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
rdsv3_send_reset(struct rdsv3_connection * conn)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
rdsv3_send_xmit(struct rdsv3_connection * conn)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
rdsv3_send_sndbuf_remove(struct rdsv3_sock * rs,struct rdsv3_message * rm)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
rdsv3_send_is_acked(struct rdsv3_message * rm,uint64_t ack,is_acked_func is_acked)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
rdsv3_send_acked_before(struct rdsv3_connection * conn,uint64_t seq)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
rdsv3_rdma_send_complete(struct rdsv3_message * rm,int status)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
__rdsv3_rdma_send_complete(struct rdsv3_sock * rs,struct rdsv3_message * rm,int status)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 *
rdsv3_send_get_message(struct rdsv3_connection * conn,struct rdsv3_rdma_op * op)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
rdsv3_send_remove_from_sock(struct list * messages,int status)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
rdsv3_send_drop_acked(struct rdsv3_connection * conn,uint64_t ack,is_acked_func is_acked)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
rdsv3_send_drop_to(struct rdsv3_sock * rs,struct sockaddr_in * dest)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
rdsv3_send_queue_rm(struct rdsv3_sock * rs,struct rdsv3_connection * conn,struct rdsv3_message * rm,uint16_be_t sport,uint16_be_t dport,int * queued)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
rdsv3_cmsg_send(struct rdsv3_sock * rs,struct rdsv3_message * rm,struct msghdr * msg,int * allocated_mr)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
rdsv3_sendmsg(struct rdsv3_sock * rs,uio_t * uio,struct nmsghdr * msg,size_t payload_len)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
rdsv3_send_pong(struct rdsv3_connection * conn,uint16_be_t dport)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