xref: /illumos-gate/usr/src/uts/common/io/ib/clients/rdsv3/send.c (revision 2bbdd445a21f9d61f4a0ca0faf05d5ceb2bd91f3)
1 /*
2  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
3  */
4 
5 /*
6  * This file contains code imported from the OFED rds source file send.c
7  * Oracle elects to have and use the contents of send.c under and governed
8  * by the OpenIB.org BSD license (see below for full license text). However,
9  * the following notice accompanied the original version of this file:
10  */
11 
12 /*
13  * Copyright (c) 2006 Oracle.  All rights reserved.
14  *
15  * This software is available to you under a choice of one of two
16  * licenses.  You may choose to be licensed under the terms of the GNU
17  * General Public License (GPL) Version 2, available from the file
18  * COPYING in the main directory of this source tree, or the
19  * OpenIB.org BSD license below:
20  *
21  *     Redistribution and use in source and binary forms, with or
22  *     without modification, are permitted provided that the following
23  *     conditions are met:
24  *
25  *      - Redistributions of source code must retain the above
26  *        copyright notice, this list of conditions and the following
27  *        disclaimer.
28  *
29  *      - Redistributions in binary form must reproduce the above
30  *        copyright notice, this list of conditions and the following
31  *        disclaimer in the documentation and/or other materials
32  *        provided with the distribution.
33  *
34  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
35  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
36  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
37  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
38  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
39  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
40  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
41  * SOFTWARE.
42  *
43  */
44 #include <sys/stropts.h>
45 #include <sys/systm.h>
46 
47 #include <sys/rds.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 
51 #include <sys/ib/clients/rdsv3/rdsv3.h>
52 #include <sys/ib/clients/rdsv3/rdma.h>
53 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
54 
55 /*
56  * When transmitting messages in rdsv3_send_xmit, we need to emerge from
57  * time to time and briefly release the CPU. Otherwise the softlock watchdog
58  * will kick our shin.
59  * Also, it seems fairer to not let one busy connection stall all the
60  * others.
61  *
62  * send_batch_count is the number of times we'll loop in send_xmit. Setting
63  * it to 0 will restore the old behavior (where we looped until we had
64  * drained the queue).
65  */
66 static int send_batch_count = 64;
67 
68 extern void rdsv3_ib_send_unmap_rdma(void *ic, struct rdsv3_rdma_op *op);
69 /*
70  * Reset the send state. Caller must hold c_send_lock when calling here.
71  */
72 void
73 rdsv3_send_reset(struct rdsv3_connection *conn)
74 {
75 	struct rdsv3_message *rm, *tmp;
76 	struct rdsv3_rdma_op *ro;
77 
78 	RDSV3_DPRINTF4("rdsv3_send_reset", "Enter(conn: %p)", conn);
79 
80 	ASSERT(MUTEX_HELD(&conn->c_send_lock));
81 
82 	if (conn->c_xmit_rm) {
83 		rm = conn->c_xmit_rm;
84 		ro = rm->m_rdma_op;
85 		if (ro && ro->r_mapped) {
86 			RDSV3_DPRINTF2("rdsv3_send_reset",
87 			    "rm %p mflg 0x%x map %d mihdl %p sgl %p",
88 			    rm, rm->m_flags, ro->r_mapped,
89 			    ro->r_rdma_sg[0].mihdl,
90 			    ro->r_rdma_sg[0].swr.wr_sgl);
91 			rdsv3_ib_send_unmap_rdma(conn->c_transport_data, ro);
92 		}
93 		/*
94 		 * Tell the user the RDMA op is no longer mapped by the
95 		 * transport. This isn't entirely true (it's flushed out
96 		 * independently) but as the connection is down, there's
97 		 * no ongoing RDMA to/from that memory
98 		 */
99 		rdsv3_message_unmapped(conn->c_xmit_rm);
100 		rdsv3_message_put(conn->c_xmit_rm);
101 		conn->c_xmit_rm = NULL;
102 	}
103 
104 	conn->c_xmit_sg = 0;
105 	conn->c_xmit_hdr_off = 0;
106 	conn->c_xmit_data_off = 0;
107 	conn->c_xmit_rdma_sent = 0;
108 	conn->c_map_queued = 0;
109 
110 	conn->c_unacked_packets = rdsv3_sysctl_max_unacked_packets;
111 	conn->c_unacked_bytes = rdsv3_sysctl_max_unacked_bytes;
112 
113 	/* Mark messages as retransmissions, and move them to the send q */
114 	mutex_enter(&conn->c_lock);
115 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_retrans, m_conn_item) {
116 		set_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags);
117 		set_bit(RDSV3_MSG_RETRANSMITTED, &rm->m_flags);
118 		if (rm->m_rdma_op && rm->m_rdma_op->r_mapped) {
119 			RDSV3_DPRINTF4("_send_reset",
120 			    "RT rm %p mflg 0x%x sgl %p",
121 			    rm, rm->m_flags,
122 			    rm->m_rdma_op->r_rdma_sg[0].swr.wr_sgl);
123 		}
124 	}
125 	list_move_tail(&conn->c_send_queue, &conn->c_retrans);
126 	mutex_exit(&conn->c_lock);
127 
128 	RDSV3_DPRINTF4("rdsv3_send_reset", "Return(conn: %p)", conn);
129 }
130 
131 /*
132  * We're making the concious trade-off here to only send one message
133  * down the connection at a time.
134  *   Pro:
135  *      - tx queueing is a simple fifo list
136  *   	- reassembly is optional and easily done by transports per conn
137  *      - no per flow rx lookup at all, straight to the socket
138  *   	- less per-frag memory and wire overhead
139  *   Con:
140  *      - queued acks can be delayed behind large messages
141  *   Depends:
142  *      - small message latency is higher behind queued large messages
143  *      - large message latency isn't starved by intervening small sends
144  */
145 int
146 rdsv3_send_xmit(struct rdsv3_connection *conn)
147 {
148 	struct rdsv3_message *rm;
149 	unsigned int tmp;
150 	unsigned int send_quota = send_batch_count;
151 	struct rdsv3_scatterlist *sg;
152 	int ret = 0;
153 	int was_empty = 0;
154 	list_t to_be_dropped;
155 
156 restart:
157 	if (!rdsv3_conn_up(conn))
158 		goto out;
159 
160 	RDSV3_DPRINTF4("rdsv3_send_xmit", "Enter(conn: %p)", conn);
161 
162 	list_create(&to_be_dropped, sizeof (struct rdsv3_message),
163 	    offsetof(struct rdsv3_message, m_conn_item));
164 
165 	/*
166 	 * sendmsg calls here after having queued its message on the send
167 	 * queue.  We only have one task feeding the connection at a time.  If
168 	 * another thread is already feeding the queue then we back off.  This
169 	 * avoids blocking the caller and trading per-connection data between
170 	 * caches per message.
171 	 */
172 	if (!mutex_tryenter(&conn->c_send_lock)) {
173 		RDSV3_DPRINTF4("rdsv3_send_xmit",
174 		    "Another thread running(conn: %p)", conn);
175 		rdsv3_stats_inc(s_send_sem_contention);
176 		ret = -ENOMEM;
177 		goto out;
178 	}
179 	atomic_inc_32(&conn->c_senders);
180 
181 	if (conn->c_trans->xmit_prepare)
182 		conn->c_trans->xmit_prepare(conn);
183 
184 	/*
185 	 * spin trying to push headers and data down the connection until
186 	 * the connection doesn't make forward progress.
187 	 */
188 	while (--send_quota) {
189 		/*
190 		 * See if need to send a congestion map update if we're
191 		 * between sending messages.  The send_sem protects our sole
192 		 * use of c_map_offset and _bytes.
193 		 * Note this is used only by transports that define a special
194 		 * xmit_cong_map function. For all others, we create allocate
195 		 * a cong_map message and treat it just like any other send.
196 		 */
197 		if (conn->c_map_bytes) {
198 			ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
199 			    conn->c_map_offset);
200 			if (ret <= 0)
201 				break;
202 
203 			conn->c_map_offset += ret;
204 			conn->c_map_bytes -= ret;
205 			if (conn->c_map_bytes)
206 				continue;
207 		}
208 
209 		/*
210 		 * If we're done sending the current message, clear the
211 		 * offset and S/G temporaries.
212 		 */
213 		rm = conn->c_xmit_rm;
214 		if (rm != NULL &&
215 		    conn->c_xmit_hdr_off == sizeof (struct rdsv3_header) &&
216 		    conn->c_xmit_sg == rm->m_nents) {
217 			conn->c_xmit_rm = NULL;
218 			conn->c_xmit_sg = 0;
219 			conn->c_xmit_hdr_off = 0;
220 			conn->c_xmit_data_off = 0;
221 			conn->c_xmit_rdma_sent = 0;
222 
223 			/* Release the reference to the previous message. */
224 			rdsv3_message_put(rm);
225 			rm = NULL;
226 		}
227 
228 		/* If we're asked to send a cong map update, do so. */
229 		if (rm == NULL && test_and_clear_bit(0, &conn->c_map_queued)) {
230 			if (conn->c_trans->xmit_cong_map != NULL) {
231 				conn->c_map_offset = 0;
232 				conn->c_map_bytes =
233 				    sizeof (struct rdsv3_header) +
234 				    RDSV3_CONG_MAP_BYTES;
235 				continue;
236 			}
237 
238 			rm = rdsv3_cong_update_alloc(conn);
239 			if (IS_ERR(rm)) {
240 				ret = PTR_ERR(rm);
241 				break;
242 			}
243 
244 			conn->c_xmit_rm = rm;
245 		}
246 
247 		/*
248 		 * Grab the next message from the send queue, if there is one.
249 		 *
250 		 * c_xmit_rm holds a ref while we're sending this message down
251 		 * the connction.  We can use this ref while holding the
252 		 * send_sem.. rdsv3_send_reset() is serialized with it.
253 		 */
254 		if (rm == NULL) {
255 			unsigned int len;
256 
257 			mutex_enter(&conn->c_lock);
258 
259 			if (!list_is_empty(&conn->c_send_queue)) {
260 				rm = list_remove_head(&conn->c_send_queue);
261 				rdsv3_message_addref(rm);
262 
263 				/*
264 				 * Move the message from the send queue to
265 				 * the retransmit
266 				 * list right away.
267 				 */
268 				list_insert_tail(&conn->c_retrans, rm);
269 			}
270 
271 			mutex_exit(&conn->c_lock);
272 
273 			if (rm == NULL) {
274 				was_empty = 1;
275 				break;
276 			}
277 
278 			/*
279 			 * Unfortunately, the way Infiniband deals with
280 			 * RDMA to a bad MR key is by moving the entire
281 			 * queue pair to error state. We cold possibly
282 			 * recover from that, but right now we drop the
283 			 * connection.
284 			 * Therefore, we never retransmit messages with
285 			 * RDMA ops.
286 			 */
287 			if (rm->m_rdma_op &&
288 			    test_bit(RDSV3_MSG_RETRANSMITTED, &rm->m_flags)) {
289 				mutex_enter(&conn->c_lock);
290 				if (test_and_clear_bit(RDSV3_MSG_ON_CONN,
291 				    &rm->m_flags))
292 					list_remove_node(&rm->m_conn_item);
293 					list_insert_tail(&to_be_dropped, rm);
294 				mutex_exit(&conn->c_lock);
295 				rdsv3_message_put(rm);
296 				continue;
297 			}
298 
299 			/* Require an ACK every once in a while */
300 			len = ntohl(rm->m_inc.i_hdr.h_len);
301 			if (conn->c_unacked_packets == 0 ||
302 			    conn->c_unacked_bytes < len) {
303 				set_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags);
304 
305 				conn->c_unacked_packets =
306 				    rdsv3_sysctl_max_unacked_packets;
307 				conn->c_unacked_bytes =
308 				    rdsv3_sysctl_max_unacked_bytes;
309 				rdsv3_stats_inc(s_send_ack_required);
310 			} else {
311 				conn->c_unacked_bytes -= len;
312 				conn->c_unacked_packets--;
313 			}
314 
315 			conn->c_xmit_rm = rm;
316 		}
317 
318 		/*
319 		 * Try and send an rdma message.  Let's see if we can
320 		 * keep this simple and require that the transport either
321 		 * send the whole rdma or none of it.
322 		 */
323 		if (rm->m_rdma_op && !conn->c_xmit_rdma_sent) {
324 			ret = conn->c_trans->xmit_rdma(conn, rm->m_rdma_op);
325 			if (ret)
326 				break;
327 			conn->c_xmit_rdma_sent = 1;
328 			/*
329 			 * The transport owns the mapped memory for now.
330 			 * You can't unmap it while it's on the send queue
331 			 */
332 			set_bit(RDSV3_MSG_MAPPED, &rm->m_flags);
333 		}
334 
335 		if (conn->c_xmit_hdr_off < sizeof (struct rdsv3_header) ||
336 		    conn->c_xmit_sg < rm->m_nents) {
337 			ret = conn->c_trans->xmit(conn, rm,
338 			    conn->c_xmit_hdr_off,
339 			    conn->c_xmit_sg,
340 			    conn->c_xmit_data_off);
341 			if (ret <= 0)
342 				break;
343 
344 			if (conn->c_xmit_hdr_off <
345 			    sizeof (struct rdsv3_header)) {
346 				tmp = min(ret,
347 				    sizeof (struct rdsv3_header) -
348 				    conn->c_xmit_hdr_off);
349 				conn->c_xmit_hdr_off += tmp;
350 				ret -= tmp;
351 			}
352 
353 			sg = &rm->m_sg[conn->c_xmit_sg];
354 			while (ret) {
355 				tmp = min(ret, rdsv3_sg_len(sg) -
356 				    conn->c_xmit_data_off);
357 				conn->c_xmit_data_off += tmp;
358 				ret -= tmp;
359 				if (conn->c_xmit_data_off == rdsv3_sg_len(sg)) {
360 					conn->c_xmit_data_off = 0;
361 					sg++;
362 					conn->c_xmit_sg++;
363 					ASSERT(!(ret != 0 &&
364 					    conn->c_xmit_sg == rm->m_nents));
365 				}
366 			}
367 		}
368 	}
369 
370 	/* Nuke any messages we decided not to retransmit. */
371 	if (!list_is_empty(&to_be_dropped))
372 		rdsv3_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
373 
374 	if (conn->c_trans->xmit_complete)
375 		conn->c_trans->xmit_complete(conn);
376 
377 	/*
378 	 * We might be racing with another sender who queued a message but
379 	 * backed off on noticing that we held the c_send_lock.  If we check
380 	 * for queued messages after dropping the sem then either we'll
381 	 * see the queued message or the queuer will get the sem.  If we
382 	 * notice the queued message then we trigger an immediate retry.
383 	 *
384 	 * We need to be careful only to do this when we stopped processing
385 	 * the send queue because it was empty.  It's the only way we
386 	 * stop processing the loop when the transport hasn't taken
387 	 * responsibility for forward progress.
388 	 */
389 	mutex_exit(&conn->c_send_lock);
390 
391 	if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) {
392 		/*
393 		 * We exhausted the send quota, but there's work left to
394 		 * do. Return and (re-)schedule the send worker.
395 		 */
396 		ret = -EAGAIN;
397 	}
398 
399 	atomic_dec_32(&conn->c_senders);
400 
401 	if (ret == 0 && was_empty) {
402 		/*
403 		 * A simple bit test would be way faster than taking the
404 		 * spin lock
405 		 */
406 		mutex_enter(&conn->c_lock);
407 		if (!list_is_empty(&conn->c_send_queue)) {
408 			rdsv3_stats_inc(s_send_sem_queue_raced);
409 			ret = -EAGAIN;
410 		}
411 		mutex_exit(&conn->c_lock);
412 	}
413 
414 out:
415 	RDSV3_DPRINTF4("rdsv3_send_xmit", "Return(conn: %p, ret: %d)",
416 	    conn, ret);
417 	return (ret);
418 }
419 
420 static void
421 rdsv3_send_sndbuf_remove(struct rdsv3_sock *rs, struct rdsv3_message *rm)
422 {
423 	uint32_t len = ntohl(rm->m_inc.i_hdr.h_len);
424 
425 	ASSERT(mutex_owned(&rs->rs_lock));
426 
427 	ASSERT(rs->rs_snd_bytes >= len);
428 	rs->rs_snd_bytes -= len;
429 
430 	if (rs->rs_snd_bytes == 0)
431 		rdsv3_stats_inc(s_send_queue_empty);
432 }
433 
434 static inline int
435 rdsv3_send_is_acked(struct rdsv3_message *rm, uint64_t ack,
436     is_acked_func is_acked)
437 {
438 	if (is_acked)
439 		return (is_acked(rm, ack));
440 	return (ntohll(rm->m_inc.i_hdr.h_sequence) <= ack);
441 }
442 
443 /*
444  * Returns true if there are no messages on the send and retransmit queues
445  * which have a sequence number greater than or equal to the given sequence
446  * number.
447  */
448 int
449 rdsv3_send_acked_before(struct rdsv3_connection *conn, uint64_t seq)
450 {
451 	struct rdsv3_message *rm;
452 	int ret = 1;
453 
454 	RDSV3_DPRINTF4("rdsv3_send_acked_before", "Enter(conn: %p)", conn);
455 
456 	mutex_enter(&conn->c_lock);
457 
458 	/* XXX - original code spits out warning */
459 	rm = list_head(&conn->c_retrans);
460 	if (ntohll(rm->m_inc.i_hdr.h_sequence) < seq)
461 		ret = 0;
462 
463 	/* XXX - original code spits out warning */
464 	rm = list_head(&conn->c_send_queue);
465 	if (ntohll(rm->m_inc.i_hdr.h_sequence) < seq)
466 		ret = 0;
467 
468 	mutex_exit(&conn->c_lock);
469 
470 	RDSV3_DPRINTF4("rdsv3_send_acked_before", "Return(conn: %p)", conn);
471 
472 	return (ret);
473 }
474 
475 /*
476  * This is pretty similar to what happens below in the ACK
477  * handling code - except that we call here as soon as we get
478  * the IB send completion on the RDMA op and the accompanying
479  * message.
480  */
481 void
482 rdsv3_rdma_send_complete(struct rdsv3_message *rm, int status)
483 {
484 	struct rdsv3_sock *rs = NULL;
485 	struct rdsv3_rdma_op *ro;
486 	struct rdsv3_notifier *notifier;
487 
488 	RDSV3_DPRINTF4("rdsv3_rdma_send_complete", "Enter(rm: %p)", rm);
489 
490 	mutex_enter(&rm->m_rs_lock);
491 
492 	ro = rm->m_rdma_op;
493 	if (test_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags) &&
494 	    ro && ro->r_notify && ro->r_notifier) {
495 		notifier = ro->r_notifier;
496 		rs = rm->m_rs;
497 		rdsv3_sk_sock_hold(rdsv3_rs_to_sk(rs));
498 
499 		notifier->n_status = status;
500 		mutex_enter(&rs->rs_lock);
501 		list_insert_tail(&rs->rs_notify_queue, notifier);
502 		mutex_exit(&rs->rs_lock);
503 		ro->r_notifier = NULL;
504 	}
505 
506 	mutex_exit(&rm->m_rs_lock);
507 
508 	if (rs) {
509 		struct rsock *sk = rdsv3_rs_to_sk(rs);
510 		int error;
511 
512 		rdsv3_wake_sk_sleep(rs);
513 
514 		/* wake up anyone waiting in poll */
515 		sk->sk_upcalls->su_recv(sk->sk_upper_handle, NULL,
516 		    0, 0, &error, NULL);
517 		if (error != 0) {
518 			RDSV3_DPRINTF2("rdsv3_recv_incoming",
519 			    "su_recv returned: %d", error);
520 		}
521 
522 		rdsv3_sk_sock_put(rdsv3_rs_to_sk(rs));
523 	}
524 
525 	RDSV3_DPRINTF4("rdsv3_rdma_send_complete", "Return(rm: %p)", rm);
526 }
527 
528 /*
529  * This is the same as rdsv3_rdma_send_complete except we
530  * don't do any locking - we have all the ingredients (message,
531  * socket, socket lock) and can just move the notifier.
532  */
533 static inline void
534 __rdsv3_rdma_send_complete(struct rdsv3_sock *rs, struct rdsv3_message *rm,
535     int status)
536 {
537 	struct rdsv3_rdma_op *ro;
538 	void *ic;
539 
540 	RDSV3_DPRINTF4("__rdsv3_rdma_send_complete",
541 	    "Enter(rs: %p, rm: %p)", rs, rm);
542 
543 	ro = rm->m_rdma_op;
544 	if (ro && ro->r_notify && ro->r_notifier) {
545 		ro->r_notifier->n_status = status;
546 		list_insert_tail(&rs->rs_notify_queue, ro->r_notifier);
547 		ro->r_notifier = NULL;
548 	}
549 
550 	/* No need to wake the app - caller does this */
551 }
552 
553 /*
554  * This is called from the IB send completion when we detect
555  * a RDMA operation that failed with remote access error.
556  * So speed is not an issue here.
557  */
558 struct rdsv3_message *
559 rdsv3_send_get_message(struct rdsv3_connection *conn,
560     struct rdsv3_rdma_op *op)
561 {
562 	struct rdsv3_message *rm, *tmp, *found = NULL;
563 
564 	RDSV3_DPRINTF4("rdsv3_send_get_message", "Enter(conn: %p)", conn);
565 
566 	mutex_enter(&conn->c_lock);
567 
568 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_retrans, m_conn_item) {
569 		if (rm->m_rdma_op == op) {
570 			atomic_inc_32(&rm->m_refcount);
571 			found = rm;
572 			goto out;
573 		}
574 	}
575 
576 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_send_queue,
577 	    m_conn_item) {
578 		if (rm->m_rdma_op == op) {
579 			atomic_inc_32(&rm->m_refcount);
580 			found = rm;
581 			break;
582 		}
583 	}
584 
585 out:
586 	mutex_exit(&conn->c_lock);
587 
588 	return (found);
589 }
590 
591 /*
592  * This removes messages from the socket's list if they're on it.  The list
593  * argument must be private to the caller, we must be able to modify it
594  * without locks.  The messages must have a reference held for their
595  * position on the list.  This function will drop that reference after
596  * removing the messages from the 'messages' list regardless of if it found
597  * the messages on the socket list or not.
598  */
599 void
600 rdsv3_send_remove_from_sock(struct list *messages, int status)
601 {
602 	struct rdsv3_sock *rs = NULL;
603 	struct rdsv3_message *rm;
604 
605 	RDSV3_DPRINTF4("rdsv3_send_remove_from_sock", "Enter");
606 
607 	while (!list_is_empty(messages)) {
608 		int was_on_sock = 0;
609 		rm = list_remove_head(messages);
610 
611 		/*
612 		 * If we see this flag cleared then we're *sure* that someone
613 		 * else beat us to removing it from the sock.  If we race
614 		 * with their flag update we'll get the lock and then really
615 		 * see that the flag has been cleared.
616 		 *
617 		 * The message spinlock makes sure nobody clears rm->m_rs
618 		 * while we're messing with it. It does not prevent the
619 		 * message from being removed from the socket, though.
620 		 */
621 		mutex_enter(&rm->m_rs_lock);
622 		if (!test_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags))
623 			goto unlock_and_drop;
624 
625 		if (rs != rm->m_rs) {
626 			if (rs) {
627 				rdsv3_wake_sk_sleep(rs);
628 				rdsv3_sk_sock_put(rdsv3_rs_to_sk(rs));
629 			}
630 			rs = rm->m_rs;
631 			rdsv3_sk_sock_hold(rdsv3_rs_to_sk(rs));
632 		}
633 
634 		mutex_enter(&rs->rs_lock);
635 		if (test_and_clear_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags)) {
636 			struct rdsv3_rdma_op *ro = rm->m_rdma_op;
637 			struct rdsv3_notifier *notifier;
638 
639 			list_remove_node(&rm->m_sock_item);
640 			rdsv3_send_sndbuf_remove(rs, rm);
641 			if (ro && ro->r_notifier &&
642 			    (status || ro->r_notify)) {
643 				notifier = ro->r_notifier;
644 				list_insert_tail(&rs->rs_notify_queue,
645 				    notifier);
646 				if (!notifier->n_status)
647 					notifier->n_status = status;
648 				rm->m_rdma_op->r_notifier = NULL;
649 			}
650 			was_on_sock = 1;
651 			rm->m_rs = NULL;
652 		}
653 		mutex_exit(&rs->rs_lock);
654 
655 unlock_and_drop:
656 		mutex_exit(&rm->m_rs_lock);
657 		rdsv3_message_put(rm);
658 		if (was_on_sock)
659 			rdsv3_message_put(rm);
660 	}
661 
662 	if (rs) {
663 		rdsv3_wake_sk_sleep(rs);
664 		rdsv3_sk_sock_put(rdsv3_rs_to_sk(rs));
665 	}
666 
667 	RDSV3_DPRINTF4("rdsv3_send_remove_from_sock", "Return");
668 }
669 
670 /*
671  * Transports call here when they've determined that the receiver queued
672  * messages up to, and including, the given sequence number.  Messages are
673  * moved to the retrans queue when rdsv3_send_xmit picks them off the send
674  * queue. This means that in the TCP case, the message may not have been
675  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
676  * checks the RDSV3_MSG_HAS_ACK_SEQ bit.
677  *
678  * XXX It's not clear to me how this is safely serialized with socket
679  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
680  */
681 void
682 rdsv3_send_drop_acked(struct rdsv3_connection *conn, uint64_t ack,
683     is_acked_func is_acked)
684 {
685 	struct rdsv3_message *rm, *tmp;
686 	list_t list;
687 
688 	RDSV3_DPRINTF4("rdsv3_send_drop_acked", "Enter(conn: %p)", conn);
689 
690 	list_create(&list, sizeof (struct rdsv3_message),
691 	    offsetof(struct rdsv3_message, m_conn_item));
692 
693 	mutex_enter(&conn->c_lock);
694 
695 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &conn->c_retrans, m_conn_item) {
696 		if (!rdsv3_send_is_acked(rm, ack, is_acked))
697 			break;
698 
699 		list_remove_node(&rm->m_conn_item);
700 		list_insert_tail(&list, rm);
701 		clear_bit(RDSV3_MSG_ON_CONN, &rm->m_flags);
702 	}
703 
704 #if 0
705 XXX
706 	/* order flag updates with spin locks */
707 	if (!list_is_empty(&list))
708 		smp_mb__after_clear_bit();
709 #endif
710 
711 	mutex_exit(&conn->c_lock);
712 
713 	/* now remove the messages from the sock list as needed */
714 	rdsv3_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
715 
716 	RDSV3_DPRINTF4("rdsv3_send_drop_acked", "Return(conn: %p)", conn);
717 }
718 
719 void
720 rdsv3_send_drop_to(struct rdsv3_sock *rs, struct sockaddr_in *dest)
721 {
722 	struct rdsv3_message *rm, *tmp;
723 	struct rdsv3_connection *conn;
724 	list_t list;
725 	int wake = 0;
726 
727 	RDSV3_DPRINTF4("rdsv3_send_drop_to", "Enter(rs: %p)", rs);
728 
729 	list_create(&list, sizeof (struct rdsv3_message),
730 	    offsetof(struct rdsv3_message, m_sock_item));
731 
732 	/* get all the messages we're dropping under the rs lock */
733 	mutex_enter(&rs->rs_lock);
734 
735 	RDSV3_FOR_EACH_LIST_NODE_SAFE(rm, tmp, &rs->rs_send_queue,
736 	    m_sock_item) {
737 		if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
738 		    dest->sin_port != rm->m_inc.i_hdr.h_dport))
739 			continue;
740 		wake = 1;
741 		list_remove(&rs->rs_send_queue, rm);
742 		list_insert_tail(&list, rm);
743 		rdsv3_send_sndbuf_remove(rs, rm);
744 		clear_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags);
745 	}
746 
747 	mutex_exit(&rs->rs_lock);
748 
749 	conn = NULL;
750 
751 	/* now remove the messages from the conn list as needed */
752 	RDSV3_FOR_EACH_LIST_NODE(rm, &list, m_sock_item) {
753 		/*
754 		 * We do this here rather than in the loop above, so that
755 		 * we don't have to nest m_rs_lock under rs->rs_lock
756 		 */
757 		mutex_enter(&rm->m_rs_lock);
758 		/* If this is a RDMA operation, notify the app. */
759 		__rdsv3_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
760 		rm->m_rs = NULL;
761 		mutex_exit(&rm->m_rs_lock);
762 
763 		/*
764 		 * If we see this flag cleared then we're *sure* that someone
765 		 * else beat us to removing it from the conn.  If we race
766 		 * with their flag update we'll get the lock and then really
767 		 * see that the flag has been cleared.
768 		 */
769 		if (!test_bit(RDSV3_MSG_ON_CONN, &rm->m_flags))
770 			continue;
771 
772 		if (conn != rm->m_inc.i_conn) {
773 			if (conn)
774 				mutex_exit(&conn->c_lock);
775 			conn = rm->m_inc.i_conn;
776 			mutex_enter(&conn->c_lock);
777 		}
778 
779 		if (test_and_clear_bit(RDSV3_MSG_ON_CONN, &rm->m_flags)) {
780 			list_remove_node(&rm->m_conn_item);
781 			rdsv3_message_put(rm);
782 		}
783 	}
784 
785 	if (conn)
786 		mutex_exit(&conn->c_lock);
787 
788 	if (wake)
789 		rdsv3_wake_sk_sleep(rs);
790 
791 	while (!list_is_empty(&list)) {
792 		rm = list_remove_head(&list);
793 
794 		rdsv3_message_wait(rm);
795 		rdsv3_message_put(rm);
796 	}
797 
798 	RDSV3_DPRINTF4("rdsv3_send_drop_to", "Return(rs: %p)", rs);
799 }
800 
801 /*
802  * we only want this to fire once so we use the callers 'queued'.  It's
803  * possible that another thread can race with us and remove the
804  * message from the flow with RDSV3_CANCEL_SENT_TO.
805  */
806 static int
807 rdsv3_send_queue_rm(struct rdsv3_sock *rs, struct rdsv3_connection *conn,
808     struct rdsv3_message *rm, uint16_be_t sport,
809     uint16_be_t dport, int *queued)
810 {
811 	uint32_t len;
812 
813 	RDSV3_DPRINTF4("rdsv3_send_queue_rm", "Enter(rs: %p, rm: %p)", rs, rm);
814 
815 	if (*queued)
816 		goto out;
817 
818 	len = ntohl(rm->m_inc.i_hdr.h_len);
819 
820 	/*
821 	 * this is the only place which holds both the socket's rs_lock
822 	 * and the connection's c_lock
823 	 */
824 	mutex_enter(&rs->rs_lock);
825 
826 	/*
827 	 * If there is a little space in sndbuf, we don't queue anything,
828 	 * and userspace gets -EAGAIN. But poll() indicates there's send
829 	 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
830 	 * freed up by incoming acks. So we check the *old* value of
831 	 * rs_snd_bytes here to allow the last msg to exceed the buffer,
832 	 * and poll() now knows no more data can be sent.
833 	 */
834 	if (rs->rs_snd_bytes < rdsv3_sk_sndbuf(rs)) {
835 		rs->rs_snd_bytes += len;
836 
837 		/*
838 		 * let recv side know we are close to send space exhaustion.
839 		 * This is probably not the optimal way to do it, as this
840 		 * means we set the flag on *all* messages as soon as our
841 		 * throughput hits a certain threshold.
842 		 */
843 		if (rs->rs_snd_bytes >= rdsv3_sk_sndbuf(rs) / 2)
844 			set_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags);
845 
846 		list_insert_tail(&rs->rs_send_queue, rm);
847 		set_bit(RDSV3_MSG_ON_SOCK, &rm->m_flags);
848 
849 		rdsv3_message_addref(rm);
850 		rm->m_rs = rs;
851 
852 		/*
853 		 * The code ordering is a little weird, but we're
854 		 * trying to minimize the time we hold c_lock
855 		 */
856 		rdsv3_message_populate_header(&rm->m_inc.i_hdr, sport,
857 		    dport, 0);
858 		rm->m_inc.i_conn = conn;
859 		rdsv3_message_addref(rm);	/* XXX - called twice */
860 
861 		mutex_enter(&conn->c_lock);
862 		rm->m_inc.i_hdr.h_sequence = htonll(conn->c_next_tx_seq++);
863 		list_insert_tail(&conn->c_send_queue, rm);
864 		set_bit(RDSV3_MSG_ON_CONN, &rm->m_flags);
865 		mutex_exit(&conn->c_lock);
866 
867 		RDSV3_DPRINTF5("rdsv3_send_queue_rm",
868 		    "queued msg %p len %d, rs %p bytes %d seq %llu",
869 		    rm, len, rs, rs->rs_snd_bytes,
870 		    (unsigned long long)ntohll(
871 		    rm->m_inc.i_hdr.h_sequence));
872 
873 		*queued = 1;
874 	}
875 
876 	mutex_exit(&rs->rs_lock);
877 
878 	RDSV3_DPRINTF4("rdsv3_send_queue_rm", "Return(rs: %p)", rs);
879 out:
880 	return (*queued);
881 }
882 
883 static int
884 rdsv3_cmsg_send(struct rdsv3_sock *rs, struct rdsv3_message *rm,
885     struct msghdr *msg, int *allocated_mr)
886 {
887 	struct cmsghdr *cmsg;
888 	int ret = 0;
889 
890 	RDSV3_DPRINTF4("rdsv3_cmsg_send", "Enter(rs: %p)", rs);
891 
892 	for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
893 
894 		if (cmsg->cmsg_level != SOL_RDS)
895 			continue;
896 
897 		RDSV3_DPRINTF4("rdsv3_cmsg_send", "cmsg(%p, %p) type %d",
898 		    cmsg, rm, cmsg->cmsg_type);
899 		/*
900 		 * As a side effect, RDMA_DEST and RDMA_MAP will set
901 		 * rm->m_rdma_cookie and rm->m_rdma_mr.
902 		 */
903 		switch (cmsg->cmsg_type) {
904 		case RDS_CMSG_RDMA_ARGS:
905 			ret = rdsv3_cmsg_rdma_args(rs, rm, cmsg);
906 			break;
907 
908 		case RDS_CMSG_RDMA_DEST:
909 			ret = rdsv3_cmsg_rdma_dest(rs, rm, cmsg);
910 			break;
911 
912 		case RDS_CMSG_RDMA_MAP:
913 			ret = rdsv3_cmsg_rdma_map(rs, rm, cmsg);
914 			if (ret)
915 				*allocated_mr = 1;
916 			break;
917 
918 		default:
919 			return (-EINVAL);
920 		}
921 
922 		if (ret)
923 			break;
924 	}
925 
926 	RDSV3_DPRINTF4("rdsv3_cmsg_send", "Return(rs: %p)", rs);
927 
928 	return (ret);
929 }
930 
931 extern unsigned long rdsv3_max_bcopy_size;
932 
933 int
934 rdsv3_sendmsg(struct rdsv3_sock *rs, uio_t *uio, struct nmsghdr *msg,
935     size_t payload_len)
936 {
937 	struct rsock *sk = rdsv3_rs_to_sk(rs);
938 	struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
939 	uint32_be_t daddr;
940 	uint16_be_t dport;
941 	struct rdsv3_message *rm = NULL;
942 	struct rdsv3_connection *conn;
943 	int ret = 0;
944 	int queued = 0, allocated_mr = 0;
945 	int nonblock = msg->msg_flags & MSG_DONTWAIT;
946 	long timeo = rdsv3_sndtimeo(sk, nonblock);
947 
948 	RDSV3_DPRINTF4("rdsv3_sendmsg", "Enter(rs: %p)", rs);
949 
950 	if (msg->msg_namelen) {
951 		/* XXX fail non-unicast destination IPs? */
952 		if (msg->msg_namelen < sizeof (*usin) ||
953 		    usin->sin_family != AF_INET_OFFLOAD) {
954 			ret = -EINVAL;
955 			RDSV3_DPRINTF2("rdsv3_sendmsg", "returning: %d", -ret);
956 			goto out;
957 		}
958 		daddr = usin->sin_addr.s_addr;
959 		dport = usin->sin_port;
960 	} else {
961 		/* We only care about consistency with ->connect() */
962 		mutex_enter(&sk->sk_lock);
963 		daddr = rs->rs_conn_addr;
964 		dport = rs->rs_conn_port;
965 		mutex_exit(&sk->sk_lock);
966 	}
967 
968 	/* racing with another thread binding seems ok here */
969 	if (daddr == 0 || rs->rs_bound_addr == 0) {
970 		ret = -ENOTCONN; /* XXX not a great errno */
971 		RDSV3_DPRINTF2("rdsv3_sendmsg", "returning: %d", -ret);
972 		goto out;
973 	}
974 
975 	if (payload_len > rdsv3_max_bcopy_size) {
976 		RDSV3_DPRINTF2("rdsv3_sendmsg", "Message too large: %d",
977 		    payload_len);
978 		ret = -EMSGSIZE;
979 		goto out;
980 	}
981 
982 	rm = rdsv3_message_copy_from_user(uio, payload_len);
983 	if (IS_ERR(rm)) {
984 		ret = PTR_ERR(rm);
985 		RDSV3_DPRINTF2("rdsv3_sendmsg",
986 		    "rdsv3_message_copy_from_user failed %d", -ret);
987 		rm = NULL;
988 		goto out;
989 	}
990 
991 	rm->m_daddr = daddr;
992 
993 	/* Parse any control messages the user may have included. */
994 	ret = rdsv3_cmsg_send(rs, rm, msg, &allocated_mr);
995 	if (ret) {
996 		RDSV3_DPRINTF2("rdsv3_sendmsg",
997 		    "rdsv3_cmsg_send(rs: %p rm: %p msg: %p) returned: %d",
998 		    rs, rm, msg, ret);
999 		goto out;
1000 	}
1001 
1002 	/*
1003 	 * rdsv3_conn_create has a spinlock that runs with IRQ off.
1004 	 * Caching the conn in the socket helps a lot.
1005 	 */
1006 	mutex_enter(&rs->rs_conn_lock);
1007 	if (rs->rs_conn && rs->rs_conn->c_faddr == daddr) {
1008 		conn = rs->rs_conn;
1009 	} else {
1010 		conn = rdsv3_conn_create_outgoing(rs->rs_bound_addr,
1011 		    daddr, rs->rs_transport, KM_NOSLEEP);
1012 		if (IS_ERR(conn)) {
1013 			mutex_exit(&rs->rs_conn_lock);
1014 			ret = PTR_ERR(conn);
1015 			RDSV3_DPRINTF2("rdsv3_sendmsg",
1016 			    "rdsv3_conn_create_outgoing failed %d",
1017 			    -ret);
1018 			goto out;
1019 		}
1020 		rs->rs_conn = conn;
1021 	}
1022 	mutex_exit(&rs->rs_conn_lock);
1023 
1024 	if ((rm->m_rdma_cookie || rm->m_rdma_op) &&
1025 	    conn->c_trans->xmit_rdma == NULL) {
1026 		RDSV3_DPRINTF2("rdsv3_sendmsg", "rdma_op %p conn xmit_rdma %p",
1027 		    rm->m_rdma_op, conn->c_trans->xmit_rdma);
1028 		ret = -EOPNOTSUPP;
1029 		goto out;
1030 	}
1031 
1032 	/*
1033 	 * If the connection is down, trigger a connect. We may
1034 	 * have scheduled a delayed reconnect however - in this case
1035 	 * we should not interfere.
1036 	 */
1037 	if (rdsv3_conn_state(conn) == RDSV3_CONN_DOWN &&
1038 	    !test_and_set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags))
1039 		rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
1040 
1041 	ret = rdsv3_cong_wait(conn->c_fcong, dport, nonblock, rs);
1042 	if (ret) {
1043 		mutex_enter(&rs->rs_congested_lock);
1044 		rs->rs_seen_congestion = 1;
1045 		cv_signal(&rs->rs_congested_cv);
1046 		mutex_exit(&rs->rs_congested_lock);
1047 
1048 		RDSV3_DPRINTF2("rdsv3_sendmsg",
1049 		    "rdsv3_cong_wait (dport: %d) returned: %d", dport, ret);
1050 		goto out;
1051 	}
1052 
1053 	(void) rdsv3_send_queue_rm(rs, conn, rm, rs->rs_bound_port, dport,
1054 	    &queued);
1055 	if (!queued) {
1056 		/* rdsv3_stats_inc(s_send_queue_full); */
1057 		/* XXX make sure this is reasonable */
1058 		if (payload_len > rdsv3_sk_sndbuf(rs)) {
1059 			ret = -EMSGSIZE;
1060 			RDSV3_DPRINTF2("rdsv3_sendmsg",
1061 			    "msgsize(%d) too big, returning: %d",
1062 			    payload_len, -ret);
1063 			goto out;
1064 		}
1065 		if (nonblock) {
1066 			ret = -EAGAIN;
1067 			RDSV3_DPRINTF3("rdsv3_sendmsg",
1068 			    "send queue full (%d), returning: %d",
1069 			    payload_len, -ret);
1070 			goto out;
1071 		}
1072 
1073 #if 0
1074 		ret = rdsv3_wait_sig(sk->sk_sleep,
1075 		    (rdsv3_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1076 		    dport, &queued)));
1077 		if (ret == 0) {
1078 			/* signal/timeout pending */
1079 			RDSV3_DPRINTF2("rdsv3_sendmsg",
1080 			    "woke due to signal: %d", ret);
1081 			ret = -ERESTART;
1082 			goto out;
1083 		}
1084 #else
1085 		mutex_enter(&sk->sk_sleep->waitq_mutex);
1086 		sk->sk_sleep->waitq_waiters++;
1087 		while (!rdsv3_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1088 		    dport, &queued)) {
1089 			ret = cv_wait_sig(&sk->sk_sleep->waitq_cv,
1090 			    &sk->sk_sleep->waitq_mutex);
1091 			if (ret == 0) {
1092 				/* signal/timeout pending */
1093 				RDSV3_DPRINTF2("rdsv3_sendmsg",
1094 				    "woke due to signal: %d", ret);
1095 				ret = -EINTR;
1096 				sk->sk_sleep->waitq_waiters--;
1097 				mutex_exit(&sk->sk_sleep->waitq_mutex);
1098 				goto out;
1099 			}
1100 		}
1101 		sk->sk_sleep->waitq_waiters--;
1102 		mutex_exit(&sk->sk_sleep->waitq_mutex);
1103 #endif
1104 
1105 		RDSV3_DPRINTF5("rdsv3_sendmsg", "sendmsg woke queued %d",
1106 		    queued);
1107 
1108 		ASSERT(queued);
1109 		ret = 0;
1110 	}
1111 
1112 	/*
1113 	 * By now we've committed to the send.  We reuse rdsv3_send_worker()
1114 	 * to retry sends in the rds thread if the transport asks us to.
1115 	 */
1116 	rdsv3_stats_inc(s_send_queued);
1117 
1118 	if (!test_bit(RDSV3_LL_SEND_FULL, &conn->c_flags))
1119 		(void) rdsv3_send_worker(&conn->c_send_w.work);
1120 
1121 	rdsv3_message_put(rm);
1122 	RDSV3_DPRINTF4("rdsv3_sendmsg", "Return(rs: %p, len: %d)",
1123 	    rs, payload_len);
1124 	return (payload_len);
1125 
1126 out:
1127 	/*
1128 	 * If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1129 	 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1130 	 * or in any other way, we need to destroy the MR again
1131 	 */
1132 	if (allocated_mr)
1133 		rdsv3_rdma_unuse(rs, rdsv3_rdma_cookie_key(rm->m_rdma_cookie),
1134 		    1);
1135 
1136 	if (rm)
1137 		rdsv3_message_put(rm);
1138 	return (ret);
1139 }
1140 
1141 /*
1142  * Reply to a ping packet.
1143  */
1144 int
1145 rdsv3_send_pong(struct rdsv3_connection *conn, uint16_be_t dport)
1146 {
1147 	struct rdsv3_message *rm;
1148 	int ret = 0;
1149 
1150 	RDSV3_DPRINTF4("rdsv3_send_pong", "Enter(conn: %p)", conn);
1151 
1152 	rm = rdsv3_message_alloc(0, KM_NOSLEEP);
1153 	if (!rm) {
1154 		ret = -ENOMEM;
1155 		goto out;
1156 	}
1157 
1158 	rm->m_daddr = conn->c_faddr;
1159 
1160 	/*
1161 	 * If the connection is down, trigger a connect. We may
1162 	 * have scheduled a delayed reconnect however - in this case
1163 	 * we should not interfere.
1164 	 */
1165 	if (rdsv3_conn_state(conn) == RDSV3_CONN_DOWN &&
1166 	    !test_and_set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags))
1167 		rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
1168 
1169 	ret = rdsv3_cong_wait(conn->c_fcong, dport, 1, NULL);
1170 	if (ret)
1171 		goto out;
1172 
1173 	mutex_enter(&conn->c_lock);
1174 	list_insert_tail(&conn->c_send_queue, rm);
1175 	set_bit(RDSV3_MSG_ON_CONN, &rm->m_flags);
1176 	rdsv3_message_addref(rm);
1177 	rm->m_inc.i_conn = conn;
1178 
1179 	rdsv3_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1180 	    conn->c_next_tx_seq);
1181 	conn->c_next_tx_seq++;
1182 	mutex_exit(&conn->c_lock);
1183 
1184 	rdsv3_stats_inc(s_send_queued);
1185 	rdsv3_stats_inc(s_send_pong);
1186 
1187 	if (!test_bit(RDSV3_LL_SEND_FULL, &conn->c_flags))
1188 		(void) rdsv3_send_xmit(conn);
1189 
1190 	rdsv3_message_put(rm);
1191 
1192 	RDSV3_DPRINTF4("rdsv3_send_pong", "Return(conn: %p)", conn);
1193 	return (0);
1194 
1195 out:
1196 	if (rm)
1197 		rdsv3_message_put(rm);
1198 	return (ret);
1199 }
1200