Loading...
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/moduleparam.h>
35#include <linux/gfp.h>
36#include <net/sock.h>
37#include <linux/in.h>
38#include <linux/list.h>
39#include <linux/ratelimit.h>
40#include <linux/export.h>
41#include <linux/sizes.h>
42
43#include "rds.h"
44
45/* When transmitting messages in rds_send_xmit, we need to emerge from
46 * time to time and briefly release the CPU. Otherwise the softlock watchdog
47 * will kick our shin.
48 * Also, it seems fairer to not let one busy connection stall all the
49 * others.
50 *
51 * send_batch_count is the number of times we'll loop in send_xmit. Setting
52 * it to 0 will restore the old behavior (where we looped until we had
53 * drained the queue).
54 */
55static int send_batch_count = SZ_1K;
56module_param(send_batch_count, int, 0444);
57MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
58
59static void rds_send_remove_from_sock(struct list_head *messages, int status);
60
61/*
62 * Reset the send state. Callers must ensure that this doesn't race with
63 * rds_send_xmit().
64 */
65void rds_send_reset(struct rds_connection *conn)
66{
67 struct rds_message *rm, *tmp;
68 unsigned long flags;
69
70 if (conn->c_xmit_rm) {
71 rm = conn->c_xmit_rm;
72 conn->c_xmit_rm = NULL;
73 /* Tell the user the RDMA op is no longer mapped by the
74 * transport. This isn't entirely true (it's flushed out
75 * independently) but as the connection is down, there's
76 * no ongoing RDMA to/from that memory */
77 rds_message_unmapped(rm);
78 rds_message_put(rm);
79 }
80
81 conn->c_xmit_sg = 0;
82 conn->c_xmit_hdr_off = 0;
83 conn->c_xmit_data_off = 0;
84 conn->c_xmit_atomic_sent = 0;
85 conn->c_xmit_rdma_sent = 0;
86 conn->c_xmit_data_sent = 0;
87
88 conn->c_map_queued = 0;
89
90 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
91 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
92
93 /* Mark messages as retransmissions, and move them to the send q */
94 spin_lock_irqsave(&conn->c_lock, flags);
95 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
96 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
97 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
98 }
99 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
100 spin_unlock_irqrestore(&conn->c_lock, flags);
101}
102
103static int acquire_in_xmit(struct rds_connection *conn)
104{
105 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
106}
107
108static void release_in_xmit(struct rds_connection *conn)
109{
110 clear_bit(RDS_IN_XMIT, &conn->c_flags);
111 smp_mb__after_atomic();
112 /*
113 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
114 * hot path and finding waiters is very rare. We don't want to walk
115 * the system-wide hashed waitqueue buckets in the fast path only to
116 * almost never find waiters.
117 */
118 if (waitqueue_active(&conn->c_waitq))
119 wake_up_all(&conn->c_waitq);
120}
121
122/*
123 * We're making the conscious trade-off here to only send one message
124 * down the connection at a time.
125 * Pro:
126 * - tx queueing is a simple fifo list
127 * - reassembly is optional and easily done by transports per conn
128 * - no per flow rx lookup at all, straight to the socket
129 * - less per-frag memory and wire overhead
130 * Con:
131 * - queued acks can be delayed behind large messages
132 * Depends:
133 * - small message latency is higher behind queued large messages
134 * - large message latency isn't starved by intervening small sends
135 */
136int rds_send_xmit(struct rds_connection *conn)
137{
138 struct rds_message *rm;
139 unsigned long flags;
140 unsigned int tmp;
141 struct scatterlist *sg;
142 int ret = 0;
143 LIST_HEAD(to_be_dropped);
144 int batch_count;
145 unsigned long send_gen = 0;
146
147restart:
148 batch_count = 0;
149
150 /*
151 * sendmsg calls here after having queued its message on the send
152 * queue. We only have one task feeding the connection at a time. If
153 * another thread is already feeding the queue then we back off. This
154 * avoids blocking the caller and trading per-connection data between
155 * caches per message.
156 */
157 if (!acquire_in_xmit(conn)) {
158 rds_stats_inc(s_send_lock_contention);
159 ret = -ENOMEM;
160 goto out;
161 }
162
163 /*
164 * we record the send generation after doing the xmit acquire.
165 * if someone else manages to jump in and do some work, we'll use
166 * this to avoid a goto restart farther down.
167 *
168 * The acquire_in_xmit() check above ensures that only one
169 * caller can increment c_send_gen at any time.
170 */
171 conn->c_send_gen++;
172 send_gen = conn->c_send_gen;
173
174 /*
175 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
176 * we do the opposite to avoid races.
177 */
178 if (!rds_conn_up(conn)) {
179 release_in_xmit(conn);
180 ret = 0;
181 goto out;
182 }
183
184 if (conn->c_trans->xmit_prepare)
185 conn->c_trans->xmit_prepare(conn);
186
187 /*
188 * spin trying to push headers and data down the connection until
189 * the connection doesn't make forward progress.
190 */
191 while (1) {
192
193 rm = conn->c_xmit_rm;
194
195 /*
196 * If between sending messages, we can send a pending congestion
197 * map update.
198 */
199 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
200 rm = rds_cong_update_alloc(conn);
201 if (IS_ERR(rm)) {
202 ret = PTR_ERR(rm);
203 break;
204 }
205 rm->data.op_active = 1;
206
207 conn->c_xmit_rm = rm;
208 }
209
210 /*
211 * If not already working on one, grab the next message.
212 *
213 * c_xmit_rm holds a ref while we're sending this message down
214 * the connction. We can use this ref while holding the
215 * send_sem.. rds_send_reset() is serialized with it.
216 */
217 if (!rm) {
218 unsigned int len;
219
220 batch_count++;
221
222 /* we want to process as big a batch as we can, but
223 * we also want to avoid softlockups. If we've been
224 * through a lot of messages, lets back off and see
225 * if anyone else jumps in
226 */
227 if (batch_count >= send_batch_count)
228 goto over_batch;
229
230 spin_lock_irqsave(&conn->c_lock, flags);
231
232 if (!list_empty(&conn->c_send_queue)) {
233 rm = list_entry(conn->c_send_queue.next,
234 struct rds_message,
235 m_conn_item);
236 rds_message_addref(rm);
237
238 /*
239 * Move the message from the send queue to the retransmit
240 * list right away.
241 */
242 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
243 }
244
245 spin_unlock_irqrestore(&conn->c_lock, flags);
246
247 if (!rm)
248 break;
249
250 /* Unfortunately, the way Infiniband deals with
251 * RDMA to a bad MR key is by moving the entire
252 * queue pair to error state. We cold possibly
253 * recover from that, but right now we drop the
254 * connection.
255 * Therefore, we never retransmit messages with RDMA ops.
256 */
257 if (rm->rdma.op_active &&
258 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
259 spin_lock_irqsave(&conn->c_lock, flags);
260 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
261 list_move(&rm->m_conn_item, &to_be_dropped);
262 spin_unlock_irqrestore(&conn->c_lock, flags);
263 continue;
264 }
265
266 /* Require an ACK every once in a while */
267 len = ntohl(rm->m_inc.i_hdr.h_len);
268 if (conn->c_unacked_packets == 0 ||
269 conn->c_unacked_bytes < len) {
270 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
271
272 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
273 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
274 rds_stats_inc(s_send_ack_required);
275 } else {
276 conn->c_unacked_bytes -= len;
277 conn->c_unacked_packets--;
278 }
279
280 conn->c_xmit_rm = rm;
281 }
282
283 /* The transport either sends the whole rdma or none of it */
284 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
285 rm->m_final_op = &rm->rdma;
286 /* The transport owns the mapped memory for now.
287 * You can't unmap it while it's on the send queue
288 */
289 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
290 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
291 if (ret) {
292 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
293 wake_up_interruptible(&rm->m_flush_wait);
294 break;
295 }
296 conn->c_xmit_rdma_sent = 1;
297
298 }
299
300 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
301 rm->m_final_op = &rm->atomic;
302 /* The transport owns the mapped memory for now.
303 * You can't unmap it while it's on the send queue
304 */
305 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
306 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
307 if (ret) {
308 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
309 wake_up_interruptible(&rm->m_flush_wait);
310 break;
311 }
312 conn->c_xmit_atomic_sent = 1;
313
314 }
315
316 /*
317 * A number of cases require an RDS header to be sent
318 * even if there is no data.
319 * We permit 0-byte sends; rds-ping depends on this.
320 * However, if there are exclusively attached silent ops,
321 * we skip the hdr/data send, to enable silent operation.
322 */
323 if (rm->data.op_nents == 0) {
324 int ops_present;
325 int all_ops_are_silent = 1;
326
327 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
328 if (rm->atomic.op_active && !rm->atomic.op_silent)
329 all_ops_are_silent = 0;
330 if (rm->rdma.op_active && !rm->rdma.op_silent)
331 all_ops_are_silent = 0;
332
333 if (ops_present && all_ops_are_silent
334 && !rm->m_rdma_cookie)
335 rm->data.op_active = 0;
336 }
337
338 if (rm->data.op_active && !conn->c_xmit_data_sent) {
339 rm->m_final_op = &rm->data;
340 ret = conn->c_trans->xmit(conn, rm,
341 conn->c_xmit_hdr_off,
342 conn->c_xmit_sg,
343 conn->c_xmit_data_off);
344 if (ret <= 0)
345 break;
346
347 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
348 tmp = min_t(int, ret,
349 sizeof(struct rds_header) -
350 conn->c_xmit_hdr_off);
351 conn->c_xmit_hdr_off += tmp;
352 ret -= tmp;
353 }
354
355 sg = &rm->data.op_sg[conn->c_xmit_sg];
356 while (ret) {
357 tmp = min_t(int, ret, sg->length -
358 conn->c_xmit_data_off);
359 conn->c_xmit_data_off += tmp;
360 ret -= tmp;
361 if (conn->c_xmit_data_off == sg->length) {
362 conn->c_xmit_data_off = 0;
363 sg++;
364 conn->c_xmit_sg++;
365 BUG_ON(ret != 0 &&
366 conn->c_xmit_sg == rm->data.op_nents);
367 }
368 }
369
370 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
371 (conn->c_xmit_sg == rm->data.op_nents))
372 conn->c_xmit_data_sent = 1;
373 }
374
375 /*
376 * A rm will only take multiple times through this loop
377 * if there is a data op. Thus, if the data is sent (or there was
378 * none), then we're done with the rm.
379 */
380 if (!rm->data.op_active || conn->c_xmit_data_sent) {
381 conn->c_xmit_rm = NULL;
382 conn->c_xmit_sg = 0;
383 conn->c_xmit_hdr_off = 0;
384 conn->c_xmit_data_off = 0;
385 conn->c_xmit_rdma_sent = 0;
386 conn->c_xmit_atomic_sent = 0;
387 conn->c_xmit_data_sent = 0;
388
389 rds_message_put(rm);
390 }
391 }
392
393over_batch:
394 if (conn->c_trans->xmit_complete)
395 conn->c_trans->xmit_complete(conn);
396 release_in_xmit(conn);
397
398 /* Nuke any messages we decided not to retransmit. */
399 if (!list_empty(&to_be_dropped)) {
400 /* irqs on here, so we can put(), unlike above */
401 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
402 rds_message_put(rm);
403 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
404 }
405
406 /*
407 * Other senders can queue a message after we last test the send queue
408 * but before we clear RDS_IN_XMIT. In that case they'd back off and
409 * not try and send their newly queued message. We need to check the
410 * send queue after having cleared RDS_IN_XMIT so that their message
411 * doesn't get stuck on the send queue.
412 *
413 * If the transport cannot continue (i.e ret != 0), then it must
414 * call us when more room is available, such as from the tx
415 * completion handler.
416 *
417 * We have an extra generation check here so that if someone manages
418 * to jump in after our release_in_xmit, we'll see that they have done
419 * some work and we will skip our goto
420 */
421 if (ret == 0) {
422 smp_mb();
423 if ((test_bit(0, &conn->c_map_queued) ||
424 !list_empty(&conn->c_send_queue)) &&
425 send_gen == conn->c_send_gen) {
426 rds_stats_inc(s_send_lock_queue_raced);
427 if (batch_count < send_batch_count)
428 goto restart;
429 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
430 }
431 }
432out:
433 return ret;
434}
435EXPORT_SYMBOL_GPL(rds_send_xmit);
436
437static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
438{
439 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
440
441 assert_spin_locked(&rs->rs_lock);
442
443 BUG_ON(rs->rs_snd_bytes < len);
444 rs->rs_snd_bytes -= len;
445
446 if (rs->rs_snd_bytes == 0)
447 rds_stats_inc(s_send_queue_empty);
448}
449
450static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
451 is_acked_func is_acked)
452{
453 if (is_acked)
454 return is_acked(rm, ack);
455 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
456}
457
458/*
459 * This is pretty similar to what happens below in the ACK
460 * handling code - except that we call here as soon as we get
461 * the IB send completion on the RDMA op and the accompanying
462 * message.
463 */
464void rds_rdma_send_complete(struct rds_message *rm, int status)
465{
466 struct rds_sock *rs = NULL;
467 struct rm_rdma_op *ro;
468 struct rds_notifier *notifier;
469 unsigned long flags;
470
471 spin_lock_irqsave(&rm->m_rs_lock, flags);
472
473 ro = &rm->rdma;
474 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
475 ro->op_active && ro->op_notify && ro->op_notifier) {
476 notifier = ro->op_notifier;
477 rs = rm->m_rs;
478 sock_hold(rds_rs_to_sk(rs));
479
480 notifier->n_status = status;
481 spin_lock(&rs->rs_lock);
482 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
483 spin_unlock(&rs->rs_lock);
484
485 ro->op_notifier = NULL;
486 }
487
488 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
489
490 if (rs) {
491 rds_wake_sk_sleep(rs);
492 sock_put(rds_rs_to_sk(rs));
493 }
494}
495EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
496
497/*
498 * Just like above, except looks at atomic op
499 */
500void rds_atomic_send_complete(struct rds_message *rm, int status)
501{
502 struct rds_sock *rs = NULL;
503 struct rm_atomic_op *ao;
504 struct rds_notifier *notifier;
505 unsigned long flags;
506
507 spin_lock_irqsave(&rm->m_rs_lock, flags);
508
509 ao = &rm->atomic;
510 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
511 && ao->op_active && ao->op_notify && ao->op_notifier) {
512 notifier = ao->op_notifier;
513 rs = rm->m_rs;
514 sock_hold(rds_rs_to_sk(rs));
515
516 notifier->n_status = status;
517 spin_lock(&rs->rs_lock);
518 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
519 spin_unlock(&rs->rs_lock);
520
521 ao->op_notifier = NULL;
522 }
523
524 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
525
526 if (rs) {
527 rds_wake_sk_sleep(rs);
528 sock_put(rds_rs_to_sk(rs));
529 }
530}
531EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
532
533/*
534 * This is the same as rds_rdma_send_complete except we
535 * don't do any locking - we have all the ingredients (message,
536 * socket, socket lock) and can just move the notifier.
537 */
538static inline void
539__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
540{
541 struct rm_rdma_op *ro;
542 struct rm_atomic_op *ao;
543
544 ro = &rm->rdma;
545 if (ro->op_active && ro->op_notify && ro->op_notifier) {
546 ro->op_notifier->n_status = status;
547 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
548 ro->op_notifier = NULL;
549 }
550
551 ao = &rm->atomic;
552 if (ao->op_active && ao->op_notify && ao->op_notifier) {
553 ao->op_notifier->n_status = status;
554 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
555 ao->op_notifier = NULL;
556 }
557
558 /* No need to wake the app - caller does this */
559}
560
561/*
562 * This is called from the IB send completion when we detect
563 * a RDMA operation that failed with remote access error.
564 * So speed is not an issue here.
565 */
566struct rds_message *rds_send_get_message(struct rds_connection *conn,
567 struct rm_rdma_op *op)
568{
569 struct rds_message *rm, *tmp, *found = NULL;
570 unsigned long flags;
571
572 spin_lock_irqsave(&conn->c_lock, flags);
573
574 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
575 if (&rm->rdma == op) {
576 atomic_inc(&rm->m_refcount);
577 found = rm;
578 goto out;
579 }
580 }
581
582 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
583 if (&rm->rdma == op) {
584 atomic_inc(&rm->m_refcount);
585 found = rm;
586 break;
587 }
588 }
589
590out:
591 spin_unlock_irqrestore(&conn->c_lock, flags);
592
593 return found;
594}
595EXPORT_SYMBOL_GPL(rds_send_get_message);
596
597/*
598 * This removes messages from the socket's list if they're on it. The list
599 * argument must be private to the caller, we must be able to modify it
600 * without locks. The messages must have a reference held for their
601 * position on the list. This function will drop that reference after
602 * removing the messages from the 'messages' list regardless of if it found
603 * the messages on the socket list or not.
604 */
605static void rds_send_remove_from_sock(struct list_head *messages, int status)
606{
607 unsigned long flags;
608 struct rds_sock *rs = NULL;
609 struct rds_message *rm;
610
611 while (!list_empty(messages)) {
612 int was_on_sock = 0;
613
614 rm = list_entry(messages->next, struct rds_message,
615 m_conn_item);
616 list_del_init(&rm->m_conn_item);
617
618 /*
619 * If we see this flag cleared then we're *sure* that someone
620 * else beat us to removing it from the sock. If we race
621 * with their flag update we'll get the lock and then really
622 * see that the flag has been cleared.
623 *
624 * The message spinlock makes sure nobody clears rm->m_rs
625 * while we're messing with it. It does not prevent the
626 * message from being removed from the socket, though.
627 */
628 spin_lock_irqsave(&rm->m_rs_lock, flags);
629 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
630 goto unlock_and_drop;
631
632 if (rs != rm->m_rs) {
633 if (rs) {
634 rds_wake_sk_sleep(rs);
635 sock_put(rds_rs_to_sk(rs));
636 }
637 rs = rm->m_rs;
638 if (rs)
639 sock_hold(rds_rs_to_sk(rs));
640 }
641 if (!rs)
642 goto unlock_and_drop;
643 spin_lock(&rs->rs_lock);
644
645 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
646 struct rm_rdma_op *ro = &rm->rdma;
647 struct rds_notifier *notifier;
648
649 list_del_init(&rm->m_sock_item);
650 rds_send_sndbuf_remove(rs, rm);
651
652 if (ro->op_active && ro->op_notifier &&
653 (ro->op_notify || (ro->op_recverr && status))) {
654 notifier = ro->op_notifier;
655 list_add_tail(¬ifier->n_list,
656 &rs->rs_notify_queue);
657 if (!notifier->n_status)
658 notifier->n_status = status;
659 rm->rdma.op_notifier = NULL;
660 }
661 was_on_sock = 1;
662 rm->m_rs = NULL;
663 }
664 spin_unlock(&rs->rs_lock);
665
666unlock_and_drop:
667 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
668 rds_message_put(rm);
669 if (was_on_sock)
670 rds_message_put(rm);
671 }
672
673 if (rs) {
674 rds_wake_sk_sleep(rs);
675 sock_put(rds_rs_to_sk(rs));
676 }
677}
678
679/*
680 * Transports call here when they've determined that the receiver queued
681 * messages up to, and including, the given sequence number. Messages are
682 * moved to the retrans queue when rds_send_xmit picks them off the send
683 * queue. This means that in the TCP case, the message may not have been
684 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
685 * checks the RDS_MSG_HAS_ACK_SEQ bit.
686 */
687void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
688 is_acked_func is_acked)
689{
690 struct rds_message *rm, *tmp;
691 unsigned long flags;
692 LIST_HEAD(list);
693
694 spin_lock_irqsave(&conn->c_lock, flags);
695
696 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
697 if (!rds_send_is_acked(rm, ack, is_acked))
698 break;
699
700 list_move(&rm->m_conn_item, &list);
701 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
702 }
703
704 /* order flag updates with spin locks */
705 if (!list_empty(&list))
706 smp_mb__after_atomic();
707
708 spin_unlock_irqrestore(&conn->c_lock, flags);
709
710 /* now remove the messages from the sock list as needed */
711 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
712}
713EXPORT_SYMBOL_GPL(rds_send_drop_acked);
714
715void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
716{
717 struct rds_message *rm, *tmp;
718 struct rds_connection *conn;
719 unsigned long flags;
720 LIST_HEAD(list);
721
722 /* get all the messages we're dropping under the rs lock */
723 spin_lock_irqsave(&rs->rs_lock, flags);
724
725 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
726 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
727 dest->sin_port != rm->m_inc.i_hdr.h_dport))
728 continue;
729
730 list_move(&rm->m_sock_item, &list);
731 rds_send_sndbuf_remove(rs, rm);
732 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
733 }
734
735 /* order flag updates with the rs lock */
736 smp_mb__after_atomic();
737
738 spin_unlock_irqrestore(&rs->rs_lock, flags);
739
740 if (list_empty(&list))
741 return;
742
743 /* Remove the messages from the conn */
744 list_for_each_entry(rm, &list, m_sock_item) {
745
746 conn = rm->m_inc.i_conn;
747
748 spin_lock_irqsave(&conn->c_lock, flags);
749 /*
750 * Maybe someone else beat us to removing rm from the conn.
751 * If we race with their flag update we'll get the lock and
752 * then really see that the flag has been cleared.
753 */
754 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
755 spin_unlock_irqrestore(&conn->c_lock, flags);
756 spin_lock_irqsave(&rm->m_rs_lock, flags);
757 rm->m_rs = NULL;
758 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
759 continue;
760 }
761 list_del_init(&rm->m_conn_item);
762 spin_unlock_irqrestore(&conn->c_lock, flags);
763
764 /*
765 * Couldn't grab m_rs_lock in top loop (lock ordering),
766 * but we can now.
767 */
768 spin_lock_irqsave(&rm->m_rs_lock, flags);
769
770 spin_lock(&rs->rs_lock);
771 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
772 spin_unlock(&rs->rs_lock);
773
774 rm->m_rs = NULL;
775 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
776
777 rds_message_put(rm);
778 }
779
780 rds_wake_sk_sleep(rs);
781
782 while (!list_empty(&list)) {
783 rm = list_entry(list.next, struct rds_message, m_sock_item);
784 list_del_init(&rm->m_sock_item);
785 rds_message_wait(rm);
786
787 /* just in case the code above skipped this message
788 * because RDS_MSG_ON_CONN wasn't set, run it again here
789 * taking m_rs_lock is the only thing that keeps us
790 * from racing with ack processing.
791 */
792 spin_lock_irqsave(&rm->m_rs_lock, flags);
793
794 spin_lock(&rs->rs_lock);
795 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
796 spin_unlock(&rs->rs_lock);
797
798 rm->m_rs = NULL;
799 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
800
801 rds_message_put(rm);
802 }
803}
804
805/*
806 * we only want this to fire once so we use the callers 'queued'. It's
807 * possible that another thread can race with us and remove the
808 * message from the flow with RDS_CANCEL_SENT_TO.
809 */
810static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
811 struct rds_message *rm, __be16 sport,
812 __be16 dport, int *queued)
813{
814 unsigned long flags;
815 u32 len;
816
817 if (*queued)
818 goto out;
819
820 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
821
822 /* this is the only place which holds both the socket's rs_lock
823 * and the connection's c_lock */
824 spin_lock_irqsave(&rs->rs_lock, flags);
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 < rds_sk_sndbuf(rs)) {
835 rs->rs_snd_bytes += len;
836
837 /* let recv side know we are close to send space exhaustion.
838 * This is probably not the optimal way to do it, as this
839 * means we set the flag on *all* messages as soon as our
840 * throughput hits a certain threshold.
841 */
842 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
843 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
844
845 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
846 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
847 rds_message_addref(rm);
848 rm->m_rs = rs;
849
850 /* The code ordering is a little weird, but we're
851 trying to minimize the time we hold c_lock */
852 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
853 rm->m_inc.i_conn = conn;
854 rds_message_addref(rm);
855
856 spin_lock(&conn->c_lock);
857 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
858 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
859 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
860 spin_unlock(&conn->c_lock);
861
862 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
863 rm, len, rs, rs->rs_snd_bytes,
864 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
865
866 *queued = 1;
867 }
868
869 spin_unlock_irqrestore(&rs->rs_lock, flags);
870out:
871 return *queued;
872}
873
874/*
875 * rds_message is getting to be quite complicated, and we'd like to allocate
876 * it all in one go. This figures out how big it needs to be up front.
877 */
878static int rds_rm_size(struct msghdr *msg, int data_len)
879{
880 struct cmsghdr *cmsg;
881 int size = 0;
882 int cmsg_groups = 0;
883 int retval;
884
885 for_each_cmsghdr(cmsg, msg) {
886 if (!CMSG_OK(msg, cmsg))
887 return -EINVAL;
888
889 if (cmsg->cmsg_level != SOL_RDS)
890 continue;
891
892 switch (cmsg->cmsg_type) {
893 case RDS_CMSG_RDMA_ARGS:
894 cmsg_groups |= 1;
895 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
896 if (retval < 0)
897 return retval;
898 size += retval;
899
900 break;
901
902 case RDS_CMSG_RDMA_DEST:
903 case RDS_CMSG_RDMA_MAP:
904 cmsg_groups |= 2;
905 /* these are valid but do no add any size */
906 break;
907
908 case RDS_CMSG_ATOMIC_CSWP:
909 case RDS_CMSG_ATOMIC_FADD:
910 case RDS_CMSG_MASKED_ATOMIC_CSWP:
911 case RDS_CMSG_MASKED_ATOMIC_FADD:
912 cmsg_groups |= 1;
913 size += sizeof(struct scatterlist);
914 break;
915
916 default:
917 return -EINVAL;
918 }
919
920 }
921
922 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
923
924 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
925 if (cmsg_groups == 3)
926 return -EINVAL;
927
928 return size;
929}
930
931static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
932 struct msghdr *msg, int *allocated_mr)
933{
934 struct cmsghdr *cmsg;
935 int ret = 0;
936
937 for_each_cmsghdr(cmsg, msg) {
938 if (!CMSG_OK(msg, cmsg))
939 return -EINVAL;
940
941 if (cmsg->cmsg_level != SOL_RDS)
942 continue;
943
944 /* As a side effect, RDMA_DEST and RDMA_MAP will set
945 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
946 */
947 switch (cmsg->cmsg_type) {
948 case RDS_CMSG_RDMA_ARGS:
949 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
950 break;
951
952 case RDS_CMSG_RDMA_DEST:
953 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
954 break;
955
956 case RDS_CMSG_RDMA_MAP:
957 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
958 if (!ret)
959 *allocated_mr = 1;
960 break;
961 case RDS_CMSG_ATOMIC_CSWP:
962 case RDS_CMSG_ATOMIC_FADD:
963 case RDS_CMSG_MASKED_ATOMIC_CSWP:
964 case RDS_CMSG_MASKED_ATOMIC_FADD:
965 ret = rds_cmsg_atomic(rs, rm, cmsg);
966 break;
967
968 default:
969 return -EINVAL;
970 }
971
972 if (ret)
973 break;
974 }
975
976 return ret;
977}
978
979int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
980{
981 struct sock *sk = sock->sk;
982 struct rds_sock *rs = rds_sk_to_rs(sk);
983 DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
984 __be32 daddr;
985 __be16 dport;
986 struct rds_message *rm = NULL;
987 struct rds_connection *conn;
988 int ret = 0;
989 int queued = 0, allocated_mr = 0;
990 int nonblock = msg->msg_flags & MSG_DONTWAIT;
991 long timeo = sock_sndtimeo(sk, nonblock);
992
993 /* Mirror Linux UDP mirror of BSD error message compatibility */
994 /* XXX: Perhaps MSG_MORE someday */
995 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
996 ret = -EOPNOTSUPP;
997 goto out;
998 }
999
1000 if (msg->msg_namelen) {
1001 /* XXX fail non-unicast destination IPs? */
1002 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1003 ret = -EINVAL;
1004 goto out;
1005 }
1006 daddr = usin->sin_addr.s_addr;
1007 dport = usin->sin_port;
1008 } else {
1009 /* We only care about consistency with ->connect() */
1010 lock_sock(sk);
1011 daddr = rs->rs_conn_addr;
1012 dport = rs->rs_conn_port;
1013 release_sock(sk);
1014 }
1015
1016 lock_sock(sk);
1017 if (daddr == 0 || rs->rs_bound_addr == 0) {
1018 release_sock(sk);
1019 ret = -ENOTCONN; /* XXX not a great errno */
1020 goto out;
1021 }
1022 release_sock(sk);
1023
1024 if (payload_len > rds_sk_sndbuf(rs)) {
1025 ret = -EMSGSIZE;
1026 goto out;
1027 }
1028
1029 /* size of rm including all sgs */
1030 ret = rds_rm_size(msg, payload_len);
1031 if (ret < 0)
1032 goto out;
1033
1034 rm = rds_message_alloc(ret, GFP_KERNEL);
1035 if (!rm) {
1036 ret = -ENOMEM;
1037 goto out;
1038 }
1039
1040 /* Attach data to the rm */
1041 if (payload_len) {
1042 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1043 if (!rm->data.op_sg) {
1044 ret = -ENOMEM;
1045 goto out;
1046 }
1047 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1048 if (ret)
1049 goto out;
1050 }
1051 rm->data.op_active = 1;
1052
1053 rm->m_daddr = daddr;
1054
1055 /* rds_conn_create has a spinlock that runs with IRQ off.
1056 * Caching the conn in the socket helps a lot. */
1057 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1058 conn = rs->rs_conn;
1059 else {
1060 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1061 rs->rs_bound_addr, daddr,
1062 rs->rs_transport,
1063 sock->sk->sk_allocation);
1064 if (IS_ERR(conn)) {
1065 ret = PTR_ERR(conn);
1066 goto out;
1067 }
1068 rs->rs_conn = conn;
1069 }
1070
1071 /* Parse any control messages the user may have included. */
1072 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1073 if (ret)
1074 goto out;
1075
1076 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1077 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1078 &rm->rdma, conn->c_trans->xmit_rdma);
1079 ret = -EOPNOTSUPP;
1080 goto out;
1081 }
1082
1083 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1084 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1085 &rm->atomic, conn->c_trans->xmit_atomic);
1086 ret = -EOPNOTSUPP;
1087 goto out;
1088 }
1089
1090 rds_conn_connect_if_down(conn);
1091
1092 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1093 if (ret) {
1094 rs->rs_seen_congestion = 1;
1095 goto out;
1096 }
1097
1098 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1099 dport, &queued)) {
1100 rds_stats_inc(s_send_queue_full);
1101
1102 if (nonblock) {
1103 ret = -EAGAIN;
1104 goto out;
1105 }
1106
1107 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1108 rds_send_queue_rm(rs, conn, rm,
1109 rs->rs_bound_port,
1110 dport,
1111 &queued),
1112 timeo);
1113 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1114 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1115 continue;
1116
1117 ret = timeo;
1118 if (ret == 0)
1119 ret = -ETIMEDOUT;
1120 goto out;
1121 }
1122
1123 /*
1124 * By now we've committed to the send. We reuse rds_send_worker()
1125 * to retry sends in the rds thread if the transport asks us to.
1126 */
1127 rds_stats_inc(s_send_queued);
1128
1129 ret = rds_send_xmit(conn);
1130 if (ret == -ENOMEM || ret == -EAGAIN)
1131 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1132
1133 rds_message_put(rm);
1134 return payload_len;
1135
1136out:
1137 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1138 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1139 * or in any other way, we need to destroy the MR again */
1140 if (allocated_mr)
1141 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1142
1143 if (rm)
1144 rds_message_put(rm);
1145 return ret;
1146}
1147
1148/*
1149 * Reply to a ping packet.
1150 */
1151int
1152rds_send_pong(struct rds_connection *conn, __be16 dport)
1153{
1154 struct rds_message *rm;
1155 unsigned long flags;
1156 int ret = 0;
1157
1158 rm = rds_message_alloc(0, GFP_ATOMIC);
1159 if (!rm) {
1160 ret = -ENOMEM;
1161 goto out;
1162 }
1163
1164 rm->m_daddr = conn->c_faddr;
1165 rm->data.op_active = 1;
1166
1167 rds_conn_connect_if_down(conn);
1168
1169 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1170 if (ret)
1171 goto out;
1172
1173 spin_lock_irqsave(&conn->c_lock, flags);
1174 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1175 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1176 rds_message_addref(rm);
1177 rm->m_inc.i_conn = conn;
1178
1179 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1180 conn->c_next_tx_seq);
1181 conn->c_next_tx_seq++;
1182 spin_unlock_irqrestore(&conn->c_lock, flags);
1183
1184 rds_stats_inc(s_send_queued);
1185 rds_stats_inc(s_send_pong);
1186
1187 /* schedule the send work on rds_wq */
1188 queue_delayed_work(rds_wq, &conn->c_send_w, 1);
1189
1190 rds_message_put(rm);
1191 return 0;
1192
1193out:
1194 if (rm)
1195 rds_message_put(rm);
1196 return ret;
1197}
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/gfp.h>
35#include <net/sock.h>
36#include <linux/in.h>
37#include <linux/list.h>
38#include <linux/ratelimit.h>
39
40#include "rds.h"
41
42/* When transmitting messages in rds_send_xmit, we need to emerge from
43 * time to time and briefly release the CPU. Otherwise the softlock watchdog
44 * will kick our shin.
45 * Also, it seems fairer to not let one busy connection stall all the
46 * others.
47 *
48 * send_batch_count is the number of times we'll loop in send_xmit. Setting
49 * it to 0 will restore the old behavior (where we looped until we had
50 * drained the queue).
51 */
52static int send_batch_count = 64;
53module_param(send_batch_count, int, 0444);
54MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
55
56static void rds_send_remove_from_sock(struct list_head *messages, int status);
57
58/*
59 * Reset the send state. Callers must ensure that this doesn't race with
60 * rds_send_xmit().
61 */
62void rds_send_reset(struct rds_connection *conn)
63{
64 struct rds_message *rm, *tmp;
65 unsigned long flags;
66
67 if (conn->c_xmit_rm) {
68 rm = conn->c_xmit_rm;
69 conn->c_xmit_rm = NULL;
70 /* Tell the user the RDMA op is no longer mapped by the
71 * transport. This isn't entirely true (it's flushed out
72 * independently) but as the connection is down, there's
73 * no ongoing RDMA to/from that memory */
74 rds_message_unmapped(rm);
75 rds_message_put(rm);
76 }
77
78 conn->c_xmit_sg = 0;
79 conn->c_xmit_hdr_off = 0;
80 conn->c_xmit_data_off = 0;
81 conn->c_xmit_atomic_sent = 0;
82 conn->c_xmit_rdma_sent = 0;
83 conn->c_xmit_data_sent = 0;
84
85 conn->c_map_queued = 0;
86
87 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
88 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
89
90 /* Mark messages as retransmissions, and move them to the send q */
91 spin_lock_irqsave(&conn->c_lock, flags);
92 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
93 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
94 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
95 }
96 list_splice_init(&conn->c_retrans, &conn->c_send_queue);
97 spin_unlock_irqrestore(&conn->c_lock, flags);
98}
99
100static int acquire_in_xmit(struct rds_connection *conn)
101{
102 return test_and_set_bit(RDS_IN_XMIT, &conn->c_flags) == 0;
103}
104
105static void release_in_xmit(struct rds_connection *conn)
106{
107 clear_bit(RDS_IN_XMIT, &conn->c_flags);
108 smp_mb__after_clear_bit();
109 /*
110 * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
111 * hot path and finding waiters is very rare. We don't want to walk
112 * the system-wide hashed waitqueue buckets in the fast path only to
113 * almost never find waiters.
114 */
115 if (waitqueue_active(&conn->c_waitq))
116 wake_up_all(&conn->c_waitq);
117}
118
119/*
120 * We're making the conscious trade-off here to only send one message
121 * down the connection at a time.
122 * Pro:
123 * - tx queueing is a simple fifo list
124 * - reassembly is optional and easily done by transports per conn
125 * - no per flow rx lookup at all, straight to the socket
126 * - less per-frag memory and wire overhead
127 * Con:
128 * - queued acks can be delayed behind large messages
129 * Depends:
130 * - small message latency is higher behind queued large messages
131 * - large message latency isn't starved by intervening small sends
132 */
133int rds_send_xmit(struct rds_connection *conn)
134{
135 struct rds_message *rm;
136 unsigned long flags;
137 unsigned int tmp;
138 struct scatterlist *sg;
139 int ret = 0;
140 LIST_HEAD(to_be_dropped);
141
142restart:
143
144 /*
145 * sendmsg calls here after having queued its message on the send
146 * queue. We only have one task feeding the connection at a time. If
147 * another thread is already feeding the queue then we back off. This
148 * avoids blocking the caller and trading per-connection data between
149 * caches per message.
150 */
151 if (!acquire_in_xmit(conn)) {
152 rds_stats_inc(s_send_lock_contention);
153 ret = -ENOMEM;
154 goto out;
155 }
156
157 /*
158 * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
159 * we do the opposite to avoid races.
160 */
161 if (!rds_conn_up(conn)) {
162 release_in_xmit(conn);
163 ret = 0;
164 goto out;
165 }
166
167 if (conn->c_trans->xmit_prepare)
168 conn->c_trans->xmit_prepare(conn);
169
170 /*
171 * spin trying to push headers and data down the connection until
172 * the connection doesn't make forward progress.
173 */
174 while (1) {
175
176 rm = conn->c_xmit_rm;
177
178 /*
179 * If between sending messages, we can send a pending congestion
180 * map update.
181 */
182 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
183 rm = rds_cong_update_alloc(conn);
184 if (IS_ERR(rm)) {
185 ret = PTR_ERR(rm);
186 break;
187 }
188 rm->data.op_active = 1;
189
190 conn->c_xmit_rm = rm;
191 }
192
193 /*
194 * If not already working on one, grab the next message.
195 *
196 * c_xmit_rm holds a ref while we're sending this message down
197 * the connction. We can use this ref while holding the
198 * send_sem.. rds_send_reset() is serialized with it.
199 */
200 if (!rm) {
201 unsigned int len;
202
203 spin_lock_irqsave(&conn->c_lock, flags);
204
205 if (!list_empty(&conn->c_send_queue)) {
206 rm = list_entry(conn->c_send_queue.next,
207 struct rds_message,
208 m_conn_item);
209 rds_message_addref(rm);
210
211 /*
212 * Move the message from the send queue to the retransmit
213 * list right away.
214 */
215 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
216 }
217
218 spin_unlock_irqrestore(&conn->c_lock, flags);
219
220 if (!rm)
221 break;
222
223 /* Unfortunately, the way Infiniband deals with
224 * RDMA to a bad MR key is by moving the entire
225 * queue pair to error state. We cold possibly
226 * recover from that, but right now we drop the
227 * connection.
228 * Therefore, we never retransmit messages with RDMA ops.
229 */
230 if (rm->rdma.op_active &&
231 test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
232 spin_lock_irqsave(&conn->c_lock, flags);
233 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
234 list_move(&rm->m_conn_item, &to_be_dropped);
235 spin_unlock_irqrestore(&conn->c_lock, flags);
236 continue;
237 }
238
239 /* Require an ACK every once in a while */
240 len = ntohl(rm->m_inc.i_hdr.h_len);
241 if (conn->c_unacked_packets == 0 ||
242 conn->c_unacked_bytes < len) {
243 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
244
245 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
246 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
247 rds_stats_inc(s_send_ack_required);
248 } else {
249 conn->c_unacked_bytes -= len;
250 conn->c_unacked_packets--;
251 }
252
253 conn->c_xmit_rm = rm;
254 }
255
256 /* The transport either sends the whole rdma or none of it */
257 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
258 rm->m_final_op = &rm->rdma;
259 ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
260 if (ret)
261 break;
262 conn->c_xmit_rdma_sent = 1;
263
264 /* The transport owns the mapped memory for now.
265 * You can't unmap it while it's on the send queue */
266 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
267 }
268
269 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
270 rm->m_final_op = &rm->atomic;
271 ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
272 if (ret)
273 break;
274 conn->c_xmit_atomic_sent = 1;
275
276 /* The transport owns the mapped memory for now.
277 * You can't unmap it while it's on the send queue */
278 set_bit(RDS_MSG_MAPPED, &rm->m_flags);
279 }
280
281 /*
282 * A number of cases require an RDS header to be sent
283 * even if there is no data.
284 * We permit 0-byte sends; rds-ping depends on this.
285 * However, if there are exclusively attached silent ops,
286 * we skip the hdr/data send, to enable silent operation.
287 */
288 if (rm->data.op_nents == 0) {
289 int ops_present;
290 int all_ops_are_silent = 1;
291
292 ops_present = (rm->atomic.op_active || rm->rdma.op_active);
293 if (rm->atomic.op_active && !rm->atomic.op_silent)
294 all_ops_are_silent = 0;
295 if (rm->rdma.op_active && !rm->rdma.op_silent)
296 all_ops_are_silent = 0;
297
298 if (ops_present && all_ops_are_silent
299 && !rm->m_rdma_cookie)
300 rm->data.op_active = 0;
301 }
302
303 if (rm->data.op_active && !conn->c_xmit_data_sent) {
304 rm->m_final_op = &rm->data;
305 ret = conn->c_trans->xmit(conn, rm,
306 conn->c_xmit_hdr_off,
307 conn->c_xmit_sg,
308 conn->c_xmit_data_off);
309 if (ret <= 0)
310 break;
311
312 if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
313 tmp = min_t(int, ret,
314 sizeof(struct rds_header) -
315 conn->c_xmit_hdr_off);
316 conn->c_xmit_hdr_off += tmp;
317 ret -= tmp;
318 }
319
320 sg = &rm->data.op_sg[conn->c_xmit_sg];
321 while (ret) {
322 tmp = min_t(int, ret, sg->length -
323 conn->c_xmit_data_off);
324 conn->c_xmit_data_off += tmp;
325 ret -= tmp;
326 if (conn->c_xmit_data_off == sg->length) {
327 conn->c_xmit_data_off = 0;
328 sg++;
329 conn->c_xmit_sg++;
330 BUG_ON(ret != 0 &&
331 conn->c_xmit_sg == rm->data.op_nents);
332 }
333 }
334
335 if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
336 (conn->c_xmit_sg == rm->data.op_nents))
337 conn->c_xmit_data_sent = 1;
338 }
339
340 /*
341 * A rm will only take multiple times through this loop
342 * if there is a data op. Thus, if the data is sent (or there was
343 * none), then we're done with the rm.
344 */
345 if (!rm->data.op_active || conn->c_xmit_data_sent) {
346 conn->c_xmit_rm = NULL;
347 conn->c_xmit_sg = 0;
348 conn->c_xmit_hdr_off = 0;
349 conn->c_xmit_data_off = 0;
350 conn->c_xmit_rdma_sent = 0;
351 conn->c_xmit_atomic_sent = 0;
352 conn->c_xmit_data_sent = 0;
353
354 rds_message_put(rm);
355 }
356 }
357
358 if (conn->c_trans->xmit_complete)
359 conn->c_trans->xmit_complete(conn);
360
361 release_in_xmit(conn);
362
363 /* Nuke any messages we decided not to retransmit. */
364 if (!list_empty(&to_be_dropped)) {
365 /* irqs on here, so we can put(), unlike above */
366 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
367 rds_message_put(rm);
368 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
369 }
370
371 /*
372 * Other senders can queue a message after we last test the send queue
373 * but before we clear RDS_IN_XMIT. In that case they'd back off and
374 * not try and send their newly queued message. We need to check the
375 * send queue after having cleared RDS_IN_XMIT so that their message
376 * doesn't get stuck on the send queue.
377 *
378 * If the transport cannot continue (i.e ret != 0), then it must
379 * call us when more room is available, such as from the tx
380 * completion handler.
381 */
382 if (ret == 0) {
383 smp_mb();
384 if (!list_empty(&conn->c_send_queue)) {
385 rds_stats_inc(s_send_lock_queue_raced);
386 goto restart;
387 }
388 }
389out:
390 return ret;
391}
392
393static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
394{
395 u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
396
397 assert_spin_locked(&rs->rs_lock);
398
399 BUG_ON(rs->rs_snd_bytes < len);
400 rs->rs_snd_bytes -= len;
401
402 if (rs->rs_snd_bytes == 0)
403 rds_stats_inc(s_send_queue_empty);
404}
405
406static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
407 is_acked_func is_acked)
408{
409 if (is_acked)
410 return is_acked(rm, ack);
411 return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
412}
413
414/*
415 * This is pretty similar to what happens below in the ACK
416 * handling code - except that we call here as soon as we get
417 * the IB send completion on the RDMA op and the accompanying
418 * message.
419 */
420void rds_rdma_send_complete(struct rds_message *rm, int status)
421{
422 struct rds_sock *rs = NULL;
423 struct rm_rdma_op *ro;
424 struct rds_notifier *notifier;
425 unsigned long flags;
426
427 spin_lock_irqsave(&rm->m_rs_lock, flags);
428
429 ro = &rm->rdma;
430 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
431 ro->op_active && ro->op_notify && ro->op_notifier) {
432 notifier = ro->op_notifier;
433 rs = rm->m_rs;
434 sock_hold(rds_rs_to_sk(rs));
435
436 notifier->n_status = status;
437 spin_lock(&rs->rs_lock);
438 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
439 spin_unlock(&rs->rs_lock);
440
441 ro->op_notifier = NULL;
442 }
443
444 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
445
446 if (rs) {
447 rds_wake_sk_sleep(rs);
448 sock_put(rds_rs_to_sk(rs));
449 }
450}
451EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
452
453/*
454 * Just like above, except looks at atomic op
455 */
456void rds_atomic_send_complete(struct rds_message *rm, int status)
457{
458 struct rds_sock *rs = NULL;
459 struct rm_atomic_op *ao;
460 struct rds_notifier *notifier;
461 unsigned long flags;
462
463 spin_lock_irqsave(&rm->m_rs_lock, flags);
464
465 ao = &rm->atomic;
466 if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
467 && ao->op_active && ao->op_notify && ao->op_notifier) {
468 notifier = ao->op_notifier;
469 rs = rm->m_rs;
470 sock_hold(rds_rs_to_sk(rs));
471
472 notifier->n_status = status;
473 spin_lock(&rs->rs_lock);
474 list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
475 spin_unlock(&rs->rs_lock);
476
477 ao->op_notifier = NULL;
478 }
479
480 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
481
482 if (rs) {
483 rds_wake_sk_sleep(rs);
484 sock_put(rds_rs_to_sk(rs));
485 }
486}
487EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
488
489/*
490 * This is the same as rds_rdma_send_complete except we
491 * don't do any locking - we have all the ingredients (message,
492 * socket, socket lock) and can just move the notifier.
493 */
494static inline void
495__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
496{
497 struct rm_rdma_op *ro;
498 struct rm_atomic_op *ao;
499
500 ro = &rm->rdma;
501 if (ro->op_active && ro->op_notify && ro->op_notifier) {
502 ro->op_notifier->n_status = status;
503 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
504 ro->op_notifier = NULL;
505 }
506
507 ao = &rm->atomic;
508 if (ao->op_active && ao->op_notify && ao->op_notifier) {
509 ao->op_notifier->n_status = status;
510 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
511 ao->op_notifier = NULL;
512 }
513
514 /* No need to wake the app - caller does this */
515}
516
517/*
518 * This is called from the IB send completion when we detect
519 * a RDMA operation that failed with remote access error.
520 * So speed is not an issue here.
521 */
522struct rds_message *rds_send_get_message(struct rds_connection *conn,
523 struct rm_rdma_op *op)
524{
525 struct rds_message *rm, *tmp, *found = NULL;
526 unsigned long flags;
527
528 spin_lock_irqsave(&conn->c_lock, flags);
529
530 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
531 if (&rm->rdma == op) {
532 atomic_inc(&rm->m_refcount);
533 found = rm;
534 goto out;
535 }
536 }
537
538 list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
539 if (&rm->rdma == op) {
540 atomic_inc(&rm->m_refcount);
541 found = rm;
542 break;
543 }
544 }
545
546out:
547 spin_unlock_irqrestore(&conn->c_lock, flags);
548
549 return found;
550}
551EXPORT_SYMBOL_GPL(rds_send_get_message);
552
553/*
554 * This removes messages from the socket's list if they're on it. The list
555 * argument must be private to the caller, we must be able to modify it
556 * without locks. The messages must have a reference held for their
557 * position on the list. This function will drop that reference after
558 * removing the messages from the 'messages' list regardless of if it found
559 * the messages on the socket list or not.
560 */
561static void rds_send_remove_from_sock(struct list_head *messages, int status)
562{
563 unsigned long flags;
564 struct rds_sock *rs = NULL;
565 struct rds_message *rm;
566
567 while (!list_empty(messages)) {
568 int was_on_sock = 0;
569
570 rm = list_entry(messages->next, struct rds_message,
571 m_conn_item);
572 list_del_init(&rm->m_conn_item);
573
574 /*
575 * If we see this flag cleared then we're *sure* that someone
576 * else beat us to removing it from the sock. If we race
577 * with their flag update we'll get the lock and then really
578 * see that the flag has been cleared.
579 *
580 * The message spinlock makes sure nobody clears rm->m_rs
581 * while we're messing with it. It does not prevent the
582 * message from being removed from the socket, though.
583 */
584 spin_lock_irqsave(&rm->m_rs_lock, flags);
585 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
586 goto unlock_and_drop;
587
588 if (rs != rm->m_rs) {
589 if (rs) {
590 rds_wake_sk_sleep(rs);
591 sock_put(rds_rs_to_sk(rs));
592 }
593 rs = rm->m_rs;
594 sock_hold(rds_rs_to_sk(rs));
595 }
596 spin_lock(&rs->rs_lock);
597
598 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
599 struct rm_rdma_op *ro = &rm->rdma;
600 struct rds_notifier *notifier;
601
602 list_del_init(&rm->m_sock_item);
603 rds_send_sndbuf_remove(rs, rm);
604
605 if (ro->op_active && ro->op_notifier &&
606 (ro->op_notify || (ro->op_recverr && status))) {
607 notifier = ro->op_notifier;
608 list_add_tail(¬ifier->n_list,
609 &rs->rs_notify_queue);
610 if (!notifier->n_status)
611 notifier->n_status = status;
612 rm->rdma.op_notifier = NULL;
613 }
614 was_on_sock = 1;
615 rm->m_rs = NULL;
616 }
617 spin_unlock(&rs->rs_lock);
618
619unlock_and_drop:
620 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
621 rds_message_put(rm);
622 if (was_on_sock)
623 rds_message_put(rm);
624 }
625
626 if (rs) {
627 rds_wake_sk_sleep(rs);
628 sock_put(rds_rs_to_sk(rs));
629 }
630}
631
632/*
633 * Transports call here when they've determined that the receiver queued
634 * messages up to, and including, the given sequence number. Messages are
635 * moved to the retrans queue when rds_send_xmit picks them off the send
636 * queue. This means that in the TCP case, the message may not have been
637 * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
638 * checks the RDS_MSG_HAS_ACK_SEQ bit.
639 *
640 * XXX It's not clear to me how this is safely serialized with socket
641 * destruction. Maybe it should bail if it sees SOCK_DEAD.
642 */
643void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
644 is_acked_func is_acked)
645{
646 struct rds_message *rm, *tmp;
647 unsigned long flags;
648 LIST_HEAD(list);
649
650 spin_lock_irqsave(&conn->c_lock, flags);
651
652 list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
653 if (!rds_send_is_acked(rm, ack, is_acked))
654 break;
655
656 list_move(&rm->m_conn_item, &list);
657 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
658 }
659
660 /* order flag updates with spin locks */
661 if (!list_empty(&list))
662 smp_mb__after_clear_bit();
663
664 spin_unlock_irqrestore(&conn->c_lock, flags);
665
666 /* now remove the messages from the sock list as needed */
667 rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
668}
669EXPORT_SYMBOL_GPL(rds_send_drop_acked);
670
671void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
672{
673 struct rds_message *rm, *tmp;
674 struct rds_connection *conn;
675 unsigned long flags;
676 LIST_HEAD(list);
677
678 /* get all the messages we're dropping under the rs lock */
679 spin_lock_irqsave(&rs->rs_lock, flags);
680
681 list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
682 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
683 dest->sin_port != rm->m_inc.i_hdr.h_dport))
684 continue;
685
686 list_move(&rm->m_sock_item, &list);
687 rds_send_sndbuf_remove(rs, rm);
688 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
689 }
690
691 /* order flag updates with the rs lock */
692 smp_mb__after_clear_bit();
693
694 spin_unlock_irqrestore(&rs->rs_lock, flags);
695
696 if (list_empty(&list))
697 return;
698
699 /* Remove the messages from the conn */
700 list_for_each_entry(rm, &list, m_sock_item) {
701
702 conn = rm->m_inc.i_conn;
703
704 spin_lock_irqsave(&conn->c_lock, flags);
705 /*
706 * Maybe someone else beat us to removing rm from the conn.
707 * If we race with their flag update we'll get the lock and
708 * then really see that the flag has been cleared.
709 */
710 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
711 spin_unlock_irqrestore(&conn->c_lock, flags);
712 continue;
713 }
714 list_del_init(&rm->m_conn_item);
715 spin_unlock_irqrestore(&conn->c_lock, flags);
716
717 /*
718 * Couldn't grab m_rs_lock in top loop (lock ordering),
719 * but we can now.
720 */
721 spin_lock_irqsave(&rm->m_rs_lock, flags);
722
723 spin_lock(&rs->rs_lock);
724 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
725 spin_unlock(&rs->rs_lock);
726
727 rm->m_rs = NULL;
728 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
729
730 rds_message_put(rm);
731 }
732
733 rds_wake_sk_sleep(rs);
734
735 while (!list_empty(&list)) {
736 rm = list_entry(list.next, struct rds_message, m_sock_item);
737 list_del_init(&rm->m_sock_item);
738
739 rds_message_wait(rm);
740 rds_message_put(rm);
741 }
742}
743
744/*
745 * we only want this to fire once so we use the callers 'queued'. It's
746 * possible that another thread can race with us and remove the
747 * message from the flow with RDS_CANCEL_SENT_TO.
748 */
749static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
750 struct rds_message *rm, __be16 sport,
751 __be16 dport, int *queued)
752{
753 unsigned long flags;
754 u32 len;
755
756 if (*queued)
757 goto out;
758
759 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
760
761 /* this is the only place which holds both the socket's rs_lock
762 * and the connection's c_lock */
763 spin_lock_irqsave(&rs->rs_lock, flags);
764
765 /*
766 * If there is a little space in sndbuf, we don't queue anything,
767 * and userspace gets -EAGAIN. But poll() indicates there's send
768 * room. This can lead to bad behavior (spinning) if snd_bytes isn't
769 * freed up by incoming acks. So we check the *old* value of
770 * rs_snd_bytes here to allow the last msg to exceed the buffer,
771 * and poll() now knows no more data can be sent.
772 */
773 if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
774 rs->rs_snd_bytes += len;
775
776 /* let recv side know we are close to send space exhaustion.
777 * This is probably not the optimal way to do it, as this
778 * means we set the flag on *all* messages as soon as our
779 * throughput hits a certain threshold.
780 */
781 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
782 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
783
784 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
785 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
786 rds_message_addref(rm);
787 rm->m_rs = rs;
788
789 /* The code ordering is a little weird, but we're
790 trying to minimize the time we hold c_lock */
791 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
792 rm->m_inc.i_conn = conn;
793 rds_message_addref(rm);
794
795 spin_lock(&conn->c_lock);
796 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
797 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
798 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
799 spin_unlock(&conn->c_lock);
800
801 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
802 rm, len, rs, rs->rs_snd_bytes,
803 (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
804
805 *queued = 1;
806 }
807
808 spin_unlock_irqrestore(&rs->rs_lock, flags);
809out:
810 return *queued;
811}
812
813/*
814 * rds_message is getting to be quite complicated, and we'd like to allocate
815 * it all in one go. This figures out how big it needs to be up front.
816 */
817static int rds_rm_size(struct msghdr *msg, int data_len)
818{
819 struct cmsghdr *cmsg;
820 int size = 0;
821 int cmsg_groups = 0;
822 int retval;
823
824 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
825 if (!CMSG_OK(msg, cmsg))
826 return -EINVAL;
827
828 if (cmsg->cmsg_level != SOL_RDS)
829 continue;
830
831 switch (cmsg->cmsg_type) {
832 case RDS_CMSG_RDMA_ARGS:
833 cmsg_groups |= 1;
834 retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
835 if (retval < 0)
836 return retval;
837 size += retval;
838
839 break;
840
841 case RDS_CMSG_RDMA_DEST:
842 case RDS_CMSG_RDMA_MAP:
843 cmsg_groups |= 2;
844 /* these are valid but do no add any size */
845 break;
846
847 case RDS_CMSG_ATOMIC_CSWP:
848 case RDS_CMSG_ATOMIC_FADD:
849 case RDS_CMSG_MASKED_ATOMIC_CSWP:
850 case RDS_CMSG_MASKED_ATOMIC_FADD:
851 cmsg_groups |= 1;
852 size += sizeof(struct scatterlist);
853 break;
854
855 default:
856 return -EINVAL;
857 }
858
859 }
860
861 size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
862
863 /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
864 if (cmsg_groups == 3)
865 return -EINVAL;
866
867 return size;
868}
869
870static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
871 struct msghdr *msg, int *allocated_mr)
872{
873 struct cmsghdr *cmsg;
874 int ret = 0;
875
876 for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
877 if (!CMSG_OK(msg, cmsg))
878 return -EINVAL;
879
880 if (cmsg->cmsg_level != SOL_RDS)
881 continue;
882
883 /* As a side effect, RDMA_DEST and RDMA_MAP will set
884 * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
885 */
886 switch (cmsg->cmsg_type) {
887 case RDS_CMSG_RDMA_ARGS:
888 ret = rds_cmsg_rdma_args(rs, rm, cmsg);
889 break;
890
891 case RDS_CMSG_RDMA_DEST:
892 ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
893 break;
894
895 case RDS_CMSG_RDMA_MAP:
896 ret = rds_cmsg_rdma_map(rs, rm, cmsg);
897 if (!ret)
898 *allocated_mr = 1;
899 break;
900 case RDS_CMSG_ATOMIC_CSWP:
901 case RDS_CMSG_ATOMIC_FADD:
902 case RDS_CMSG_MASKED_ATOMIC_CSWP:
903 case RDS_CMSG_MASKED_ATOMIC_FADD:
904 ret = rds_cmsg_atomic(rs, rm, cmsg);
905 break;
906
907 default:
908 return -EINVAL;
909 }
910
911 if (ret)
912 break;
913 }
914
915 return ret;
916}
917
918int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
919 size_t payload_len)
920{
921 struct sock *sk = sock->sk;
922 struct rds_sock *rs = rds_sk_to_rs(sk);
923 struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
924 __be32 daddr;
925 __be16 dport;
926 struct rds_message *rm = NULL;
927 struct rds_connection *conn;
928 int ret = 0;
929 int queued = 0, allocated_mr = 0;
930 int nonblock = msg->msg_flags & MSG_DONTWAIT;
931 long timeo = sock_sndtimeo(sk, nonblock);
932
933 /* Mirror Linux UDP mirror of BSD error message compatibility */
934 /* XXX: Perhaps MSG_MORE someday */
935 if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
936 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
937 ret = -EOPNOTSUPP;
938 goto out;
939 }
940
941 if (msg->msg_namelen) {
942 /* XXX fail non-unicast destination IPs? */
943 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
944 ret = -EINVAL;
945 goto out;
946 }
947 daddr = usin->sin_addr.s_addr;
948 dport = usin->sin_port;
949 } else {
950 /* We only care about consistency with ->connect() */
951 lock_sock(sk);
952 daddr = rs->rs_conn_addr;
953 dport = rs->rs_conn_port;
954 release_sock(sk);
955 }
956
957 /* racing with another thread binding seems ok here */
958 if (daddr == 0 || rs->rs_bound_addr == 0) {
959 ret = -ENOTCONN; /* XXX not a great errno */
960 goto out;
961 }
962
963 /* size of rm including all sgs */
964 ret = rds_rm_size(msg, payload_len);
965 if (ret < 0)
966 goto out;
967
968 rm = rds_message_alloc(ret, GFP_KERNEL);
969 if (!rm) {
970 ret = -ENOMEM;
971 goto out;
972 }
973
974 /* Attach data to the rm */
975 if (payload_len) {
976 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
977 if (!rm->data.op_sg) {
978 ret = -ENOMEM;
979 goto out;
980 }
981 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
982 if (ret)
983 goto out;
984 }
985 rm->data.op_active = 1;
986
987 rm->m_daddr = daddr;
988
989 /* rds_conn_create has a spinlock that runs with IRQ off.
990 * Caching the conn in the socket helps a lot. */
991 if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
992 conn = rs->rs_conn;
993 else {
994 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
995 rs->rs_transport,
996 sock->sk->sk_allocation);
997 if (IS_ERR(conn)) {
998 ret = PTR_ERR(conn);
999 goto out;
1000 }
1001 rs->rs_conn = conn;
1002 }
1003
1004 /* Parse any control messages the user may have included. */
1005 ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1006 if (ret)
1007 goto out;
1008
1009 if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1010 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1011 &rm->rdma, conn->c_trans->xmit_rdma);
1012 ret = -EOPNOTSUPP;
1013 goto out;
1014 }
1015
1016 if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1017 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1018 &rm->atomic, conn->c_trans->xmit_atomic);
1019 ret = -EOPNOTSUPP;
1020 goto out;
1021 }
1022
1023 rds_conn_connect_if_down(conn);
1024
1025 ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1026 if (ret) {
1027 rs->rs_seen_congestion = 1;
1028 goto out;
1029 }
1030
1031 while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1032 dport, &queued)) {
1033 rds_stats_inc(s_send_queue_full);
1034 /* XXX make sure this is reasonable */
1035 if (payload_len > rds_sk_sndbuf(rs)) {
1036 ret = -EMSGSIZE;
1037 goto out;
1038 }
1039 if (nonblock) {
1040 ret = -EAGAIN;
1041 goto out;
1042 }
1043
1044 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1045 rds_send_queue_rm(rs, conn, rm,
1046 rs->rs_bound_port,
1047 dport,
1048 &queued),
1049 timeo);
1050 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1051 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1052 continue;
1053
1054 ret = timeo;
1055 if (ret == 0)
1056 ret = -ETIMEDOUT;
1057 goto out;
1058 }
1059
1060 /*
1061 * By now we've committed to the send. We reuse rds_send_worker()
1062 * to retry sends in the rds thread if the transport asks us to.
1063 */
1064 rds_stats_inc(s_send_queued);
1065
1066 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1067 rds_send_xmit(conn);
1068
1069 rds_message_put(rm);
1070 return payload_len;
1071
1072out:
1073 /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1074 * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1075 * or in any other way, we need to destroy the MR again */
1076 if (allocated_mr)
1077 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1078
1079 if (rm)
1080 rds_message_put(rm);
1081 return ret;
1082}
1083
1084/*
1085 * Reply to a ping packet.
1086 */
1087int
1088rds_send_pong(struct rds_connection *conn, __be16 dport)
1089{
1090 struct rds_message *rm;
1091 unsigned long flags;
1092 int ret = 0;
1093
1094 rm = rds_message_alloc(0, GFP_ATOMIC);
1095 if (!rm) {
1096 ret = -ENOMEM;
1097 goto out;
1098 }
1099
1100 rm->m_daddr = conn->c_faddr;
1101 rm->data.op_active = 1;
1102
1103 rds_conn_connect_if_down(conn);
1104
1105 ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1106 if (ret)
1107 goto out;
1108
1109 spin_lock_irqsave(&conn->c_lock, flags);
1110 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1111 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1112 rds_message_addref(rm);
1113 rm->m_inc.i_conn = conn;
1114
1115 rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1116 conn->c_next_tx_seq);
1117 conn->c_next_tx_seq++;
1118 spin_unlock_irqrestore(&conn->c_lock, flags);
1119
1120 rds_stats_inc(s_send_queued);
1121 rds_stats_inc(s_send_pong);
1122
1123 if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1124 rds_send_xmit(conn);
1125
1126 rds_message_put(rm);
1127 return 0;
1128
1129out:
1130 if (rm)
1131 rds_message_put(rm);
1132 return ret;
1133}