tp_subr.c

早期freebsd实现

		idleticks = tpcb->tp_inact_ticks - tpcb->tp_timer[TM_inact];
		if (idleticks > tpcb->tp_dt_ticks)
			/*
			 * We have been idle for "a while" and no acks are
			 * expected to clock out any data we send --
			 * slow start to get ack "clock" running again.
			 */
			tpcb->tp_cong_win = tpcb->tp_l_tpdusize;
	}

	cong_win = tpcb->tp_cong_win;
	highseq = SEQ(tpcb, tpcb->tp_fcredit + tpcb->tp_snduna);
	if (tpcb->tp_Xsnd.sb_mb)
		highseq = SEQ_MIN(tpcb, highseq, tpcb->tp_sndnew);
		
	IFDEBUG(D_DATA)
		printf("tp_send enter tpcb 0x%x nxt 0x%x win %d high 0x%x\n",
				tpcb, tpcb->tp_sndnxt, cong_win, highseq);
	ENDDEBUG
	IFTRACE(D_DATA)
		tptraceTPCB( TPPTmisc, "tp_send sndnew snduna", 
			tpcb->tp_sndnew,  tpcb->tp_snduna, 0, 0);
		tptraceTPCB( TPPTmisc, "tp_send tpcb->tp_sndnxt win fcredit congwin", 
			tpcb->tp_sndnxt, cong_win, tpcb->tp_fcredit, tpcb->tp_cong_win);
	ENDTRACE
	IFTRACE(D_DATA)
		tptraceTPCB( TPPTmisc, "tp_send 2 nxt high fcredit congwin", 
			tpcb->tp_sndnxt, highseq, tpcb->tp_fcredit, cong_win);
	ENDTRACE

	if (tpcb->tp_sndnxt_m)
		m = tpcb->tp_sndnxt_m;
	else {
		off = SEQ_SUB(tpcb, tpcb->tp_sndnxt, tpcb->tp_snduna);
		for (m = sb->sb_mb; m && off > 0; m = m->m_next)
			off--;
	}
send:
	/*
	 * Avoid silly window syndrome here . . . figure out how!
	 */
	checkseq = tpcb->tp_sndnum;
	if (idle && SEQ_LT(tpcb, tpcb->tp_sndnum, highseq))
		checkseq = highseq; /* i.e. DON'T retain highest assigned packet */

	while ((SEQ_LT(tpcb, tpcb->tp_sndnxt, highseq)) && m && cong_win > 0) {

		eotsdu = (m->m_flags & M_EOR) != 0;
		len = m->m_pkthdr.len;
		if (tpcb->tp_sndnxt == checkseq && eotsdu == 0 &&
			len < (tpcb->tp_l_tpdusize / 2))
				break;  /* Nagle . . . . . */
		cong_win -= len;
		/* make a copy - mb goes into the retransmission list 
		 * while m gets emitted.  m_copy won't copy a zero-length mbuf.
		 */
		mb = m;
		m = m_copy(mb, 0, M_COPYALL);
		if (m == MNULL)
				break;
		IFTRACE(D_STASH)
			tptraceTPCB( TPPTmisc, 
				"tp_send mcopy nxt high eotsdu len", 
				tpcb->tp_sndnxt, highseq, eotsdu, len);
		ENDTRACE

		IFDEBUG(D_DATA)
			printf("tp_sending tpcb 0x%x nxt 0x%x\n",
				tpcb, tpcb->tp_sndnxt);
		ENDDEBUG
		/* when headers are precomputed, may need to fill
			   in checksum here */
		if (tpcb->tp_sock->so_error =
			tp_emit(DT_TPDU_type, tpcb, tpcb->tp_sndnxt, eotsdu, m)) {
			/* error */
			break;
		}
		m = mb->m_nextpkt;
		tpcb->tp_sndnxt_m = m;
		if (tpcb->tp_sndnxt == tpcb->tp_sndnew) {
			SEQ_INC(tpcb, tpcb->tp_sndnew);
			/*
			 * Time this transmission if not a retransmission and
			 * not currently timing anything.
			 */
			if (tpcb->tp_rttemit == 0) {
				tpcb->tp_rttemit = ticks;
				tpcb->tp_rttseq = tpcb->tp_sndnxt;
			}
			tpcb->tp_sndnxt = tpcb->tp_sndnew;
		} else
			SEQ_INC(tpcb, tpcb->tp_sndnxt);
		/*
		 * Set retransmit timer if not currently set.
		 * Initial value for retransmit timer is smoothed
		 * round-trip time + 2 * round-trip time variance.
		 * Initialize shift counter which is used for backoff
		 * of retransmit time.
		 */
		if (tpcb->tp_timer[TM_data_retrans] == 0 &&
			tpcb->tp_class != TP_CLASS_0) {
			tpcb->tp_timer[TM_data_retrans] = tpcb->tp_dt_ticks;
			tpcb->tp_timer[TM_sendack] = tpcb->tp_keepalive_ticks;
			tpcb->tp_rxtshift = 0;
		}
	}
	if (SEQ_GT(tpcb, tpcb->tp_sndnew, tpcb->tp_sndnum))
		tpcb->tp_oktonagle = 0;
#ifdef TP_PERF_MEAS
	IFPERF(tpcb)
		{
			register int npkts;
			int	 elapsed = ticks - send_start_time, *t;
			struct timeval now;

			npkts = SEQ_SUB(tpcb, tpcb->tp_sndnxt, oldnxt);

			if (npkts > 0) 
				tpcb->tp_Nwindow++;

			if (npkts > TP_PM_MAX) 
				npkts = TP_PM_MAX; 

			t = &(tpcb->tp_p_meas->tps_sendtime[npkts]);
			*t += (t - elapsed) >> TP_RTT_ALPHA;

			if (mb == 0) {
				IncPStat(tpcb, tps_win_lim_by_data[npkts] );
			} else {
				IncPStat(tpcb, tps_win_lim_by_cdt[npkts] );
				/* not true with congestion-window being used */
			}
			now.tv_sec = elapsed / hz;
			now.tv_usec = (elapsed - (hz * now.tv_sec)) * 1000000 / hz;
			tpmeas( tpcb->tp_lref, 
					TPsbsend, &elapsed, newseq, tpcb->tp_Nwindow, npkts);
		}
	ENDPERF
#endif /* TP_PERF_MEAS */


	IFTRACE(D_DATA)
		tptraceTPCB( TPPTmisc, 
			"tp_send at end: new nxt eotsdu error",
			tpcb->tp_sndnew, tpcb->tp_sndnxt, eotsdu, tpcb->tp_sock->so_error);
		
	ENDTRACE
}

int TPNagleok;
int TPNagled;

tp_packetize(tpcb, m, eotsdu)
register struct tp_pcb *tpcb;
register struct mbuf *m;
int eotsdu;
{
	register struct mbuf *n;
	register struct sockbuf *sb = &tpcb->tp_sock->so_snd;
	int	maxsize = tpcb->tp_l_tpdusize 
			- tp_headersize(DT_TPDU_type, tpcb)
			- (tpcb->tp_use_checksum?4:0) ;
	int totlen = m->m_pkthdr.len;
	struct mbuf *m_split();
	/*
	 * Pre-packetize the data in the sockbuf
	 * according to negotiated mtu.  Do it here
	 * where we can safely wait for mbufs.
	 *
	 * This presumes knowledge of sockbuf conventions.
	 * TODO: allocate space for header and fill it in (once!).
	 */
	IFDEBUG(D_DATA)
		printf("SEND BF: maxsize %d totlen %d eotsdu %d sndnum 0x%x\n",
			maxsize, totlen, eotsdu, tpcb->tp_sndnum);
	ENDTRACE
	if (tpcb->tp_oktonagle) {
		if ((n = sb->sb_mb) == 0)
			panic("tp_packetize");
		while (n->m_act)
			n = n->m_act;
		if (n->m_flags & M_EOR)
			panic("tp_packetize 2");
		SEQ_INC(tpcb, tpcb->tp_sndnum);
		if (totlen + n->m_pkthdr.len < maxsize) {
			/* There is an unsent packet with space, combine data */
			struct mbuf *old_n = n;
			tpsbcheck(tpcb,3);
			n->m_pkthdr.len += totlen;
			while (n->m_next)
				n = n->m_next;
			sbcompress(sb, m, n);
			tpsbcheck(tpcb,4);
			n = old_n;
			TPNagled++;
			goto out;
		}
	}
	while (m) {
		n = m;
		if (totlen > maxsize) {
			if ((m = m_split(n, maxsize, M_WAIT)) == 0)
				panic("tp_packetize");
		} else
			m = 0;
		totlen -= maxsize;
		tpsbcheck(tpcb, 5);
		sbappendrecord(sb, n);
		tpsbcheck(tpcb, 6);
		SEQ_INC(tpcb, tpcb->tp_sndnum);
	}
out:
	if (eotsdu) {
		n->m_flags |= M_EOR;  /* XXX belongs at end */
		tpcb->tp_oktonagle = 0;
	} else {
		SEQ_DEC(tpcb, tpcb->tp_sndnum);
		tpcb->tp_oktonagle = 1;
		TPNagleok++;
	}
	IFDEBUG(D_DATA)
		printf("SEND out: oktonagle %d sndnum 0x%x\n",
			tpcb->tp_oktonagle, tpcb->tp_sndnum);
	ENDTRACE
	return 0;
}


/*
 * NAME: tp_stash()
 * CALLED FROM:
 *	tp.trans on arrival of a DT tpdu
 * FUNCTION, ARGUMENTS, and RETURN VALUE:
 * 	Returns 1 if 
 *		a) something new arrived and it's got eotsdu_reached bit on,
 * 		b) this arrival was caused other out-of-sequence things to be
 *    	accepted, or
 * 		c) this arrival is the highest seq # for which we last gave credit
 *   	(sender just sent a whole window)
 *  In other words, returns 1 if tp should send an ack immediately, 0 if 
 *  the ack can wait a while.
 *
 * Note: this implementation no longer renegs on credit, (except
 * when debugging option D_RENEG is on, for the purpose of testing
 * ack subsequencing), so we don't  need to check for incoming tpdus 
 * being in a reneged portion of the window.
 */

tp_stash(tpcb, e)
	register struct tp_pcb		*tpcb;
	register struct tp_event	*e;
{
	register int		ack_reason= tpcb->tp_ack_strat & ACK_STRAT_EACH;
									/* 0--> delay acks until full window */
									/* 1--> ack each tpdu */
#ifndef lint
#define E e->ATTR(DT_TPDU)
#else /* lint */
#define E e->ev_union.EV_DT_TPDU
#endif /* lint */

	if ( E.e_eot ) {
		register struct mbuf *n = E.e_data;
		n->m_flags |= M_EOR;
		n->m_act = 0;
	}
		IFDEBUG(D_STASH)
			dump_mbuf(tpcb->tp_sock->so_rcv.sb_mb, 
				"stash: so_rcv before appending");
			dump_mbuf(E.e_data,
				"stash: e_data before appending");
		ENDDEBUG

	IFPERF(tpcb)
		PStat(tpcb, Nb_from_ll) += E.e_datalen;
		tpmeas(tpcb->tp_lref, TPtime_from_ll, &e->e_time,
			E.e_seq, (u_int)PStat(tpcb, Nb_from_ll), (u_int)E.e_datalen);
	ENDPERF

	if (E.e_seq == tpcb->tp_rcvnxt) {

		IFDEBUG(D_STASH)
			printf("stash EQ: seq 0x%x datalen 0x%x eot 0x%x\n", 
			E.e_seq, E.e_datalen, E.e_eot);
		ENDDEBUG

		IFTRACE(D_STASH)
			tptraceTPCB(TPPTmisc, "stash EQ: seq len eot", 
			E.e_seq, E.e_datalen, E.e_eot, 0);
		ENDTRACE

		SET_DELACK(tpcb);

		sbappend(&tpcb->tp_sock->so_rcv, E.e_data);

		SEQ_INC( tpcb, tpcb->tp_rcvnxt );
		/* 
		 * move chains from the reassembly queue to the socket buffer
		 */
		if (tpcb->tp_rsycnt) {
			register struct mbuf **mp;
			struct mbuf **mplim;

			mp = tpcb->tp_rsyq + (tpcb->tp_rcvnxt % tpcb->tp_maxlcredit);
			mplim = tpcb->tp_rsyq + tpcb->tp_maxlcredit;

			while (tpcb->tp_rsycnt && *mp) {
				sbappend(&tpcb->tp_sock->so_rcv, *mp);
				tpcb->tp_rsycnt--;
				*mp = 0;
				SEQ_INC(tpcb, tpcb->tp_rcvnxt);
				ack_reason |= ACK_REORDER;
				if (++mp == mplim)
					mp = tpcb->tp_rsyq;
			}
		}
		IFDEBUG(D_STASH)
			dump_mbuf(tpcb->tp_sock->so_rcv.sb_mb, 
				"stash: so_rcv after appending");
		ENDDEBUG

	} else {
		register struct mbuf **mp;
		SeqNum uwe;

		IFTRACE(D_STASH)
			tptraceTPCB(TPPTmisc, "stash Reseq: seq rcvnxt lcdt", 
			E.e_seq, tpcb->tp_rcvnxt, tpcb->tp_lcredit, 0);
		ENDTRACE

		if (tpcb->tp_rsyq == 0)
			tp_rsyset(tpcb);
		uwe = SEQ(tpcb, tpcb->tp_rcvnxt + tpcb->tp_maxlcredit);
		if (tpcb->tp_rsyq == 0 ||
						!IN_RWINDOW(tpcb, E.e_seq, tpcb->tp_rcvnxt, uwe)) {
			ack_reason = ACK_DONT;
			m_freem(E.e_data);
		} else if (*(mp = tpcb->tp_rsyq + (E.e_seq % tpcb->tp_maxlcredit))) {
			IFDEBUG(D_STASH)
				printf("tp_stash - drop & ack\n");
			ENDDEBUG

			/* retransmission - drop it and force an ack */
			IncStat(ts_dt_dup);
			IFPERF(tpcb)
				IncPStat(tpcb, tps_n_ack_cuz_dup);
			ENDPERF

			m_freem(E.e_data);
			ack_reason |= ACK_DUP;
		} else {
			*mp = E.e_data;
			tpcb->tp_rsycnt++;
			ack_reason = ACK_DONT;
		}
	}
	/* there were some comments of historical interest here. */
	{
		LOCAL_CREDIT(tpcb);

		if ( E.e_seq ==  tpcb->tp_sent_uwe )
			ack_reason |= ACK_STRAT_FULLWIN;

		IFTRACE(D_STASH)
			tptraceTPCB(TPPTmisc, 
				"end of stash, eot, ack_reason, sent_uwe ",
				E.e_eot, ack_reason, tpcb->tp_sent_uwe, 0); 
		ENDTRACE

		if ( ack_reason == ACK_DONT ) {
			IncStat( ts_ackreason[ACK_DONT] );
			return 0;
		} else {
			IFPERF(tpcb)
				if(ack_reason & ACK_STRAT_EACH) {
					IncPStat(tpcb, tps_n_ack_cuz_strat);
				} else if(ack_reason & ACK_STRAT_FULLWIN) {
					IncPStat(tpcb, tps_n_ack_cuz_fullwin);
				} else if(ack_reason & ACK_REORDER) {
					IncPStat(tpcb, tps_n_ack_cuz_reorder);
				}
				tpmeas(tpcb->tp_lref, TPtime_ack_sent, 0, 
							SEQ_ADD(tpcb, E.e_seq, 1), 0, 0);
			ENDPERF
			{
				register int i;

				/* keep track of all reasons that apply */
				for( i=1; i<_ACK_NUM_REASONS_ ;i++) {
					if( ack_reason & (1<<i) ) 
						IncStat( ts_ackreason[i] );
				}
			}
			return 1;
		}
	}
}

/*
 * tp_rsyflush - drop all the packets on the reassembly queue.
 * Do this when closing the socket, or when somebody has changed
 * the space avaible in the receive socket (XXX).
 */
tp_rsyflush(tpcb)
register struct tp_pcb *tpcb;
{
	register struct mbuf *m, **mp;
	if (tpcb->tp_rsycnt) {
		for (mp == tpcb->tp_rsyq + tpcb->tp_maxlcredit;
									 --mp >= tpcb->tp_rsyq; )
			if (*mp) {
				tpcb->tp_rsycnt--;
				m_freem(*mp);
			}
		if (tpcb->tp_rsycnt) {
			printf("tp_rsyflush %x\n", tpcb);
			tpcb->tp_rsycnt = 0;
		}
	}
	free((caddr_t)tpcb->tp_rsyq, M_PCB);
	tpcb->tp_rsyq = 0;
}

tp_rsyset(tpcb)
register struct tp_pcb *tpcb;
{
	register struct socket *so = tpcb->tp_sock;
	int maxcredit  = tpcb->tp_xtd_format ? 0xffff : 0xf;
	int old_credit = tpcb->tp_maxlcredit;
	caddr_t	rsyq;

	tpcb->tp_maxlcredit = maxcredit = min(maxcredit,
		  (so->so_rcv.sb_hiwat + tpcb->tp_l_tpdusize)/ tpcb->tp_l_tpdusize);

	if (old_credit == tpcb->tp_maxlcredit && tpcb->tp_rsyq != 0)
		return;
	maxcredit *= sizeof(struct mbuf *);
	if (tpcb->tp_rsyq)
		tp_rsyflush(tpcb);
	if (rsyq = (caddr_t)malloc(maxcredit, M_PCB, M_NOWAIT))
		bzero(rsyq, maxcredit);
	tpcb->tp_rsyq = (struct mbuf **)rsyq;
}

tpsbcheck(tpcb, i)
struct tp_pcb *tpcb;
{
	register struct mbuf *n, *m;
	register int len = 0, mbcnt = 0, pktlen;
	struct sockbuf *sb = &tpcb->tp_sock->so_snd;

	for (n = sb->sb_mb; n; n = n->m_nextpkt) {
		if ((n->m_flags & M_PKTHDR) == 0)
			panic("tpsbcheck nohdr");
		pktlen = len + n->m_pkthdr.len;
	    for (m = n; m; m = m->m_next) {
			len += m->m_len;
			mbcnt += MSIZE;
			if (m->m_flags & M_EXT)
				mbcnt += m->m_ext.ext_size;
		}
		if (len != pktlen) {
			printf("test %d; len %d != pktlen %d on mbuf 0x%x\n",
				i, len, pktlen, n);
			panic("tpsbcheck short");
		}
	}
	if (len != sb->sb_cc || mbcnt != sb->sb_mbcnt) {
		printf("test %d: cc %d != %d || mbcnt %d != %d\n", i, len, sb->sb_cc,
		    mbcnt, sb->sb_mbcnt);
		panic("tpsbcheck");
	}
}