_itimer_done = FALSE; static void tcp_remove_timer(struct tcpcb *tp); static void tcp_sched_timerlist(uint32_t offset); -static uint32_t tcp_run_conn_timer(struct tcpcb *tp, uint16_t *next_index); +static u_int32_t tcp_run_conn_timer(struct tcpcb *tp, u_int16_t *mode, + u_int16_t probe_if_index); static void tcp_sched_timers(struct tcpcb *tp); static inline void tcp_set_lotimer_index(struct tcpcb *); -static void tcp_rexmt_save_state(struct tcpcb *tp); -void tcp_remove_from_time_wait(struct inpcb *inp); +__private_extern__ void tcp_remove_from_time_wait(struct inpcb *inp); +static inline void tcp_update_mss_core(struct tcpcb *tp, struct ifnet *ifp); +__private_extern__ void tcp_report_stats(void); -/* Macro to compare two timers. If there is a reset of the sign bit, it is - * safe to assume that the timer has wrapped around. By doing signed comparision, - * we take care of wrap around such that the value with the sign bit reset is - * actually ahead of the other. - */ +static u_int64_t tcp_last_report_time; -static inline int32_t -timer_diff(uint32_t t1, uint32_t toff1, uint32_t t2, uint32_t toff2) { - return (int32_t)((t1 + toff1) - (t2 + toff2)); +/* + * Structure to store previously reported stats so that we can send + * incremental changes in each report interval. + */ +struct tcp_last_report_stats { + u_int32_t tcps_connattempt; + u_int32_t tcps_accepts; + u_int32_t tcps_ecn_client_setup; + u_int32_t tcps_ecn_server_setup; + u_int32_t tcps_ecn_client_success; + u_int32_t tcps_ecn_server_success; + u_int32_t tcps_ecn_not_supported; + u_int32_t tcps_ecn_lost_syn; + u_int32_t tcps_ecn_lost_synack; + u_int32_t tcps_ecn_recv_ce; + u_int32_t tcps_ecn_recv_ece; + u_int32_t tcps_ecn_sent_ece; + u_int32_t tcps_ecn_conn_recv_ce; + u_int32_t tcps_ecn_conn_recv_ece; + u_int32_t tcps_ecn_conn_plnoce; + u_int32_t tcps_ecn_conn_pl_ce; + u_int32_t tcps_ecn_conn_nopl_ce; + u_int32_t tcps_ecn_fallback_synloss; + u_int32_t tcps_ecn_fallback_reorder; + u_int32_t tcps_ecn_fallback_ce; + + /* TFO-related statistics */ + u_int32_t tcps_tfo_syn_data_rcv; + u_int32_t tcps_tfo_cookie_req_rcv; + u_int32_t tcps_tfo_cookie_sent; + u_int32_t tcps_tfo_cookie_invalid; + u_int32_t tcps_tfo_cookie_req; + u_int32_t tcps_tfo_cookie_rcv; + u_int32_t tcps_tfo_syn_data_sent; + u_int32_t tcps_tfo_syn_data_acked; + u_int32_t tcps_tfo_syn_loss; + u_int32_t tcps_tfo_blackhole; + u_int32_t tcps_tfo_cookie_wrong; + u_int32_t tcps_tfo_no_cookie_rcv; + u_int32_t tcps_tfo_heuristics_disable; + u_int32_t tcps_tfo_sndblackhole; }; + /* Returns true if the timer is on the timer list */ #define TIMER_IS_ON_LIST(tp) ((tp)->t_flags & TF_TIMER_ONLIST) +/* Run the TCP timerlist atleast once every hour */ +#define TCP_TIMERLIST_MAX_OFFSET (60 * 60 * TCP_RETRANSHZ) -static void add_to_time_wait_locked(struct tcpcb *tp, uint32_t delay); -void add_to_time_wait(struct tcpcb *tp, uint32_t delay) ; +static void add_to_time_wait_locked(struct tcpcb *tp, uint32_t delay); static boolean_t tcp_garbage_collect(struct inpcb *, int); +#define TIMERENTRY_TO_TP(te) ((struct tcpcb *)((uintptr_t)te - offsetof(struct tcpcb, tentry.le.le_next))) + +#define VERIFY_NEXT_LINK(elm,field) do { \ + if (LIST_NEXT((elm),field) != NULL && \ + LIST_NEXT((elm),field)->field.le_prev != \ + &((elm)->field.le_next)) \ + panic("Bad link elm %p next->prev != elm", (elm)); \ +} while(0) + +#define VERIFY_PREV_LINK(elm,field) do { \ + if (*(elm)->field.le_prev != (elm)) \ + panic("Bad link elm %p prev->next != elm", (elm)); \ +} while(0) + +#define TCP_SET_TIMER_MODE(mode, i) do { \ + if (IS_TIMER_HZ_10MS(i)) \ + (mode) |= TCP_TIMERLIST_10MS_MODE; \ + else if (IS_TIMER_HZ_100MS(i)) \ + (mode) |= TCP_TIMERLIST_100MS_MODE; \ + else \ + (mode) |= TCP_TIMERLIST_500MS_MODE; \ +} while(0) + +#if (DEVELOPMENT || DEBUG) +SYSCTL_UINT(_net_inet_tcp, OID_AUTO, mss_rec_medium, + CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_mss_rec_medium, 0, + "Medium MSS based on recommendation in link status report"); +SYSCTL_UINT(_net_inet_tcp, OID_AUTO, mss_rec_low, + CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_mss_rec_low, 0, + "Low MSS based on recommendation in link status report"); + +static int32_t tcp_change_mss_recommended = 0; +static int +sysctl_change_mss_recommended SYSCTL_HANDLER_ARGS +{ +#pragma unused(oidp, arg1, arg2) + int i, err = 0, changed = 0; + struct ifnet *ifp; + struct if_link_status ifsr; + struct if_cellular_status_v1 *new_cell_sr; + err = sysctl_io_number(req, tcp_change_mss_recommended, + sizeof (int32_t), &i, &changed); + if (changed) { + ifnet_head_lock_shared(); + TAILQ_FOREACH(ifp, &ifnet_head, if_link) { + if (IFNET_IS_CELLULAR(ifp)) { + bzero(&ifsr, sizeof (ifsr)); + new_cell_sr = &ifsr.ifsr_u.ifsr_cell.if_cell_u.if_status_v1; + ifsr.ifsr_version = IF_CELLULAR_STATUS_REPORT_CURRENT_VERSION; + ifsr.ifsr_len = sizeof(*new_cell_sr); + + /* Set MSS recommended */ + new_cell_sr->valid_bitmask |= IF_CELL_UL_MSS_RECOMMENDED_VALID; + new_cell_sr->mss_recommended = i; + err = ifnet_link_status_report(ifp, new_cell_sr, sizeof (new_cell_sr)); + if (err == 0) { + tcp_change_mss_recommended = i; + } else { + break; + } + } + } + ifnet_head_done(); + } + return (err); +} + +SYSCTL_PROC(_net_inet_tcp, OID_AUTO, change_mss_recommended, + CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_change_mss_recommended, + 0, sysctl_change_mss_recommended, "IU", "Change MSS recommended"); + +SYSCTL_INT(_net_inet_tcp, OID_AUTO, report_stats_interval, + CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_report_stats_interval, 0, + "Report stats interval"); +#endif /* (DEVELOPMENT || DEBUG) */ + +/* + * Macro to compare two timers. If there is a reset of the sign bit, + * it is safe to assume that the timer has wrapped around. By doing + * signed comparision, we take care of wrap around such that the value + * with the sign bit reset is actually ahead of the other. + */ +inline int32_t +timer_diff(uint32_t t1, uint32_t toff1, uint32_t t2, uint32_t toff2) { + return (int32_t)((t1 + toff1) - (t2 + toff2)); +}; + /* * Add to tcp timewait list, delay is given in milliseconds. */ @@ -305,7 +428,7 @@ add_to_time_wait_locked(struct tcpcb *tp, uint32_t delay) if (!(inp->inp_flags2 & INP2_TIMEWAIT)) { pcbinfo->ipi_twcount++; inp->inp_flags2 |= INP2_TIMEWAIT; - + /* Remove from global inp list */ LIST_REMOVE(inp, inp_list); } else { @@ -314,7 +437,7 @@ add_to_time_wait_locked(struct tcpcb *tp, uint32_t delay) /* Compute the time at which this socket can be closed */ timer = tcp_now + delay; - + /* We will use the TCPT_2MSL timer for tracking this delay */ if (TIMER_IS_ON_LIST(tp)) @@ -328,6 +451,11 @@ void add_to_time_wait(struct tcpcb *tp, uint32_t delay) { struct inpcbinfo *pcbinfo = &tcbinfo; + if (tp->t_inpcb->inp_socket->so_options & SO_NOWAKEFROMSLEEP) + socket_post_kev_msg_closed(tp->t_inpcb->inp_socket); + + /* 19182803: Notify nstat that connection is closing before waiting. */ + nstat_pcb_detach(tp->t_inpcb); if (!lck_rw_try_lock_exclusive(pcbinfo->ipi_lock)) { tcp_unlock(tp->t_inpcb->inp_socket, 0, 0); @@ -422,13 +550,13 @@ tcp_garbage_collect(struct inpcb *inp, int istimewait) } /* - * We get here because the PCB is no longer searchable - * (WNT_STOPUSING); detach (if needed) and dispose if it is dead - * (usecount is 0). This covers all cases, including overflow - * sockets and those that are considered as "embryonic", - * i.e. created by sonewconn() in TCP input path, and have + * We get here because the PCB is no longer searchable + * (WNT_STOPUSING); detach (if needed) and dispose if it is dead + * (usecount is 0). This covers all cases, including overflow + * sockets and those that are considered as "embryonic", + * i.e. created by sonewconn() in TCP input path, and have * not yet been committed. For the former, we reduce the usecount - * to 0 as done by the code above. For the latter, the usecount + * to 0 as done by the code above. For the latter, the usecount * would have reduced to 0 as part calling soabort() when the * socket is dropped at the end of tcp_input(). */ @@ -439,7 +567,7 @@ tcp_garbage_collect(struct inpcb *inp, int istimewait) lck_mtx_convert_spin(&inp->inpcb_mtx); /* - * If this tp still happens to be on the timer list, + * If this tp still happens to be on the timer list, * take it out */ if (TIMER_IS_ON_LIST(tp)) { @@ -516,16 +644,19 @@ tcp_gc(struct inpcbinfo *ipi) /* Now cleanup the time wait ones */ TAILQ_FOREACH_SAFE(tw_tp, &tcp_tw_tailq, t_twentry, tw_ntp) { /* - * We check the timestamp here without holding the + * We check the timestamp here without holding the * socket lock for better performance. If there are * any pcbs in time-wait, the timer will get rescheduled. * Hence some error in this check can be tolerated. + * + * Sometimes a socket on time-wait queue can be closed if + * 2MSL timer expired but the application still has a + * usecount on it. */ - if (TSTMP_GEQ(tcp_now, tw_tp->t_timer[TCPT_2MSL])) { + if (tw_tp->t_state == TCPS_CLOSED || + TSTMP_GEQ(tcp_now, tw_tp->t_timer[TCPT_2MSL])) { if (tcp_garbage_collect(tw_tp->t_inpcb, 1)) atomic_add_32(&ipi->ipi_gc_req.intimer_lazy, 1); - } else { - break; } } @@ -548,10 +679,9 @@ tcp_gc(struct inpcbinfo *ipi) * Cancel all timers for TCP tp. */ void -tcp_canceltimers(tp) - struct tcpcb *tp; +tcp_canceltimers(struct tcpcb *tp) { - register int i; + int i; tcp_remove_timer(tp); for (i = 0; i < TCPT_NTIMERS; i++) @@ -568,12 +698,13 @@ int tcp_backoff[TCP_MAXRXTSHIFT + 1] = static int tcp_totbackoff = 511; /* sum of tcp_backoff[] */ -static void tcp_rexmt_save_state(struct tcpcb *tp) +void +tcp_rexmt_save_state(struct tcpcb *tp) { u_int32_t fsize; if (TSTMP_SUPPORTED(tp)) { /* - * Since timestamps are supported on the connection, + * Since timestamps are supported on the connection, * we can do recovery as described in rfc 4015. */ fsize = tp->snd_max - tp->snd_una; @@ -584,9 +715,9 @@ static void tcp_rexmt_save_state(struct tcpcb *tp) * Timestamp option is not supported on this connection. * Record ssthresh and cwnd so they can * be recovered if this turns out to be a "bad" retransmit. - * A retransmit is considered "bad" if an ACK for this + * A retransmit is considered "bad" if an ACK for this * segment is received within RTT/2 interval; the assumption - * here is that the ACK was already in flight. See + * here is that the ACK was already in flight. See * "On Estimating End-to-End Network Path Properties" by * Allman and Paxson for more details. */ @@ -603,20 +734,40 @@ static void tcp_rexmt_save_state(struct tcpcb *tp) tp->t_flagsext &= ~(TF_RECOMPUTE_RTT); } +/* + * Revert to the older segment size if there is an indication that PMTU + * blackhole detection was not needed. + */ +void +tcp_pmtud_revert_segment_size(struct tcpcb *tp) +{ + int32_t optlen; + + VERIFY(tp->t_pmtud_saved_maxopd > 0); + tp->t_flags |= TF_PMTUD; + tp->t_flags &= ~TF_BLACKHOLE; + optlen = tp->t_maxopd - tp->t_maxseg; + tp->t_maxopd = tp->t_pmtud_saved_maxopd; + tp->t_maxseg = tp->t_maxopd - optlen; + /* + * Reset the slow-start flight size as it + * may depend on the new MSS + */ + if (CC_ALGO(tp)->cwnd_init != NULL) + CC_ALGO(tp)->cwnd_init(tp); + tp->t_pmtud_start_ts = 0; + tcpstat.tcps_pmtudbh_reverted++; +} + /* * TCP timer processing. */ struct tcpcb * -tcp_timers(tp, timer) - register struct tcpcb *tp; - int timer; +tcp_timers(struct tcpcb *tp, int timer) { - register int rexmt; + int32_t rexmt, optlen = 0, idle_time = 0; struct socket *so; struct tcptemp *t_template; - int optlen = 0; - int idle_time = 0; - #if TCPDEBUG int ostate; #endif @@ -624,6 +775,8 @@ tcp_timers(tp, timer) #if INET6 int isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV4) == 0; #endif /* INET6 */ + u_int64_t accsleep_ms; + u_int32_t last_sleep_ms = 0; so = tp->t_inpcb->inp_socket; idle_time = tcp_now - tp->t_rcvtime; @@ -643,7 +796,7 @@ tcp_timers(tp, timer) if (tp->t_state != TCPS_TIME_WAIT && tp->t_state != TCPS_FIN_WAIT_2 && ((idle_time > 0) && (idle_time < TCP_CONN_MAXIDLE(tp)))) { - tp->t_timer[TCPT_2MSL] = OFFSET_FROM_START(tp, + tp->t_timer[TCPT_2MSL] = OFFSET_FROM_START(tp, (u_int32_t)TCP_CONN_KEEPINTVL(tp)); } else { tp = tcp_close(tp); @@ -657,29 +810,53 @@ tcp_timers(tp, timer) * to a longer retransmit interval and retransmit one segment. */ case TCPT_REXMT: - /* Drop a connection in the retransmit timer - * 1. If we have retransmitted more than TCP_MAXRXTSHIFT times - * 2. If the time spent in this retransmission episode is more than - * the time limit set with TCP_RXT_CONNDROPTIME socket option - * 3. If TCP_RXT_FINDROP socket option was set and we have already - * retransmitted the FIN 3 times without receiving an ack + absolutetime_to_nanoseconds(mach_absolutetime_asleep, + &accsleep_ms); + accsleep_ms = accsleep_ms / 1000000UL; + if (accsleep_ms > tp->t_accsleep_ms) + last_sleep_ms = accsleep_ms - tp->t_accsleep_ms; + /* + * Drop a connection in the retransmit timer + * 1. If we have retransmitted more than TCP_MAXRXTSHIFT + * times + * 2. If the time spent in this retransmission episode is + * more than the time limit set with TCP_RXT_CONNDROPTIME + * socket option + * 3. If TCP_RXT_FINDROP socket option was set and + * we have already retransmitted the FIN 3 times without + * receiving an ack */ if (++tp->t_rxtshift > TCP_MAXRXTSHIFT || - (tp->t_rxt_conndroptime > 0 && tp->t_rxtstart > 0 && - (tcp_now - tp->t_rxtstart) >= tp->t_rxt_conndroptime) || - ((tp->t_flagsext & TF_RXTFINDROP) != 0 && - (tp->t_flags & TF_SENTFIN) != 0 && - tp->t_rxtshift >= 4)) { - + (tp->t_rxt_conndroptime > 0 && tp->t_rxtstart > 0 && + (tcp_now - tp->t_rxtstart) >= tp->t_rxt_conndroptime) || + ((tp->t_flagsext & TF_RXTFINDROP) != 0 && + (tp->t_flags & TF_SENTFIN) != 0 && tp->t_rxtshift >= 4) || + (tp->t_rxtshift > 4 && last_sleep_ms >= TCP_SLEEP_TOO_LONG)) { if ((tp->t_flagsext & TF_RXTFINDROP) != 0) { tcpstat.tcps_rxtfindrop++; + } else if (last_sleep_ms >= TCP_SLEEP_TOO_LONG) { + tcpstat.tcps_drop_after_sleep++; } else { tcpstat.tcps_timeoutdrop++; } + if (tp->t_rxtshift >= TCP_MAXRXTSHIFT) { + if (TCP_ECN_ENABLED(tp)) { + INP_INC_IFNET_STAT(tp->t_inpcb, + ecn_on.rxmit_drop); + } else { + INP_INC_IFNET_STAT(tp->t_inpcb, + ecn_off.rxmit_drop); + } + } tp->t_rxtshift = TCP_MAXRXTSHIFT; - postevent(so, 0, EV_TIMEOUT); - soevent(so, + postevent(so, 0, EV_TIMEOUT); + soevent(so, (SO_FILT_HINT_LOCKED|SO_FILT_HINT_TIMEOUT)); + + if (TCP_ECN_ENABLED(tp) && + tp->t_state == TCPS_ESTABLISHED) + tcp_heuristic_ecn_droprxmt(tp); + tp = tcp_drop(tp, tp->t_softerror ? tp->t_softerror : ETIMEDOUT); @@ -687,21 +864,22 @@ tcp_timers(tp, timer) } tcpstat.tcps_rexmttimeo++; + tp->t_accsleep_ms = accsleep_ms; - if (tp->t_rxtshift == 1 && + if (tp->t_rxtshift == 1 && tp->t_state == TCPS_ESTABLISHED) { /* Set the time at which retransmission started. */ tp->t_rxtstart = tcp_now; - /* + /* * if this is the first retransmit timeout, save * the state so that we can recover if the timeout * is spurious. - */ + */ tcp_rexmt_save_state(tp); } #if MPTCP - if ((tp->t_rxtshift == mptcp_fail_thresh) && + if ((tp->t_rxtshift >= mptcp_fail_thresh) && (tp->t_state == TCPS_ESTABLISHED) && (tp->t_mpflags & TMPF_MPTCP_TRUE)) { mptcp_act_on_txfail(so); @@ -718,14 +896,67 @@ tcp_timers(tp, timer) SO_FILT_HINT_ADAPTIVE_WTIMO)); } + /* + * If this is a retransmit timeout after PTO, the PTO + * was not effective + */ + if (tp->t_flagsext & TF_SENT_TLPROBE) { + tp->t_flagsext &= ~(TF_SENT_TLPROBE); + tcpstat.tcps_rto_after_pto++; + } + + if (tp->t_flagsext & TF_DELAY_RECOVERY) { + /* + * Retransmit timer fired before entering recovery + * on a connection with packet re-ordering. This + * suggests that the reordering metrics computed + * are not accurate. + */ + tp->t_reorderwin = 0; + tp->t_timer[TCPT_DELAYFR] = 0; + tp->t_flagsext &= ~(TF_DELAY_RECOVERY); + } + + if (tp->t_state == TCPS_SYN_RECEIVED) + tcp_disable_tfo(tp); + + if ((tp->t_tfo_stats & TFO_S_SYN_DATA_SENT) && + !(tp->t_tfo_flags & TFO_F_NO_SNDPROBING) && + ((tp->t_state != TCPS_SYN_SENT && tp->t_rxtshift > 1) || + tp->t_rxtshift > 2)) { + /* + * For regular retransmissions, a first one is being + * done for tail-loss probe. + * Thus, if rxtshift > 1, this means we have sent the segment + * a total of 3 times. + * + * If we are in SYN-SENT state, then there is no tail-loss + * probe thus we have to let rxtshift go up to 3. + */ + tcp_heuristic_tfo_middlebox(tp); + + so->so_error = ENODATA; + sorwakeup(so); + sowwakeup(so); + + tp->t_tfo_stats |= TFO_S_SEND_BLACKHOLE; + tcpstat.tcps_tfo_sndblackhole++; + } + if (tp->t_state == TCPS_SYN_SENT) { rexmt = TCP_REXMTVAL(tp) * tcp_syn_backoff[tp->t_rxtshift]; tp->t_stat.synrxtshift = tp->t_rxtshift; - } - else + + /* When retransmitting, disable TFO */ + if (tfo_enabled(tp)) { + tp->t_flagsext &= ~TF_FASTOPEN; + tp->t_tfo_flags |= TFO_F_SYN_LOSS; + } + } else { rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; - TCPT_RANGESET(tp->t_rxtcur, rexmt, - tp->t_rttmin, TCPTV_REXMTMAX, + } + + TCPT_RANGESET(tp->t_rxtcur, rexmt, tp->t_rttmin, TCPTV_REXMTMAX, TCP_ADD_REXMTSLOP(tp)); tp->t_timer[TCPT_REXMT] = OFFSET_FROM_START(tp, tp->t_rxtcur); @@ -734,16 +965,20 @@ tcp_timers(tp, timer) tcp_free_sackholes(tp); /* - * Check for potential Path MTU Discovery Black Hole + * Check for potential Path MTU Discovery Black Hole */ - - if (tcp_pmtud_black_hole_detect && (tp->t_state == TCPS_ESTABLISHED)) { - if (((tp->t_flags & (TF_PMTUD|TF_MAXSEGSNT)) == (TF_PMTUD|TF_MAXSEGSNT)) && - (tp->t_rxtshift == 2)) { - /* + if (tcp_pmtud_black_hole_detect && + !(tp->t_flagsext & TF_NOBLACKHOLE_DETECTION) && + (tp->t_state == TCPS_ESTABLISHED)) { + if ((tp->t_flags & TF_PMTUD) && + ((tp->t_flags & TF_MAXSEGSNT) + || tp->t_pmtud_lastseg_size > tcp_pmtud_black_hole_mss) && + tp->t_rxtshift == 2) { + /* * Enter Path MTU Black-hole Detection mechanism: * - Disable Path MTU Discovery (IP "DF" bit). - * - Reduce MTU to lower value than what we negociated with peer. + * - Reduce MTU to lower value than what we + * negotiated with the peer. */ /* Disable Path MTU Discovery for now */ tp->t_flags &= ~TF_PMTUD; @@ -752,6 +987,9 @@ tcp_timers(tp, timer) optlen = tp->t_maxopd - tp->t_maxseg; /* Keep track of previous MSS */ tp->t_pmtud_saved_maxopd = tp->t_maxopd; + tp->t_pmtud_start_ts = tcp_now; + if (tp->t_pmtud_start_ts == 0) + tp->t_pmtud_start_ts++; /* Reduce the MSS to intermediary value */ if (tp->t_maxopd > tcp_pmtud_black_hole_mss) { tp->t_maxopd = tcp_pmtud_black_hole_mss; @@ -765,49 +1003,41 @@ tcp_timers(tp, timer) tp->t_maxseg = tp->t_maxopd - optlen; /* - * Reset the slow-start flight size + * Reset the slow-start flight size * as it may depend on the new MSS */ if (CC_ALGO(tp)->cwnd_init != NULL) CC_ALGO(tp)->cwnd_init(tp); + tp->snd_cwnd = tp->t_maxseg; } /* - * If further retransmissions are still unsuccessful with a lowered MTU, - * maybe this isn't a Black Hole and we restore the previous MSS and - * blackhole detection flags. + * If further retransmissions are still + * unsuccessful with a lowered MTU, maybe this + * isn't a Black Hole and we restore the previous + * MSS and blackhole detection flags. */ else { - - if ((tp->t_flags & TF_BLACKHOLE) && (tp->t_rxtshift > 4)) { - tp->t_flags |= TF_PMTUD; - tp->t_flags &= ~TF_BLACKHOLE; - optlen = tp->t_maxopd - tp->t_maxseg; - tp->t_maxopd = tp->t_pmtud_saved_maxopd; - tp->t_maxseg = tp->t_maxopd - optlen; - /* - * Reset the slow-start flight size as it - * may depend on the new MSS - */ - if (CC_ALGO(tp)->cwnd_init != NULL) - CC_ALGO(tp)->cwnd_init(tp); + + if ((tp->t_flags & TF_BLACKHOLE) && + (tp->t_rxtshift > 4)) { + tcp_pmtud_revert_segment_size(tp); + tp->snd_cwnd = tp->t_maxseg; } } } /* - * Disable rfc1323 and rfc1644 if we haven't got any response to - * our SYN (after we reach the threshold) to work-around some - * broken terminal servers (most of which have hopefully been - * retired) that have bad VJ header compression code which - * trashes TCP segments containing unknown-to-them TCP options. + * Disable rfc1323 and rfc1644 if we haven't got any + * response to our SYN (after we reach the threshold) + * to work-around some broken terminal servers (most of + * which have hopefully been retired) that have bad VJ + * header compression code which trashes TCP segments + * containing unknown-to-them TCP options. * Do this only on non-local connections. */ if (tp->t_state == TCPS_SYN_SENT && - ((!(tp->t_flags & TF_LOCAL) && - tp->t_rxtshift == tcp_broken_peer_syn_rxmit_thres) || - ((tp->t_flags & TF_LOCAL) && - tp->t_rxtshift == tcp_broken_peer_syn_rxmit_thres_local))) + tp->t_rxtshift == tcp_broken_peer_syn_rxmit_thres) tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_REQ_CC); /* @@ -838,33 +1068,40 @@ tcp_timers(tp, timer) * Force a segment to be sent. */ tp->t_flags |= TF_ACKNOW; - /* - * If timing a segment in this window, stop the timer. - */ + + /* If timing a segment in this window, stop the timer */ tp->t_rtttime = 0; - EXIT_FASTRECOVERY(tp); + if (!IN_FASTRECOVERY(tp) && tp->t_rxtshift == 1) + tcpstat.tcps_tailloss_rto++; - /* RFC 5681 says: when a TCP sender detects segment loss + + /* + * RFC 5681 says: when a TCP sender detects segment loss * using retransmit timer and the given segment has already * been retransmitted by way of the retransmission timer at * least once, the value of ssthresh is held constant */ - if (tp->t_rxtshift == 1 && - CC_ALGO(tp)->after_timeout != NULL) + if (tp->t_rxtshift == 1 && + CC_ALGO(tp)->after_timeout != NULL) { CC_ALGO(tp)->after_timeout(tp); + /* + * CWR notifications are to be sent on new data + * right after Fast Retransmits and ECE + * notification receipts. + */ + if (TCP_ECN_ENABLED(tp)) + tp->ecn_flags |= TE_SENDCWR; + } + + EXIT_FASTRECOVERY(tp); + + /* Exit cwnd non validated phase */ + tp->t_flagsext &= ~TF_CWND_NONVALIDATED; - /* CWR notifications are to be sent on new data right after - * RTOs, Fast Retransmits and ECE notification receipts. - */ - if ((tp->ecn_flags & TE_ECN_ON) == TE_ECN_ON) { - tp->ecn_flags |= TE_SENDCWR; - } fc_output: - DTRACE_TCP5(cc, void, NULL, struct inpcb *, tp->t_inpcb, - struct tcpcb *, tp, struct tcphdr *, NULL, - int32_t, TCP_CC_REXMT_TIMEOUT); + tcp_ccdbg_trace(tp, NULL, TCP_CC_REXMT_TIMEOUT); (void) tcp_output(tp); break; @@ -881,15 +1118,15 @@ fc_output: * backoff, drop the connection if the idle time * (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. - * - * Drop the connection if we reached the maximum allowed time for - * Zero Window Probes without a non-zero update from the peer. + * + * Drop the connection if we reached the maximum allowed time for + * Zero Window Probes without a non-zero update from the peer. * See rdar://5805356 */ if ((tp->t_rxtshift == TCP_MAXRXTSHIFT && (idle_time >= tcp_maxpersistidle || - idle_time >= TCP_REXMTVAL(tp) * tcp_totbackoff)) || - ((tp->t_persist_stop != 0) && + idle_time >= TCP_REXMTVAL(tp) * tcp_totbackoff)) || + ((tp->t_persist_stop != 0) && TSTMP_LEQ(tp->t_persist_stop, tcp_now))) { tcpstat.tcps_persistdrop++; postevent(so, 0, EV_TIMEOUT); @@ -899,9 +1136,9 @@ fc_output: break; } tcp_setpersist(tp); - tp->t_force = 1; + tp->t_flagsext |= TF_FORCE; (void) tcp_output(tp); - tp->t_force = 0; + tp->t_flagsext &= ~TF_FORCE; break; /* @@ -916,8 +1153,8 @@ fc_output: * MPTCP must not also, after sending Data FINs. */ struct mptcb *mp_tp = tp->t_mptcb; - if ((tp->t_mpflags & TMPF_MPTCP_TRUE) && - (mp_tp == NULL)) { + if ((tp->t_mpflags & TMPF_MPTCP_TRUE) && + (tp->t_state > TCPS_ESTABLISHED)) { goto dropit; } else if (mp_tp != NULL) { if ((mptcp_ok_to_keepalive(mp_tp) == 0)) @@ -928,7 +1165,8 @@ fc_output: goto dropit; if ((always_keepalive || (tp->t_inpcb->inp_socket->so_options & SO_KEEPALIVE) || - (tp->t_flagsext & TF_DETECT_READSTALL)) && + (tp->t_flagsext & TF_DETECT_READSTALL) || + (tp->t_tfo_probe_state == TFO_PROBE_PROBING)) && (tp->t_state <= TCPS_CLOSING || tp->t_state == TCPS_FIN_WAIT_2)) { if (idle_time >= TCP_CONN_KEEPIDLE(tp) + TCP_CONN_MAXIDLE(tp)) goto dropit; @@ -947,53 +1185,90 @@ fc_output: tcpstat.tcps_keepprobe++; t_template = tcp_maketemplate(tp); if (t_template) { - unsigned int ifscope, nocell = 0; - + struct inpcb *inp = tp->t_inpcb; + struct tcp_respond_args tra; + + bzero(&tra, sizeof(tra)); + tra.nocell = INP_NO_CELLULAR(inp); + tra.noexpensive = INP_NO_EXPENSIVE(inp); + tra.awdl_unrestricted = INP_AWDL_UNRESTRICTED(inp); + tra.intcoproc_allowed = INP_INTCOPROC_ALLOWED(inp); if (tp->t_inpcb->inp_flags & INP_BOUND_IF) - ifscope = tp->t_inpcb->inp_boundifp->if_index; + tra.ifscope = tp->t_inpcb->inp_boundifp->if_index; else - ifscope = IFSCOPE_NONE; - - /* - * If the socket isn't allowed to use the - * cellular interface, indicate it as such. - */ - if (tp->t_inpcb->inp_flags & INP_NO_IFT_CELLULAR) - nocell = 1; - + tra.ifscope = IFSCOPE_NONE; tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, - tp->rcv_nxt, tp->snd_una - 1, 0, ifscope, - nocell); + tp->rcv_nxt, tp->snd_una - 1, 0, &tra); (void) m_free(dtom(t_template)); if (tp->t_flagsext & TF_DETECT_READSTALL) tp->t_rtimo_probes++; } tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, - TCP_CONN_KEEPINTVL(tp)); + TCP_CONN_KEEPINTVL(tp)); } else { tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, - TCP_CONN_KEEPIDLE(tp)); + TCP_CONN_KEEPIDLE(tp)); } if (tp->t_flagsext & TF_DETECT_READSTALL) { - /* + struct ifnet *outifp = tp->t_inpcb->inp_last_outifp; + bool reenable_probe = false; + /* * The keep alive packets sent to detect a read - * stall did not get a response from the + * stall did not get a response from the * peer. Generate more keep-alives to confirm this. * If the number of probes sent reaches the limit, * generate an event. */ - if (tp->t_rtimo_probes > tp->t_adaptive_rtimo) { - /* Generate an event */ - soevent(so, - (SO_FILT_HINT_LOCKED| - SO_FILT_HINT_ADAPTIVE_RTIMO)); - tcp_keepalive_reset(tp); + if (tp->t_adaptive_rtimo > 0) { + if (tp->t_rtimo_probes > tp->t_adaptive_rtimo) { + /* Generate an event */ + soevent(so, + (SO_FILT_HINT_LOCKED | + SO_FILT_HINT_ADAPTIVE_RTIMO)); + tcp_keepalive_reset(tp); + } else { + reenable_probe = true; + } + } else if (outifp != NULL && + (outifp->if_eflags & IFEF_PROBE_CONNECTIVITY) && + tp->t_rtimo_probes <= TCP_CONNECTIVITY_PROBES_MAX) { + reenable_probe = true; } else { + tp->t_flagsext &= ~TF_DETECT_READSTALL; + } + if (reenable_probe) { + int ind = min(tp->t_rtimo_probes, + TCP_MAXRXTSHIFT); tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START( - tp, TCP_REXMTVAL(tp)); + tp, tcp_backoff[ind] * TCP_REXMTVAL(tp)); } } + if (tp->t_tfo_probe_state == TFO_PROBE_PROBING) { + int ind; + + tp->t_tfo_probes++; + ind = min(tp->t_tfo_probes, TCP_MAXRXTSHIFT); + + /* + * We take the minimum among the time set by true + * keepalive (see above) and the backoff'd RTO. That + * way we backoff in case of packet-loss but will never + * timeout slower than regular keepalive due to the + * backing off. + */ + tp->t_timer[TCPT_KEEP] = min(OFFSET_FROM_START( + tp, tcp_backoff[ind] * TCP_REXMTVAL(tp)), + tp->t_timer[TCPT_KEEP]); + } else if (tp->t_tfo_probe_state == TFO_PROBE_WAIT_DATA) { + /* Still no data! Let's assume a TFO-error and err out... */ + tcp_heuristic_tfo_middlebox(tp); + + so->so_error = ENODATA; + sorwakeup(so); + tp->t_tfo_stats |= TFO_S_RECV_BLACKHOLE; + tcpstat.tcps_tfo_blackhole++; + } break; case TCPT_DELACK: if (tcp_delack_enabled && (tp->t_flags & TF_DELACK)) { @@ -1001,16 +1276,40 @@ fc_output: tp->t_timer[TCPT_DELACK] = 0; tp->t_flags |= TF_ACKNOW; - /* If delayed ack timer fired while stretching acks - * go back to acking every other packet + /* + * If delayed ack timer fired while stretching + * acks, count the number of times the streaming + * detection was not correct. If this exceeds a + * threshold, disable strech ack on this + * connection + * + * Also, go back to acking every other packet. */ - if ((tp->t_flags & TF_STRETCHACK) != 0) + if ((tp->t_flags & TF_STRETCHACK)) { + if (tp->t_unacksegs > 1 && + tp->t_unacksegs < maxseg_unacked) + tp->t_stretchack_delayed++; + + if (tp->t_stretchack_delayed > + TCP_STRETCHACK_DELAY_THRESHOLD) { + tp->t_flagsext |= TF_DISABLE_STRETCHACK; + /* + * Note the time at which stretch + * ack was disabled automatically + */ + tp->rcv_nostrack_ts = tcp_now; + tcpstat.tcps_nostretchack++; + tp->t_stretchack_delayed = 0; + tp->rcv_nostrack_pkts = 0; + } tcp_reset_stretch_ack(tp); + } - /* If we are measuring inter packet arrival jitter for - * throttling a connection, this delayed ack might be - * the reason for accumulating some jitter. So let's - * restart the measurement. + /* + * If we are measuring inter packet arrival jitter + * for throttling a connection, this delayed ack + * might be the reason for accumulating some + * jitter. So let's restart the measurement. */ CLEAR_IAJ_STATE(tp); @@ -1027,7 +1326,7 @@ fc_output: if (++tp->t_mprxtshift > TCP_MAXRXTSHIFT) { tcpstat.tcps_timeoutdrop++; postevent(so, 0, EV_TIMEOUT); - soevent(so, + soevent(so, (SO_FILT_HINT_LOCKED| SO_FILT_HINT_TIMEOUT)); tp = tcp_drop(tp, tp->t_softerror ? @@ -1038,7 +1337,7 @@ fc_output: tp->t_flags |= TF_ACKNOW; /* - * No backoff is implemented for simplicity for this + * No backoff is implemented for simplicity for this * corner case. */ (void) tcp_output(tp); @@ -1046,11 +1345,86 @@ fc_output: break; #endif /* MPTCP */ -#if TCPDEBUG - if (tp->t_inpcb->inp_socket->so_options & SO_DEBUG) - tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, - PRU_SLOWTIMO); -#endif + case TCPT_PTO: + { + int32_t snd_len; + tp->t_flagsext &= ~(TF_SENT_TLPROBE); + + /* + * Check if the connection is in the right state to + * send a probe + */ + if (tp->t_state != TCPS_ESTABLISHED || + (tp->t_rxtshift > 0 && !(tp->t_flagsext & TF_PROBING)) + || tp->snd_max == tp->snd_una || + !SACK_ENABLED(tp) || !TAILQ_EMPTY(&tp->snd_holes) || + IN_FASTRECOVERY(tp)) + break; + + /* + * If there is no new data to send or if the + * connection is limited by receive window then + * retransmit the last segment, otherwise send + * new data. + */ + snd_len = min(so->so_snd.sb_cc, tp->snd_wnd) + - (tp->snd_max - tp->snd_una); + if (snd_len > 0) { + tp->snd_nxt = tp->snd_max; + } else { + snd_len = min((tp->snd_max - tp->snd_una), + tp->t_maxseg); + tp->snd_nxt = tp->snd_max - snd_len; + } + + tcpstat.tcps_pto++; + if (tp->t_flagsext & TF_PROBING) + tcpstat.tcps_probe_if++; + + /* If timing a segment in this window, stop the timer */ + tp->t_rtttime = 0; + /* Note that tail loss probe is being sent */ + tp->t_flagsext |= TF_SENT_TLPROBE; + tp->t_tlpstart = tcp_now; + + tp->snd_cwnd += tp->t_maxseg; + (void )tcp_output(tp); + tp->snd_cwnd -= tp->t_maxseg; + + tp->t_tlphighrxt = tp->snd_nxt; + break; + } + case TCPT_DELAYFR: + tp->t_flagsext &= ~TF_DELAY_RECOVERY; + + /* + * Don't do anything if one of the following is true: + * - the connection is already in recovery + * - sequence until snd_recover has been acknowledged. + * - retransmit timeout has fired + */ + if (IN_FASTRECOVERY(tp) || + SEQ_GEQ(tp->snd_una, tp->snd_recover) || + tp->t_rxtshift > 0) + break; + + VERIFY(SACK_ENABLED(tp)); + tcp_rexmt_save_state(tp); + if (CC_ALGO(tp)->pre_fr != NULL) { + CC_ALGO(tp)->pre_fr(tp); + if (TCP_ECN_ENABLED(tp)) + tp->ecn_flags |= TE_SENDCWR; + } + ENTER_FASTRECOVERY(tp); + + tp->t_timer[TCPT_REXMT] = 0; + tcpstat.tcps_sack_recovery_episode++; + tp->t_sack_recovery_episode++; + tp->sack_newdata = tp->snd_nxt; + tp->snd_cwnd = tp->t_maxseg; + tcp_ccdbg_trace(tp, NULL, TCP_CC_ENTER_FASTRECOVERY); + (void) tcp_output(tp); + break; dropit: tcpstat.tcps_keepdrops++; postevent(so, 0, EV_TIMEOUT); @@ -1059,6 +1433,11 @@ fc_output: tp = tcp_drop(tp, ETIMEDOUT); break; } +#if TCPDEBUG + if (tp->t_inpcb->inp_socket->so_options & SO_DEBUG) + tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, + PRU_SLOWTIMO); +#endif return (tp); } @@ -1073,13 +1452,13 @@ tcp_remove_timer(struct tcpcb *tp) return; } lck_mtx_lock(listp->mtx); - + /* Check if pcb is on timer list again after acquiring the lock */ if (!(TIMER_IS_ON_LIST(tp))) { lck_mtx_unlock(listp->mtx); return; } - + if (listp->next_te != NULL && listp->next_te == &tp->tentry) listp->next_te = LIST_NEXT(&tp->tentry, le); @@ -1093,109 +1472,136 @@ tcp_remove_timer(struct tcpcb *tp) lck_mtx_unlock(listp->mtx); } -/* Function to check if the timerlist needs to be rescheduled to run +/* + * Function to check if the timerlist needs to be rescheduled to run * the timer entry correctly. Basically, this is to check if we can avoid * taking the list lock. */ static boolean_t -need_to_resched_timerlist(uint32_t runtime, uint16_t index) { +need_to_resched_timerlist(u_int32_t runtime, u_int16_t mode) +{ struct tcptimerlist *listp = &tcp_timer_list; int32_t diff; - boolean_t is_fast; - if (runtime == 0 || index == TCPT_NONE) - return FALSE; - is_fast = !(IS_TIMER_SLOW(index)); - - /* If the list is being processed then the state of the list is in flux. - * In this case always acquire the lock and set the state correctly. + /* + * If the list is being processed then the state of the list is + * in flux. In this case always acquire the lock and set the state + * correctly. */ - if (listp->running) { - return TRUE; - } + if (listp->running) + return (TRUE); + + if (!listp->scheduled) + return (TRUE); diff = timer_diff(listp->runtime, 0, runtime, 0); if (diff <= 0) { /* The list is going to run before this timer */ - return FALSE; + return (FALSE); } else { - if (is_fast) { - if (diff <= listp->fast_quantum) - return FALSE; + if (mode & TCP_TIMERLIST_10MS_MODE) { + if (diff <= TCP_TIMER_10MS_QUANTUM) + return (FALSE); + } else if (mode & TCP_TIMERLIST_100MS_MODE) { + if (diff <= TCP_TIMER_100MS_QUANTUM) + return (FALSE); } else { - if (diff <= listp->slow_quantum) - return FALSE; + if (diff <= TCP_TIMER_500MS_QUANTUM) + return (FALSE); } } - return TRUE; + return (TRUE); } void -tcp_sched_timerlist(uint32_t offset) +tcp_sched_timerlist(uint32_t offset) { - uint64_t deadline = 0; struct tcptimerlist *listp = &tcp_timer_list; lck_mtx_assert(listp->mtx, LCK_MTX_ASSERT_OWNED); + offset = min(offset, TCP_TIMERLIST_MAX_OFFSET); listp->runtime = tcp_now + offset; + if (listp->runtime == 0) { + listp->runtime++; + offset++; + } - clock_interval_to_deadline(offset, NSEC_PER_SEC / TCP_RETRANSHZ, - &deadline); + clock_interval_to_deadline(offset, USEC_PER_SEC, &deadline); thread_call_enter_delayed(listp->call, deadline); + listp->scheduled = TRUE; } -/* Function to run the timers for a connection. +/* + * Function to run the timers for a connection. * - * Returns the offset of next timer to be run for this connection which + * Returns the offset of next timer to be run for this connection which * can be used to reschedule the timerlist. + * + * te_mode is an out parameter that indicates the modes of active + * timers for this connection. */ -uint32_t -tcp_run_conn_timer(struct tcpcb *tp, uint16_t *next_index) { - - struct socket *so; - uint16_t i = 0, index = TCPT_NONE, lo_index = TCPT_NONE; - uint32_t timer_val, offset = 0, lo_timer = 0; +u_int32_t +tcp_run_conn_timer(struct tcpcb *tp, u_int16_t *te_mode, + u_int16_t probe_if_index) +{ + struct socket *so; + u_int16_t i = 0, index = TCPT_NONE, lo_index = TCPT_NONE; + u_int32_t timer_val, offset = 0, lo_timer = 0; int32_t diff; boolean_t needtorun[TCPT_NTIMERS]; int count = 0; - VERIFY(tp != NULL); - bzero(needtorun, sizeof(needtorun)); + VERIFY(tp != NULL); + bzero(needtorun, sizeof(needtorun)); + *te_mode = 0; - tcp_lock(tp->t_inpcb->inp_socket, 1, 0); + tcp_lock(tp->t_inpcb->inp_socket, 1, 0); - so = tp->t_inpcb->inp_socket; - /* Release the want count on inp */ - if (in_pcb_checkstate(tp->t_inpcb, WNT_RELEASE, 1) == WNT_STOPUSING) { + so = tp->t_inpcb->inp_socket; + /* Release the want count on inp */ + if (in_pcb_checkstate(tp->t_inpcb, WNT_RELEASE, 1) + == WNT_STOPUSING) { if (TIMER_IS_ON_LIST(tp)) { tcp_remove_timer(tp); } - /* Looks like the TCP connection got closed while we + /* Looks like the TCP connection got closed while we * were waiting for the lock.. Done */ goto done; } - /* Since the timer thread needs to wait for tcp lock, it may race - * with another thread that can cancel or reschedule the timer that is - * about to run. Check if we need to run anything. - */ - index = tp->tentry.index; - timer_val = tp->t_timer[index]; + /* + * If this connection is over an interface that needs to + * be probed, send probe packets to reinitiate communication. + */ + if (probe_if_index > 0 && tp->t_inpcb->inp_last_outifp != NULL && + tp->t_inpcb->inp_last_outifp->if_index == probe_if_index) { + tp->t_flagsext |= TF_PROBING; + tcp_timers(tp, TCPT_PTO); + tp->t_timer[TCPT_PTO] = 0; + tp->t_flagsext &= ~TF_PROBING; + } - if (index == TCPT_NONE || tp->tentry.runtime == 0) + /* + * Since the timer thread needs to wait for tcp lock, it may race + * with another thread that can cancel or reschedule the timer + * that is about to run. Check if we need to run anything. + */ + if ((index = tp->tentry.index) == TCPT_NONE) goto done; + timer_val = tp->t_timer[index]; + diff = timer_diff(tp->tentry.runtime, 0, tcp_now, 0); if (diff > 0) { if (tp->tentry.index != TCPT_NONE) { offset = diff; - *(next_index) = tp->tentry.index; + *(te_mode) = tp->tentry.mode; } goto done; } @@ -1203,18 +1609,20 @@ tcp_run_conn_timer(struct tcpcb *tp, uint16_t *next_index) { tp->t_timer[index] = 0; if (timer_val > 0) { tp = tcp_timers(tp, index); - if (tp == NULL) + if (tp == NULL) goto done; } - - /* Check if there are any other timers that need to be run. While doing it, - * adjust the timer values wrt tcp_now. + + /* + * Check if there are any other timers that need to be run. + * While doing it, adjust the timer values wrt tcp_now. */ + tp->tentry.mode = 0; for (i = 0; i < TCPT_NTIMERS; ++i) { if (tp->t_timer[i] != 0) { - diff = timer_diff(tp->tentry.timer_start, tp->t_timer[i], tcp_now, 0); + diff = timer_diff(tp->tentry.timer_start, + tp->t_timer[i], tcp_now, 0); if (diff <= 0) { - tp->t_timer[i] = 0; needtorun[i] = TRUE; count++; } else { @@ -1224,26 +1632,33 @@ tcp_run_conn_timer(struct tcpcb *tp, uint16_t *next_index) { lo_timer = diff; lo_index = i; } + TCP_SET_TIMER_MODE(tp->tentry.mode, i); } } } - + tp->tentry.timer_start = tcp_now; tp->tentry.index = lo_index; - if (lo_index != TCPT_NONE) { - tp->tentry.runtime = tp->tentry.timer_start + tp->t_timer[lo_index]; - } else { - tp->tentry.runtime = 0; + VERIFY(tp->tentry.index == TCPT_NONE || tp->tentry.mode > 0); + + if (tp->tentry.index != TCPT_NONE) { + tp->tentry.runtime = tp->tentry.timer_start + + tp->t_timer[tp->tentry.index]; + if (tp->tentry.runtime == 0) + tp->tentry.runtime++; } if (count > 0) { - /* run any other timers that are also outstanding at this time. */ + /* run any other timers outstanding at this time. */ for (i = 0; i < TCPT_NTIMERS; ++i) { if (needtorun[i]) { tp->t_timer[i] = 0; tp = tcp_timers(tp, i); - if (tp == NULL) + if (tp == NULL) { + offset = 0; + *(te_mode) = 0; goto done; + } } } tcp_set_lotimer_index(tp); @@ -1251,58 +1666,63 @@ tcp_run_conn_timer(struct tcpcb *tp, uint16_t *next_index) { if (tp->tentry.index < TCPT_NONE) { offset = tp->t_timer[tp->tentry.index]; - *(next_index) = tp->tentry.index; + *(te_mode) = tp->tentry.mode; } done: if (tp != NULL && tp->tentry.index == TCPT_NONE) { tcp_remove_timer(tp); + offset = 0; } - tcp_unlock(so, 1, 0); - return offset; + + tcp_unlock(so, 1, 0); + return(offset); } void -tcp_run_timerlist(void * arg1, void * arg2) { - +tcp_run_timerlist(void * arg1, void * arg2) +{ #pragma unused(arg1, arg2) - struct tcptimerentry *te, *next_te; struct tcptimerlist *listp = &tcp_timer_list; struct tcpcb *tp; - uint32_t next_timer = 0; - uint16_t index = TCPT_NONE; - boolean_t need_fast = FALSE; + uint32_t next_timer = 0; /* offset of the next timer on the list */ + u_int16_t te_mode = 0; /* modes of all active timers in a tcpcb */ + u_int16_t list_mode = 0; /* cumulative of modes of all tcpcbs */ uint32_t active_count = 0; - uint32_t mode = TCP_TIMERLIST_FASTMODE; calculate_tcp_clock(); lck_mtx_lock(listp->mtx); listp->running = TRUE; - + LIST_FOREACH_SAFE(te, &listp->lhead, le, next_te) { uint32_t offset = 0; uint32_t runtime = te->runtime; - if (TSTMP_GT(runtime, tcp_now)) { + if (te->index < TCPT_NONE && TSTMP_GT(runtime, tcp_now)) { offset = timer_diff(runtime, 0, tcp_now, 0); if (next_timer == 0 || offset < next_timer) { next_timer = offset; } + list_mode |= te->mode; continue; } - active_count++; tp = TIMERENTRY_TO_TP(te); - /* Acquire an inp wantcnt on the inpcb so that the socket won't get - * detached even if tcp_close is called + /* + * Acquire an inp wantcnt on the inpcb so that the socket + * won't get detached even if tcp_close is called */ - if (in_pcb_checkstate(tp->t_inpcb, WNT_ACQUIRE, 0) == WNT_STOPUSING) { - /* Some how this pcb went into dead state while on the timer list, - * just take it off the list. Since the timer list entry pointers - * are protected by the timer list lock, we can do it here + if (in_pcb_checkstate(tp->t_inpcb, WNT_ACQUIRE, 0) + == WNT_STOPUSING) { + /* + * Some how this pcb went into dead state while + * on the timer list, just take it off the list. + * Since the timer list entry pointers are + * protected by the timer list lock, we can + * do it here without the socket lock. */ if (TIMER_IS_ON_LIST(tp)) { tp->t_flags &= ~(TF_TIMER_ONLIST); @@ -1314,94 +1734,106 @@ tcp_run_timerlist(void * arg1, void * arg2) { } continue; } + active_count++; - /* Store the next timerentry pointer before releasing the list lock. - * If that entry has to be removed when we release the lock, this - * pointer will be updated to the element after that. + /* + * Store the next timerentry pointer before releasing the + * list lock. If that entry has to be removed when we + * release the lock, this pointer will be updated to the + * element after that. */ - listp->next_te = next_te; + listp->next_te = next_te; VERIFY_NEXT_LINK(&tp->tentry, le); VERIFY_PREV_LINK(&tp->tentry, le); lck_mtx_unlock(listp->mtx); - index = TCPT_NONE; - offset = tcp_run_conn_timer(tp, &index); - + offset = tcp_run_conn_timer(tp, &te_mode, + listp->probe_if_index); + lck_mtx_lock(listp->mtx); next_te = listp->next_te; listp->next_te = NULL; - if (offset > 0) { - if (index < TCPT_NONE) { - /* Check if this is a fast_timer. */ - if (!need_fast && !(IS_TIMER_SLOW(index))) { - need_fast = TRUE; - } + if (offset > 0 && te_mode != 0) { + list_mode |= te_mode; - if (next_timer == 0 || offset < next_timer) { - next_timer = offset; - } - } + if (next_timer == 0 || offset < next_timer) + next_timer = offset; } } if (!LIST_EMPTY(&listp->lhead)) { - if (listp->mode == TCP_TIMERLIST_FASTMODE) { - if (need_fast || active_count > 0 || - listp->pref_mode == TCP_TIMERLIST_FASTMODE) { - listp->idlegen = 0; - } else { - listp->idlegen++; - if (listp->idlegen > timer_fastmode_idlemax) { - mode = TCP_TIMERLIST_SLOWMODE; - listp->idlegen = 0; - } - } - } else { - if (!need_fast) { - mode = TCP_TIMERLIST_SLOWMODE; - } - } + u_int16_t next_mode = 0; + if ((list_mode & TCP_TIMERLIST_10MS_MODE) || + (listp->pref_mode & TCP_TIMERLIST_10MS_MODE)) + next_mode = TCP_TIMERLIST_10MS_MODE; + else if ((list_mode & TCP_TIMERLIST_100MS_MODE) || + (listp->pref_mode & TCP_TIMERLIST_100MS_MODE)) + next_mode = TCP_TIMERLIST_100MS_MODE; + else + next_mode = TCP_TIMERLIST_500MS_MODE; - if (mode == TCP_TIMERLIST_FASTMODE || - listp->pref_mode == TCP_TIMERLIST_FASTMODE) { - next_timer = listp->fast_quantum; + if (next_mode != TCP_TIMERLIST_500MS_MODE) { + listp->idleruns = 0; } else { - if (listp->pref_offset != 0 && - listp->pref_offset < next_timer) - next_timer = listp->pref_offset; - if (next_timer < listp->slow_quantum) - next_timer = listp->slow_quantum; + /* + * the next required mode is slow mode, but if + * the last one was a faster mode and we did not + * have enough idle runs, repeat the last mode. + * + * We try to keep the timer list in fast mode for + * some idle time in expectation of new data. + */ + if (listp->mode != next_mode && + listp->idleruns < timer_fastmode_idlemax) { + listp->idleruns++; + next_mode = listp->mode; + next_timer = TCP_TIMER_100MS_QUANTUM; + } else { + listp->idleruns = 0; + } } + listp->mode = next_mode; + if (listp->pref_offset != 0) + next_timer = min(listp->pref_offset, next_timer); - listp->mode = mode; + if (listp->mode == TCP_TIMERLIST_500MS_MODE) + next_timer = max(next_timer, + TCP_TIMER_500MS_QUANTUM); tcp_sched_timerlist(next_timer); } else { - /* No need to reschedule this timer */ - listp->runtime = 0; + /* + * No need to reschedule this timer, but always run + * periodically at a much higher granularity. + */ + tcp_sched_timerlist(TCP_TIMERLIST_MAX_OFFSET); } listp->running = FALSE; listp->pref_mode = 0; listp->pref_offset = 0; + listp->probe_if_index = 0; lck_mtx_unlock(listp->mtx); } -/* Function to verify if a change in timer state is required for a connection */ -void -tcp_sched_timers(struct tcpcb *tp) +/* + * Function to check if the timerlist needs to be rescheduled to run this + * connection's timers correctly. + */ +void +tcp_sched_timers(struct tcpcb *tp) { struct tcptimerentry *te = &tp->tentry; - uint16_t index = te->index; + u_int16_t index = te->index; + u_int16_t mode = te->mode; struct tcptimerlist *listp = &tcp_timer_list; - uint32_t offset = 0; - boolean_t is_fast; - int list_locked = 0; + int32_t offset = 0; + boolean_t list_locked = FALSE; if (tp->t_inpcb->inp_state == INPCB_STATE_DEAD) { /* Just return without adding the dead pcb to the list */ @@ -1412,64 +1844,74 @@ tcp_sched_timers(struct tcpcb *tp) } if (index == TCPT_NONE) { + /* Nothing to run */ tcp_remove_timer(tp); return; } - is_fast = !(IS_TIMER_SLOW(index)); - offset = te->runtime - tcp_now; - if (offset == 0) { + /* + * compute the offset at which the next timer for this connection + * has to run. + */ + offset = timer_diff(te->runtime, 0, tcp_now, 0); + if (offset <= 0) { offset = 1; tcp_timer_advanced++; } - if (is_fast) - offset = listp->fast_quantum; if (!TIMER_IS_ON_LIST(tp)) { if (!list_locked) { lck_mtx_lock(listp->mtx); - list_locked = 1; + list_locked = TRUE; } LIST_INSERT_HEAD(&listp->lhead, te, le); tp->t_flags |= TF_TIMER_ONLIST; - listp->entries++; - if (listp->entries > listp->maxentries) - listp->maxentries = listp->entries; + listp->entries++; + if (listp->entries > listp->maxentries) + listp->maxentries = listp->entries; /* if the list is not scheduled, just schedule it */ - if (listp->runtime == 0) + if (!listp->scheduled) goto schedule; - } - /* timer entry is currently on the list */ - if (need_to_resched_timerlist(te->runtime, index)) { + /* + * Timer entry is currently on the list, check if the list needs + * to be rescheduled. + */ + if (need_to_resched_timerlist(te->runtime, mode)) { tcp_resched_timerlist++; - + if (!list_locked) { lck_mtx_lock(listp->mtx); - list_locked = 1; + list_locked = TRUE; } VERIFY_NEXT_LINK(te, le); VERIFY_PREV_LINK(te, le); if (listp->running) { - if (is_fast) { - listp->pref_mode = TCP_TIMERLIST_FASTMODE; - } else if (listp->pref_offset == 0 || - ((int)offset) < listp->pref_offset) { + listp->pref_mode |= mode; + if (listp->pref_offset == 0 || + offset < listp->pref_offset) { listp->pref_offset = offset; } } else { - int32_t diff; - diff = timer_diff(listp->runtime, 0, tcp_now, offset); - if (diff <= 0) { - /* The list is going to run before this timer */ - goto done; + /* + * The list could have got rescheduled while + * this thread was waiting for the lock + */ + if (listp->scheduled) { + int32_t diff; + diff = timer_diff(listp->runtime, 0, + tcp_now, offset); + if (diff <= 0) + goto done; + else + goto schedule; } else { goto schedule; } @@ -1478,9 +1920,20 @@ tcp_sched_timers(struct tcpcb *tp) goto done; schedule: - if (is_fast) { - listp->mode = TCP_TIMERLIST_FASTMODE; - listp->idlegen = 0; + /* + * Since a connection with timers is getting scheduled, the timer + * list moves from idle to active state and that is why idlegen is + * reset + */ + if (mode & TCP_TIMERLIST_10MS_MODE) { + listp->mode = TCP_TIMERLIST_10MS_MODE; + listp->idleruns = 0; + offset = min(offset, TCP_TIMER_10MS_QUANTUM); + } else if (mode & TCP_TIMERLIST_100MS_MODE) { + if (listp->mode > TCP_TIMERLIST_100MS_MODE) + listp->mode = TCP_TIMERLIST_100MS_MODE; + listp->idleruns = 0; + offset = min(offset, TCP_TIMER_100MS_QUANTUM); } tcp_sched_timerlist(offset); @@ -1490,33 +1943,506 @@ done: return; } - -void -tcp_set_lotimer_index(struct tcpcb *tp) { - uint16_t i, lo_index = TCPT_NONE; + +static inline void +tcp_set_lotimer_index(struct tcpcb *tp) +{ + uint16_t i, lo_index = TCPT_NONE, mode = 0; uint32_t lo_timer = 0; for (i = 0; i < TCPT_NTIMERS; ++i) { - if (tp->t_timer[i] != 0 && - (lo_timer == 0 || tp->t_timer[i] < lo_timer)) { - lo_timer = tp->t_timer[i]; - lo_index = i; + if (tp->t_timer[i] != 0) { + TCP_SET_TIMER_MODE(mode, i); + if (lo_timer == 0 || tp->t_timer[i] < lo_timer) { + lo_timer = tp->t_timer[i]; + lo_index = i; + } } } tp->tentry.index = lo_index; - if (lo_index != TCPT_NONE) { - tp->tentry.runtime = tp->tentry.timer_start + tp->t_timer[lo_index]; - } else { - tp->tentry.runtime = 0; + tp->tentry.mode = mode; + VERIFY(tp->tentry.index == TCPT_NONE || tp->tentry.mode > 0); + + if (tp->tentry.index != TCPT_NONE) { + tp->tentry.runtime = tp->tentry.timer_start + + tp->t_timer[tp->tentry.index]; + if (tp->tentry.runtime == 0) + tp->tentry.runtime++; } } void -tcp_check_timer_state(struct tcpcb *tp) { - +tcp_check_timer_state(struct tcpcb *tp) +{ lck_mtx_assert(&tp->t_inpcb->inpcb_mtx, LCK_MTX_ASSERT_OWNED); + if (tp->t_inpcb->inp_flags2 & INP2_TIMEWAIT) + return; + tcp_set_lotimer_index(tp); tcp_sched_timers(tp); return; } + +static inline void +tcp_cumulative_stat(u_int32_t cur, u_int32_t *prev, u_int32_t *dest) +{ + /* handle wrap around */ + int32_t diff = (int32_t) (cur - *prev); + if (diff > 0) + *dest = diff; + else + *dest = 0; + *prev = cur; + return; +} + +__private_extern__ void +tcp_report_stats(void) +{ + struct nstat_sysinfo_data data; + struct sockaddr_in dst; + struct sockaddr_in6 dst6; + struct rtentry *rt = NULL; + static struct tcp_last_report_stats prev; + u_int64_t var, uptime; + +#define stat data.u.tcp_stats + if (((uptime = net_uptime()) - tcp_last_report_time) < + tcp_report_stats_interval) + return; + + tcp_last_report_time = uptime; + + bzero(&data, sizeof(data)); + data.flags = NSTAT_SYSINFO_TCP_STATS; + + bzero(&dst, sizeof(dst)); + dst.sin_len = sizeof(dst); + dst.sin_family = AF_INET; + + /* ipv4 avg rtt */ + lck_mtx_lock(rnh_lock); + rt = rt_lookup(TRUE, (struct sockaddr *)&dst, NULL, + rt_tables[AF_INET], IFSCOPE_NONE); + lck_mtx_unlock(rnh_lock); + if (rt != NULL) { + RT_LOCK(rt); + if (rt_primary_default(rt, rt_key(rt)) && + rt->rt_stats != NULL) { + stat.ipv4_avgrtt = rt->rt_stats->nstat_avg_rtt; + } + RT_UNLOCK(rt); + rtfree(rt); + rt = NULL; + } + + /* ipv6 avg rtt */ + bzero(&dst6, sizeof(dst6)); + dst6.sin6_len = sizeof(dst6); + dst6.sin6_family = AF_INET6; + + lck_mtx_lock(rnh_lock); + rt = rt_lookup(TRUE,(struct sockaddr *)&dst6, NULL, + rt_tables[AF_INET6], IFSCOPE_NONE); + lck_mtx_unlock(rnh_lock); + if (rt != NULL) { + RT_LOCK(rt); + if (rt_primary_default(rt, rt_key(rt)) && + rt->rt_stats != NULL) { + stat.ipv6_avgrtt = rt->rt_stats->nstat_avg_rtt; + } + RT_UNLOCK(rt); + rtfree(rt); + rt = NULL; + } + + /* send packet loss rate, shift by 10 for precision */ + if (tcpstat.tcps_sndpack > 0 && tcpstat.tcps_sndrexmitpack > 0) { + var = tcpstat.tcps_sndrexmitpack << 10; + stat.send_plr = (var * 100) / tcpstat.tcps_sndpack; + } + + /* recv packet loss rate, shift by 10 for precision */ + if (tcpstat.tcps_rcvpack > 0 && tcpstat.tcps_recovered_pkts > 0) { + var = tcpstat.tcps_recovered_pkts << 10; + stat.recv_plr = (var * 100) / tcpstat.tcps_rcvpack; + } + + /* RTO after tail loss, shift by 10 for precision */ + if (tcpstat.tcps_sndrexmitpack > 0 + && tcpstat.tcps_tailloss_rto > 0) { + var = tcpstat.tcps_tailloss_rto << 10; + stat.send_tlrto_rate = + (var * 100) / tcpstat.tcps_sndrexmitpack; + } + + /* packet reordering */ + if (tcpstat.tcps_sndpack > 0 && tcpstat.tcps_reordered_pkts > 0) { + var = tcpstat.tcps_reordered_pkts << 10; + stat.send_reorder_rate = + (var * 100) / tcpstat.tcps_sndpack; + } + + if (tcp_ecn_outbound == 1) + stat.ecn_client_enabled = 1; + if (tcp_ecn_inbound == 1) + stat.ecn_server_enabled = 1; + tcp_cumulative_stat(tcpstat.tcps_connattempt, + &prev.tcps_connattempt, &stat.connection_attempts); + tcp_cumulative_stat(tcpstat.tcps_accepts, + &prev.tcps_accepts, &stat.connection_accepts); + tcp_cumulative_stat(tcpstat.tcps_ecn_client_setup, + &prev.tcps_ecn_client_setup, &stat.ecn_client_setup); + tcp_cumulative_stat(tcpstat.tcps_ecn_server_setup, + &prev.tcps_ecn_server_setup, &stat.ecn_server_setup); + tcp_cumulative_stat(tcpstat.tcps_ecn_client_success, + &prev.tcps_ecn_client_success, &stat.ecn_client_success); + tcp_cumulative_stat(tcpstat.tcps_ecn_server_success, + &prev.tcps_ecn_server_success, &stat.ecn_server_success); + tcp_cumulative_stat(tcpstat.tcps_ecn_not_supported, + &prev.tcps_ecn_not_supported, &stat.ecn_not_supported); + tcp_cumulative_stat(tcpstat.tcps_ecn_lost_syn, + &prev.tcps_ecn_lost_syn, &stat.ecn_lost_syn); + tcp_cumulative_stat(tcpstat.tcps_ecn_lost_synack, + &prev.tcps_ecn_lost_synack, &stat.ecn_lost_synack); + tcp_cumulative_stat(tcpstat.tcps_ecn_recv_ce, + &prev.tcps_ecn_recv_ce, &stat.ecn_recv_ce); + tcp_cumulative_stat(tcpstat.tcps_ecn_recv_ece, + &prev.tcps_ecn_recv_ece, &stat.ecn_recv_ece); + tcp_cumulative_stat(tcpstat.tcps_ecn_recv_ece, + &prev.tcps_ecn_recv_ece, &stat.ecn_recv_ece); + tcp_cumulative_stat(tcpstat.tcps_ecn_sent_ece, + &prev.tcps_ecn_sent_ece, &stat.ecn_sent_ece); + tcp_cumulative_stat(tcpstat.tcps_ecn_sent_ece, + &prev.tcps_ecn_sent_ece, &stat.ecn_sent_ece); + tcp_cumulative_stat(tcpstat.tcps_ecn_conn_recv_ce, + &prev.tcps_ecn_conn_recv_ce, &stat.ecn_conn_recv_ce); + tcp_cumulative_stat(tcpstat.tcps_ecn_conn_recv_ece, + &prev.tcps_ecn_conn_recv_ece, &stat.ecn_conn_recv_ece); + tcp_cumulative_stat(tcpstat.tcps_ecn_conn_plnoce, + &prev.tcps_ecn_conn_plnoce, &stat.ecn_conn_plnoce); + tcp_cumulative_stat(tcpstat.tcps_ecn_conn_pl_ce, + &prev.tcps_ecn_conn_pl_ce, &stat.ecn_conn_pl_ce); + tcp_cumulative_stat(tcpstat.tcps_ecn_conn_nopl_ce, + &prev.tcps_ecn_conn_nopl_ce, &stat.ecn_conn_nopl_ce); + tcp_cumulative_stat(tcpstat.tcps_ecn_fallback_synloss, + &prev.tcps_ecn_fallback_synloss, &stat.ecn_fallback_synloss); + tcp_cumulative_stat(tcpstat.tcps_ecn_fallback_reorder, + &prev.tcps_ecn_fallback_reorder, &stat.ecn_fallback_reorder); + tcp_cumulative_stat(tcpstat.tcps_ecn_fallback_ce, + &prev.tcps_ecn_fallback_ce, &stat.ecn_fallback_ce); + tcp_cumulative_stat(tcpstat.tcps_tfo_syn_data_rcv, + &prev.tcps_tfo_syn_data_rcv, &stat.tfo_syn_data_rcv); + tcp_cumulative_stat(tcpstat.tcps_tfo_cookie_req_rcv, + &prev.tcps_tfo_cookie_req_rcv, &stat.tfo_cookie_req_rcv); + tcp_cumulative_stat(tcpstat.tcps_tfo_cookie_sent, + &prev.tcps_tfo_cookie_sent, &stat.tfo_cookie_sent); + tcp_cumulative_stat(tcpstat.tcps_tfo_cookie_invalid, + &prev.tcps_tfo_cookie_invalid, &stat.tfo_cookie_invalid); + tcp_cumulative_stat(tcpstat.tcps_tfo_cookie_req, + &prev.tcps_tfo_cookie_req, &stat.tfo_cookie_req); + tcp_cumulative_stat(tcpstat.tcps_tfo_cookie_rcv, + &prev.tcps_tfo_cookie_rcv, &stat.tfo_cookie_rcv); + tcp_cumulative_stat(tcpstat.tcps_tfo_syn_data_sent, + &prev.tcps_tfo_syn_data_sent, &stat.tfo_syn_data_sent); + tcp_cumulative_stat(tcpstat.tcps_tfo_syn_data_acked, + &prev.tcps_tfo_syn_data_acked, &stat.tfo_syn_data_acked); + tcp_cumulative_stat(tcpstat.tcps_tfo_syn_loss, + &prev.tcps_tfo_syn_loss, &stat.tfo_syn_loss); + tcp_cumulative_stat(tcpstat.tcps_tfo_blackhole, + &prev.tcps_tfo_blackhole, &stat.tfo_blackhole); + tcp_cumulative_stat(tcpstat.tcps_tfo_cookie_wrong, + &prev.tcps_tfo_cookie_wrong, &stat.tfo_cookie_wrong); + tcp_cumulative_stat(tcpstat.tcps_tfo_no_cookie_rcv, + &prev.tcps_tfo_no_cookie_rcv, &stat.tfo_no_cookie_rcv); + tcp_cumulative_stat(tcpstat.tcps_tfo_heuristics_disable, + &prev.tcps_tfo_heuristics_disable, &stat.tfo_heuristics_disable); + tcp_cumulative_stat(tcpstat.tcps_tfo_sndblackhole, + &prev.tcps_tfo_sndblackhole, &stat.tfo_sndblackhole); + + + + + nstat_sysinfo_send_data(&data); + +#undef stat +} + +void +tcp_interface_send_probe(u_int16_t probe_if_index) +{ + int32_t offset = 0; + struct tcptimerlist *listp = &tcp_timer_list; + + /* Make sure TCP clock is up to date */ + calculate_tcp_clock(); + + lck_mtx_lock(listp->mtx); + if (listp->probe_if_index > 0) { + tcpstat.tcps_probe_if_conflict++; + goto done; + } + + listp->probe_if_index = probe_if_index; + if (listp->running) + goto done; + + /* + * Reschedule the timerlist to run within the next 10ms, which is + * the fastest that we can do. + */ + offset = TCP_TIMER_10MS_QUANTUM; + if (listp->scheduled) { + int32_t diff; + diff = timer_diff(listp->runtime, 0, tcp_now, offset); + if (diff <= 0) { + /* The timer will fire sooner than what's needed */ + goto done; + } + } + listp->mode = TCP_TIMERLIST_10MS_MODE; + listp->idleruns = 0; + + tcp_sched_timerlist(offset); + +done: + lck_mtx_unlock(listp->mtx); + return; +} + +/* + * Enable read probes on this connection, if: + * - it is in established state + * - doesn't have any data outstanding + * - the outgoing ifp matches + * - we have not already sent any read probes + */ +static void +tcp_enable_read_probe(struct tcpcb *tp, struct ifnet *ifp) +{ + if (tp->t_state == TCPS_ESTABLISHED && + tp->snd_max == tp->snd_una && + tp->t_inpcb->inp_last_outifp == ifp && + !(tp->t_flagsext & TF_DETECT_READSTALL) && + tp->t_rtimo_probes == 0) { + tp->t_flagsext |= TF_DETECT_READSTALL; + tp->t_rtimo_probes = 0; + tp->t_timer[TCPT_KEEP] = OFFSET_FROM_START(tp, + TCP_TIMER_10MS_QUANTUM); + if (tp->tentry.index == TCPT_NONE) { + tp->tentry.index = TCPT_KEEP; + tp->tentry.runtime = tcp_now + + TCP_TIMER_10MS_QUANTUM; + } else { + int32_t diff = 0; + + /* Reset runtime to be in next 10ms */ + diff = timer_diff(tp->tentry.runtime, 0, + tcp_now, TCP_TIMER_10MS_QUANTUM); + if (diff > 0) { + tp->tentry.index = TCPT_KEEP; + tp->tentry.runtime = tcp_now + + TCP_TIMER_10MS_QUANTUM; + if (tp->tentry.runtime == 0) + tp->tentry.runtime++; + } + } + } +} + +/* + * Disable read probe and reset the keep alive timer + */ +static void +tcp_disable_read_probe(struct tcpcb *tp) +{ + if (tp->t_adaptive_rtimo == 0 && + ((tp->t_flagsext & TF_DETECT_READSTALL) || + tp->t_rtimo_probes > 0)) { + tcp_keepalive_reset(tp); + } +} + +/* + * Reschedule the tcp timerlist in the next 10ms to re-enable read/write + * probes on connections going over a particular interface. + */ +void +tcp_probe_connectivity(struct ifnet *ifp, u_int32_t enable) +{ + int32_t offset; + struct tcptimerlist *listp = &tcp_timer_list; + struct inpcbinfo *pcbinfo = &tcbinfo; + struct inpcb *inp, *nxt; + + if (ifp == NULL) + return; + + /* update clock */ + calculate_tcp_clock(); + + /* + * Enable keep alive timer on all connections that are + * active/established on this interface. + */ + lck_rw_lock_shared(pcbinfo->ipi_lock); + + LIST_FOREACH_SAFE(inp, pcbinfo->ipi_listhead, inp_list, nxt) { + struct tcpcb *tp = NULL; + if (in_pcb_checkstate(inp, WNT_ACQUIRE, 0) == + WNT_STOPUSING) + continue; + + /* Acquire lock to look at the state of the connection */ + tcp_lock(inp->inp_socket, 1, 0); + + /* Release the want count */ + if (inp->inp_ppcb == NULL || + (in_pcb_checkstate(inp, WNT_RELEASE, 1) == WNT_STOPUSING)) { + tcp_unlock(inp->inp_socket, 1, 0); + continue; + } + tp = intotcpcb(inp); + if (enable) + tcp_enable_read_probe(tp, ifp); + else + tcp_disable_read_probe(tp); + + tcp_unlock(inp->inp_socket, 1, 0); + } + lck_rw_done(pcbinfo->ipi_lock); + + lck_mtx_lock(listp->mtx); + if (listp->running) { + listp->pref_mode |= TCP_TIMERLIST_10MS_MODE; + goto done; + } + + /* Reschedule within the next 10ms */ + offset = TCP_TIMER_10MS_QUANTUM; + if (listp->scheduled) { + int32_t diff; + diff = timer_diff(listp->runtime, 0, tcp_now, offset); + if (diff <= 0) { + /* The timer will fire sooner than what's needed */ + goto done; + } + } + listp->mode = TCP_TIMERLIST_10MS_MODE; + listp->idleruns = 0; + + tcp_sched_timerlist(offset); +done: + lck_mtx_unlock(listp->mtx); + return; +} + +inline void +tcp_update_mss_core(struct tcpcb *tp, struct ifnet *ifp) +{ + struct if_cellular_status_v1 *ifsr; + u_int32_t optlen; + ifsr = &ifp->if_link_status->ifsr_u.ifsr_cell.if_cell_u.if_status_v1; + if (ifsr->valid_bitmask & IF_CELL_UL_MSS_RECOMMENDED_VALID) { + optlen = tp->t_maxopd - tp->t_maxseg; + + if (ifsr->mss_recommended == + IF_CELL_UL_MSS_RECOMMENDED_NONE && + tp->t_cached_maxopd > 0 && + tp->t_maxopd < tp->t_cached_maxopd) { + tp->t_maxopd = tp->t_cached_maxopd; + tcpstat.tcps_mss_to_default++; + } else if (ifsr->mss_recommended == + IF_CELL_UL_MSS_RECOMMENDED_MEDIUM && + tp->t_maxopd > tcp_mss_rec_medium) { + tp->t_cached_maxopd = tp->t_maxopd; + tp->t_maxopd = tcp_mss_rec_medium; + tcpstat.tcps_mss_to_medium++; + } else if (ifsr->mss_recommended == + IF_CELL_UL_MSS_RECOMMENDED_LOW && + tp->t_maxopd > tcp_mss_rec_low) { + tp->t_cached_maxopd = tp->t_maxopd; + tp->t_maxopd = tcp_mss_rec_low; + tcpstat.tcps_mss_to_low++; + } + tp->t_maxseg = tp->t_maxopd - optlen; + + /* + * clear the cached value if it is same as the current + */ + if (tp->t_maxopd == tp->t_cached_maxopd) + tp->t_cached_maxopd = 0; + } +} + +void +tcp_update_mss_locked(struct socket *so, struct ifnet *ifp) +{ + struct inpcb *inp = sotoinpcb(so); + struct tcpcb *tp = intotcpcb(inp); + + if (ifp == NULL && inp->inp_last_outifp == NULL) + return; + + if (ifp == NULL) + ifp = inp->inp_last_outifp; + + if (!IFNET_IS_CELLULAR(ifp)) { + /* + * This optimization is implemented for cellular + * networks only + */ + return; + } + if ( tp->t_state <= TCPS_CLOSE_WAIT) { + /* + * If the connection is currently doing or has done PMTU + * blackhole detection, do not change the MSS + */ + if (tp->t_flags & TF_BLACKHOLE) + return; + if (ifp->if_link_status == NULL) + return; + tcp_update_mss_core(tp, ifp); + } +} + +void +tcp_itimer(struct inpcbinfo *ipi) +{ + struct inpcb *inp, *nxt; + + if (lck_rw_try_lock_exclusive(ipi->ipi_lock) == FALSE) { + if (tcp_itimer_done == TRUE) { + tcp_itimer_done = FALSE; + atomic_add_32(&ipi->ipi_timer_req.intimer_fast, 1); + return; + } + /* Upgrade failed, lost lock now take it again exclusive */ + lck_rw_lock_exclusive(ipi->ipi_lock); + } + tcp_itimer_done = TRUE; + + LIST_FOREACH_SAFE(inp, &tcb, inp_list, nxt) { + struct socket *so; + + if (inp->inp_ppcb == NULL || + in_pcb_checkstate(inp, WNT_ACQUIRE, 0) == WNT_STOPUSING) + continue; + so = inp->inp_socket; + tcp_lock(so, 1, 0); + if (in_pcb_checkstate(inp, WNT_RELEASE, 1) == WNT_STOPUSING) { + tcp_unlock(so, 1, 0); + continue; + } + so_check_extended_bk_idle_time(so); + if (ipi->ipi_flags & INPCBINFO_UPDATE_MSS) { + tcp_update_mss_locked(so, NULL); + } + tcp_unlock(so, 1, 0); + } + + ipi->ipi_flags &= ~INPCBINFO_UPDATE_MSS; + lck_rw_done(ipi->ipi_lock); +}