xref: /illumos-gate/usr/src/uts/common/io/ib/clients/rdsv3/send.c (revision 02b4e56ca3a4e4a4fe9e52fca9c2972101f0e57f)
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) 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*
26  * Copyright (c) 2006 Oracle.  All rights reserved.
27  *
28  * This software is available to you under a choice of one of two
29  * licenses.  You may choose to be licensed under the terms of the GNU
30  * General Public License (GPL) Version 2, available from the file
31  * COPYING in the main directory of this source tree, or the
32  * OpenIB.org BSD license below:
33  *
34  *     Redistribution and use in source and binary forms, with or
35  *     without modification, are permitted provided that the following
36  *     conditions are met:
37  *
38  *      - Redistributions of source code must retain the above
39  *        copyright notice, this list of conditions and the following
40  *        disclaimer.
41  *
42  *      - Redistributions in binary form must reproduce the above
43  *        copyright notice, this list of conditions and the following
44  *        disclaimer in the documentation and/or other materials
45  *        provided with the distribution.
46  *
47  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
48  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
49  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
50  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
51  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
52  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
53  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
54  * SOFTWARE.
55  *
56  */
57 #include <sys/stropts.h>
58 #include <sys/systm.h>
59 
60 #include <sys/rds.h>
61 #include <sys/socket.h>
62 #include <sys/socketvar.h>
63 
64 #include <sys/ib/clients/rdsv3/rdsv3.h>
65 #include <sys/ib/clients/rdsv3/rdma.h>
66 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
67 
68 /*
69  * When transmitting messages in rdsv3_send_xmit, we need to emerge from
70  * time to time and briefly release the CPU. Otherwise the softlock watchdog
71  * will kick our shin.
72  * Also, it seems fairer to not let one busy connection stall all the
73  * others.
74  *
75  * send_batch_count is the number of times we'll loop in send_xmit. Setting
76  * it to 0 will restore the old behavior (where we looped until we had
77  * drained the queue).
78  */
79 static int send_batch_count = 64;
80 
81 extern void rdsv3_ib_send_unmap_rdma(void *ic, struct rdsv3_rdma_op *op);
82 /*
83  * Reset the send state. Caller must hold c_send_lock when calling here.
84  */
85 void
86 rdsv3_send_reset(struct rdsv3_connection *conn)
87 {
88 	struct rdsv3_message *rm, *tmp;
89 	struct rdsv3_rdma_op *ro;
90 
91 	RDSV3_DPRINTF4("rdsv3_send_reset", "Enter(conn: %p)", conn);
92 
93 	if (conn->c_xmit_rm) {
94 		rm = conn->c_xmit_rm;
95 		ro = rm->m_rdma_op;
96 		if (ro && ro->r_mapped) {
97 			RDSV3_DPRINTF2("rdsv3_send_reset",
98 			    "rm %p mflg 0x%x map %d mihdl %p sgl %p",
99 			    rm, rm->m_flags, ro->r_mapped,
100 			    ro->r_rdma_sg[0].mihdl,
101 			    ro->r_rdma_sg[0].swr.wr_sgl);
102 			rdsv3_ib_send_unmap_rdma(conn->c_transport_data, ro);
103 		}
104 		/*
105 		 * Tell the user the RDMA op is no longer mapped by the
106 		 * transport. This isn't entirely true (it's flushed out
107 		 * independently) but as the connection is down, there's
108 		 * no ongoing RDMA to/from that memory
109 		 */
110 		rdsv3_message_unmapped(conn->c_xmit_rm);
111 		rdsv3_message_put(conn->c_xmit_rm);
112 		conn->c_xmit_rm = NULL;
113 	}
114 	conn->c_xmit_sg = 0;
115 	conn->c_xmit_hdr_off = 0;
116 	conn->c_xmit_data_off = 0;
117 	conn->c_xmit_rdma_sent = 0;
118 
119 	conn->c_map_queued = 0;
120 
121 	conn->c_unacked_packets = rdsv3_sysctl_max_unacked_packets;
122 	conn->c_unacked_bytes = rdsv3_sysctl_max_unacked_bytes;
123 
124 	/* Mark messages as retransmissions, and move them to the send q */
125 	mutex_enter(&conn->c_lock);
126 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_retrans, m_conn_item) {
127 		set_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags);
128 		set_bit(RDSV3_MSG_RETRANSMITTED, &rm->m_flags);
129 		if (rm->m_rdma_op && rm->m_rdma_op->r_mapped) {
130 			RDSV3_DPRINTF4("_send_reset",
131 			    "RT rm %p mflg 0x%x sgl %p",
132 			    rm, rm->m_flags,
133 			    rm->m_rdma_op->r_rdma_sg[0].swr.wr_sgl);
134 		}
135 	}
136 	list_move_tail(&conn->c_send_queue, &conn->c_retrans);
137 	mutex_exit(&conn->c_lock);
138 
139 	RDSV3_DPRINTF4("rdsv3_send_reset", "Return(conn: %p)", conn);
140 }
141 
142 /*
143  * We're making the concious trade-off here to only send one message
144  * down the connection at a time.
145  *   Pro:
146  *      - tx queueing is a simple fifo list
147  *   	- reassembly is optional and easily done by transports per conn
148  *      - no per flow rx lookup at all, straight to the socket
149  *   	- less per-frag memory and wire overhead
150  *   Con:
151  *      - queued acks can be delayed behind large messages
152  *   Depends:
153  *      - small message latency is higher behind queued large messages
154  *      - large message latency isn't starved by intervening small sends
155  */
156 int
157 rdsv3_send_xmit(struct rdsv3_connection *conn)
158 {
159 	struct rdsv3_message *rm;
160 	unsigned int tmp;
161 	unsigned int send_quota = send_batch_count;
162 	struct rdsv3_scatterlist *sg;
163 	int ret = 0;
164 	int was_empty = 0;
165 	list_t to_be_dropped;
166 
167 	RDSV3_DPRINTF4("rdsv3_send_xmit", "Enter(conn: %p)", conn);
168 
169 	list_create(&to_be_dropped, sizeof (struct rdsv3_message),
170 	    offsetof(struct rdsv3_message, m_conn_item));
171 
172 	/*
173 	 * sendmsg calls here after having queued its message on the send
174 	 * queue.  We only have one task feeding the connection at a time.  If
175 	 * another thread is already feeding the queue then we back off.  This
176 	 * avoids blocking the caller and trading per-connection data between
177 	 * caches per message.
178 	 *
179 	 * The sem holder will issue a retry if they notice that someone queued
180 	 * a message after they stopped walking the send queue but before they
181 	 * dropped the sem.
182 	 */
183 	if (!mutex_tryenter(&conn->c_send_lock)) {
184 		RDSV3_DPRINTF4("rdsv3_send_xmit",
185 		    "Another thread running(conn: %p)", conn);
186 		rdsv3_stats_inc(s_send_sem_contention);
187 		ret = -ENOMEM;
188 		goto out;
189 	}
190 
191 	if (conn->c_trans->xmit_prepare)
192 		conn->c_trans->xmit_prepare(conn);
193 
194 	/*
195 	 * spin trying to push headers and data down the connection until
196 	 * the connection doens't make forward progress.
197 	 */
198 	while (--send_quota) {
199 		/*
200 		 * See if need to send a congestion map update if we're
201 		 * between sending messages.  The send_sem protects our sole
202 		 * use of c_map_offset and _bytes.
203 		 * Note this is used only by transports that define a special
204 		 * xmit_cong_map function. For all others, we create allocate
205 		 * a cong_map message and treat it just like any other send.
206 		 */
207 		if (conn->c_map_bytes) {
208 			ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
209 			    conn->c_map_offset);
210 			if (ret <= 0)
211 				break;
212 
213 			conn->c_map_offset += ret;
214 			conn->c_map_bytes -= ret;
215 			if (conn->c_map_bytes)
216 				continue;
217 		}
218 
219 		/*
220 		 * If we're done sending the current message, clear the
221 		 * offset and S/G temporaries.
222 		 */
223 		rm = conn->c_xmit_rm;
224 		if (rm != NULL &&
225 		    conn->c_xmit_hdr_off == sizeof (struct rdsv3_header) &&
226 		    conn->c_xmit_sg == rm->m_nents) {
227 			conn->c_xmit_rm = NULL;
228 			conn->c_xmit_sg = 0;
229 			conn->c_xmit_hdr_off = 0;
230 			conn->c_xmit_data_off = 0;
231 			conn->c_xmit_rdma_sent = 0;
232 
233 			/* Release the reference to the previous message. */
234 			rdsv3_message_put(rm);
235 			rm = NULL;
236 		}
237 
238 		/* If we're asked to send a cong map update, do so. */
239 		if (rm == NULL && test_and_clear_bit(0, &conn->c_map_queued)) {
240 			if (conn->c_trans->xmit_cong_map != NULL) {
241 				conn->c_map_offset = 0;
242 				conn->c_map_bytes =
243 				    sizeof (struct rdsv3_header) +
244 				    RDSV3_CONG_MAP_BYTES;
245 				continue;
246 			}
247 
248 			rm = rdsv3_cong_update_alloc(conn);
249 			if (IS_ERR(rm)) {
250 				ret = PTR_ERR(rm);
251 				break;
252 			}
253 
254 			conn->c_xmit_rm = rm;
255 		}
256 
257 		/*
258 		 * Grab the next message from the send queue, if there is one.
259 		 *
260 		 * c_xmit_rm holds a ref while we're sending this message down
261 		 * the connction.  We can use this ref while holding the
262 		 * send_sem.. rdsv3_send_reset() is serialized with it.
263 		 */
264 		if (rm == NULL) {
265 			unsigned int len;
266 
267 			mutex_enter(&conn->c_lock);
268 
269 			if (!list_is_empty(&conn->c_send_queue)) {
270 				rm = list_remove_head(&conn->c_send_queue);
271 				rdsv3_message_addref(rm);
272 
273 				/*
274 				 * Move the message from the send queue to
275 				 * the retransmit
276 				 * list right away.
277 				 */
278 				list_insert_tail(&conn->c_retrans, rm);
279 			}
280 
281 			mutex_exit(&conn->c_lock);
282 
283 			if (rm == NULL) {
284 				was_empty = 1;
285 				break;
286 			}
287 
288 			/*
289 			 * Unfortunately, the way Infiniband deals with
290 			 * RDMA to a bad MR key is by moving the entire
291 			 * queue pair to error state. We cold possibly
292 			 * recover from that, but right now we drop the
293 			 * connection.
294 			 * Therefore, we never retransmit messages with
295 			 * RDMA ops.
296 			 */
297 			if (rm->m_rdma_op &&
298 			    test_bit(RDSV3_MSG_RETRANSMITTED, &rm->m_flags)) {
299 				mutex_enter(&conn->c_lock);
300 				if (test_and_clear_bit(RDSV3_MSG_ON_CONN,
301 				    &rm->m_flags))
302 					list_remove_node(&rm->m_conn_item);
303 					list_insert_tail(&to_be_dropped, rm);
304 				mutex_exit(&conn->c_lock);
305 				rdsv3_message_put(rm);
306 				continue;
307 			}
308 
309 			/* Require an ACK every once in a while */
310 			len = ntohl(rm->m_inc.i_hdr.h_len);
311 			if (conn->c_unacked_packets == 0 ||
312 			    conn->c_unacked_bytes < len) {
313 				set_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags);
314 
315 				conn->c_unacked_packets =
316 				    rdsv3_sysctl_max_unacked_packets;
317 				conn->c_unacked_bytes =
318 				    rdsv3_sysctl_max_unacked_bytes;
319 				rdsv3_stats_inc(s_send_ack_required);
320 			} else {
321 				conn->c_unacked_bytes -= len;
322 				conn->c_unacked_packets--;
323 			}
324 
325 			conn->c_xmit_rm = rm;
326 		}
327 
328 		/*
329 		 * Try and send an rdma message.  Let's see if we can
330 		 * keep this simple and require that the transport either
331 		 * send the whole rdma or none of it.
332 		 */
333 		if (rm->m_rdma_op && !conn->c_xmit_rdma_sent) {
334 			ret = conn->c_trans->xmit_rdma(conn, rm->m_rdma_op);
335 			if (ret)
336 				break;
337 			conn->c_xmit_rdma_sent = 1;
338 			/*
339 			 * The transport owns the mapped memory for now.
340 			 * You can't unmap it while it's on the send queue
341 			 */
342 			set_bit(RDSV3_MSG_MAPPED, &rm->m_flags);
343 		}
344 
345 		if (conn->c_xmit_hdr_off < sizeof (struct rdsv3_header) ||
346 		    conn->c_xmit_sg < rm->m_nents) {
347 			ret = conn->c_trans->xmit(conn, rm,
348 			    conn->c_xmit_hdr_off,
349 			    conn->c_xmit_sg,
350 			    conn->c_xmit_data_off);
351 			if (ret <= 0)
352 				break;
353 
354 			if (conn->c_xmit_hdr_off <
355 			    sizeof (struct rdsv3_header)) {
356 				tmp = min(ret,
357 				    sizeof (struct rdsv3_header) -
358 				    conn->c_xmit_hdr_off);
359 				conn->c_xmit_hdr_off += tmp;
360 				ret -= tmp;
361 			}
362 
363 			sg = &rm->m_sg[conn->c_xmit_sg];
364 			while (ret) {
365 				tmp = min(ret, rdsv3_sg_len(sg) -
366 				    conn->c_xmit_data_off);
367 				conn->c_xmit_data_off += tmp;
368 				ret -= tmp;
369 				if (conn->c_xmit_data_off == rdsv3_sg_len(sg)) {
370 					conn->c_xmit_data_off = 0;
371 					sg++;
372 					conn->c_xmit_sg++;
373 					ASSERT(!(ret != 0 &&
374 					    conn->c_xmit_sg == rm->m_nents));
375 				}
376 			}
377 		}
378 	}
379 
380 	/* Nuke any messages we decided not to retransmit. */
381 	if (!list_is_empty(&to_be_dropped))
382 		rdsv3_send_remove_from_sock(&to_be_dropped, RDSV3_RDMA_DROPPED);
383 
384 	if (conn->c_trans->xmit_complete)
385 		conn->c_trans->xmit_complete(conn);
386 
387 	/*
388 	 * We might be racing with another sender who queued a message but
389 	 * backed off on noticing that we held the c_send_lock.  If we check
390 	 * for queued messages after dropping the sem then either we'll
391 	 * see the queued message or the queuer will get the sem.  If we
392 	 * notice the queued message then we trigger an immediate retry.
393 	 *
394 	 * We need to be careful only to do this when we stopped processing
395 	 * the send queue because it was empty.  It's the only way we
396 	 * stop processing the loop when the transport hasn't taken
397 	 * responsibility for forward progress.
398 	 */
399 	mutex_exit(&conn->c_send_lock);
400 
401 	if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) {
402 		/*
403 		 * We exhausted the send quota, but there's work left to
404 		 * do. Return and (re-)schedule the send worker.
405 		 */
406 		ret = -EAGAIN;
407 	}
408 
409 	if (ret == 0 && was_empty) {
410 		/*
411 		 * A simple bit test would be way faster than taking the
412 		 * spin lock
413 		 */
414 		mutex_enter(&conn->c_lock);
415 		if (!list_is_empty(&conn->c_send_queue)) {
416 			rdsv3_stats_inc(s_send_sem_queue_raced);
417 			ret = -EAGAIN;
418 		}
419 		mutex_exit(&conn->c_lock);
420 	}
421 
422 out:
423 	RDSV3_DPRINTF4("rdsv3_send_xmit", "Return(conn: %p, ret: %d)",
424 	    conn, ret);
425 	return (ret);
426 }
427 
428 static void
429 rdsv3_send_sndbuf_remove(struct rdsv3_sock *rs, struct rdsv3_message *rm)
430 {
431 	uint32_t len = ntohl(rm->m_inc.i_hdr.h_len);
432 
433 	ASSERT(mutex_owned(&rs->rs_lock));
434 
435 	ASSERT(rs->rs_snd_bytes >= len);
436 	rs->rs_snd_bytes -= len;
437 
438 	if (rs->rs_snd_bytes == 0)
439 		rdsv3_stats_inc(s_send_queue_empty);
440 }
441 
442 static inline int
443 rdsv3_send_is_acked(struct rdsv3_message *rm, uint64_t ack,
444     is_acked_func is_acked)
445 {
446 	if (is_acked)
447 		return (is_acked(rm, ack));
448 	return (ntohll(rm->m_inc.i_hdr.h_sequence) <= ack);
449 }
450 
451 /*
452  * Returns true if there are no messages on the send and retransmit queues
453  * which have a sequence number greater than or equal to the given sequence
454  * number.
455  */
456 int
457 rdsv3_send_acked_before(struct rdsv3_connection *conn, uint64_t seq)
458 {
459 	struct rdsv3_message *rm;
460 	int ret = 1;
461 
462 	RDSV3_DPRINTF4("rdsv3_send_acked_before", "Enter(conn: %p)", conn);
463 
464 	mutex_enter(&conn->c_lock);
465 
466 	/* XXX - original code spits out warning */
467 	rm = list_head(&conn->c_retrans);
468 	if (ntohll(rm->m_inc.i_hdr.h_sequence) < seq)
469 		ret = 0;
470 
471 	/* XXX - original code spits out warning */
472 	rm = list_head(&conn->c_send_queue);
473 	if (ntohll(rm->m_inc.i_hdr.h_sequence) < seq)
474 		ret = 0;
475 
476 	mutex_exit(&conn->c_lock);
477 
478 	RDSV3_DPRINTF4("rdsv3_send_acked_before", "Return(conn: %p)", conn);
479 
480 	return (ret);
481 }
482 
483 /*
484  * This is pretty similar to what happens below in the ACK
485  * handling code - except that we call here as soon as we get
486  * the IB send completion on the RDMA op and the accompanying
487  * message.
488  */
489 void
490 rdsv3_rdma_send_complete(struct rdsv3_message *rm, int status)
491 {
492 	struct rdsv3_sock *rs = NULL;
493 	struct rdsv3_rdma_op *ro;
494 	struct rdsv3_notifier *notifier;
495 
496 	RDSV3_DPRINTF4("rdsv3_rdma_send_complete", "Enter(rm: %p)", rm);
497 
498 	mutex_enter(&rm->m_rs_lock);
499 
500 	ro = rm->m_rdma_op;
501 	if (test_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags) &&
502 	    ro && ro->r_notify &&
503 	    (notifier = ro->r_notifier) != NULL) {
504 		ro->r_notifier = NULL;
505 		rs = rm->m_rs;
506 		rdsv3_sk_sock_hold(rdsv3_rs_to_sk(rs));
507 
508 		notifier->n_status = status;
509 		mutex_enter(&rs->rs_lock);
510 		list_insert_tail(&rs->rs_notify_queue, notifier);
511 		mutex_exit(&rs->rs_lock);
512 	}
513 
514 	mutex_exit(&rm->m_rs_lock);
515 
516 	if (rs) {
517 		rdsv3_wake_sk_sleep(rs);
518 		rdsv3_sk_sock_put(rdsv3_rs_to_sk(rs));
519 	}
520 
521 	RDSV3_DPRINTF4("rdsv3_rdma_send_complete", "Return(rm: %p)", rm);
522 }
523 
524 /*
525  * This is the same as rdsv3_rdma_send_complete except we
526  * don't do any locking - we have all the ingredients (message,
527  * socket, socket lock) and can just move the notifier.
528  */
529 static inline void
530 __rdsv3_rdma_send_complete(struct rdsv3_sock *rs, struct rdsv3_message *rm,
531     int status)
532 {
533 	struct rdsv3_rdma_op *ro;
534 	void *ic;
535 
536 	RDSV3_DPRINTF4("__rdsv3_rdma_send_complete",
537 	    "Enter(rs: %p, rm: %p)", rs, rm);
538 
539 	ro = rm->m_rdma_op;
540 	if (ro && ro->r_notify && ro->r_notifier) {
541 		ro->r_notifier->n_status = status;
542 		list_insert_tail(&rs->rs_notify_queue, ro->r_notifier);
543 		ro->r_notifier = NULL;
544 	}
545 
546 	/* No need to wake the app - caller does this */
547 }
548 
549 /*
550  * This is called from the IB send completion when we detect
551  * a RDMA operation that failed with remote access error.
552  * So speed is not an issue here.
553  */
554 struct rdsv3_message *
555 rdsv3_send_get_message(struct rdsv3_connection *conn,
556     struct rdsv3_rdma_op *op)
557 {
558 	struct rdsv3_message *rm, *tmp, *found = NULL;
559 
560 	RDSV3_DPRINTF4("rdsv3_send_get_message", "Enter(conn: %p)", conn);
561 
562 	mutex_enter(&conn->c_lock);
563 
564 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_retrans, m_conn_item) {
565 		if (rm->m_rdma_op == op) {
566 			atomic_add_32(&rm->m_refcount, 1);
567 			found = rm;
568 			goto out;
569 		}
570 	}
571 
572 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_send_queue,
573 	    m_conn_item) {
574 		if (rm->m_rdma_op == op) {
575 			atomic_add_32(&rm->m_refcount, 1);
576 			found = rm;
577 			break;
578 		}
579 	}
580 
581 out:
582 	mutex_exit(&conn->c_lock);
583 
584 	return (found);
585 }
586 
587 /*
588  * This removes messages from the socket's list if they're on it.  The list
589  * argument must be private to the caller, we must be able to modify it
590  * without locks.  The messages must have a reference held for their
591  * position on the list.  This function will drop that reference after
592  * removing the messages from the 'messages' list regardless of if it found
593  * the messages on the socket list or not.
594  */
595 void
596 rdsv3_send_remove_from_sock(struct list *messages, int status)
597 {
598 	struct rdsv3_sock *rs = NULL;
599 	struct rdsv3_message *rm;
600 
601 	RDSV3_DPRINTF4("rdsv3_send_remove_from_sock", "Enter");
602 
603 	while (!list_is_empty(messages)) {
604 		rm = list_remove_head(messages);
605 
606 		/*
607 		 * If we see this flag cleared then we're *sure* that someone
608 		 * else beat us to removing it from the sock.  If we race
609 		 * with their flag update we'll get the lock and then really
610 		 * see that the flag has been cleared.
611 		 *
612 		 * The message spinlock makes sure nobody clears rm->m_rs
613 		 * while we're messing with it. It does not prevent the
614 		 * message from being removed from the socket, though.
615 		 */
616 		mutex_enter(&rm->m_rs_lock);
617 		if (!test_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags))
618 			goto unlock_and_drop;
619 
620 		if (rs != rm->m_rs) {
621 			if (rs) {
622 				rdsv3_wake_sk_sleep(rs);
623 				rdsv3_sk_sock_put(rdsv3_rs_to_sk(rs));
624 			}
625 			rs = rm->m_rs;
626 			rdsv3_sk_sock_hold(rdsv3_rs_to_sk(rs));
627 		}
628 
629 		mutex_enter(&rs->rs_lock);
630 		if (test_and_clear_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags)) {
631 			struct rdsv3_rdma_op *ro = rm->m_rdma_op;
632 			struct rdsv3_notifier *notifier;
633 
634 			list_remove_node(&rm->m_sock_item);
635 			rdsv3_send_sndbuf_remove(rs, rm);
636 
637 			if (ro &&
638 			    (notifier = ro->r_notifier) != NULL &&
639 			    (status || ro->r_notify)) {
640 				list_insert_tail(&rs->rs_notify_queue,
641 				    notifier);
642 				if (!notifier->n_status)
643 					notifier->n_status = status;
644 				rm->m_rdma_op->r_notifier = NULL;
645 			}
646 			rdsv3_message_put(rm);
647 			rm->m_rs = NULL;
648 		}
649 		mutex_exit(&rs->rs_lock);
650 
651 unlock_and_drop:
652 		mutex_exit(&rm->m_rs_lock);
653 		rdsv3_message_put(rm);
654 	}
655 
656 	if (rs) {
657 		rdsv3_wake_sk_sleep(rs);
658 		rdsv3_sk_sock_put(rdsv3_rs_to_sk(rs));
659 	}
660 
661 	RDSV3_DPRINTF4("rdsv3_send_remove_from_sock", "Return");
662 }
663 
664 /*
665  * Transports call here when they've determined that the receiver queued
666  * messages up to, and including, the given sequence number.  Messages are
667  * moved to the retrans queue when rdsv3_send_xmit picks them off the send
668  * queue. This means that in the TCP case, the message may not have been
669  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
670  * checks the RDSV3_MSG_HAS_ACK_SEQ bit.
671  *
672  * XXX It's not clear to me how this is safely serialized with socket
673  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
674  */
675 void
676 rdsv3_send_drop_acked(struct rdsv3_connection *conn, uint64_t ack,
677     is_acked_func is_acked)
678 {
679 	struct rdsv3_message *rm, *tmp;
680 	list_t list;
681 
682 	RDSV3_DPRINTF4("rdsv3_send_drop_acked", "Enter(conn: %p)", conn);
683 
684 	list_create(&list, sizeof (struct rdsv3_message),
685 	    offsetof(struct rdsv3_message, m_conn_item));
686 
687 	mutex_enter(&conn->c_lock);
688 
689 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_retrans, m_conn_item) {
690 		if (!rdsv3_send_is_acked(rm, ack, is_acked))
691 			break;
692 
693 		list_remove_node(&rm->m_conn_item);
694 		list_insert_tail(&list, rm);
695 		clear_bit(RDSV3_MSG_ON_CONN, &rm->m_flags);
696 	}
697 
698 #if 0
699 XXX
700 	/* order flag updates with spin locks */
701 	if (!list_is_empty(&list))
702 		smp_mb__after_clear_bit();
703 #endif
704 
705 	mutex_exit(&conn->c_lock);
706 
707 	/* now remove the messages from the sock list as needed */
708 	rdsv3_send_remove_from_sock(&list, RDSV3_RDMA_SUCCESS);
709 
710 	RDSV3_DPRINTF4("rdsv3_send_drop_acked", "Return(conn: %p)", conn);
711 }
712 
713 void
714 rdsv3_send_drop_to(struct rdsv3_sock *rs, struct sockaddr_in *dest)
715 {
716 	struct rdsv3_message *rm, *tmp;
717 	struct rdsv3_connection *conn;
718 	list_t list;
719 	int wake = 0;
720 
721 	RDSV3_DPRINTF4("rdsv3_send_drop_to", "Enter(rs: %p)", rs);
722 
723 	list_create(&list, sizeof (struct rdsv3_message),
724 	    offsetof(struct rdsv3_message, m_sock_item));
725 
726 	/* get all the messages we're dropping under the rs lock */
727 	mutex_enter(&rs->rs_lock);
728 
729 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &rs->rs_send_queue,
730 	    m_sock_item) {
731 		if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
732 		    dest->sin_port != rm->m_inc.i_hdr.h_dport))
733 			continue;
734 
735 		wake = 1;
736 		list_remove(&rs->rs_send_queue, rm);
737 		list_insert_tail(&list, rm);
738 		rdsv3_send_sndbuf_remove(rs, rm);
739 		clear_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags);
740 	}
741 
742 	mutex_exit(&rs->rs_lock);
743 
744 	conn = NULL;
745 
746 	/* now remove the messages from the conn list as needed */
747 	RDSV3_FOR_EACH_LIST_NODE(rm, &list, m_sock_item) {
748 		/*
749 		 * We do this here rather than in the loop above, so that
750 		 * we don't have to nest m_rs_lock under rs->rs_lock
751 		 */
752 		mutex_enter(&rm->m_rs_lock);
753 		/* If this is a RDMA operation, notify the app. */
754 		__rdsv3_rdma_send_complete(rs, rm, RDSV3_RDMA_CANCELED);
755 		rm->m_rs = NULL;
756 		mutex_exit(&rm->m_rs_lock);
757 
758 		/*
759 		 * If we see this flag cleared then we're *sure* that someone
760 		 * else beat us to removing it from the conn.  If we race
761 		 * with their flag update we'll get the lock and then really
762 		 * see that the flag has been cleared.
763 		 */
764 		if (!test_bit(RDSV3_MSG_ON_CONN, &rm->m_flags))
765 			continue;
766 
767 		if (conn != rm->m_inc.i_conn) {
768 			if (conn)
769 				mutex_exit(&conn->c_lock);
770 			conn = rm->m_inc.i_conn;
771 			mutex_enter(&conn->c_lock);
772 		}
773 
774 		if (test_and_clear_bit(RDSV3_MSG_ON_CONN, &rm->m_flags)) {
775 			list_remove_node(&rm->m_conn_item);
776 			rdsv3_message_put(rm);
777 		}
778 	}
779 
780 	if (conn)
781 		mutex_exit(&conn->c_lock);
782 
783 	if (wake)
784 		rdsv3_wake_sk_sleep(rs);
785 
786 	while (!list_is_empty(&list)) {
787 		rm = list_remove_head(&list);
788 
789 		rdsv3_message_wait(rm);
790 		rdsv3_message_put(rm);
791 	}
792 
793 	RDSV3_DPRINTF4("rdsv3_send_drop_to", "Return(rs: %p)", rs);
794 }
795 
796 /*
797  * we only want this to fire once so we use the callers 'queued'.  It's
798  * possible that another thread can race with us and remove the
799  * message from the flow with RDSV3_CANCEL_SENT_TO.
800  */
801 static int
802 rdsv3_send_queue_rm(struct rdsv3_sock *rs, struct rdsv3_connection *conn,
803     struct rdsv3_message *rm, uint16_be_t sport,
804     uint16_be_t dport, int *queued)
805 {
806 	uint32_t len;
807 
808 	RDSV3_DPRINTF4("rdsv3_send_queue_rm", "Enter(rs: %p, rm: %p)", rs, rm);
809 
810 	if (*queued)
811 		goto out;
812 
813 	len = ntohl(rm->m_inc.i_hdr.h_len);
814 
815 	/*
816 	 * this is the only place which holds both the socket's rs_lock
817 	 * and the connection's c_lock
818 	 */
819 	mutex_enter(&rs->rs_lock);
820 
821 	/*
822 	 * If there is a little space in sndbuf, we don't queue anything,
823 	 * and userspace gets -EAGAIN. But poll() indicates there's send
824 	 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
825 	 * freed up by incoming acks. So we check the *old* value of
826 	 * rs_snd_bytes here to allow the last msg to exceed the buffer,
827 	 * and poll() now knows no more data can be sent.
828 	 */
829 	if (rs->rs_snd_bytes < rdsv3_sk_sndbuf(rs)) {
830 		rs->rs_snd_bytes += len;
831 
832 		/*
833 		 * let recv side know we are close to send space exhaustion.
834 		 * This is probably not the optimal way to do it, as this
835 		 * means we set the flag on *all* messages as soon as our
836 		 * throughput hits a certain threshold.
837 		 */
838 		if (rs->rs_snd_bytes >= rdsv3_sk_sndbuf(rs) / 2)
839 			set_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags);
840 
841 		list_insert_tail(&rs->rs_send_queue, rm);
842 		set_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags);
843 
844 		rdsv3_message_addref(rm);
845 		rm->m_rs = rs;
846 
847 		/*
848 		 * The code ordering is a little weird, but we're
849 		 * trying to minimize the time we hold c_lock
850 		 */
851 		rdsv3_message_populate_header(&rm->m_inc.i_hdr, sport,
852 		    dport, 0);
853 		rm->m_inc.i_conn = conn;
854 		rdsv3_message_addref(rm);	/* XXX - called twice */
855 
856 		mutex_enter(&conn->c_lock);
857 		rm->m_inc.i_hdr.h_sequence = htonll(conn->c_next_tx_seq++);
858 		list_insert_tail(&conn->c_send_queue, rm);
859 		set_bit(RDSV3_MSG_ON_CONN, &rm->m_flags);
860 		mutex_exit(&conn->c_lock);
861 
862 		RDSV3_DPRINTF5("rdsv3_send_queue_rm",
863 		    "queued msg %p len %d, rs %p bytes %d seq %llu",
864 		    rm, len, rs, rs->rs_snd_bytes,
865 		    (unsigned long long)ntohll(
866 		    rm->m_inc.i_hdr.h_sequence));
867 
868 		*queued = 1;
869 	}
870 
871 	mutex_exit(&rs->rs_lock);
872 
873 	RDSV3_DPRINTF4("rdsv3_send_queue_rm", "Return(rs: %p)", rs);
874 out:
875 	return (*queued);
876 }
877 
878 static int
879 rdsv3_cmsg_send(struct rdsv3_sock *rs, struct rdsv3_message *rm,
880     struct msghdr *msg, int *allocated_mr)
881 {
882 	struct cmsghdr *cmsg;
883 	int ret = 0;
884 
885 	RDSV3_DPRINTF4("rdsv3_cmsg_send", "Enter(rs: %p)", rs);
886 
887 	for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
888 
889 		if (cmsg->cmsg_level != SOL_RDS)
890 			continue;
891 
892 		RDSV3_DPRINTF4("rdsv3_cmsg_send", "cmsg(%p, %p) type %d",
893 		    cmsg, rm, cmsg->cmsg_type);
894 		/*
895 		 * As a side effect, RDMA_DEST and RDMA_MAP will set
896 		 * rm->m_rdma_cookie and rm->m_rdma_mr.
897 		 */
898 		switch (cmsg->cmsg_type) {
899 		case RDSV3_CMSG_RDMA_ARGS:
900 			ret = rdsv3_cmsg_rdma_args(rs, rm, cmsg);
901 			break;
902 
903 		case RDSV3_CMSG_RDMA_DEST:
904 			ret = rdsv3_cmsg_rdma_dest(rs, rm, cmsg);
905 			break;
906 
907 		case RDSV3_CMSG_RDMA_MAP:
908 			ret = rdsv3_cmsg_rdma_map(rs, rm, cmsg);
909 			if (ret)
910 				*allocated_mr = 1;
911 			break;
912 
913 		default:
914 			return (-EINVAL);
915 		}
916 
917 		if (ret)
918 			break;
919 	}
920 
921 	RDSV3_DPRINTF4("rdsv3_cmsg_send", "Return(rs: %p)", rs);
922 
923 	return (ret);
924 }
925 
926 int
927 rdsv3_sendmsg(struct rdsv3_sock *rs, uio_t *uio, struct nmsghdr *msg,
928     size_t payload_len)
929 {
930 	struct rsock *sk = rdsv3_rs_to_sk(rs);
931 	struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
932 	uint32_be_t daddr;
933 	uint16_be_t dport;
934 	struct rdsv3_message *rm = NULL;
935 	struct rdsv3_connection *conn;
936 	int ret = 0;
937 	int queued = 0, allocated_mr = 0;
938 	int nonblock = msg->msg_flags & MSG_DONTWAIT;
939 	long timeo = rdsv3_rcvtimeo(sk, nonblock);
940 
941 	RDSV3_DPRINTF4("rdsv3_sendmsg", "Enter(rs: %p)", rs);
942 
943 	if (msg->msg_namelen) {
944 		/* XXX fail non-unicast destination IPs? */
945 		if (msg->msg_namelen < sizeof (*usin) ||
946 		    usin->sin_family != AF_INET_OFFLOAD) {
947 			ret = -EINVAL;
948 			RDSV3_DPRINTF2("rdsv3_sendmsg", "returning: %d", -ret);
949 			goto out;
950 		}
951 		daddr = usin->sin_addr.s_addr;
952 		dport = usin->sin_port;
953 	} else {
954 		/* We only care about consistency with ->connect() */
955 		mutex_enter(&sk->sk_lock);
956 		daddr = rs->rs_conn_addr;
957 		dport = rs->rs_conn_port;
958 		mutex_exit(&sk->sk_lock);
959 	}
960 
961 	/* racing with another thread binding seems ok here */
962 	if (daddr == 0 || rs->rs_bound_addr == 0) {
963 		ret = -ENOTCONN; /* XXX not a great errno */
964 		RDSV3_DPRINTF2("rdsv3_sendmsg", "returning: %d", -ret);
965 		goto out;
966 	}
967 
968 	rm = rdsv3_message_copy_from_user(uio, payload_len);
969 	if (IS_ERR(rm)) {
970 		ret = PTR_ERR(rm);
971 		RDSV3_DPRINTF2("rdsv3_sendmsg",
972 		    "rdsv3_message_copy_from_user failed %d", -ret);
973 		rm = NULL;
974 		goto out;
975 	}
976 
977 	rm->m_daddr = daddr;
978 
979 	/*
980 	 * rdsv3_conn_create has a spinlock that runs with IRQ off.
981 	 * Caching the conn in the socket helps a lot.
982 	 */
983 	mutex_enter(&rs->rs_conn_lock);
984 	if (rs->rs_conn && rs->rs_conn->c_faddr == daddr) {
985 		conn = rs->rs_conn;
986 	} else {
987 		conn = rdsv3_conn_create_outgoing(rs->rs_bound_addr,
988 		    daddr, rs->rs_transport, KM_NOSLEEP);
989 		if (IS_ERR(conn)) {
990 			mutex_exit(&rs->rs_conn_lock);
991 			ret = PTR_ERR(conn);
992 			RDSV3_DPRINTF2("rdsv3_sendmsg",
993 			    "rdsv3_conn_create_outgoing failed %d",
994 			    -ret);
995 			goto out;
996 		}
997 		rs->rs_conn = conn;
998 	}
999 	mutex_exit(&rs->rs_conn_lock);
1000 
1001 	/* Parse any control messages the user may have included. */
1002 	ret = rdsv3_cmsg_send(rs, rm, msg, &allocated_mr);
1003 	if (ret) {
1004 		RDSV3_DPRINTF2("rdsv3_sendmsg",
1005 		    "rdsv3_cmsg_send(rs: %p rm: %p msg: %p) returned: %d",
1006 		    rs, rm, msg, ret);
1007 		goto out;
1008 	}
1009 
1010 	if ((rm->m_rdma_cookie || rm->m_rdma_op) &&
1011 	    conn->c_trans->xmit_rdma == NULL) {
1012 		RDSV3_DPRINTF0("rdsv3_sendmsg", "rdma_op %p conn xmit_rdma %p",
1013 		    rm->m_rdma_op, conn->c_trans->xmit_rdma);
1014 		ret = -EOPNOTSUPP;
1015 		goto out;
1016 	}
1017 
1018 	/*
1019 	 * If the connection is down, trigger a connect. We may
1020 	 * have scheduled a delayed reconnect however - in this case
1021 	 * we should not interfere.
1022 	 */
1023 	if (rdsv3_conn_state(conn) == RDSV3_CONN_DOWN &&
1024 	    !test_and_set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags))
1025 		rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
1026 
1027 	ret = rdsv3_cong_wait(conn->c_fcong, dport, nonblock, rs);
1028 	if (ret) {
1029 		RDSV3_DPRINTF2("rdsv3_sendmsg",
1030 		    "rdsv3_cong_wait (dport: %d) returned: %d", dport, ret);
1031 		goto out;
1032 	}
1033 
1034 	(void) rdsv3_send_queue_rm(rs, conn, rm, rs->rs_bound_port, dport,
1035 	    &queued);
1036 	if (!queued) {
1037 		/* rdsv3_stats_inc(s_send_queue_full); */
1038 		/* XXX make sure this is reasonable */
1039 		if (payload_len > rdsv3_sk_sndbuf(rs)) {
1040 			ret = -EMSGSIZE;
1041 			RDSV3_DPRINTF2("rdsv3_sendmsg",
1042 			    "msgsize(%d) too big, returning: %d",
1043 			    payload_len, -ret);
1044 			goto out;
1045 		}
1046 		if (nonblock) {
1047 			ret = -EAGAIN;
1048 			RDSV3_DPRINTF3("rdsv3_sendmsg",
1049 			    "send queue full (%d), returning: %d",
1050 			    payload_len, -ret);
1051 			goto out;
1052 		}
1053 
1054 		mutex_enter(&sk->sk_sleep->waitq_mutex);
1055 		while (!rdsv3_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1056 		    dport, &queued)) {
1057 #if 0
1058 			ret = cv_timedwait_sig(&sk->sk_sleep->waitq_cv,
1059 			    &sk->sk_sleep->waitq_mutex,
1060 			    timeo * drv_usectohz(1000000) + ddi_get_lbolt());
1061 			if (ret <= 0) {
1062 				/* signal/timeout pending */
1063 				RDSV3_DPRINTF2("rdsv3_sendmsg",
1064 				    "woke due to signal/timeout: %d",
1065 				    ret);
1066 				ret = (ret == 0) ? -ERESTART : -ETIMEDOUT;
1067 				mutex_exit(&sk->sk_sleep->waitq_mutex);
1068 				goto out;
1069 			}
1070 #else
1071 			ret = cv_wait_sig(&sk->sk_sleep->waitq_cv,
1072 			    &sk->sk_sleep->waitq_mutex);
1073 			if (ret == 0) {
1074 				/* signal/timeout pending */
1075 				RDSV3_DPRINTF2("rdsv3_sendmsg",
1076 				    "woke due to signal: %d",
1077 				    ret);
1078 				ret = -ERESTART;
1079 				mutex_exit(&sk->sk_sleep->waitq_mutex);
1080 				goto out;
1081 			}
1082 #endif
1083 		}
1084 		mutex_exit(&sk->sk_sleep->waitq_mutex);
1085 
1086 		RDSV3_DPRINTF5("rdsv3_sendmsg", "sendmsg woke queued %d",
1087 		    queued);
1088 
1089 		ASSERT(queued);
1090 		ret = 0;
1091 	}
1092 
1093 	/*
1094 	 * By now we've committed to the send.  We reuse rdsv3_send_worker()
1095 	 * to retry sends in the rds thread if the transport asks us to.
1096 	 */
1097 	rdsv3_stats_inc(s_send_queued);
1098 
1099 	if (!test_bit(RDSV3_LL_SEND_FULL, &conn->c_flags))
1100 		rdsv3_send_worker(&conn->c_send_w.work);
1101 
1102 	rdsv3_message_put(rm);
1103 	RDSV3_DPRINTF4("rdsv3_sendmsg", "Return(rs: %p, len: %d)",
1104 	    rs, payload_len);
1105 	return (payload_len);
1106 
1107 out:
1108 	/*
1109 	 * If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1110 	 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1111 	 * or in any other way, we need to destroy the MR again
1112 	 */
1113 	if (allocated_mr)
1114 		rdsv3_rdma_unuse(rs, rdsv3_rdma_cookie_key(rm->m_rdma_cookie),
1115 		    1);
1116 
1117 	if (rm)
1118 		rdsv3_message_put(rm);
1119 	return (ret);
1120 }
1121 
1122 /*
1123  * Reply to a ping packet.
1124  */
1125 int
1126 rdsv3_send_pong(struct rdsv3_connection *conn, uint16_be_t dport)
1127 {
1128 	struct rdsv3_message *rm;
1129 	int ret = 0;
1130 
1131 	RDSV3_DPRINTF4("rdsv3_send_pong", "Enter(conn: %p)", conn);
1132 
1133 	rm = rdsv3_message_alloc(0, KM_NOSLEEP);
1134 	if (rm == NULL) {
1135 		ret = -ENOMEM;
1136 		goto out;
1137 	}
1138 
1139 	rm->m_daddr = conn->c_faddr;
1140 
1141 	/*
1142 	 * If the connection is down, trigger a connect. We may
1143 	 * have scheduled a delayed reconnect however - in this case
1144 	 * we should not interfere.
1145 	 */
1146 	if (rdsv3_conn_state(conn) == RDSV3_CONN_DOWN &&
1147 	    !test_and_set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags))
1148 		rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
1149 
1150 	ret = rdsv3_cong_wait(conn->c_fcong, dport, 1, NULL);
1151 	if (ret)
1152 		goto out;
1153 
1154 	mutex_enter(&conn->c_lock);
1155 	list_insert_tail(&conn->c_send_queue, rm);
1156 	set_bit(RDSV3_MSG_ON_CONN, &rm->m_flags);
1157 	rdsv3_message_addref(rm);
1158 	rm->m_inc.i_conn = conn;
1159 
1160 	rdsv3_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1161 	    conn->c_next_tx_seq);
1162 	conn->c_next_tx_seq++;
1163 	mutex_exit(&conn->c_lock);
1164 
1165 	rdsv3_stats_inc(s_send_queued);
1166 	rdsv3_stats_inc(s_send_pong);
1167 
1168 	rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
1169 	rdsv3_message_put(rm);
1170 
1171 	RDSV3_DPRINTF4("rdsv3_send_pong", "Return(conn: %p)", conn);
1172 	return (0);
1173 
1174 out:
1175 	if (rm)
1176 		rdsv3_message_put(rm);
1177 	return (ret);
1178 }
1179