[apple/xnu.git] / bsd / netinet / tcp_cc.c

/*
 * Copyright (c) 2013-2014 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
 * compliance with the License. The rights granted to you under the License
 * may not be used to create, or enable the creation or redistribution of,
 * 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,
 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 * 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@
 */

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/syslog.h>
#include <sys/protosw.h>
#include <sys/socketvar.h>
#include <sys/kern_control.h>
#include <sys/domain.h>

#include <netinet/in.h>
#include <netinet/tcp.h>
#include <netinet/tcp_var.h>
#include <netinet/tcp_cc.h>
#include <mach/sdt.h>
#include <libkern/OSAtomic.h>

struct tcp_cc_debug_state {
	u_int64_t ccd_tsns;
	char ccd_srcaddr[INET6_ADDRSTRLEN];
	uint16_t ccd_srcport;
	char ccd_destaddr[INET6_ADDRSTRLEN];
	uint16_t ccd_destport;
	uint32_t ccd_snd_cwnd;
	uint32_t ccd_snd_wnd;
	uint32_t ccd_snd_ssthresh;
	uint32_t ccd_pipeack;
	uint32_t ccd_rttcur;
	uint32_t ccd_rxtcur;
	uint32_t ccd_srtt;
	uint32_t ccd_event;
	uint32_t ccd_sndcc;
	uint32_t ccd_sndhiwat;
	uint32_t ccd_bytes_acked;
	u_int8_t ccd_cc_index;
	u_int8_t ccd_unused_1__;
	u_int16_t ccd_unused_2__;
	union {
		struct {
			uint32_t ccd_last_max;
			uint32_t ccd_tcp_win;
			uint32_t ccd_target_win;
			uint32_t ccd_avg_lastmax;
			uint32_t ccd_mean_deviation;
		} cubic_state;
		struct {
			u_int32_t led_base_rtt;
		} ledbat_state;
	} u;
};

int tcp_cc_debug = 0;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, cc_debug, CTLFLAG_RW | CTLFLAG_LOCKED,
	&tcp_cc_debug, 0, "Enable debug data collection");

extern struct tcp_cc_algo tcp_cc_newreno;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno_sockets,
	CTLFLAG_RD | CTLFLAG_LOCKED, &tcp_cc_newreno.num_sockets,
	0, "Number of sockets using newreno");

extern struct tcp_cc_algo tcp_cc_ledbat;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, background_sockets,
	CTLFLAG_RD | CTLFLAG_LOCKED, &tcp_cc_ledbat.num_sockets,
	0, "Number of sockets using background transport");

extern struct tcp_cc_algo tcp_cc_cubic;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, cubic_sockets,
	CTLFLAG_RD | CTLFLAG_LOCKED,&tcp_cc_cubic.num_sockets, 
	0, "Number of sockets using cubic");

int tcp_use_newreno = 0;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, use_newreno,
	CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_use_newreno, 0, 
	"Use TCP NewReno by default");

static int tcp_check_cwnd_nonvalidated = 1;
#if (DEBUG || DEVELOPMENT)
SYSCTL_INT(_net_inet_tcp, OID_AUTO, cwnd_nonvalidated,
    CTLFLAG_RW | CTLFLAG_LOCKED, &tcp_check_cwnd_nonvalidated, 0,
    "Check if congestion window is non-validated");
#endif /* (DEBUG || DEVELOPMENT) */

 #define SET_SNDSB_IDEAL_SIZE(sndsb, size) \
	sndsb->sb_idealsize = min(max(tcp_sendspace, tp->snd_ssthresh), \
	tcp_autosndbuf_max);

/* Array containing pointers to currently implemented TCP CC algorithms */
struct tcp_cc_algo* tcp_cc_algo_list[TCP_CC_ALGO_COUNT];
struct zone *tcp_cc_zone;

/* Information for colelcting TCP debug information using control socket */
#define TCP_CCDEBUG_CONTROL_NAME "com.apple.network.tcp_ccdebug"
#define TCP_CCDBG_NOUNIT 0xffffffff
static kern_ctl_ref tcp_ccdbg_ctlref = NULL;
volatile UInt32 tcp_ccdbg_unit = TCP_CCDBG_NOUNIT;

void tcp_cc_init(void);
static void tcp_cc_control_register(void);
static errno_t tcp_ccdbg_control_connect(kern_ctl_ref kctl,
	struct sockaddr_ctl *sac, void **uinfo);
static errno_t tcp_ccdbg_control_disconnect(kern_ctl_ref kctl,
	u_int32_t unit, void *uinfo);
static struct tcp_cc_algo tcp_cc_algo_none;
/*
 * Initialize TCP congestion control algorithms.
 */
 
void
tcp_cc_init(void)
{
	bzero(&tcp_cc_algo_list, sizeof(tcp_cc_algo_list));
	bzero(&tcp_cc_algo_none, sizeof(tcp_cc_algo_none));

	tcp_cc_algo_list[TCP_CC_ALGO_NONE] = &tcp_cc_algo_none;
	tcp_cc_algo_list[TCP_CC_ALGO_NEWRENO_INDEX] = &tcp_cc_newreno;
	tcp_cc_algo_list[TCP_CC_ALGO_BACKGROUND_INDEX] = &tcp_cc_ledbat;
	tcp_cc_algo_list[TCP_CC_ALGO_CUBIC_INDEX] = &tcp_cc_cubic;

	tcp_cc_control_register();
}

static void
tcp_cc_control_register(void)
{
	struct kern_ctl_reg ccdbg_control;
	errno_t err;

	bzero(&ccdbg_control, sizeof(ccdbg_control));
	strlcpy(ccdbg_control.ctl_name, TCP_CCDEBUG_CONTROL_NAME,
	    sizeof(ccdbg_control.ctl_name));
	ccdbg_control.ctl_connect = tcp_ccdbg_control_connect;
	ccdbg_control.ctl_disconnect = tcp_ccdbg_control_disconnect;
	ccdbg_control.ctl_flags |= CTL_FLAG_PRIVILEGED;
	ccdbg_control.ctl_flags |= CTL_FLAG_REG_SOCK_STREAM;

	err = ctl_register(&ccdbg_control, &tcp_ccdbg_ctlref);
	if (err != 0) {
		log(LOG_ERR, "failed to register tcp_cc debug control");
	}
}

/* Allow only one socket to connect at any time for debugging */
static errno_t
tcp_ccdbg_control_connect(kern_ctl_ref kctl, struct sockaddr_ctl *sac,
	void **uinfo)
{
#pragma unused(kctl)
#pragma unused(uinfo)

	UInt32 old_value = TCP_CCDBG_NOUNIT;
	UInt32 new_value = sac->sc_unit;

	if (tcp_ccdbg_unit != old_value)
		return (EALREADY);

	if (OSCompareAndSwap(old_value, new_value, &tcp_ccdbg_unit))
		return (0);
	else
		return (EALREADY);
}

static errno_t
tcp_ccdbg_control_disconnect(kern_ctl_ref kctl, u_int32_t unit, void *uinfo)
{
#pragma unused(kctl, unit, uinfo)

	if (unit == tcp_ccdbg_unit) {
		UInt32 old_value = tcp_ccdbg_unit;
		UInt32 new_value = TCP_CCDBG_NOUNIT;
		if (tcp_ccdbg_unit == new_value)
			return (0);

		if (!OSCompareAndSwap(old_value, new_value,
			&tcp_ccdbg_unit))
			log(LOG_DEBUG, 
			    "failed to disconnect tcp_cc debug control");
	}
	return (0);
}

inline void
tcp_ccdbg_trace(struct tcpcb *tp, struct tcphdr *th, int32_t event)
{
#if !CONFIG_DTRACE
#pragma unused(th)
#endif /* !CONFIG_DTRACE */
	struct inpcb *inp = tp->t_inpcb;

	if (tcp_cc_debug && tcp_ccdbg_unit > 0) {
		struct tcp_cc_debug_state dbg_state;
		struct timespec tv;

		bzero(&dbg_state, sizeof(dbg_state));
		
		nanotime(&tv);
		/* Take time in seconds */
		dbg_state.ccd_tsns = (tv.tv_sec * 1000000000) + tv.tv_nsec;
		inet_ntop(SOCK_DOM(inp->inp_socket), 
		    ((SOCK_DOM(inp->inp_socket) == PF_INET) ?
		    (void *)&inp->inp_laddr.s_addr :
		    (void *)&inp->in6p_laddr), dbg_state.ccd_srcaddr,
		    sizeof(dbg_state.ccd_srcaddr));
		dbg_state.ccd_srcport = ntohs(inp->inp_lport);
		inet_ntop(SOCK_DOM(inp->inp_socket),
		    ((SOCK_DOM(inp->inp_socket) == PF_INET) ?
		    (void *)&inp->inp_faddr.s_addr :
		    (void *)&inp->in6p_faddr), dbg_state.ccd_destaddr,
		    sizeof(dbg_state.ccd_destaddr));
		dbg_state.ccd_destport = ntohs(inp->inp_fport);

		dbg_state.ccd_snd_cwnd = tp->snd_cwnd;
		dbg_state.ccd_snd_wnd = tp->snd_wnd;
		dbg_state.ccd_snd_ssthresh = tp->snd_ssthresh;
		dbg_state.ccd_pipeack = tp->t_pipeack;
		dbg_state.ccd_rttcur = tp->t_rttcur;
		dbg_state.ccd_rxtcur = tp->t_rxtcur;
		dbg_state.ccd_srtt = tp->t_srtt >> TCP_RTT_SHIFT;
		dbg_state.ccd_event = event;
		dbg_state.ccd_sndcc = inp->inp_socket->so_snd.sb_cc;
		dbg_state.ccd_sndhiwat = inp->inp_socket->so_snd.sb_hiwat;
		dbg_state.ccd_bytes_acked = tp->t_bytes_acked;
		dbg_state.ccd_cc_index = tp->tcp_cc_index;
		switch (tp->tcp_cc_index) {
		    case TCP_CC_ALGO_CUBIC_INDEX:
			dbg_state.u.cubic_state.ccd_last_max =
			    tp->t_ccstate->cub_last_max;
			dbg_state.u.cubic_state.ccd_tcp_win =
			    tp->t_ccstate->cub_tcp_win;
			dbg_state.u.cubic_state.ccd_target_win =
			    tp->t_ccstate->cub_target_win;
			dbg_state.u.cubic_state.ccd_avg_lastmax =
			    tp->t_ccstate->cub_avg_lastmax;
			dbg_state.u.cubic_state.ccd_mean_deviation =
			    tp->t_ccstate->cub_mean_dev;
			break;
		    case TCP_CC_ALGO_BACKGROUND_INDEX:
			dbg_state.u.ledbat_state.led_base_rtt =
			    get_base_rtt(tp);
			break;
		    default:
			break;
		}

		ctl_enqueuedata(tcp_ccdbg_ctlref, tcp_ccdbg_unit,
			&dbg_state, sizeof(dbg_state), 0);
	}
	DTRACE_TCP5(cc, void, NULL, struct inpcb *, inp,
		struct tcpcb *, tp, struct tcphdr *, th, int32_t, event);
}

void tcp_cc_resize_sndbuf(struct tcpcb *tp)
{
	struct sockbuf *sb;
	/*
	 * If the send socket buffer size is bigger than ssthresh,
	 * it is time to trim it because we do not want to hold
	 * too many mbufs in the socket buffer
	 */
	sb = &tp->t_inpcb->inp_socket->so_snd;
	if (sb->sb_hiwat > tp->snd_ssthresh &&
		(sb->sb_flags & SB_AUTOSIZE)) {
		if (sb->sb_idealsize > tp->snd_ssthresh) {
			SET_SNDSB_IDEAL_SIZE(sb, tp->snd_ssthresh);
		}
		sb->sb_flags |= SB_TRIM;
	}
}

void tcp_bad_rexmt_fix_sndbuf(struct tcpcb *tp)
{
	struct sockbuf *sb;
	sb = &tp->t_inpcb->inp_socket->so_snd;
	if ((sb->sb_flags & (SB_TRIM|SB_AUTOSIZE)) == (SB_TRIM|SB_AUTOSIZE)) {
		/*
		 * If there was a retransmission that was not necessary 
		 * then the size of socket buffer can be restored to
		 * what it was before
		 */
		SET_SNDSB_IDEAL_SIZE(sb, tp->snd_ssthresh);
		if (sb->sb_hiwat <= sb->sb_idealsize) {
			sbreserve(sb, sb->sb_idealsize);
			sb->sb_flags &= ~SB_TRIM;
		}
	}
}

/*
 * Calculate initial cwnd according to RFC3390.
 *
 * Keep the old ss_fltsz sysctl for ABI compabitility issues.
 * but it will be overriden if tcp_do_rfc3390 sysctl when it is set.
 */
void
tcp_cc_cwnd_init_or_reset(struct tcpcb *tp)
{
	if (tp->t_flags & TF_LOCAL) {
		tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
	} else {
		/* initial congestion window according to RFC 3390 */
		if (tcp_do_rfc3390)
			tp->snd_cwnd = min(4 * tp->t_maxseg,
				max(2 * tp->t_maxseg, TCP_CC_CWND_INIT_BYTES));
		else
			tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
	}
}

/*
 * Indicate whether this ack should be delayed.
 * Here is the explanation for different settings of tcp_delack_enabled:
 *  - when set to 1, the bhavior is same as when set to 2. We kept this 
 *    for binary compatibility.
 *  - when set to 2, will "ack every other packet"
 *      - if our last ack wasn't a 0-sized window.
 *      - if the peer hasn't sent us a TH_PUSH data packet (radar 3649245). 
 *              If TH_PUSH is set, 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
 *  - when set to 3,  will do "streaming detection" 
 *      - if we receive more than "maxseg_unacked" full packets 
 *        in the last 100ms
 *      - if the connection is not in slow-start or idle or 
 *        loss/recovery states
 *      - if those criteria aren't met, it will ack every other packet.
 */
int
tcp_cc_delay_ack(struct tcpcb *tp, struct tcphdr *th)
{
	switch (tcp_delack_enabled) {
	    case 1:
	    case 2:
		if ((tp->t_flags & TF_RXWIN0SENT) == 0 &&
		    (th->th_flags & TH_PUSH) == 0 &&
		    (tp->t_unacksegs == 1))
			return(1);
		break;  
	    case 3:
		if ((tp->t_flags & TF_RXWIN0SENT) == 0 &&
		    (th->th_flags & TH_PUSH) == 0 &&
		    ((tp->t_unacksegs == 1) ||
		    ((tp->t_flags & TF_STRETCHACK) != 0 &&
			tp->t_unacksegs < (maxseg_unacked))))
			return(1);
		break;
	}
	return(0);
}

void
tcp_cc_allocate_state(struct tcpcb *tp)
{
	if (tp->tcp_cc_index == TCP_CC_ALGO_CUBIC_INDEX &&
		tp->t_ccstate == NULL) {
		tp->t_ccstate = (struct tcp_ccstate *)zalloc(tcp_cc_zone);

		/*
		 * If we could not allocate memory for congestion control
		 * state, revert to using TCP NewReno as it does not
		 * require any state
		 */
		if (tp->t_ccstate == NULL)
			tp->tcp_cc_index = TCP_CC_ALGO_NEWRENO_INDEX;
		else
			bzero(tp->t_ccstate, sizeof(*tp->t_ccstate));
	}
}

/*
 * If stretch ack was disabled automatically on long standing connections, 
 * re-evaluate the situation after 15 minutes to enable it.
 */
#define	TCP_STRETCHACK_DISABLE_WIN	(15 * 60 * TCP_RETRANSHZ)
void
tcp_cc_after_idle_stretchack(struct tcpcb *tp)
{
	int32_t tdiff;

	if (!(tp->t_flagsext & TF_DISABLE_STRETCHACK))
		return;

	tdiff = timer_diff(tcp_now, 0, tp->rcv_nostrack_ts, 0);
	if (tdiff < 0)
		tdiff = -tdiff;

	if (tdiff > TCP_STRETCHACK_DISABLE_WIN) {
		tp->t_flagsext &= ~TF_DISABLE_STRETCHACK;
		tp->t_stretchack_delayed = 0;

		tcp_reset_stretch_ack(tp);
	}
}

/*
 * Detect if the congestion window is non-vlidated according to
 * draft-ietf-tcpm-newcwv-07
 */

inline uint32_t
tcp_cc_is_cwnd_nonvalidated(struct tcpcb *tp)
{
	if (tp->t_pipeack == 0 || tcp_check_cwnd_nonvalidated == 0) {
		tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
		return (0);
	}
	if (tp->t_pipeack >= (tp->snd_cwnd) >> 1)
		tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
	else
		tp->t_flagsext |= TF_CWND_NONVALIDATED;
	return (tp->t_flagsext & TF_CWND_NONVALIDATED);
}

/*
 * Adjust congestion window in response to congestion in non-validated
 * phase.
 */
inline void
tcp_cc_adjust_nonvalidated_cwnd(struct tcpcb *tp)
{
	tp->t_pipeack = tcp_get_max_pipeack(tp);
	tcp_clear_pipeack_state(tp);
	tp->snd_cwnd = (max(tp->t_pipeack, tp->t_lossflightsize) >> 1);
	tp->snd_cwnd = max(tp->snd_cwnd, TCP_CC_CWND_INIT_BYTES);
	tp->snd_cwnd += tp->t_maxseg * tcprexmtthresh;
	tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
}

/*
 * Return maximum of all the pipeack samples. Since the number of samples
 * TCP_PIPEACK_SAMPLE_COUNT is 3 at this time, it will be simpler to do
 * a comparision. We should change ths if the number of samples increases.
 */
inline u_int32_t
tcp_get_max_pipeack(struct tcpcb *tp)
{
	u_int32_t max_pipeack = 0;
	max_pipeack = (tp->t_pipeack_sample[0] > tp->t_pipeack_sample[1]) ?
	    tp->t_pipeack_sample[0] : tp->t_pipeack_sample[1];
	max_pipeack = (tp->t_pipeack_sample[2] > max_pipeack) ?
	    tp->t_pipeack_sample[2] : max_pipeack;

	return (max_pipeack);
}

inline void
tcp_clear_pipeack_state(struct tcpcb *tp)
{
	bzero(tp->t_pipeack_sample, sizeof(tp->t_pipeack_sample));
	tp->t_pipeack_ind = 0;
	tp->t_lossflightsize = 0;
}
Commit	Line	Data
fe8ab488 A	1	/*
	2	* Copyright (c) 2013-2014 Apple Inc. All rights reserved.
	3	*
	4	* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
	5	*
	6	* This file contains Original Code and/or Modifications of Original Code
	7	* as defined in and that are subject to the Apple Public Source License
	8	* Version 2.0 (the 'License'). You may not use this file except in
	9	* compliance with the License. The rights granted to you under the License
	10	* may not be used to create, or enable the creation or redistribution of,
	11	* unlawful or unlicensed copies of an Apple operating system, or to
	12	* circumvent, violate, or enable the circumvention or violation of, any
	13	* terms of an Apple operating system software license agreement.
	14	*
	15	* Please obtain a copy of the License at
	16	* http://www.opensource.apple.com/apsl/ and read it before using this file.
	17	*
	18	* The Original Code and all software distributed under the License are
	19	* distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
	20	* EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
	21	* INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
	22	* FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
	23	* Please see the License for the specific language governing rights and
	24	* limitations under the License.
	25	*
	26	* @APPLE_OSREFERENCE_LICENSE_HEADER_END@
	27	*/
	28
	29	#include <sys/param.h>
	30	#include <sys/systm.h>
	31	#include <sys/kernel.h>
	32	#include <sys/syslog.h>
	33	#include <sys/protosw.h>
	34	#include <sys/socketvar.h>
	35	#include <sys/kern_control.h>
	36	#include <sys/domain.h>
	37
	38	#include <netinet/in.h>
	39	#include <netinet/tcp.h>
	40	#include <netinet/tcp_var.h>
	41	#include <netinet/tcp_cc.h>
	42	#include <mach/sdt.h>
	43	#include <libkern/OSAtomic.h>
	44
	45	struct tcp_cc_debug_state {
	46	u_int64_t ccd_tsns;
	47	char ccd_srcaddr[INET6_ADDRSTRLEN];
	48	uint16_t ccd_srcport;
	49	char ccd_destaddr[INET6_ADDRSTRLEN];
	50	uint16_t ccd_destport;
	51	uint32_t ccd_snd_cwnd;
	52	uint32_t ccd_snd_wnd;
	53	uint32_t ccd_snd_ssthresh;
3e170ce0	54	uint32_t ccd_pipeack;
fe8ab488 A	55	uint32_t ccd_rttcur;
	56	uint32_t ccd_rxtcur;
	57	uint32_t ccd_srtt;
	58	uint32_t ccd_event;
	59	uint32_t ccd_sndcc;
	60	uint32_t ccd_sndhiwat;
	61	uint32_t ccd_bytes_acked;
ecc0ceb4 A	62	u_int8_t ccd_cc_index;
	63	u_int8_t ccd_unused_1__;
	64	u_int16_t ccd_unused_2__;
fe8ab488 A	65	union {
	66	struct {
	67	uint32_t ccd_last_max;
	68	uint32_t ccd_tcp_win;
	69	uint32_t ccd_target_win;
	70	uint32_t ccd_avg_lastmax;
	71	uint32_t ccd_mean_deviation;
	72	} cubic_state;
ecc0ceb4 A	73	struct {
	74	u_int32_t led_base_rtt;
	75	} ledbat_state;
fe8ab488 A	76	} u;
	77	};
	78
	79	int tcp_cc_debug = 0;
	80	SYSCTL_INT(_net_inet_tcp, OID_AUTO, cc_debug, CTLFLAG_RW \| CTLFLAG_LOCKED,
	81	&tcp_cc_debug, 0, "Enable debug data collection");
	82
	83	extern struct tcp_cc_algo tcp_cc_newreno;
	84	SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno_sockets,
	85	CTLFLAG_RD \| CTLFLAG_LOCKED, &tcp_cc_newreno.num_sockets,
	86	0, "Number of sockets using newreno");
	87
	88	extern struct tcp_cc_algo tcp_cc_ledbat;
	89	SYSCTL_INT(_net_inet_tcp, OID_AUTO, background_sockets,
	90	CTLFLAG_RD \| CTLFLAG_LOCKED, &tcp_cc_ledbat.num_sockets,
	91	0, "Number of sockets using background transport");
	92
	93	extern struct tcp_cc_algo tcp_cc_cubic;
	94	SYSCTL_INT(_net_inet_tcp, OID_AUTO, cubic_sockets,
	95	CTLFLAG_RD \| CTLFLAG_LOCKED,&tcp_cc_cubic.num_sockets,
	96	0, "Number of sockets using cubic");
	97
	98	int tcp_use_newreno = 0;
	99	SYSCTL_INT(_net_inet_tcp, OID_AUTO, use_newreno,
	100	CTLFLAG_RW \| CTLFLAG_LOCKED, &tcp_use_newreno, 0,
	101	"Use TCP NewReno by default");
	102
3e170ce0 A	103	static int tcp_check_cwnd_nonvalidated = 1;
	104	#if (DEBUG \|\| DEVELOPMENT)
	105	SYSCTL_INT(_net_inet_tcp, OID_AUTO, cwnd_nonvalidated,
	106	CTLFLAG_RW \| CTLFLAG_LOCKED, &tcp_check_cwnd_nonvalidated, 0,
	107	"Check if congestion window is non-validated");
	108	#endif /* (DEBUG \|\| DEVELOPMENT) */
	109
fe8ab488 A	110	#define SET_SNDSB_IDEAL_SIZE(sndsb, size) \
	111	sndsb->sb_idealsize = min(max(tcp_sendspace, tp->snd_ssthresh), \
	112	tcp_autosndbuf_max);
	113
	114	/* Array containing pointers to currently implemented TCP CC algorithms */
	115	struct tcp_cc_algo* tcp_cc_algo_list[TCP_CC_ALGO_COUNT];
	116	struct zone *tcp_cc_zone;
	117
	118	/* Information for colelcting TCP debug information using control socket */
	119	#define TCP_CCDEBUG_CONTROL_NAME "com.apple.network.tcp_ccdebug"
	120	#define TCP_CCDBG_NOUNIT 0xffffffff
	121	static kern_ctl_ref tcp_ccdbg_ctlref = NULL;
	122	volatile UInt32 tcp_ccdbg_unit = TCP_CCDBG_NOUNIT;
	123
	124	void tcp_cc_init(void);
	125	static void tcp_cc_control_register(void);
	126	static errno_t tcp_ccdbg_control_connect(kern_ctl_ref kctl,
	127	struct sockaddr_ctl sac, void *uinfo);
	128	static errno_t tcp_ccdbg_control_disconnect(kern_ctl_ref kctl,
	129	u_int32_t unit, void *uinfo);
	130	static struct tcp_cc_algo tcp_cc_algo_none;
	131	/*
	132	* Initialize TCP congestion control algorithms.
	133	*/
	134
	135	void
	136	tcp_cc_init(void)
	137	{
	138	bzero(&tcp_cc_algo_list, sizeof(tcp_cc_algo_list));
	139	bzero(&tcp_cc_algo_none, sizeof(tcp_cc_algo_none));
	140
	141	tcp_cc_algo_list[TCP_CC_ALGO_NONE] = &tcp_cc_algo_none;
	142	tcp_cc_algo_list[TCP_CC_ALGO_NEWRENO_INDEX] = &tcp_cc_newreno;
	143	tcp_cc_algo_list[TCP_CC_ALGO_BACKGROUND_INDEX] = &tcp_cc_ledbat;
	144	tcp_cc_algo_list[TCP_CC_ALGO_CUBIC_INDEX] = &tcp_cc_cubic;
	145
	146	tcp_cc_control_register();
	147	}
	148
	149	static void
	150	tcp_cc_control_register(void)
	151	{
	152	struct kern_ctl_reg ccdbg_control;
	153	errno_t err;
	154
	155	bzero(&ccdbg_control, sizeof(ccdbg_control));
	156	strlcpy(ccdbg_control.ctl_name, TCP_CCDEBUG_CONTROL_NAME,
	157	sizeof(ccdbg_control.ctl_name));
	158	ccdbg_control.ctl_connect = tcp_ccdbg_control_connect;
	159	ccdbg_control.ctl_disconnect = tcp_ccdbg_control_disconnect;
	160	ccdbg_control.ctl_flags \|= CTL_FLAG_PRIVILEGED;
	161	ccdbg_control.ctl_flags \|= CTL_FLAG_REG_SOCK_STREAM;
	162
	163	err = ctl_register(&ccdbg_control, &tcp_ccdbg_ctlref);
	164	if (err != 0) {
	165	log(LOG_ERR, "failed to register tcp_cc debug control");
	166	}
	167	}
	168
	169	/* Allow only one socket to connect at any time for debugging */
	170	static errno_t
	171	tcp_ccdbg_control_connect(kern_ctl_ref kctl, struct sockaddr_ctl *sac,
	172	void **uinfo)
	173	{
174	#pragma unused(kctl)
175	#pragma unused(uinfo)
176
177	UInt32 old_value = TCP_CCDBG_NOUNIT;
178	UInt32 new_value = sac->sc_unit;
179
180	if (tcp_ccdbg_unit != old_value)
181	return (EALREADY);
182
183	if (OSCompareAndSwap(old_value, new_value, &tcp_ccdbg_unit))
184	return (0);
185	else
186	return (EALREADY);
187	}
188
189	static errno_t
190	tcp_ccdbg_control_disconnect(kern_ctl_ref kctl, u_int32_t unit, void *uinfo)
191	{
192	#pragma unused(kctl, unit, uinfo)
193
194	if (unit == tcp_ccdbg_unit) {
195	UInt32 old_value = tcp_ccdbg_unit;
196	UInt32 new_value = TCP_CCDBG_NOUNIT;
197	if (tcp_ccdbg_unit == new_value)
198	return (0);
199
200	if (!OSCompareAndSwap(old_value, new_value,
201	&tcp_ccdbg_unit))
202	log(LOG_DEBUG,
203	"failed to disconnect tcp_cc debug control");
204	}
205	return (0);
206	}
207
208	inline void
209	tcp_ccdbg_trace(struct tcpcb tp, struct tcphdr th, int32_t event)
210	{
211	#if !CONFIG_DTRACE
212	#pragma unused(th)
213	#endif /* !CONFIG_DTRACE */
214	struct inpcb *inp = tp->t_inpcb;
215
216	if (tcp_cc_debug && tcp_ccdbg_unit > 0) {
217	struct tcp_cc_debug_state dbg_state;
218	struct timespec tv;
219
220	bzero(&dbg_state, sizeof(dbg_state));
221
222	nanotime(&tv);
223	/* Take time in seconds */
224	dbg_state.ccd_tsns = (tv.tv_sec * 1000000000) + tv.tv_nsec;
225	inet_ntop(SOCK_DOM(inp->inp_socket),
226	((SOCK_DOM(inp->inp_socket) == PF_INET) ?
227	(void *)&inp->inp_laddr.s_addr :
228	(void *)&inp->in6p_laddr), dbg_state.ccd_srcaddr,
229	sizeof(dbg_state.ccd_srcaddr));
230	dbg_state.ccd_srcport = ntohs(inp->inp_lport);
231	inet_ntop(SOCK_DOM(inp->inp_socket),
232	((SOCK_DOM(inp->inp_socket) == PF_INET) ?
233	(void *)&inp->inp_faddr.s_addr :
234	(void *)&inp->in6p_faddr), dbg_state.ccd_destaddr,
235	sizeof(dbg_state.ccd_destaddr));
236	dbg_state.ccd_destport = ntohs(inp->inp_fport);
237
238	dbg_state.ccd_snd_cwnd = tp->snd_cwnd;
239	dbg_state.ccd_snd_wnd = tp->snd_wnd;
240	dbg_state.ccd_snd_ssthresh = tp->snd_ssthresh;
3e170ce0	241	dbg_state.ccd_pipeack = tp->t_pipeack;
fe8ab488 A	242	dbg_state.ccd_rttcur = tp->t_rttcur;
	243	dbg_state.ccd_rxtcur = tp->t_rxtcur;
	244	dbg_state.ccd_srtt = tp->t_srtt >> TCP_RTT_SHIFT;
	245	dbg_state.ccd_event = event;
	246	dbg_state.ccd_sndcc = inp->inp_socket->so_snd.sb_cc;
	247	dbg_state.ccd_sndhiwat = inp->inp_socket->so_snd.sb_hiwat;
	248	dbg_state.ccd_bytes_acked = tp->t_bytes_acked;
ecc0ceb4	249	dbg_state.ccd_cc_index = tp->tcp_cc_index;
fe8ab488 A	250	switch (tp->tcp_cc_index) {
	251	case TCP_CC_ALGO_CUBIC_INDEX:
	252	dbg_state.u.cubic_state.ccd_last_max =
	253	tp->t_ccstate->cub_last_max;
	254	dbg_state.u.cubic_state.ccd_tcp_win =
	255	tp->t_ccstate->cub_tcp_win;
	256	dbg_state.u.cubic_state.ccd_target_win =
	257	tp->t_ccstate->cub_target_win;
	258	dbg_state.u.cubic_state.ccd_avg_lastmax =
	259	tp->t_ccstate->cub_avg_lastmax;
	260	dbg_state.u.cubic_state.ccd_mean_deviation =
	261	tp->t_ccstate->cub_mean_dev;
	262	break;
ecc0ceb4 A	263	case TCP_CC_ALGO_BACKGROUND_INDEX:
	264	dbg_state.u.ledbat_state.led_base_rtt =
	265	get_base_rtt(tp);
	266	break;
fe8ab488 A	267	default:
	268	break;
	269	}
	270
	271	ctl_enqueuedata(tcp_ccdbg_ctlref, tcp_ccdbg_unit,
	272	&dbg_state, sizeof(dbg_state), 0);
	273	}
	274	DTRACE_TCP5(cc, void, NULL, struct inpcb *, inp,
	275	struct tcpcb , tp, struct tcphdr , th, int32_t, event);
	276	}
	277
	278	void tcp_cc_resize_sndbuf(struct tcpcb *tp)
	279	{
	280	struct sockbuf *sb;
	281	/*
	282	* If the send socket buffer size is bigger than ssthresh,
	283	* it is time to trim it because we do not want to hold
	284	* too many mbufs in the socket buffer
	285	*/
	286	sb = &tp->t_inpcb->inp_socket->so_snd;
	287	if (sb->sb_hiwat > tp->snd_ssthresh &&
	288	(sb->sb_flags & SB_AUTOSIZE)) {
	289	if (sb->sb_idealsize > tp->snd_ssthresh) {
	290	SET_SNDSB_IDEAL_SIZE(sb, tp->snd_ssthresh);
	291	}
	292	sb->sb_flags \|= SB_TRIM;
	293	}
	294	}
	295
	296	void tcp_bad_rexmt_fix_sndbuf(struct tcpcb *tp)
	297	{
	298	struct sockbuf *sb;
	299	sb = &tp->t_inpcb->inp_socket->so_snd;
	300	if ((sb->sb_flags & (SB_TRIM\|SB_AUTOSIZE)) == (SB_TRIM\|SB_AUTOSIZE)) {
	301	/*
	302	* If there was a retransmission that was not necessary
	303	* then the size of socket buffer can be restored to
	304	* what it was before
	305	*/
	306	SET_SNDSB_IDEAL_SIZE(sb, tp->snd_ssthresh);
	307	if (sb->sb_hiwat <= sb->sb_idealsize) {
	308	sbreserve(sb, sb->sb_idealsize);
	309	sb->sb_flags &= ~SB_TRIM;
	310	}
	311	}
	312	}
	313
	314	/*
	315	* Calculate initial cwnd according to RFC3390.
	316	*
	317	* Keep the old ss_fltsz sysctl for ABI compabitility issues.
	318	* but it will be overriden if tcp_do_rfc3390 sysctl when it is set.
	319	*/
	320	void
	321	tcp_cc_cwnd_init_or_reset(struct tcpcb *tp)
	322	{
	323	if (tp->t_flags & TF_LOCAL) {
	324	tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
	325	} else {
	326	/* initial congestion window according to RFC 3390 */
	327	if (tcp_do_rfc3390)
	328	tp->snd_cwnd = min(4 * tp->t_maxseg,
	329	max(2 * tp->t_maxseg, TCP_CC_CWND_INIT_BYTES));
	330	else
331	tp->snd_cwnd = tp->t_maxseg * ss_fltsz;
332	}
333	}
334
335	/*
336	* Indicate whether this ack should be delayed.
337	* Here is the explanation for different settings of tcp_delack_enabled:
338	* - when set to 1, the bhavior is same as when set to 2. We kept this
339	* for binary compatibility.
340	* - when set to 2, will "ack every other packet"
341	* - if our last ack wasn't a 0-sized window.
342	* - if the peer hasn't sent us a TH_PUSH data packet (radar 3649245).
343	* If TH_PUSH is set, take this as a clue that we need to ACK
344	* with no delay. This helps higher level protocols who
345	* won't send us more data even if the window is open
346	* because their last "segment" hasn't been ACKed
347	* - when set to 3, will do "streaming detection"
348	* - if we receive more than "maxseg_unacked" full packets
349	* in the last 100ms
350	* - if the connection is not in slow-start or idle or
351	* loss/recovery states
352	* - if those criteria aren't met, it will ack every other packet.
353	*/
354	int
355	tcp_cc_delay_ack(struct tcpcb tp, struct tcphdr th)
356	{
fe8ab488 A	357	switch (tcp_delack_enabled) {
	358	case 1:
	359	case 2:
	360	if ((tp->t_flags & TF_RXWIN0SENT) == 0 &&
	361	(th->th_flags & TH_PUSH) == 0 &&
	362	(tp->t_unacksegs == 1))
	363	return(1);
	364	break;
	365	case 3:
	366	if ((tp->t_flags & TF_RXWIN0SENT) == 0 &&
	367	(th->th_flags & TH_PUSH) == 0 &&
	368	((tp->t_unacksegs == 1) \|\|
	369	((tp->t_flags & TF_STRETCHACK) != 0 &&
	370	tp->t_unacksegs < (maxseg_unacked))))
	371	return(1);
	372	break;
	373	}
	374	return(0);
	375	}
	376
	377	void
	378	tcp_cc_allocate_state(struct tcpcb *tp)
	379	{
	380	if (tp->tcp_cc_index == TCP_CC_ALGO_CUBIC_INDEX &&
	381	tp->t_ccstate == NULL) {
	382	tp->t_ccstate = (struct tcp_ccstate *)zalloc(tcp_cc_zone);
	383
	384	/*
	385	* If we could not allocate memory for congestion control
	386	* state, revert to using TCP NewReno as it does not
	387	* require any state
	388	*/
	389	if (tp->t_ccstate == NULL)
	390	tp->tcp_cc_index = TCP_CC_ALGO_NEWRENO_INDEX;
	391	else
	392	bzero(tp->t_ccstate, sizeof(*tp->t_ccstate));
	393	}
	394	}
	395
	396	/*
	397	* If stretch ack was disabled automatically on long standing connections,
	398	* re-evaluate the situation after 15 minutes to enable it.
	399	*/
	400	#define TCP_STRETCHACK_DISABLE_WIN (15 * 60 * TCP_RETRANSHZ)
	401	void
	402	tcp_cc_after_idle_stretchack(struct tcpcb *tp)
	403	{
	404	int32_t tdiff;
	405
	406	if (!(tp->t_flagsext & TF_DISABLE_STRETCHACK))
	407	return;
	408
	409	tdiff = timer_diff(tcp_now, 0, tp->rcv_nostrack_ts, 0);
	410	if (tdiff < 0)
	411	tdiff = -tdiff;
	412
	413	if (tdiff > TCP_STRETCHACK_DISABLE_WIN) {
	414	tp->t_flagsext &= ~TF_DISABLE_STRETCHACK;
	415	tp->t_stretchack_delayed = 0;
	416
	417	tcp_reset_stretch_ack(tp);
	418	}
	419	}
3e170ce0 A	420
	421	/*
	422	* Detect if the congestion window is non-vlidated according to
	423	* draft-ietf-tcpm-newcwv-07
	424	*/
	425
	426	inline uint32_t
	427	tcp_cc_is_cwnd_nonvalidated(struct tcpcb *tp)
	428	{
	429	if (tp->t_pipeack == 0 \|\| tcp_check_cwnd_nonvalidated == 0) {
	430	tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
	431	return (0);
	432	}
	433	if (tp->t_pipeack >= (tp->snd_cwnd) >> 1)
	434	tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
	435	else
	436	tp->t_flagsext \|= TF_CWND_NONVALIDATED;
	437	return (tp->t_flagsext & TF_CWND_NONVALIDATED);
	438	}
	439
	440	/*
	441	* Adjust congestion window in response to congestion in non-validated
	442	* phase.
	443	*/
	444	inline void
	445	tcp_cc_adjust_nonvalidated_cwnd(struct tcpcb *tp)
	446	{
	447	tp->t_pipeack = tcp_get_max_pipeack(tp);
	448	tcp_clear_pipeack_state(tp);
	449	tp->snd_cwnd = (max(tp->t_pipeack, tp->t_lossflightsize) >> 1);
	450	tp->snd_cwnd = max(tp->snd_cwnd, TCP_CC_CWND_INIT_BYTES);
	451	tp->snd_cwnd += tp->t_maxseg * tcprexmtthresh;
	452	tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
	453	}
	454
	455	/*
	456	* Return maximum of all the pipeack samples. Since the number of samples
	457	* TCP_PIPEACK_SAMPLE_COUNT is 3 at this time, it will be simpler to do
	458	* a comparision. We should change ths if the number of samples increases.
	459	*/
	460	inline u_int32_t
	461	tcp_get_max_pipeack(struct tcpcb *tp)
	462	{
	463	u_int32_t max_pipeack = 0;
	464	max_pipeack = (tp->t_pipeack_sample[0] > tp->t_pipeack_sample[1]) ?
	465	tp->t_pipeack_sample[0] : tp->t_pipeack_sample[1];
	466	max_pipeack = (tp->t_pipeack_sample[2] > max_pipeack) ?
	467	tp->t_pipeack_sample[2] : max_pipeack;
	468
	469	return (max_pipeack);
	470	}
	471
	472	inline void
	473	tcp_clear_pipeack_state(struct tcpcb *tp)
	474	{
	475	bzero(tp->t_pipeack_sample, sizeof(tp->t_pipeack_sample));
	476	tp->t_pipeack_ind = 0;
	477	tp->t_lossflightsize = 0;
	478	}