tcp_send.c

minix3的源代码

				tmp_pack->acc_next=
					bf_cut(tcp_conn->tc_send_data, 
					(unsigned)(seg_lo_data-queue_lo_data), 
					(unsigned) (seg_hi_data-seg_lo_data));
			}
			seg_flags |= THF_ACK;
		}

after_data:
		if (!(seg_flags & THF_ACK))
		{
			if (pack2write)
				bf_afree(pack2write);
			return NULL;
		}

		tcp_hdr->th_seq_nr= htonl(seg_seq);
		tcp_hdr->th_ack_nr= htonl(tcp_conn->tc_RCV_NXT);
		tcp_hdr->th_flags= seg_flags;
		tcp_hdr->th_window= htons(tcp_conn->tc_RCV_HI -
			tcp_conn->tc_RCV_NXT);
		tcp_hdr->th_urgptr= htons(seg_up);

		pack_size= bf_bufsize(pack2write);
		ip_hdr->ih_length= htons(pack_size);

		pack2write->acc_linkC++;
		ip_hdr_len= (ip_hdr->ih_vers_ihl & IH_IHL_MASK) << 2;
		tcp_pack= bf_delhead(pack2write, ip_hdr_len);
		tcp_hdr->th_chksum= ~tcp_pack_oneCsum(ip_hdr, tcp_pack);
		bf_afree(tcp_pack);

		new_dis= curr_time + 2*HZ*tcp_conn->tc_ttl;
		if (new_dis > tcp_conn->tc_senddis)
			tcp_conn->tc_senddis= new_dis;

		return pack2write;
	default:
		DBLOCK(1, tcp_print_conn(tcp_conn); printf("\n"));
		ip_panic(( "Illegal state" ));
	}
	assert(0);
	return NULL;
}

/*
tcp_release_retrans
*/

PUBLIC void tcp_release_retrans(tcp_conn, seg_ack, new_win)
tcp_conn_t *tcp_conn;
u32_t seg_ack;
u16_t new_win;
{
	tcp_fd_t *tcp_fd;
	size_t size, offset;
	acc_t *pack;
	clock_t retrans_time, curr_time, rtt, artt, drtt, srtt;
	u32_t queue_lo, queue_hi;
	u16_t mss, cthresh;
	unsigned window;

	DBLOCK(0x10, printf("tcp_release_retrans, conn[%d]: ack %lu, win %u\n",
		tcp_conn-tcp_conn_table, (unsigned long)seg_ack, new_win););

	assert(tcp_conn->tc_busy);
	assert (tcp_GEmod4G(seg_ack, tcp_conn->tc_SND_UNA));
	assert (tcp_LEmod4G(seg_ack, tcp_conn->tc_SND_NXT));

	tcp_conn->tc_snd_dack= 0;
	mss= tcp_conn->tc_mtu-IP_TCP_MIN_HDR_SIZE;

	curr_time= get_time();
	if (tcp_conn->tc_rt_seq != 0 && 
		tcp_Gmod4G(seg_ack, tcp_conn->tc_rt_seq))
	{
		assert(curr_time >= tcp_conn->tc_rt_time);
		retrans_time= curr_time-tcp_conn->tc_rt_time;
		rtt= tcp_conn->tc_rtt;

		tcp_conn->tc_rt_seq= 0;

		if (rtt == TCP_RTT_GRAN*CLOCK_GRAN &&
			retrans_time <= TCP_RTT_GRAN*CLOCK_GRAN)
		{
			/* Common in fast networks. Nothing to do. */
		}
		else
		{
			srtt= retrans_time * TCP_RTT_SCALE;

			artt= tcp_conn->tc_artt;
			artt= ((TCP_RTT_SMOOTH-1)*artt+srtt)/TCP_RTT_SMOOTH;

			srtt -= artt;
			if (srtt < 0)
				srtt= -srtt;
			drtt= tcp_conn->tc_drtt;
			drtt= ((TCP_RTT_SMOOTH-1)*drtt+srtt)/TCP_RTT_SMOOTH;

			rtt= (artt+TCP_DRTT_MULT*drtt-1)/TCP_RTT_SCALE+1;
			if (rtt < TCP_RTT_GRAN*CLOCK_GRAN)
			{
				rtt= TCP_RTT_GRAN*CLOCK_GRAN;
			}
			else if (rtt > TCP_RTT_MAX)
			{
#if DEBUG
				static int warned /* = 0 */;

				if (!warned)
				{
					printf(
"tcp_release_retrans: warning retransmission time is limited to %d ms\n",
						TCP_RTT_MAX*1000/HZ);
					warned= 1;
				}
#endif
				rtt= TCP_RTT_MAX;
			}
			DBLOCK(0x10, printf(
	"tcp_release_retrans, conn[%d]: retrans_time= %ld ms, rtt = %ld ms\n",
				tcp_conn-tcp_conn_table,
				retrans_time*1000/HZ,
				rtt*1000/HZ));

			DBLOCK(0x10, printf(
	"tcp_release_retrans: artt= %ld -> %ld, drtt= %ld -> %ld\n",
				tcp_conn->tc_artt, artt,
				tcp_conn->tc_drtt, drtt));

			tcp_conn->tc_artt= artt;
			tcp_conn->tc_drtt= drtt;
			tcp_conn->tc_rtt= rtt;
		}

		if (tcp_conn->tc_mtu != tcp_conn->tc_max_mtu &&
			curr_time > tcp_conn->tc_mtutim+TCP_PMTU_INCR_IV)
		{
			tcp_mtu_incr(tcp_conn);
		}
	}

	/* Update the current window. */
	window= tcp_conn->tc_snd_cwnd-tcp_conn->tc_SND_UNA;
	assert(seg_ack != tcp_conn->tc_SND_UNA);

	/* For every real ACK we try to increase the current window
	 * with 1 mss.
	 */
	window += mss;

	/* If the window becomes larger than the current threshold,
	 * increment the threshold by a small amount and set the
	 * window to the threshold.
	 */
	cthresh= tcp_conn->tc_snd_cthresh;
	if (window > cthresh)
	{
		cthresh += tcp_conn->tc_snd_cinc;
		tcp_conn->tc_snd_cthresh= cthresh;
		window= cthresh;
	}

	/* If the window is larger than the window advertised by the
	 * receiver, set the window size to the advertisement.
	 */
	if (window > new_win)
		window= new_win;

	tcp_conn->tc_snd_cwnd= seg_ack+window;

	/* Release data queued for retransmissions. */
	queue_lo= tcp_conn->tc_SND_UNA;
	queue_hi= tcp_conn->tc_SND_NXT;

	tcp_conn->tc_SND_UNA= seg_ack;
	if (tcp_Lmod4G(tcp_conn->tc_SND_TRM, seg_ack))
	{
		tcp_conn->tc_SND_TRM= seg_ack;
	}
	assert(tcp_GEmod4G(tcp_conn->tc_snd_cwnd, seg_ack));

	/* Advance ISS every 0.5GB to avoid problem with wrap around */
	if (tcp_conn->tc_SND_UNA - tcp_conn->tc_ISS > 0x40000000)
	{
		tcp_conn->tc_ISS += 0x20000000;
		DBLOCK(1, printf(
			"tcp_release_retrans: updating ISS to 0x%lx\n",
			(unsigned long)tcp_conn->tc_ISS););
		if (tcp_Lmod4G(tcp_conn->tc_SND_UP, tcp_conn->tc_ISS))
		{
			tcp_conn->tc_SND_UP= tcp_conn->tc_ISS;
			DBLOCK(1, printf(
			"tcp_release_retrans: updating SND_UP to 0x%lx\n",
				(unsigned long)tcp_conn->tc_SND_UP););
		}
	}

	if (queue_lo == tcp_conn->tc_ISS)
		queue_lo++;

	if (tcp_conn->tc_flags & TCF_FIN_SENT)
	{
		if (seg_ack == queue_hi)
			seg_ack--;
		if (queue_lo == queue_hi)
			queue_lo--;
		queue_hi--;
	}

	offset= seg_ack - queue_lo;
	size= queue_hi - seg_ack;
	pack= tcp_conn->tc_send_data;
	tcp_conn->tc_send_data= 0;

	if (!size)
	{
		bf_afree(pack);
	}
	else
	{
		pack= bf_delhead(pack, offset);
		tcp_conn->tc_send_data= pack;
	}

	if (tcp_Gmod4G(tcp_conn->tc_SND_TRM, tcp_conn->tc_snd_cwnd))
		tcp_conn->tc_SND_TRM= tcp_conn->tc_snd_cwnd;

	/* Copy in new data if an ioctl is pending or if a write request is
	 * pending and either the write can be completed or at least one
	 * mss buffer space is available.
	 */
	tcp_fd= tcp_conn->tc_fd;
	if (tcp_fd)
	{
		if (tcp_fd->tf_flags & TFF_IOCTL_IP) 
		{
			tcp_fd_write(tcp_conn);
		}
		if ((tcp_fd->tf_flags & TFF_WRITE_IP) &&
			(size+tcp_fd->tf_write_count <= TCP_MAX_SND_WND_SIZE ||
			size <= TCP_MAX_SND_WND_SIZE-mss))
		{
			tcp_fd_write(tcp_conn);
		}
		if (tcp_fd->tf_flags & TFF_SEL_WRITE) 
			tcp_rsel_write(tcp_conn);
	}
	else
	{
		if (tcp_conn->tc_SND_UNA == tcp_conn->tc_SND_NXT)
		{
			assert(tcp_conn->tc_state == TCS_CLOSING);
			DBLOCK(0x10,
			printf("all data sent in abondoned connection\n"));
			tcp_close_connection(tcp_conn, ENOTCONN);
			return;
		}
	}

	if (!size && !tcp_conn->tc_send_data)
	{
		/* Reset window if a write is completed */
		tcp_conn->tc_snd_cwnd= tcp_conn->tc_SND_UNA + mss;
	}

	DIFBLOCK(2, (tcp_conn->tc_snd_cwnd == tcp_conn->tc_SND_TRM),
		printf("not sending: zero window\n"));

	if (tcp_conn->tc_snd_cwnd != tcp_conn->tc_SND_TRM &&
		tcp_conn->tc_SND_NXT != tcp_conn->tc_SND_TRM)
	{
		tcp_conn_write(tcp_conn, 1);
	}

}

/*
tcp_fast_retrans
*/

PUBLIC void tcp_fast_retrans(tcp_conn)
tcp_conn_t *tcp_conn;
{
	u16_t mss, mss2;

	/* Update threshold sequence number for retransmission calculation. */
	if (tcp_Gmod4G(tcp_conn->tc_SND_TRM, tcp_conn->tc_rt_threshold))
		tcp_conn->tc_rt_threshold= tcp_conn->tc_SND_TRM;

	tcp_conn->tc_SND_TRM= tcp_conn->tc_SND_UNA;

	mss= tcp_conn->tc_mtu-IP_TCP_MIN_HDR_SIZE;
	mss2= 2*mss;

	if (tcp_conn->tc_snd_cwnd == tcp_conn->tc_SND_UNA)
		tcp_conn->tc_snd_cwnd++;
	if (tcp_Gmod4G(tcp_conn->tc_snd_cwnd, tcp_conn->tc_SND_UNA + mss2))
	{
		tcp_conn->tc_snd_cwnd= tcp_conn->tc_SND_UNA + mss2;
		if (tcp_Gmod4G(tcp_conn->tc_SND_TRM, tcp_conn->tc_snd_cwnd))
			tcp_conn->tc_SND_TRM= tcp_conn->tc_snd_cwnd;

		tcp_conn->tc_snd_cthresh /= 2;
		if (tcp_conn->tc_snd_cthresh < mss2)
			tcp_conn->tc_snd_cthresh= mss2;
	}

	tcp_conn_write(tcp_conn, 1);
}

#if 0
PUBLIC void do_tcp_timeout(tcp_conn)
tcp_conn_t *tcp_conn;
{
	tcp_send_timeout(tcp_conn-tcp_conn_table,
		&tcp_conn->tc_transmit_timer);
}
#endif

/*
tcp_send_timeout
*/

PRIVATE void tcp_send_timeout(conn, timer)
int conn;
struct timer *timer;
{
	tcp_conn_t *tcp_conn;
	u16_t mss, mss2;
	u32_t snd_una, snd_nxt;
	clock_t curr_time, rtt, stt, timeout;
	acc_t *pkt;
	int new_ttl, no_push;

	DBLOCK(0x20, printf("tcp_send_timeout: conn[%d]\n", conn));

	curr_time= get_time();

	tcp_conn= &tcp_conn_table[conn];
	assert(tcp_conn->tc_flags & TCF_INUSE);
	assert(tcp_conn->tc_state != TCS_CLOSED);
	assert(tcp_conn->tc_state != TCS_LISTEN);

	snd_una= tcp_conn->tc_SND_UNA;
	snd_nxt= tcp_conn->tc_SND_NXT;
	no_push= (tcp_conn->tc_flags & TCF_NO_PUSH);
	if (snd_nxt == snd_una || no_push)
	{
		/* Nothing more to send */
		assert(tcp_conn->tc_SND_TRM == snd_una || no_push);

		/* A new write sets the timer if tc_transmit_seq == SND_UNA */
		tcp_conn->tc_transmit_seq= tcp_conn->tc_SND_UNA;
		tcp_conn->tc_stt= 0;
		tcp_conn->tc_0wnd_to= 0;
		assert(!tcp_conn->tc_fd ||
			!(tcp_conn->tc_fd->tf_flags & TFF_WRITE_IP) ||
			(tcp_print_conn(tcp_conn), printf("\n"), 0));

		if (snd_nxt != snd_una)
		{
			assert(no_push);
			DBLOCK(1, printf("not setting keepalive timer\n"););

			/* No point in setting the keepalive timer if we
			 * still have to send more data.
			 */
			return;
		}

		assert(tcp_conn->tc_send_data == NULL);
		DBLOCK(0x20, printf("keep alive timer\n"));
		if (tcp_conn->tc_ka_snd != tcp_conn->tc_SND_NXT ||
			tcp_conn->tc_ka_rcv != tcp_conn->tc_RCV_NXT)
		{
			tcp_conn->tc_ka_snd= tcp_conn->tc_SND_NXT;
			tcp_conn->tc_ka_rcv= tcp_conn->tc_RCV_NXT;
			DBLOCK(0x20, printf(
"tcp_send_timeout: conn[%d] setting keepalive timer (+%ld ms)\n",
				tcp_conn-tcp_conn_table,
				tcp_conn->tc_ka_time*1000/HZ));
			clck_timer(&tcp_conn->tc_transmit_timer,
				curr_time+tcp_conn->tc_ka_time,
				tcp_send_timeout,
				tcp_conn-tcp_conn_table);
			return;
		}
		DBLOCK(0x10, printf(
		"tcp_send_timeout, conn[%d]: triggering keep alive probe\n",
			tcp_conn-tcp_conn_table));
		tcp_conn->tc_ka_snd--;
		if (!(tcp_conn->tc_flags & TCF_FIN_SENT))
		{
			pkt= bf_memreq(1);
			*ptr2acc_data(pkt)= '\xff';	/* a random char */
			tcp_conn->tc_send_data= pkt; pkt= NULL;
		}
		tcp_conn->tc_SND_UNA--;
		if (tcp_conn->tc_SND_UNA == tcp_conn->tc_ISS)
		{
			/* We didn't send anything so far. Retrying the
			 * SYN is too hard. Decrement ISS and hope
			 * that the other side doesn't care.
			 */
			tcp_conn->tc_ISS--;
		}

		/* Set tc_transmit_seq and tc_stt to trigger packet */
		tcp_conn->tc_transmit_seq= tcp_conn->tc_SND_UNA;
		tcp_conn->tc_stt= curr_time;

		/* Set tc_rt_seq for round trip measurements */
		tcp_conn->tc_rt_time= curr_time;
		tcp_conn->tc_rt_seq= tcp_conn->tc_SND_UNA;

		/* Set PSH to make sure that data gets sent */
		tcp_conn->tc_SND_PSH= tcp_conn->tc_SND_NXT;
		assert(tcp_check_conn(tcp_conn));

		/* Fall through */
	}

	rtt= tcp_conn->tc_rtt;

	if (tcp_conn->tc_transmit_seq != tcp_conn->tc_SND_UNA)
	{
		/* Some data has been acknowledged since the last time the
		 * timer was set, set the timer again. */
		tcp_conn->tc_transmit_seq= tcp_conn->tc_SND_UNA;
		tcp_conn->tc_stt= 0;
		tcp_conn->tc_0wnd_to= 0;

		DBLOCK(0x20, printf(
	"tcp_send_timeout: conn[%d] setting timer to %ld ms (+%ld ms)\n",
			tcp_conn-tcp_conn_table,
			(curr_time+rtt)*1000/HZ, rtt*1000/HZ));

		clck_timer(&tcp_conn->tc_transmit_timer,
			curr_time+rtt, tcp_send_timeout,
			tcp_conn-tcp_conn_table);
		return;
	}

	stt= tcp_conn->tc_stt;
	if (stt == 0)
	{
		/* Some packet arrived but did not acknowledge any data.
		 * Apparently, the other side is still alive and has a
		 * reason to transmit. We can asume a zero window.
		 */

		DBLOCK(0x10, printf("conn[%d] setting zero window timer\n",
			tcp_conn-tcp_conn_table));

		if (tcp_conn->tc_0wnd_to < TCP_0WND_MIN)
			tcp_conn->tc_0wnd_to= TCP_0WND_MIN;
		else if (tcp_conn->tc_0wnd_to < rtt)
			tcp_conn->tc_0wnd_to= rtt;
		else
		{
			tcp_conn->tc_0wnd_to *= 2;
			if (tcp_conn->tc_0wnd_to > TCP_0WND_MAX)
				tcp_conn->tc_0wnd_to= TCP_0WND_MAX;
		}
		tcp_conn->tc_stt= curr_time;
		tcp_conn->tc_rt_seq= 0;

		DBLOCK(0x10, printf(
	"tcp_send_timeout: conn[%d] setting timer to %ld ms (+%ld ms)\n",
			tcp_conn-tcp_conn_table,
			(curr_time+tcp_conn->tc_0wnd_to)*1000/HZ,
			tcp_conn->tc_0wnd_to*1000/HZ));

		clck_timer(&tcp_conn->tc_transmit_timer,
			curr_time+tcp_conn->tc_0wnd_to,
			tcp_send_timeout, tcp_conn-tcp_conn_table);
		return;
	}
	assert(stt <= curr_time);