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