net: implements ip_route_input_noref()
[linux-2.6.git] / net / rds / send.c
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
39 #include "rds.h"
40 #include "rdma.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  */
52 static int send_batch_count = 64;
53 module_param(send_batch_count, int, 0444);
54 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
55
56 /*
57  * Reset the send state. Caller must hold c_send_lock when calling here.
58  */
59 void rds_send_reset(struct rds_connection *conn)
60 {
61         struct rds_message *rm, *tmp;
62         unsigned long flags;
63
64         if (conn->c_xmit_rm) {
65                 /* Tell the user the RDMA op is no longer mapped by the
66                  * transport. This isn't entirely true (it's flushed out
67                  * independently) but as the connection is down, there's
68                  * no ongoing RDMA to/from that memory */
69                 rds_message_unmapped(conn->c_xmit_rm);
70                 rds_message_put(conn->c_xmit_rm);
71                 conn->c_xmit_rm = NULL;
72         }
73         conn->c_xmit_sg = 0;
74         conn->c_xmit_hdr_off = 0;
75         conn->c_xmit_data_off = 0;
76         conn->c_xmit_rdma_sent = 0;
77
78         conn->c_map_queued = 0;
79
80         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
81         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
82
83         /* Mark messages as retransmissions, and move them to the send q */
84         spin_lock_irqsave(&conn->c_lock, flags);
85         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
86                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
87                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
88         }
89         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
90         spin_unlock_irqrestore(&conn->c_lock, flags);
91 }
92
93 /*
94  * We're making the concious trade-off here to only send one message
95  * down the connection at a time.
96  *   Pro:
97  *      - tx queueing is a simple fifo list
98  *      - reassembly is optional and easily done by transports per conn
99  *      - no per flow rx lookup at all, straight to the socket
100  *      - less per-frag memory and wire overhead
101  *   Con:
102  *      - queued acks can be delayed behind large messages
103  *   Depends:
104  *      - small message latency is higher behind queued large messages
105  *      - large message latency isn't starved by intervening small sends
106  */
107 int rds_send_xmit(struct rds_connection *conn)
108 {
109         struct rds_message *rm;
110         unsigned long flags;
111         unsigned int tmp;
112         unsigned int send_quota = send_batch_count;
113         struct scatterlist *sg;
114         int ret = 0;
115         int was_empty = 0;
116         LIST_HEAD(to_be_dropped);
117
118         /*
119          * sendmsg calls here after having queued its message on the send
120          * queue.  We only have one task feeding the connection at a time.  If
121          * another thread is already feeding the queue then we back off.  This
122          * avoids blocking the caller and trading per-connection data between
123          * caches per message.
124          *
125          * The sem holder will issue a retry if they notice that someone queued
126          * a message after they stopped walking the send queue but before they
127          * dropped the sem.
128          */
129         if (!mutex_trylock(&conn->c_send_lock)) {
130                 rds_stats_inc(s_send_sem_contention);
131                 ret = -ENOMEM;
132                 goto out;
133         }
134
135         if (conn->c_trans->xmit_prepare)
136                 conn->c_trans->xmit_prepare(conn);
137
138         /*
139          * spin trying to push headers and data down the connection until
140          * the connection doens't make forward progress.
141          */
142         while (--send_quota) {
143                 /*
144                  * See if need to send a congestion map update if we're
145                  * between sending messages.  The send_sem protects our sole
146                  * use of c_map_offset and _bytes.
147                  * Note this is used only by transports that define a special
148                  * xmit_cong_map function. For all others, we create allocate
149                  * a cong_map message and treat it just like any other send.
150                  */
151                 if (conn->c_map_bytes) {
152                         ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
153                                                 conn->c_map_offset);
154                         if (ret <= 0)
155                                 break;
156
157                         conn->c_map_offset += ret;
158                         conn->c_map_bytes -= ret;
159                         if (conn->c_map_bytes)
160                                 continue;
161                 }
162
163                 /* If we're done sending the current message, clear the
164                  * offset and S/G temporaries.
165                  */
166                 rm = conn->c_xmit_rm;
167                 if (rm != NULL &&
168                     conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
169                     conn->c_xmit_sg == rm->m_nents) {
170                         conn->c_xmit_rm = NULL;
171                         conn->c_xmit_sg = 0;
172                         conn->c_xmit_hdr_off = 0;
173                         conn->c_xmit_data_off = 0;
174                         conn->c_xmit_rdma_sent = 0;
175
176                         /* Release the reference to the previous message. */
177                         rds_message_put(rm);
178                         rm = NULL;
179                 }
180
181                 /* If we're asked to send a cong map update, do so.
182                  */
183                 if (rm == NULL && test_and_clear_bit(0, &conn->c_map_queued)) {
184                         if (conn->c_trans->xmit_cong_map != NULL) {
185                                 conn->c_map_offset = 0;
186                                 conn->c_map_bytes = sizeof(struct rds_header) +
187                                         RDS_CONG_MAP_BYTES;
188                                 continue;
189                         }
190
191                         rm = rds_cong_update_alloc(conn);
192                         if (IS_ERR(rm)) {
193                                 ret = PTR_ERR(rm);
194                                 break;
195                         }
196
197                         conn->c_xmit_rm = rm;
198                 }
199
200                 /*
201                  * Grab the next message from the send queue, if there is one.
202                  *
203                  * c_xmit_rm holds a ref while we're sending this message down
204                  * the connction.  We can use this ref while holding the
205                  * send_sem.. rds_send_reset() is serialized with it.
206                  */
207                 if (rm == NULL) {
208                         unsigned int len;
209
210                         spin_lock_irqsave(&conn->c_lock, flags);
211
212                         if (!list_empty(&conn->c_send_queue)) {
213                                 rm = list_entry(conn->c_send_queue.next,
214                                                 struct rds_message,
215                                                 m_conn_item);
216                                 rds_message_addref(rm);
217
218                                 /*
219                                  * Move the message from the send queue to the retransmit
220                                  * list right away.
221                                  */
222                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
223                         }
224
225                         spin_unlock_irqrestore(&conn->c_lock, flags);
226
227                         if (rm == NULL) {
228                                 was_empty = 1;
229                                 break;
230                         }
231
232                         /* Unfortunately, the way Infiniband deals with
233                          * RDMA to a bad MR key is by moving the entire
234                          * queue pair to error state. We cold possibly
235                          * recover from that, but right now we drop the
236                          * connection.
237                          * Therefore, we never retransmit messages with RDMA ops.
238                          */
239                         if (rm->m_rdma_op &&
240                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
241                                 spin_lock_irqsave(&conn->c_lock, flags);
242                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
243                                         list_move(&rm->m_conn_item, &to_be_dropped);
244                                 spin_unlock_irqrestore(&conn->c_lock, flags);
245                                 rds_message_put(rm);
246                                 continue;
247                         }
248
249                         /* Require an ACK every once in a while */
250                         len = ntohl(rm->m_inc.i_hdr.h_len);
251                         if (conn->c_unacked_packets == 0 ||
252                             conn->c_unacked_bytes < len) {
253                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
254
255                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
256                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
257                                 rds_stats_inc(s_send_ack_required);
258                         } else {
259                                 conn->c_unacked_bytes -= len;
260                                 conn->c_unacked_packets--;
261                         }
262
263                         conn->c_xmit_rm = rm;
264                 }
265
266                 /*
267                  * Try and send an rdma message.  Let's see if we can
268                  * keep this simple and require that the transport either
269                  * send the whole rdma or none of it.
270                  */
271                 if (rm->m_rdma_op && !conn->c_xmit_rdma_sent) {
272                         ret = conn->c_trans->xmit_rdma(conn, rm->m_rdma_op);
273                         if (ret)
274                                 break;
275                         conn->c_xmit_rdma_sent = 1;
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                 if (conn->c_xmit_hdr_off < sizeof(struct rds_header) ||
282                     conn->c_xmit_sg < rm->m_nents) {
283                         ret = conn->c_trans->xmit(conn, rm,
284                                                   conn->c_xmit_hdr_off,
285                                                   conn->c_xmit_sg,
286                                                   conn->c_xmit_data_off);
287                         if (ret <= 0)
288                                 break;
289
290                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
291                                 tmp = min_t(int, ret,
292                                             sizeof(struct rds_header) -
293                                             conn->c_xmit_hdr_off);
294                                 conn->c_xmit_hdr_off += tmp;
295                                 ret -= tmp;
296                         }
297
298                         sg = &rm->m_sg[conn->c_xmit_sg];
299                         while (ret) {
300                                 tmp = min_t(int, ret, sg->length -
301                                                       conn->c_xmit_data_off);
302                                 conn->c_xmit_data_off += tmp;
303                                 ret -= tmp;
304                                 if (conn->c_xmit_data_off == sg->length) {
305                                         conn->c_xmit_data_off = 0;
306                                         sg++;
307                                         conn->c_xmit_sg++;
308                                         BUG_ON(ret != 0 &&
309                                                conn->c_xmit_sg == rm->m_nents);
310                                 }
311                         }
312                 }
313         }
314
315         /* Nuke any messages we decided not to retransmit. */
316         if (!list_empty(&to_be_dropped))
317                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
318
319         if (conn->c_trans->xmit_complete)
320                 conn->c_trans->xmit_complete(conn);
321
322         /*
323          * We might be racing with another sender who queued a message but
324          * backed off on noticing that we held the c_send_lock.  If we check
325          * for queued messages after dropping the sem then either we'll
326          * see the queued message or the queuer will get the sem.  If we
327          * notice the queued message then we trigger an immediate retry.
328          *
329          * We need to be careful only to do this when we stopped processing
330          * the send queue because it was empty.  It's the only way we
331          * stop processing the loop when the transport hasn't taken
332          * responsibility for forward progress.
333          */
334         mutex_unlock(&conn->c_send_lock);
335
336         if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) {
337                 /* We exhausted the send quota, but there's work left to
338                  * do. Return and (re-)schedule the send worker.
339                  */
340                 ret = -EAGAIN;
341         }
342
343         if (ret == 0 && was_empty) {
344                 /* A simple bit test would be way faster than taking the
345                  * spin lock */
346                 spin_lock_irqsave(&conn->c_lock, flags);
347                 if (!list_empty(&conn->c_send_queue)) {
348                         rds_stats_inc(s_send_sem_queue_raced);
349                         ret = -EAGAIN;
350                 }
351                 spin_unlock_irqrestore(&conn->c_lock, flags);
352         }
353 out:
354         return ret;
355 }
356
357 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
358 {
359         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
360
361         assert_spin_locked(&rs->rs_lock);
362
363         BUG_ON(rs->rs_snd_bytes < len);
364         rs->rs_snd_bytes -= len;
365
366         if (rs->rs_snd_bytes == 0)
367                 rds_stats_inc(s_send_queue_empty);
368 }
369
370 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
371                                     is_acked_func is_acked)
372 {
373         if (is_acked)
374                 return is_acked(rm, ack);
375         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
376 }
377
378 /*
379  * Returns true if there are no messages on the send and retransmit queues
380  * which have a sequence number greater than or equal to the given sequence
381  * number.
382  */
383 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
384 {
385         struct rds_message *rm, *tmp;
386         int ret = 1;
387
388         spin_lock(&conn->c_lock);
389
390         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
391                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
392                         ret = 0;
393                 break;
394         }
395
396         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
397                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
398                         ret = 0;
399                 break;
400         }
401
402         spin_unlock(&conn->c_lock);
403
404         return ret;
405 }
406
407 /*
408  * This is pretty similar to what happens below in the ACK
409  * handling code - except that we call here as soon as we get
410  * the IB send completion on the RDMA op and the accompanying
411  * message.
412  */
413 void rds_rdma_send_complete(struct rds_message *rm, int status)
414 {
415         struct rds_sock *rs = NULL;
416         struct rds_rdma_op *ro;
417         struct rds_notifier *notifier;
418
419         spin_lock(&rm->m_rs_lock);
420
421         ro = rm->m_rdma_op;
422         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
423             ro && ro->r_notify && ro->r_notifier) {
424                 notifier = ro->r_notifier;
425                 rs = rm->m_rs;
426                 sock_hold(rds_rs_to_sk(rs));
427
428                 notifier->n_status = status;
429                 spin_lock(&rs->rs_lock);
430                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
431                 spin_unlock(&rs->rs_lock);
432
433                 ro->r_notifier = NULL;
434         }
435
436         spin_unlock(&rm->m_rs_lock);
437
438         if (rs) {
439                 rds_wake_sk_sleep(rs);
440                 sock_put(rds_rs_to_sk(rs));
441         }
442 }
443 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
444
445 /*
446  * This is the same as rds_rdma_send_complete except we
447  * don't do any locking - we have all the ingredients (message,
448  * socket, socket lock) and can just move the notifier.
449  */
450 static inline void
451 __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
452 {
453         struct rds_rdma_op *ro;
454
455         ro = rm->m_rdma_op;
456         if (ro && ro->r_notify && ro->r_notifier) {
457                 ro->r_notifier->n_status = status;
458                 list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue);
459                 ro->r_notifier = NULL;
460         }
461
462         /* No need to wake the app - caller does this */
463 }
464
465 /*
466  * This is called from the IB send completion when we detect
467  * a RDMA operation that failed with remote access error.
468  * So speed is not an issue here.
469  */
470 struct rds_message *rds_send_get_message(struct rds_connection *conn,
471                                          struct rds_rdma_op *op)
472 {
473         struct rds_message *rm, *tmp, *found = NULL;
474         unsigned long flags;
475
476         spin_lock_irqsave(&conn->c_lock, flags);
477
478         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
479                 if (rm->m_rdma_op == op) {
480                         atomic_inc(&rm->m_refcount);
481                         found = rm;
482                         goto out;
483                 }
484         }
485
486         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
487                 if (rm->m_rdma_op == op) {
488                         atomic_inc(&rm->m_refcount);
489                         found = rm;
490                         break;
491                 }
492         }
493
494 out:
495         spin_unlock_irqrestore(&conn->c_lock, flags);
496
497         return found;
498 }
499 EXPORT_SYMBOL_GPL(rds_send_get_message);
500
501 /*
502  * This removes messages from the socket's list if they're on it.  The list
503  * argument must be private to the caller, we must be able to modify it
504  * without locks.  The messages must have a reference held for their
505  * position on the list.  This function will drop that reference after
506  * removing the messages from the 'messages' list regardless of if it found
507  * the messages on the socket list or not.
508  */
509 void rds_send_remove_from_sock(struct list_head *messages, int status)
510 {
511         unsigned long flags;
512         struct rds_sock *rs = NULL;
513         struct rds_message *rm;
514
515         while (!list_empty(messages)) {
516                 int was_on_sock = 0;
517
518                 rm = list_entry(messages->next, struct rds_message,
519                                 m_conn_item);
520                 list_del_init(&rm->m_conn_item);
521
522                 /*
523                  * If we see this flag cleared then we're *sure* that someone
524                  * else beat us to removing it from the sock.  If we race
525                  * with their flag update we'll get the lock and then really
526                  * see that the flag has been cleared.
527                  *
528                  * The message spinlock makes sure nobody clears rm->m_rs
529                  * while we're messing with it. It does not prevent the
530                  * message from being removed from the socket, though.
531                  */
532                 spin_lock_irqsave(&rm->m_rs_lock, flags);
533                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
534                         goto unlock_and_drop;
535
536                 if (rs != rm->m_rs) {
537                         if (rs) {
538                                 rds_wake_sk_sleep(rs);
539                                 sock_put(rds_rs_to_sk(rs));
540                         }
541                         rs = rm->m_rs;
542                         sock_hold(rds_rs_to_sk(rs));
543                 }
544                 spin_lock(&rs->rs_lock);
545
546                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
547                         struct rds_rdma_op *ro = rm->m_rdma_op;
548                         struct rds_notifier *notifier;
549
550                         list_del_init(&rm->m_sock_item);
551                         rds_send_sndbuf_remove(rs, rm);
552
553                         if (ro && ro->r_notifier && (status || ro->r_notify)) {
554                                 notifier = ro->r_notifier;
555                                 list_add_tail(&notifier->n_list,
556                                                 &rs->rs_notify_queue);
557                                 if (!notifier->n_status)
558                                         notifier->n_status = status;
559                                 rm->m_rdma_op->r_notifier = NULL;
560                         }
561                         was_on_sock = 1;
562                         rm->m_rs = NULL;
563                 }
564                 spin_unlock(&rs->rs_lock);
565
566 unlock_and_drop:
567                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
568                 rds_message_put(rm);
569                 if (was_on_sock)
570                         rds_message_put(rm);
571         }
572
573         if (rs) {
574                 rds_wake_sk_sleep(rs);
575                 sock_put(rds_rs_to_sk(rs));
576         }
577 }
578
579 /*
580  * Transports call here when they've determined that the receiver queued
581  * messages up to, and including, the given sequence number.  Messages are
582  * moved to the retrans queue when rds_send_xmit picks them off the send
583  * queue. This means that in the TCP case, the message may not have been
584  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
585  * checks the RDS_MSG_HAS_ACK_SEQ bit.
586  *
587  * XXX It's not clear to me how this is safely serialized with socket
588  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
589  */
590 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
591                          is_acked_func is_acked)
592 {
593         struct rds_message *rm, *tmp;
594         unsigned long flags;
595         LIST_HEAD(list);
596
597         spin_lock_irqsave(&conn->c_lock, flags);
598
599         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
600                 if (!rds_send_is_acked(rm, ack, is_acked))
601                         break;
602
603                 list_move(&rm->m_conn_item, &list);
604                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
605         }
606
607         /* order flag updates with spin locks */
608         if (!list_empty(&list))
609                 smp_mb__after_clear_bit();
610
611         spin_unlock_irqrestore(&conn->c_lock, flags);
612
613         /* now remove the messages from the sock list as needed */
614         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
615 }
616 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
617
618 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
619 {
620         struct rds_message *rm, *tmp;
621         struct rds_connection *conn;
622         unsigned long flags, flags2;
623         LIST_HEAD(list);
624         int wake = 0;
625
626         /* get all the messages we're dropping under the rs lock */
627         spin_lock_irqsave(&rs->rs_lock, flags);
628
629         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
630                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
631                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
632                         continue;
633
634                 wake = 1;
635                 list_move(&rm->m_sock_item, &list);
636                 rds_send_sndbuf_remove(rs, rm);
637                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
638         }
639
640         /* order flag updates with the rs lock */
641         if (wake)
642                 smp_mb__after_clear_bit();
643
644         spin_unlock_irqrestore(&rs->rs_lock, flags);
645
646         conn = NULL;
647
648         /* now remove the messages from the conn list as needed */
649         list_for_each_entry(rm, &list, m_sock_item) {
650                 /* We do this here rather than in the loop above, so that
651                  * we don't have to nest m_rs_lock under rs->rs_lock */
652                 spin_lock_irqsave(&rm->m_rs_lock, flags2);
653                 /* If this is a RDMA operation, notify the app. */
654                 spin_lock(&rs->rs_lock);
655                 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
656                 spin_unlock(&rs->rs_lock);
657                 rm->m_rs = NULL;
658                 spin_unlock_irqrestore(&rm->m_rs_lock, flags2);
659
660                 /*
661                  * If we see this flag cleared then we're *sure* that someone
662                  * else beat us to removing it from the conn.  If we race
663                  * with their flag update we'll get the lock and then really
664                  * see that the flag has been cleared.
665                  */
666                 if (!test_bit(RDS_MSG_ON_CONN, &rm->m_flags))
667                         continue;
668
669                 if (conn != rm->m_inc.i_conn) {
670                         if (conn)
671                                 spin_unlock_irqrestore(&conn->c_lock, flags);
672                         conn = rm->m_inc.i_conn;
673                         spin_lock_irqsave(&conn->c_lock, flags);
674                 }
675
676                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
677                         list_del_init(&rm->m_conn_item);
678                         rds_message_put(rm);
679                 }
680         }
681
682         if (conn)
683                 spin_unlock_irqrestore(&conn->c_lock, flags);
684
685         if (wake)
686                 rds_wake_sk_sleep(rs);
687
688         while (!list_empty(&list)) {
689                 rm = list_entry(list.next, struct rds_message, m_sock_item);
690                 list_del_init(&rm->m_sock_item);
691
692                 rds_message_wait(rm);
693                 rds_message_put(rm);
694         }
695 }
696
697 /*
698  * we only want this to fire once so we use the callers 'queued'.  It's
699  * possible that another thread can race with us and remove the
700  * message from the flow with RDS_CANCEL_SENT_TO.
701  */
702 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
703                              struct rds_message *rm, __be16 sport,
704                              __be16 dport, int *queued)
705 {
706         unsigned long flags;
707         u32 len;
708
709         if (*queued)
710                 goto out;
711
712         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
713
714         /* this is the only place which holds both the socket's rs_lock
715          * and the connection's c_lock */
716         spin_lock_irqsave(&rs->rs_lock, flags);
717
718         /*
719          * If there is a little space in sndbuf, we don't queue anything,
720          * and userspace gets -EAGAIN. But poll() indicates there's send
721          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
722          * freed up by incoming acks. So we check the *old* value of
723          * rs_snd_bytes here to allow the last msg to exceed the buffer,
724          * and poll() now knows no more data can be sent.
725          */
726         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
727                 rs->rs_snd_bytes += len;
728
729                 /* let recv side know we are close to send space exhaustion.
730                  * This is probably not the optimal way to do it, as this
731                  * means we set the flag on *all* messages as soon as our
732                  * throughput hits a certain threshold.
733                  */
734                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
735                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
736
737                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
738                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
739                 rds_message_addref(rm);
740                 rm->m_rs = rs;
741
742                 /* The code ordering is a little weird, but we're
743                    trying to minimize the time we hold c_lock */
744                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
745                 rm->m_inc.i_conn = conn;
746                 rds_message_addref(rm);
747
748                 spin_lock(&conn->c_lock);
749                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
750                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
751                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
752                 spin_unlock(&conn->c_lock);
753
754                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
755                          rm, len, rs, rs->rs_snd_bytes,
756                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
757
758                 *queued = 1;
759         }
760
761         spin_unlock_irqrestore(&rs->rs_lock, flags);
762 out:
763         return *queued;
764 }
765
766 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
767                          struct msghdr *msg, int *allocated_mr)
768 {
769         struct cmsghdr *cmsg;
770         int ret = 0;
771
772         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
773                 if (!CMSG_OK(msg, cmsg))
774                         return -EINVAL;
775
776                 if (cmsg->cmsg_level != SOL_RDS)
777                         continue;
778
779                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
780                  * rm->m_rdma_cookie and rm->m_rdma_mr.
781                  */
782                 switch (cmsg->cmsg_type) {
783                 case RDS_CMSG_RDMA_ARGS:
784                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
785                         break;
786
787                 case RDS_CMSG_RDMA_DEST:
788                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
789                         break;
790
791                 case RDS_CMSG_RDMA_MAP:
792                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
793                         if (!ret)
794                                 *allocated_mr = 1;
795                         break;
796
797                 default:
798                         return -EINVAL;
799                 }
800
801                 if (ret)
802                         break;
803         }
804
805         return ret;
806 }
807
808 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
809                 size_t payload_len)
810 {
811         struct sock *sk = sock->sk;
812         struct rds_sock *rs = rds_sk_to_rs(sk);
813         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
814         __be32 daddr;
815         __be16 dport;
816         struct rds_message *rm = NULL;
817         struct rds_connection *conn;
818         int ret = 0;
819         int queued = 0, allocated_mr = 0;
820         int nonblock = msg->msg_flags & MSG_DONTWAIT;
821         long timeo = sock_sndtimeo(sk, nonblock);
822
823         /* Mirror Linux UDP mirror of BSD error message compatibility */
824         /* XXX: Perhaps MSG_MORE someday */
825         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
826                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
827                 ret = -EOPNOTSUPP;
828                 goto out;
829         }
830
831         if (msg->msg_namelen) {
832                 /* XXX fail non-unicast destination IPs? */
833                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
834                         ret = -EINVAL;
835                         goto out;
836                 }
837                 daddr = usin->sin_addr.s_addr;
838                 dport = usin->sin_port;
839         } else {
840                 /* We only care about consistency with ->connect() */
841                 lock_sock(sk);
842                 daddr = rs->rs_conn_addr;
843                 dport = rs->rs_conn_port;
844                 release_sock(sk);
845         }
846
847         /* racing with another thread binding seems ok here */
848         if (daddr == 0 || rs->rs_bound_addr == 0) {
849                 ret = -ENOTCONN; /* XXX not a great errno */
850                 goto out;
851         }
852
853         rm = rds_message_copy_from_user(msg->msg_iov, payload_len);
854         if (IS_ERR(rm)) {
855                 ret = PTR_ERR(rm);
856                 rm = NULL;
857                 goto out;
858         }
859
860         rm->m_daddr = daddr;
861
862         /* rds_conn_create has a spinlock that runs with IRQ off.
863          * Caching the conn in the socket helps a lot. */
864         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
865                 conn = rs->rs_conn;
866         else {
867                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
868                                         rs->rs_transport,
869                                         sock->sk->sk_allocation);
870                 if (IS_ERR(conn)) {
871                         ret = PTR_ERR(conn);
872                         goto out;
873                 }
874                 rs->rs_conn = conn;
875         }
876
877         /* Parse any control messages the user may have included. */
878         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
879         if (ret)
880                 goto out;
881
882         if ((rm->m_rdma_cookie || rm->m_rdma_op) &&
883             conn->c_trans->xmit_rdma == NULL) {
884                 if (printk_ratelimit())
885                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
886                                 rm->m_rdma_op, conn->c_trans->xmit_rdma);
887                 ret = -EOPNOTSUPP;
888                 goto out;
889         }
890
891         /* If the connection is down, trigger a connect. We may
892          * have scheduled a delayed reconnect however - in this case
893          * we should not interfere.
894          */
895         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
896             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
897                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
898
899         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
900         if (ret) {
901                 rs->rs_seen_congestion = 1;
902                 goto out;
903         }
904
905         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
906                                   dport, &queued)) {
907                 rds_stats_inc(s_send_queue_full);
908                 /* XXX make sure this is reasonable */
909                 if (payload_len > rds_sk_sndbuf(rs)) {
910                         ret = -EMSGSIZE;
911                         goto out;
912                 }
913                 if (nonblock) {
914                         ret = -EAGAIN;
915                         goto out;
916                 }
917
918                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
919                                         rds_send_queue_rm(rs, conn, rm,
920                                                           rs->rs_bound_port,
921                                                           dport,
922                                                           &queued),
923                                         timeo);
924                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
925                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
926                         continue;
927
928                 ret = timeo;
929                 if (ret == 0)
930                         ret = -ETIMEDOUT;
931                 goto out;
932         }
933
934         /*
935          * By now we've committed to the send.  We reuse rds_send_worker()
936          * to retry sends in the rds thread if the transport asks us to.
937          */
938         rds_stats_inc(s_send_queued);
939
940         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
941                 rds_send_worker(&conn->c_send_w.work);
942
943         rds_message_put(rm);
944         return payload_len;
945
946 out:
947         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
948          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
949          * or in any other way, we need to destroy the MR again */
950         if (allocated_mr)
951                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
952
953         if (rm)
954                 rds_message_put(rm);
955         return ret;
956 }
957
958 /*
959  * Reply to a ping packet.
960  */
961 int
962 rds_send_pong(struct rds_connection *conn, __be16 dport)
963 {
964         struct rds_message *rm;
965         unsigned long flags;
966         int ret = 0;
967
968         rm = rds_message_alloc(0, GFP_ATOMIC);
969         if (rm == NULL) {
970                 ret = -ENOMEM;
971                 goto out;
972         }
973
974         rm->m_daddr = conn->c_faddr;
975
976         /* If the connection is down, trigger a connect. We may
977          * have scheduled a delayed reconnect however - in this case
978          * we should not interfere.
979          */
980         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
981             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
982                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
983
984         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
985         if (ret)
986                 goto out;
987
988         spin_lock_irqsave(&conn->c_lock, flags);
989         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
990         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
991         rds_message_addref(rm);
992         rm->m_inc.i_conn = conn;
993
994         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
995                                     conn->c_next_tx_seq);
996         conn->c_next_tx_seq++;
997         spin_unlock_irqrestore(&conn->c_lock, flags);
998
999         rds_stats_inc(s_send_queued);
1000         rds_stats_inc(s_send_pong);
1001
1002         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1003         rds_message_put(rm);
1004         return 0;
1005
1006 out:
1007         if (rm)
1008                 rds_message_put(rm);
1009         return ret;
1010 }