xref: /titanic_51/usr/src/uts/common/inet/tcp/tcp_fusion.c (revision dd51520e127b452179a2ce4ea3bd8dee949f9afe)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 #include <sys/types.h>
26 #include <sys/stream.h>
27 #include <sys/strsun.h>
28 #include <sys/strsubr.h>
29 #include <sys/debug.h>
30 #include <sys/sdt.h>
31 #include <sys/cmn_err.h>
32 #include <sys/tihdr.h>
33 
34 #include <inet/common.h>
35 #include <inet/optcom.h>
36 #include <inet/ip.h>
37 #include <inet/ip_if.h>
38 #include <inet/ip_impl.h>
39 #include <inet/tcp.h>
40 #include <inet/tcp_impl.h>
41 #include <inet/ipsec_impl.h>
42 #include <inet/ipclassifier.h>
43 #include <inet/ipp_common.h>
44 #include <inet/ip_if.h>
45 
46 /*
47  * This file implements TCP fusion - a protocol-less data path for TCP
48  * loopback connections.  The fusion of two local TCP endpoints occurs
49  * at connection establishment time.  Various conditions (see details
50  * in tcp_fuse()) need to be met for fusion to be successful.  If it
51  * fails, we fall back to the regular TCP data path; if it succeeds,
52  * both endpoints proceed to use tcp_fuse_output() as the transmit path.
53  * tcp_fuse_output() enqueues application data directly onto the peer's
54  * receive queue; no protocol processing is involved.
55  *
56  * Sychronization is handled by squeue and the mutex tcp_non_sq_lock.
57  * One of the requirements for fusion to succeed is that both endpoints
58  * need to be using the same squeue.  This ensures that neither side
59  * can disappear while the other side is still sending data. Flow
60  * control information is manipulated outside the squeue, so the
61  * tcp_non_sq_lock must be held when touching tcp_flow_stopped.
62  */
63 
64 /*
65  * Setting this to false means we disable fusion altogether and
66  * loopback connections would go through the protocol paths.
67  */
68 boolean_t do_tcp_fusion = B_TRUE;
69 
70 /*
71  * This routine gets called by the eager tcp upon changing state from
72  * SYN_RCVD to ESTABLISHED.  It fuses a direct path between itself
73  * and the active connect tcp such that the regular tcp processings
74  * may be bypassed under allowable circumstances.  Because the fusion
75  * requires both endpoints to be in the same squeue, it does not work
76  * for simultaneous active connects because there is no easy way to
77  * switch from one squeue to another once the connection is created.
78  * This is different from the eager tcp case where we assign it the
79  * same squeue as the one given to the active connect tcp during open.
80  */
81 void
82 tcp_fuse(tcp_t *tcp, uchar_t *iphdr, tcpha_t *tcpha)
83 {
84 	conn_t		*peer_connp, *connp = tcp->tcp_connp;
85 	tcp_t		*peer_tcp;
86 	tcp_stack_t	*tcps = tcp->tcp_tcps;
87 	netstack_t	*ns;
88 	ip_stack_t	*ipst = tcps->tcps_netstack->netstack_ip;
89 
90 	ASSERT(!tcp->tcp_fused);
91 	ASSERT(tcp->tcp_loopback);
92 	ASSERT(tcp->tcp_loopback_peer == NULL);
93 	/*
94 	 * We need to inherit conn_rcvbuf of the listener tcp,
95 	 * but we can't really use tcp_listener since we get here after
96 	 * sending up T_CONN_IND and tcp_tli_accept() may be called
97 	 * independently, at which point tcp_listener is cleared;
98 	 * this is why we use tcp_saved_listener. The listener itself
99 	 * is guaranteed to be around until tcp_accept_finish() is called
100 	 * on this eager -- this won't happen until we're done since we're
101 	 * inside the eager's perimeter now.
102 	 */
103 	ASSERT(tcp->tcp_saved_listener != NULL);
104 	/*
105 	 * Lookup peer endpoint; search for the remote endpoint having
106 	 * the reversed address-port quadruplet in ESTABLISHED state,
107 	 * which is guaranteed to be unique in the system.  Zone check
108 	 * is applied accordingly for loopback address, but not for
109 	 * local address since we want fusion to happen across Zones.
110 	 */
111 	if (connp->conn_ipversion == IPV4_VERSION) {
112 		peer_connp = ipcl_conn_tcp_lookup_reversed_ipv4(connp,
113 		    (ipha_t *)iphdr, tcpha, ipst);
114 	} else {
115 		peer_connp = ipcl_conn_tcp_lookup_reversed_ipv6(connp,
116 		    (ip6_t *)iphdr, tcpha, ipst);
117 	}
118 
119 	/*
120 	 * We can only proceed if peer exists, resides in the same squeue
121 	 * as our conn and is not raw-socket. We also restrict fusion to
122 	 * endpoints of the same type (STREAMS or non-STREAMS). The squeue
123 	 * assignment of this eager tcp was done earlier at the time of SYN
124 	 * processing in ip_fanout_tcp{_v6}.  Note that similar squeues by
125 	 * itself doesn't guarantee a safe condition to fuse, hence we perform
126 	 * additional tests below.
127 	 */
128 	ASSERT(peer_connp == NULL || peer_connp != connp);
129 	if (peer_connp == NULL || peer_connp->conn_sqp != connp->conn_sqp ||
130 	    !IPCL_IS_TCP(peer_connp) ||
131 	    IPCL_IS_NONSTR(connp) != IPCL_IS_NONSTR(peer_connp)) {
132 		if (peer_connp != NULL) {
133 			TCP_STAT(tcps, tcp_fusion_unqualified);
134 			CONN_DEC_REF(peer_connp);
135 		}
136 		return;
137 	}
138 	peer_tcp = peer_connp->conn_tcp;	/* active connect tcp */
139 
140 	ASSERT(peer_tcp != NULL && peer_tcp != tcp && !peer_tcp->tcp_fused);
141 	ASSERT(peer_tcp->tcp_loopback_peer == NULL);
142 	ASSERT(peer_connp->conn_sqp == connp->conn_sqp);
143 
144 	/*
145 	 * Due to IRE changes the peer and us might not agree on tcp_loopback.
146 	 * We bail in that case.
147 	 */
148 	if (!peer_tcp->tcp_loopback) {
149 		TCP_STAT(tcps, tcp_fusion_unqualified);
150 		CONN_DEC_REF(peer_connp);
151 		return;
152 	}
153 	/*
154 	 * Fuse the endpoints; we perform further checks against both
155 	 * tcp endpoints to ensure that a fusion is allowed to happen.
156 	 */
157 	ns = tcps->tcps_netstack;
158 	ipst = ns->netstack_ip;
159 
160 	if (!tcp->tcp_unfusable && !peer_tcp->tcp_unfusable &&
161 	    tcp->tcp_xmit_head == NULL && peer_tcp->tcp_xmit_head == NULL) {
162 		mblk_t *mp;
163 		queue_t *peer_rq = peer_connp->conn_rq;
164 
165 		ASSERT(!TCP_IS_DETACHED(peer_tcp));
166 		ASSERT(tcp->tcp_fused_sigurg_mp == NULL);
167 		ASSERT(peer_tcp->tcp_fused_sigurg_mp == NULL);
168 
169 		/*
170 		 * We need to drain data on both endpoints during unfuse.
171 		 * If we need to send up SIGURG at the time of draining,
172 		 * we want to be sure that an mblk is readily available.
173 		 * This is why we pre-allocate the M_PCSIG mblks for both
174 		 * endpoints which will only be used during/after unfuse.
175 		 * The mblk might already exist if we are doing a re-fuse.
176 		 */
177 		if (!IPCL_IS_NONSTR(tcp->tcp_connp)) {
178 			ASSERT(!IPCL_IS_NONSTR(peer_tcp->tcp_connp));
179 
180 			if (tcp->tcp_fused_sigurg_mp == NULL) {
181 				if ((mp = allocb(1, BPRI_HI)) == NULL)
182 					goto failed;
183 				tcp->tcp_fused_sigurg_mp = mp;
184 			}
185 
186 			if (peer_tcp->tcp_fused_sigurg_mp == NULL) {
187 				if ((mp = allocb(1, BPRI_HI)) == NULL)
188 					goto failed;
189 				peer_tcp->tcp_fused_sigurg_mp = mp;
190 			}
191 
192 			if ((mp = allocb(sizeof (struct stroptions),
193 			    BPRI_HI)) == NULL)
194 				goto failed;
195 		}
196 
197 		/* Fuse both endpoints */
198 		peer_tcp->tcp_loopback_peer = tcp;
199 		tcp->tcp_loopback_peer = peer_tcp;
200 		peer_tcp->tcp_fused = tcp->tcp_fused = B_TRUE;
201 
202 		/*
203 		 * We never use regular tcp paths in fusion and should
204 		 * therefore clear tcp_unsent on both endpoints.  Having
205 		 * them set to non-zero values means asking for trouble
206 		 * especially after unfuse, where we may end up sending
207 		 * through regular tcp paths which expect xmit_list and
208 		 * friends to be correctly setup.
209 		 */
210 		peer_tcp->tcp_unsent = tcp->tcp_unsent = 0;
211 
212 		tcp_timers_stop(tcp);
213 		tcp_timers_stop(peer_tcp);
214 
215 		/*
216 		 * Set receive buffer and max packet size for the
217 		 * active open tcp.
218 		 * eager's values will be set in tcp_accept_finish.
219 		 */
220 		(void) tcp_rwnd_set(peer_tcp, peer_tcp->tcp_connp->conn_rcvbuf);
221 
222 		/*
223 		 * Set the write offset value to zero since we won't
224 		 * be needing any room for TCP/IP headers.
225 		 */
226 		if (!IPCL_IS_NONSTR(peer_tcp->tcp_connp)) {
227 			struct stroptions *stropt;
228 
229 			DB_TYPE(mp) = M_SETOPTS;
230 			mp->b_wptr += sizeof (*stropt);
231 
232 			stropt = (struct stroptions *)mp->b_rptr;
233 			stropt->so_flags = SO_WROFF | SO_MAXBLK;
234 			stropt->so_wroff = 0;
235 			stropt->so_maxblk = INFPSZ;
236 
237 			/* Send the options up */
238 			putnext(peer_rq, mp);
239 		} else {
240 			struct sock_proto_props sopp;
241 
242 			/* The peer is a non-STREAMS end point */
243 			ASSERT(IPCL_IS_TCP(peer_connp));
244 
245 			sopp.sopp_flags = SOCKOPT_WROFF | SOCKOPT_MAXBLK;
246 			sopp.sopp_wroff = 0;
247 			sopp.sopp_maxblk = INFPSZ;
248 			(*peer_connp->conn_upcalls->su_set_proto_props)
249 			    (peer_connp->conn_upper_handle, &sopp);
250 		}
251 	} else {
252 		TCP_STAT(tcps, tcp_fusion_unqualified);
253 	}
254 	CONN_DEC_REF(peer_connp);
255 	return;
256 
257 failed:
258 	if (tcp->tcp_fused_sigurg_mp != NULL) {
259 		freeb(tcp->tcp_fused_sigurg_mp);
260 		tcp->tcp_fused_sigurg_mp = NULL;
261 	}
262 	if (peer_tcp->tcp_fused_sigurg_mp != NULL) {
263 		freeb(peer_tcp->tcp_fused_sigurg_mp);
264 		peer_tcp->tcp_fused_sigurg_mp = NULL;
265 	}
266 	CONN_DEC_REF(peer_connp);
267 }
268 
269 /*
270  * Unfuse a previously-fused pair of tcp loopback endpoints.
271  */
272 void
273 tcp_unfuse(tcp_t *tcp)
274 {
275 	tcp_t *peer_tcp = tcp->tcp_loopback_peer;
276 	tcp_stack_t *tcps = tcp->tcp_tcps;
277 
278 	ASSERT(tcp->tcp_fused && peer_tcp != NULL);
279 	ASSERT(peer_tcp->tcp_fused && peer_tcp->tcp_loopback_peer == tcp);
280 	ASSERT(tcp->tcp_connp->conn_sqp == peer_tcp->tcp_connp->conn_sqp);
281 	ASSERT(tcp->tcp_unsent == 0 && peer_tcp->tcp_unsent == 0);
282 
283 	/*
284 	 * Cancel any pending push timers.
285 	 */
286 	if (tcp->tcp_push_tid != 0) {
287 		(void) TCP_TIMER_CANCEL(tcp, tcp->tcp_push_tid);
288 		tcp->tcp_push_tid = 0;
289 	}
290 	if (peer_tcp->tcp_push_tid != 0) {
291 		(void) TCP_TIMER_CANCEL(peer_tcp, peer_tcp->tcp_push_tid);
292 		peer_tcp->tcp_push_tid = 0;
293 	}
294 
295 	/*
296 	 * Drain any pending data; Note that in case of a detached tcp, the
297 	 * draining will happen later after the tcp is unfused.  For non-
298 	 * urgent data, this can be handled by the regular tcp_rcv_drain().
299 	 * If we have urgent data sitting in the receive list, we will
300 	 * need to send up a SIGURG signal first before draining the data.
301 	 * All of these will be handled by the code in tcp_fuse_rcv_drain()
302 	 * when called from tcp_rcv_drain().
303 	 */
304 	if (!TCP_IS_DETACHED(tcp)) {
305 		(void) tcp_fuse_rcv_drain(tcp->tcp_connp->conn_rq, tcp,
306 		    &tcp->tcp_fused_sigurg_mp);
307 	}
308 	if (!TCP_IS_DETACHED(peer_tcp)) {
309 		(void) tcp_fuse_rcv_drain(peer_tcp->tcp_connp->conn_rq,
310 		    peer_tcp,  &peer_tcp->tcp_fused_sigurg_mp);
311 	}
312 
313 	/* Lift up any flow-control conditions */
314 	mutex_enter(&tcp->tcp_non_sq_lock);
315 	if (tcp->tcp_flow_stopped) {
316 		tcp_clrqfull(tcp);
317 		TCP_STAT(tcps, tcp_fusion_backenabled);
318 	}
319 	mutex_exit(&tcp->tcp_non_sq_lock);
320 
321 	mutex_enter(&peer_tcp->tcp_non_sq_lock);
322 	if (peer_tcp->tcp_flow_stopped) {
323 		tcp_clrqfull(peer_tcp);
324 		TCP_STAT(tcps, tcp_fusion_backenabled);
325 	}
326 	mutex_exit(&peer_tcp->tcp_non_sq_lock);
327 
328 	/*
329 	 * Update tha_seq and tha_ack in the header template
330 	 */
331 	tcp->tcp_tcpha->tha_seq = htonl(tcp->tcp_snxt);
332 	tcp->tcp_tcpha->tha_ack = htonl(tcp->tcp_rnxt);
333 	peer_tcp->tcp_tcpha->tha_seq = htonl(peer_tcp->tcp_snxt);
334 	peer_tcp->tcp_tcpha->tha_ack = htonl(peer_tcp->tcp_rnxt);
335 
336 	/* Unfuse the endpoints */
337 	peer_tcp->tcp_fused = tcp->tcp_fused = B_FALSE;
338 	peer_tcp->tcp_loopback_peer = tcp->tcp_loopback_peer = NULL;
339 }
340 
341 /*
342  * Fusion output routine used to handle urgent data sent by STREAMS based
343  * endpoints. This routine is called by tcp_fuse_output() for handling
344  * non-M_DATA mblks.
345  */
346 void
347 tcp_fuse_output_urg(tcp_t *tcp, mblk_t *mp)
348 {
349 	mblk_t *mp1;
350 	struct T_exdata_ind *tei;
351 	tcp_t *peer_tcp = tcp->tcp_loopback_peer;
352 	mblk_t *head, *prev_head = NULL;
353 	tcp_stack_t	*tcps = tcp->tcp_tcps;
354 
355 	ASSERT(tcp->tcp_fused);
356 	ASSERT(peer_tcp != NULL && peer_tcp->tcp_loopback_peer == tcp);
357 	ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
358 	ASSERT(DB_TYPE(mp) == M_PROTO || DB_TYPE(mp) == M_PCPROTO);
359 	ASSERT(mp->b_cont != NULL && DB_TYPE(mp->b_cont) == M_DATA);
360 	ASSERT(MBLKL(mp) >= sizeof (*tei) && MBLKL(mp->b_cont) > 0);
361 
362 	/*
363 	 * Urgent data arrives in the form of T_EXDATA_REQ from above.
364 	 * Each occurence denotes a new urgent pointer.  For each new
365 	 * urgent pointer we signal (SIGURG) the receiving app to indicate
366 	 * that it needs to go into urgent mode.  This is similar to the
367 	 * urgent data handling in the regular tcp.  We don't need to keep
368 	 * track of where the urgent pointer is, because each T_EXDATA_REQ
369 	 * "advances" the urgent pointer for us.
370 	 *
371 	 * The actual urgent data carried by T_EXDATA_REQ is then prepended
372 	 * by a T_EXDATA_IND before being enqueued behind any existing data
373 	 * destined for the receiving app.  There is only a single urgent
374 	 * pointer (out-of-band mark) for a given tcp.  If the new urgent
375 	 * data arrives before the receiving app reads some existing urgent
376 	 * data, the previous marker is lost.  This behavior is emulated
377 	 * accordingly below, by removing any existing T_EXDATA_IND messages
378 	 * and essentially converting old urgent data into non-urgent.
379 	 */
380 	ASSERT(tcp->tcp_valid_bits & TCP_URG_VALID);
381 	/* Let sender get out of urgent mode */
382 	tcp->tcp_valid_bits &= ~TCP_URG_VALID;
383 
384 	/*
385 	 * This flag indicates that a signal needs to be sent up.
386 	 * This flag will only get cleared once SIGURG is delivered and
387 	 * is not affected by the tcp_fused flag -- delivery will still
388 	 * happen even after an endpoint is unfused, to handle the case
389 	 * where the sending endpoint immediately closes/unfuses after
390 	 * sending urgent data and the accept is not yet finished.
391 	 */
392 	peer_tcp->tcp_fused_sigurg = B_TRUE;
393 
394 	/* Reuse T_EXDATA_REQ mblk for T_EXDATA_IND */
395 	DB_TYPE(mp) = M_PROTO;
396 	tei = (struct T_exdata_ind *)mp->b_rptr;
397 	tei->PRIM_type = T_EXDATA_IND;
398 	tei->MORE_flag = 0;
399 	mp->b_wptr = (uchar_t *)&tei[1];
400 
401 	TCP_STAT(tcps, tcp_fusion_urg);
402 	TCPS_BUMP_MIB(tcps, tcpOutUrg);
403 
404 	head = peer_tcp->tcp_rcv_list;
405 	while (head != NULL) {
406 		/*
407 		 * Remove existing T_EXDATA_IND, keep the data which follows
408 		 * it and relink our list.  Note that we don't modify the
409 		 * tcp_rcv_last_tail since it never points to T_EXDATA_IND.
410 		 */
411 		if (DB_TYPE(head) != M_DATA) {
412 			mp1 = head;
413 
414 			ASSERT(DB_TYPE(mp1->b_cont) == M_DATA);
415 			head = mp1->b_cont;
416 			mp1->b_cont = NULL;
417 			head->b_next = mp1->b_next;
418 			mp1->b_next = NULL;
419 			if (prev_head != NULL)
420 				prev_head->b_next = head;
421 			if (peer_tcp->tcp_rcv_list == mp1)
422 				peer_tcp->tcp_rcv_list = head;
423 			if (peer_tcp->tcp_rcv_last_head == mp1)
424 				peer_tcp->tcp_rcv_last_head = head;
425 			freeb(mp1);
426 		}
427 		prev_head = head;
428 		head = head->b_next;
429 	}
430 }
431 
432 /*
433  * Fusion output routine, called by tcp_output() and tcp_wput_proto().
434  * If we are modifying any member that can be changed outside the squeue,
435  * like tcp_flow_stopped, we need to take tcp_non_sq_lock.
436  */
437 boolean_t
438 tcp_fuse_output(tcp_t *tcp, mblk_t *mp, uint32_t send_size)
439 {
440 	conn_t		*connp = tcp->tcp_connp;
441 	tcp_t		*peer_tcp = tcp->tcp_loopback_peer;
442 	conn_t		*peer_connp = peer_tcp->tcp_connp;
443 	boolean_t	flow_stopped, peer_data_queued = B_FALSE;
444 	boolean_t	urgent = (DB_TYPE(mp) != M_DATA);
445 	boolean_t	push = B_TRUE;
446 	mblk_t		*mp1 = mp;
447 	uint_t		ip_hdr_len;
448 	uint32_t	recv_size = send_size;
449 	tcp_stack_t	*tcps = tcp->tcp_tcps;
450 	netstack_t	*ns = tcps->tcps_netstack;
451 	ip_stack_t	*ipst = ns->netstack_ip;
452 	ipsec_stack_t	*ipss = ns->netstack_ipsec;
453 	iaflags_t	ixaflags = connp->conn_ixa->ixa_flags;
454 	boolean_t	do_ipsec, hooks_out, hooks_in, ipobs_enabled;
455 
456 	ASSERT(tcp->tcp_fused);
457 	ASSERT(peer_tcp != NULL && peer_tcp->tcp_loopback_peer == tcp);
458 	ASSERT(connp->conn_sqp == peer_connp->conn_sqp);
459 	ASSERT(DB_TYPE(mp) == M_DATA || DB_TYPE(mp) == M_PROTO ||
460 	    DB_TYPE(mp) == M_PCPROTO);
461 
462 	if (send_size == 0) {
463 		freemsg(mp);
464 		return (B_TRUE);
465 	}
466 
467 	/*
468 	 * Handle urgent data; we either send up SIGURG to the peer now
469 	 * or do it later when we drain, in case the peer is detached
470 	 * or if we're short of memory for M_PCSIG mblk.
471 	 */
472 	if (urgent) {
473 		tcp_fuse_output_urg(tcp, mp);
474 
475 		mp1 = mp->b_cont;
476 	}
477 
478 	/*
479 	 * Check that we are still using an IRE_LOCAL or IRE_LOOPBACK before
480 	 * further processes.
481 	 */
482 	if (!ip_output_verify_local(connp->conn_ixa))
483 		goto unfuse;
484 
485 	/*
486 	 * Build IP and TCP header in case we have something that needs the
487 	 * headers. Those cases are:
488 	 * 1. IPsec
489 	 * 2. IPobs
490 	 * 3. FW_HOOKS
491 	 *
492 	 * If tcp_xmit_mp() fails to dupb() the message, unfuse the connection
493 	 * and back to regular path.
494 	 */
495 	if (ixaflags & IXAF_IS_IPV4) {
496 		do_ipsec = (ixaflags & IXAF_IPSEC_SECURE) ||
497 		    CONN_INBOUND_POLICY_PRESENT(peer_connp, ipss);
498 
499 		hooks_out = HOOKS4_INTERESTED_LOOPBACK_OUT(ipst);
500 		hooks_in = HOOKS4_INTERESTED_LOOPBACK_IN(ipst);
501 		ipobs_enabled = (ipst->ips_ip4_observe.he_interested != 0);
502 	} else {
503 		do_ipsec = (ixaflags & IXAF_IPSEC_SECURE) ||
504 		    CONN_INBOUND_POLICY_PRESENT_V6(peer_connp, ipss);
505 
506 		hooks_out = HOOKS6_INTERESTED_LOOPBACK_OUT(ipst);
507 		hooks_in = HOOKS6_INTERESTED_LOOPBACK_IN(ipst);
508 		ipobs_enabled = (ipst->ips_ip6_observe.he_interested != 0);
509 	}
510 
511 	/* We do logical 'or' for efficiency */
512 	if (ipobs_enabled | do_ipsec | hooks_in | hooks_out) {
513 		if ((mp1 = tcp_xmit_mp(tcp, mp1, tcp->tcp_mss, NULL, NULL,
514 		    tcp->tcp_snxt, B_TRUE, NULL, B_FALSE)) == NULL)
515 			/* If tcp_xmit_mp fails, use regular path */
516 			goto unfuse;
517 
518 		/*
519 		 * Leave all IP relevant processes to ip_output_process_local(),
520 		 * which handles IPsec, IPobs, and FW_HOOKS.
521 		 */
522 		mp1 = ip_output_process_local(mp1, connp->conn_ixa, hooks_out,
523 		    hooks_in, do_ipsec ? peer_connp : NULL);
524 
525 		/* If the message is dropped for any reason. */
526 		if (mp1 == NULL)
527 			goto unfuse;
528 
529 		/*
530 		 * Data length might have been changed by FW_HOOKS.
531 		 * We assume that the first mblk contains the TCP/IP headers.
532 		 */
533 		if (hooks_in || hooks_out) {
534 			tcpha_t *tcpha;
535 
536 			ip_hdr_len = (ixaflags & IXAF_IS_IPV4) ?
537 			    IPH_HDR_LENGTH((ipha_t *)mp1->b_rptr) :
538 			    ip_hdr_length_v6(mp1, (ip6_t *)mp1->b_rptr);
539 
540 			tcpha = (tcpha_t *)&mp1->b_rptr[ip_hdr_len];
541 			ASSERT((uchar_t *)tcpha + sizeof (tcpha_t) <=
542 			    mp1->b_wptr);
543 			recv_size += htonl(tcpha->tha_seq) - tcp->tcp_snxt;
544 
545 		}
546 
547 		/*
548 		 * The message duplicated by tcp_xmit_mp is freed.
549 		 * Note: the original message passed in remains unchanged.
550 		 */
551 		freemsg(mp1);
552 	}
553 
554 	/*
555 	 * Enqueue data into the peer's receive list; we may or may not
556 	 * drain the contents depending on the conditions below.
557 	 *
558 	 * For non-STREAMS sockets we normally queue data directly in the
559 	 * socket by calling the su_recv upcall. However, if the peer is
560 	 * detached we use tcp_rcv_enqueue() instead. Queued data will be
561 	 * drained when the accept completes (in tcp_accept_finish()).
562 	 */
563 	if (IPCL_IS_NONSTR(peer_connp) &&
564 	    !TCP_IS_DETACHED(peer_tcp)) {
565 		int error;
566 		int flags = 0;
567 
568 		if ((tcp->tcp_valid_bits & TCP_URG_VALID) &&
569 		    (tcp->tcp_urg == tcp->tcp_snxt)) {
570 			flags = MSG_OOB;
571 			(*peer_connp->conn_upcalls->su_signal_oob)
572 			    (peer_connp->conn_upper_handle, 0);
573 			tcp->tcp_valid_bits &= ~TCP_URG_VALID;
574 		}
575 		if ((*peer_connp->conn_upcalls->su_recv)(
576 		    peer_connp->conn_upper_handle, mp, recv_size,
577 		    flags, &error, &push) < 0) {
578 			ASSERT(error != EOPNOTSUPP);
579 			peer_data_queued = B_TRUE;
580 		}
581 	} else {
582 		if (IPCL_IS_NONSTR(peer_connp) &&
583 		    (tcp->tcp_valid_bits & TCP_URG_VALID) &&
584 		    (tcp->tcp_urg == tcp->tcp_snxt)) {
585 			/*
586 			 * Can not deal with urgent pointers
587 			 * that arrive before the connection has been
588 			 * accept()ed.
589 			 */
590 			tcp->tcp_valid_bits &= ~TCP_URG_VALID;
591 			freemsg(mp);
592 			return (B_TRUE);
593 		}
594 
595 		tcp_rcv_enqueue(peer_tcp, mp, recv_size,
596 		    tcp->tcp_connp->conn_cred);
597 
598 		/* In case it wrapped around and also to keep it constant */
599 		peer_tcp->tcp_rwnd += recv_size;
600 	}
601 
602 	/*
603 	 * Exercise flow-control when needed; we will get back-enabled
604 	 * in either tcp_accept_finish(), tcp_unfuse(), or when data is
605 	 * consumed. If peer endpoint is detached, we emulate streams flow
606 	 * control by checking the peer's queue size and high water mark;
607 	 * otherwise we simply use canputnext() to decide if we need to stop
608 	 * our flow.
609 	 *
610 	 * Since we are accessing our tcp_flow_stopped and might modify it,
611 	 * we need to take tcp->tcp_non_sq_lock.
612 	 */
613 	mutex_enter(&tcp->tcp_non_sq_lock);
614 	flow_stopped = tcp->tcp_flow_stopped;
615 	if ((TCP_IS_DETACHED(peer_tcp) &&
616 	    (peer_tcp->tcp_rcv_cnt >= peer_connp->conn_rcvbuf)) ||
617 	    (!TCP_IS_DETACHED(peer_tcp) &&
618 	    !IPCL_IS_NONSTR(peer_connp) && !canputnext(peer_connp->conn_rq))) {
619 		peer_data_queued = B_TRUE;
620 	}
621 
622 	if (!flow_stopped && (peer_data_queued ||
623 	    (TCP_UNSENT_BYTES(tcp) >= connp->conn_sndbuf))) {
624 		tcp_setqfull(tcp);
625 		flow_stopped = B_TRUE;
626 		TCP_STAT(tcps, tcp_fusion_flowctl);
627 		DTRACE_PROBE3(tcp__fuse__output__flowctl, tcp_t *, tcp,
628 		    uint_t, send_size, uint_t, peer_tcp->tcp_rcv_cnt);
629 	} else if (flow_stopped && !peer_data_queued &&
630 	    (TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat)) {
631 		tcp_clrqfull(tcp);
632 		TCP_STAT(tcps, tcp_fusion_backenabled);
633 		flow_stopped = B_FALSE;
634 	}
635 	mutex_exit(&tcp->tcp_non_sq_lock);
636 
637 	ipst->ips_loopback_packets++;
638 	tcp->tcp_last_sent_len = send_size;
639 
640 	/* Need to adjust the following SNMP MIB-related variables */
641 	tcp->tcp_snxt += send_size;
642 	tcp->tcp_suna = tcp->tcp_snxt;
643 	peer_tcp->tcp_rnxt += recv_size;
644 	peer_tcp->tcp_last_recv_len = recv_size;
645 	peer_tcp->tcp_rack = peer_tcp->tcp_rnxt;
646 
647 	TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
648 	TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, send_size);
649 
650 	TCPS_BUMP_MIB(tcps, tcpHCInSegs);
651 	TCPS_BUMP_MIB(tcps, tcpInDataInorderSegs);
652 	TCPS_UPDATE_MIB(tcps, tcpInDataInorderBytes, send_size);
653 
654 	BUMP_LOCAL(tcp->tcp_obsegs);
655 	BUMP_LOCAL(peer_tcp->tcp_ibsegs);
656 
657 	DTRACE_TCP5(send, void, NULL, ip_xmit_attr_t *, connp->conn_ixa,
658 	    __dtrace_tcp_void_ip_t *, NULL, tcp_t *, tcp,
659 	    __dtrace_tcp_tcph_t *, NULL);
660 	DTRACE_TCP5(receive, void, NULL, ip_xmit_attr_t *,
661 	    peer_connp->conn_ixa, __dtrace_tcp_void_ip_t *, NULL,
662 	    tcp_t *, peer_tcp, __dtrace_tcp_tcph_t *, NULL);
663 
664 	if (!IPCL_IS_NONSTR(peer_tcp->tcp_connp) &&
665 	    !TCP_IS_DETACHED(peer_tcp)) {
666 		/*
667 		 * Drain the peer's receive queue it has urgent data or if
668 		 * we're not flow-controlled.
669 		 */
670 		if (urgent || !flow_stopped) {
671 			ASSERT(peer_tcp->tcp_rcv_list != NULL);
672 			/*
673 			 * For TLI-based streams, a thread in tcp_accept_swap()
674 			 * can race with us.  That thread will ensure that the
675 			 * correct peer_connp->conn_rq is globally visible
676 			 * before peer_tcp->tcp_detached is visible as clear,
677 			 * but we must also ensure that the load of conn_rq
678 			 * cannot be reordered to be before the tcp_detached
679 			 * check.
680 			 */
681 			membar_consumer();
682 			(void) tcp_fuse_rcv_drain(peer_connp->conn_rq, peer_tcp,
683 			    NULL);
684 		}
685 	}
686 	return (B_TRUE);
687 unfuse:
688 	tcp_unfuse(tcp);
689 	return (B_FALSE);
690 }
691 
692 /*
693  * This routine gets called to deliver data upstream on a fused or
694  * previously fused tcp loopback endpoint; the latter happens only
695  * when there is a pending SIGURG signal plus urgent data that can't
696  * be sent upstream in the past.
697  */
698 boolean_t
699 tcp_fuse_rcv_drain(queue_t *q, tcp_t *tcp, mblk_t **sigurg_mpp)
700 {
701 	mblk_t *mp;
702 	conn_t	*connp = tcp->tcp_connp;
703 
704 #ifdef DEBUG
705 	uint_t cnt = 0;
706 #endif
707 	tcp_stack_t	*tcps = tcp->tcp_tcps;
708 	tcp_t		*peer_tcp = tcp->tcp_loopback_peer;
709 
710 	ASSERT(tcp->tcp_loopback);
711 	ASSERT(tcp->tcp_fused || tcp->tcp_fused_sigurg);
712 	ASSERT(!tcp->tcp_fused || tcp->tcp_loopback_peer != NULL);
713 	ASSERT(IPCL_IS_NONSTR(connp) || sigurg_mpp != NULL || tcp->tcp_fused);
714 
715 	/* No need for the push timer now, in case it was scheduled */
716 	if (tcp->tcp_push_tid != 0) {
717 		(void) TCP_TIMER_CANCEL(tcp, tcp->tcp_push_tid);
718 		tcp->tcp_push_tid = 0;
719 	}
720 	/*
721 	 * If there's urgent data sitting in receive list and we didn't
722 	 * get a chance to send up a SIGURG signal, make sure we send
723 	 * it first before draining in order to ensure that SIOCATMARK
724 	 * works properly.
725 	 */
726 	if (tcp->tcp_fused_sigurg) {
727 		ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
728 
729 		tcp->tcp_fused_sigurg = B_FALSE;
730 		/*
731 		 * sigurg_mpp is normally NULL, i.e. when we're still
732 		 * fused and didn't get here because of tcp_unfuse().
733 		 * In this case try hard to allocate the M_PCSIG mblk.
734 		 */
735 		if (sigurg_mpp == NULL &&
736 		    (mp = allocb(1, BPRI_HI)) == NULL &&
737 		    (mp = allocb_tryhard(1)) == NULL) {
738 			/* Alloc failed; try again next time */
739 			tcp->tcp_push_tid = TCP_TIMER(tcp,
740 			    tcp_push_timer, tcps->tcps_push_timer_interval);
741 			return (B_TRUE);
742 		} else if (sigurg_mpp != NULL) {
743 			/*
744 			 * Use the supplied M_PCSIG mblk; it means we're
745 			 * either unfused or in the process of unfusing,
746 			 * and the drain must happen now.
747 			 */
748 			mp = *sigurg_mpp;
749 			*sigurg_mpp = NULL;
750 		}
751 		ASSERT(mp != NULL);
752 
753 		/* Send up the signal */
754 		DB_TYPE(mp) = M_PCSIG;
755 		*mp->b_wptr++ = (uchar_t)SIGURG;
756 		putnext(q, mp);
757 
758 		/*
759 		 * Let the regular tcp_rcv_drain() path handle
760 		 * draining the data if we're no longer fused.
761 		 */
762 		if (!tcp->tcp_fused)
763 			return (B_FALSE);
764 	}
765 
766 	/* Drain the data */
767 	while ((mp = tcp->tcp_rcv_list) != NULL) {
768 		tcp->tcp_rcv_list = mp->b_next;
769 		mp->b_next = NULL;
770 #ifdef DEBUG
771 		cnt += msgdsize(mp);
772 #endif
773 		ASSERT(!IPCL_IS_NONSTR(connp));
774 		putnext(q, mp);
775 		TCP_STAT(tcps, tcp_fusion_putnext);
776 	}
777 
778 #ifdef DEBUG
779 	ASSERT(cnt == tcp->tcp_rcv_cnt);
780 #endif
781 	tcp->tcp_rcv_last_head = NULL;
782 	tcp->tcp_rcv_last_tail = NULL;
783 	tcp->tcp_rcv_cnt = 0;
784 	tcp->tcp_rwnd = tcp->tcp_connp->conn_rcvbuf;
785 
786 	mutex_enter(&peer_tcp->tcp_non_sq_lock);
787 	if (peer_tcp->tcp_flow_stopped && (TCP_UNSENT_BYTES(peer_tcp) <=
788 	    peer_tcp->tcp_connp->conn_sndlowat)) {
789 		tcp_clrqfull(peer_tcp);
790 		TCP_STAT(tcps, tcp_fusion_backenabled);
791 	}
792 	mutex_exit(&peer_tcp->tcp_non_sq_lock);
793 
794 	return (B_TRUE);
795 }
796 
797 /*
798  * Calculate the size of receive buffer for a fused tcp endpoint.
799  */
800 size_t
801 tcp_fuse_set_rcv_hiwat(tcp_t *tcp, size_t rwnd)
802 {
803 	tcp_stack_t	*tcps = tcp->tcp_tcps;
804 	uint32_t	max_win;
805 
806 	ASSERT(tcp->tcp_fused);
807 
808 	/* Ensure that value is within the maximum upper bound */
809 	if (rwnd > tcps->tcps_max_buf)
810 		rwnd = tcps->tcps_max_buf;
811 	/*
812 	 * Round up to system page size in case SO_RCVBUF is modified
813 	 * after SO_SNDBUF; the latter is also similarly rounded up.
814 	 */
815 	rwnd = P2ROUNDUP_TYPED(rwnd, PAGESIZE, size_t);
816 	max_win = TCP_MAXWIN << tcp->tcp_rcv_ws;
817 	if (rwnd > max_win) {
818 		rwnd = max_win - (max_win % tcp->tcp_mss);
819 		if (rwnd < tcp->tcp_mss)
820 			rwnd = max_win;
821 	}
822 
823 	/*
824 	 * Record high water mark, this is used for flow-control
825 	 * purposes in tcp_fuse_output().
826 	 */
827 	tcp->tcp_connp->conn_rcvbuf = rwnd;
828 	tcp->tcp_rwnd = rwnd;
829 	return (rwnd);
830 }
831 
832 /*
833  * Calculate the maximum outstanding unread data block for a fused tcp endpoint.
834  */
835 int
836 tcp_fuse_maxpsz(tcp_t *tcp)
837 {
838 	tcp_t *peer_tcp = tcp->tcp_loopback_peer;
839 	conn_t *connp = tcp->tcp_connp;
840 	uint_t sndbuf = connp->conn_sndbuf;
841 	uint_t maxpsz = sndbuf;
842 
843 	ASSERT(tcp->tcp_fused);
844 	ASSERT(peer_tcp != NULL);
845 	ASSERT(peer_tcp->tcp_connp->conn_rcvbuf != 0);
846 	/*
847 	 * In the fused loopback case, we want the stream head to split
848 	 * up larger writes into smaller chunks for a more accurate flow-
849 	 * control accounting.  Our maxpsz is half of the sender's send
850 	 * buffer or the receiver's receive buffer, whichever is smaller.
851 	 * We round up the buffer to system page size due to the lack of
852 	 * TCP MSS concept in Fusion.
853 	 */
854 	if (maxpsz > peer_tcp->tcp_connp->conn_rcvbuf)
855 		maxpsz = peer_tcp->tcp_connp->conn_rcvbuf;
856 	maxpsz = P2ROUNDUP_TYPED(maxpsz, PAGESIZE, uint_t) >> 1;
857 
858 	return (maxpsz);
859 }
860 
861 /*
862  * Called to release flow control.
863  */
864 void
865 tcp_fuse_backenable(tcp_t *tcp)
866 {
867 	tcp_t *peer_tcp = tcp->tcp_loopback_peer;
868 
869 	ASSERT(tcp->tcp_fused);
870 	ASSERT(peer_tcp != NULL && peer_tcp->tcp_fused);
871 	ASSERT(peer_tcp->tcp_loopback_peer == tcp);
872 	ASSERT(!TCP_IS_DETACHED(tcp));
873 	ASSERT(tcp->tcp_connp->conn_sqp ==
874 	    peer_tcp->tcp_connp->conn_sqp);
875 
876 	if (tcp->tcp_rcv_list != NULL)
877 		(void) tcp_fuse_rcv_drain(tcp->tcp_connp->conn_rq, tcp, NULL);
878 
879 	mutex_enter(&peer_tcp->tcp_non_sq_lock);
880 	if (peer_tcp->tcp_flow_stopped &&
881 	    (TCP_UNSENT_BYTES(peer_tcp) <=
882 	    peer_tcp->tcp_connp->conn_sndlowat)) {
883 		tcp_clrqfull(peer_tcp);
884 	}
885 	mutex_exit(&peer_tcp->tcp_non_sq_lock);
886 
887 	TCP_STAT(tcp->tcp_tcps, tcp_fusion_backenabled);
888 }
889