RDS/IW+IB: Allow max credit advertise window.
[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 <net/sock.h>
35 #include <linux/in.h>
36 #include <linux/list.h>
37
38 #include "rds.h"
39 #include "rdma.h"
40
41 /* When transmitting messages in rds_send_xmit, we need to emerge from
42  * time to time and briefly release the CPU. Otherwise the softlock watchdog
43  * will kick our shin.
44  * Also, it seems fairer to not let one busy connection stall all the
45  * others.
46  *
47  * send_batch_count is the number of times we'll loop in send_xmit. Setting
48  * it to 0 will restore the old behavior (where we looped until we had
49  * drained the queue).
50  */
51 static int send_batch_count = 64;
52 module_param(send_batch_count, int, 0444);
53 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
54
55 /*
56  * Reset the send state. Caller must hold c_send_lock when calling here.
57  */
58 void rds_send_reset(struct rds_connection *conn)
59 {
60         struct rds_message *rm, *tmp;
61         unsigned long flags;
62
63         if (conn->c_xmit_rm) {
64                 /* Tell the user the RDMA op is no longer mapped by the
65                  * transport. This isn't entirely true (it's flushed out
66                  * independently) but as the connection is down, there's
67                  * no ongoing RDMA to/from that memory */
68                 rds_message_unmapped(conn->c_xmit_rm);
69                 rds_message_put(conn->c_xmit_rm);
70                 conn->c_xmit_rm = NULL;
71         }
72         conn->c_xmit_sg = 0;
73         conn->c_xmit_hdr_off = 0;
74         conn->c_xmit_data_off = 0;
75         conn->c_xmit_rdma_sent = 0;
76
77         conn->c_map_queued = 0;
78
79         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
80         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
81
82         /* Mark messages as retransmissions, and move them to the send q */
83         spin_lock_irqsave(&conn->c_lock, flags);
84         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
85                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
86                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
87         }
88         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
89         spin_unlock_irqrestore(&conn->c_lock, flags);
90 }
91
92 /*
93  * We're making the concious trade-off here to only send one message
94  * down the connection at a time.
95  *   Pro:
96  *      - tx queueing is a simple fifo list
97  *      - reassembly is optional and easily done by transports per conn
98  *      - no per flow rx lookup at all, straight to the socket
99  *      - less per-frag memory and wire overhead
100  *   Con:
101  *      - queued acks can be delayed behind large messages
102  *   Depends:
103  *      - small message latency is higher behind queued large messages
104  *      - large message latency isn't starved by intervening small sends
105  */
106 int rds_send_xmit(struct rds_connection *conn)
107 {
108         struct rds_message *rm;
109         unsigned long flags;
110         unsigned int tmp;
111         unsigned int send_quota = send_batch_count;
112         struct scatterlist *sg;
113         int ret = 0;
114         int was_empty = 0;
115         LIST_HEAD(to_be_dropped);
116
117         /*
118          * sendmsg calls here after having queued its message on the send
119          * queue.  We only have one task feeding the connection at a time.  If
120          * another thread is already feeding the queue then we back off.  This
121          * avoids blocking the caller and trading per-connection data between
122          * caches per message.
123          *
124          * The sem holder will issue a retry if they notice that someone queued
125          * a message after they stopped walking the send queue but before they
126          * dropped the sem.
127          */
128         if (!mutex_trylock(&conn->c_send_lock)) {
129                 rds_stats_inc(s_send_sem_contention);
130                 ret = -ENOMEM;
131                 goto out;
132         }
133
134         if (conn->c_trans->xmit_prepare)
135                 conn->c_trans->xmit_prepare(conn);
136
137         /*
138          * spin trying to push headers and data down the connection until
139          * the connection doens't make forward progress.
140          */
141         while (--send_quota) {
142                 /*
143                  * See if need to send a congestion map update if we're
144                  * between sending messages.  The send_sem protects our sole
145                  * use of c_map_offset and _bytes.
146                  * Note this is used only by transports that define a special
147                  * xmit_cong_map function. For all others, we create allocate
148                  * a cong_map message and treat it just like any other send.
149                  */
150                 if (conn->c_map_bytes) {
151                         ret = conn->c_trans->xmit_cong_map(conn, conn->c_lcong,
152                                                 conn->c_map_offset);
153                         if (ret <= 0)
154                                 break;
155
156                         conn->c_map_offset += ret;
157                         conn->c_map_bytes -= ret;
158                         if (conn->c_map_bytes)
159                                 continue;
160                 }
161
162                 /* If we're done sending the current message, clear the
163                  * offset and S/G temporaries.
164                  */
165                 rm = conn->c_xmit_rm;
166                 if (rm != NULL &&
167                     conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
168                     conn->c_xmit_sg == rm->m_nents) {
169                         conn->c_xmit_rm = NULL;
170                         conn->c_xmit_sg = 0;
171                         conn->c_xmit_hdr_off = 0;
172                         conn->c_xmit_data_off = 0;
173                         conn->c_xmit_rdma_sent = 0;
174
175                         /* Release the reference to the previous message. */
176                         rds_message_put(rm);
177                         rm = NULL;
178                 }
179
180                 /* If we're asked to send a cong map update, do so.
181                  */
182                 if (rm == NULL && test_and_clear_bit(0, &conn->c_map_queued)) {
183                         if (conn->c_trans->xmit_cong_map != NULL) {
184                                 conn->c_map_offset = 0;
185                                 conn->c_map_bytes = sizeof(struct rds_header) +
186                                         RDS_CONG_MAP_BYTES;
187                                 continue;
188                         }
189
190                         rm = rds_cong_update_alloc(conn);
191                         if (IS_ERR(rm)) {
192                                 ret = PTR_ERR(rm);
193                                 break;
194                         }
195
196                         conn->c_xmit_rm = rm;
197                 }
198
199                 /*
200                  * Grab the next message from the send queue, if there is one.
201                  *
202                  * c_xmit_rm holds a ref while we're sending this message down
203                  * the connction.  We can use this ref while holding the
204                  * send_sem.. rds_send_reset() is serialized with it.
205                  */
206                 if (rm == NULL) {
207                         unsigned int len;
208
209                         spin_lock_irqsave(&conn->c_lock, flags);
210
211                         if (!list_empty(&conn->c_send_queue)) {
212                                 rm = list_entry(conn->c_send_queue.next,
213                                                 struct rds_message,
214                                                 m_conn_item);
215                                 rds_message_addref(rm);
216
217                                 /*
218                                  * Move the message from the send queue to the retransmit
219                                  * list right away.
220                                  */
221                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
222                         }
223
224                         spin_unlock_irqrestore(&conn->c_lock, flags);
225
226                         if (rm == NULL) {
227                                 was_empty = 1;
228                                 break;
229                         }
230
231                         /* Unfortunately, the way Infiniband deals with
232                          * RDMA to a bad MR key is by moving the entire
233                          * queue pair to error state. We cold possibly
234                          * recover from that, but right now we drop the
235                          * connection.
236                          * Therefore, we never retransmit messages with RDMA ops.
237                          */
238                         if (rm->m_rdma_op
239                          && test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
240                                 spin_lock_irqsave(&conn->c_lock, flags);
241                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
242                                         list_move(&rm->m_conn_item, &to_be_dropped);
243                                 spin_unlock_irqrestore(&conn->c_lock, flags);
244                                 rds_message_put(rm);
245                                 continue;
246                         }
247
248                         /* Require an ACK every once in a while */
249                         len = ntohl(rm->m_inc.i_hdr.h_len);
250                         if (conn->c_unacked_packets == 0
251                          || conn->c_unacked_bytes < len) {
252                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
253
254                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
255                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
256                                 rds_stats_inc(s_send_ack_required);
257                         } else {
258                                 conn->c_unacked_bytes -= len;
259                                 conn->c_unacked_packets--;
260                         }
261
262                         conn->c_xmit_rm = rm;
263                 }
264
265                 /*
266                  * Try and send an rdma message.  Let's see if we can
267                  * keep this simple and require that the transport either
268                  * send the whole rdma or none of it.
269                  */
270                 if (rm->m_rdma_op && !conn->c_xmit_rdma_sent) {
271                         ret = conn->c_trans->xmit_rdma(conn, rm->m_rdma_op);
272                         if (ret)
273                                 break;
274                         conn->c_xmit_rdma_sent = 1;
275                         /* The transport owns the mapped memory for now.
276                          * You can't unmap it while it's on the send queue */
277                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
278                 }
279
280                 if (conn->c_xmit_hdr_off < sizeof(struct rds_header) ||
281                     conn->c_xmit_sg < rm->m_nents) {
282                         ret = conn->c_trans->xmit(conn, rm,
283                                                   conn->c_xmit_hdr_off,
284                                                   conn->c_xmit_sg,
285                                                   conn->c_xmit_data_off);
286                         if (ret <= 0)
287                                 break;
288
289                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
290                                 tmp = min_t(int, ret,
291                                             sizeof(struct rds_header) -
292                                             conn->c_xmit_hdr_off);
293                                 conn->c_xmit_hdr_off += tmp;
294                                 ret -= tmp;
295                         }
296
297                         sg = &rm->m_sg[conn->c_xmit_sg];
298                         while (ret) {
299                                 tmp = min_t(int, ret, sg->length -
300                                                       conn->c_xmit_data_off);
301                                 conn->c_xmit_data_off += tmp;
302                                 ret -= tmp;
303                                 if (conn->c_xmit_data_off == sg->length) {
304                                         conn->c_xmit_data_off = 0;
305                                         sg++;
306                                         conn->c_xmit_sg++;
307                                         BUG_ON(ret != 0 &&
308                                                conn->c_xmit_sg == rm->m_nents);
309                                 }
310                         }
311                 }
312         }
313
314         /* Nuke any messages we decided not to retransmit. */
315         if (!list_empty(&to_be_dropped))
316                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
317
318         if (conn->c_trans->xmit_complete)
319                 conn->c_trans->xmit_complete(conn);
320
321         /*
322          * We might be racing with another sender who queued a message but
323          * backed off on noticing that we held the c_send_lock.  If we check
324          * for queued messages after dropping the sem then either we'll
325          * see the queued message or the queuer will get the sem.  If we
326          * notice the queued message then we trigger an immediate retry.
327          *
328          * We need to be careful only to do this when we stopped processing
329          * the send queue because it was empty.  It's the only way we
330          * stop processing the loop when the transport hasn't taken
331          * responsibility for forward progress.
332          */
333         mutex_unlock(&conn->c_send_lock);
334
335         if (conn->c_map_bytes || (send_quota == 0 && !was_empty)) {
336                 /* We exhausted the send quota, but there's work left to
337                  * do. Return and (re-)schedule the send worker.
338                  */
339                 ret = -EAGAIN;
340         }
341
342         if (ret == 0 && was_empty) {
343                 /* A simple bit test would be way faster than taking the
344                  * spin lock */
345                 spin_lock_irqsave(&conn->c_lock, flags);
346                 if (!list_empty(&conn->c_send_queue)) {
347                         rds_stats_inc(s_send_sem_queue_raced);
348                         ret = -EAGAIN;
349                 }
350                 spin_unlock_irqrestore(&conn->c_lock, flags);
351         }
352 out:
353         return ret;
354 }
355
356 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
357 {
358         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
359
360         assert_spin_locked(&rs->rs_lock);
361
362         BUG_ON(rs->rs_snd_bytes < len);
363         rs->rs_snd_bytes -= len;
364
365         if (rs->rs_snd_bytes == 0)
366                 rds_stats_inc(s_send_queue_empty);
367 }
368
369 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
370                                     is_acked_func is_acked)
371 {
372         if (is_acked)
373                 return is_acked(rm, ack);
374         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
375 }
376
377 /*
378  * Returns true if there are no messages on the send and retransmit queues
379  * which have a sequence number greater than or equal to the given sequence
380  * number.
381  */
382 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
383 {
384         struct rds_message *rm, *tmp;
385         int ret = 1;
386
387         spin_lock(&conn->c_lock);
388
389         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
390                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
391                         ret = 0;
392                 break;
393         }
394
395         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
396                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
397                         ret = 0;
398                 break;
399         }
400
401         spin_unlock(&conn->c_lock);
402
403         return ret;
404 }
405
406 /*
407  * This is pretty similar to what happens below in the ACK
408  * handling code - except that we call here as soon as we get
409  * the IB send completion on the RDMA op and the accompanying
410  * message.
411  */
412 void rds_rdma_send_complete(struct rds_message *rm, int status)
413 {
414         struct rds_sock *rs = NULL;
415         struct rds_rdma_op *ro;
416         struct rds_notifier *notifier;
417
418         spin_lock(&rm->m_rs_lock);
419
420         ro = rm->m_rdma_op;
421         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
422          && ro && ro->r_notify && ro->r_notifier) {
423                 notifier = ro->r_notifier;
424                 rs = rm->m_rs;
425                 sock_hold(rds_rs_to_sk(rs));
426
427                 notifier->n_status = status;
428                 spin_lock(&rs->rs_lock);
429                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
430                 spin_unlock(&rs->rs_lock);
431
432                 ro->r_notifier = NULL;
433         }
434
435         spin_unlock(&rm->m_rs_lock);
436
437         if (rs) {
438                 rds_wake_sk_sleep(rs);
439                 sock_put(rds_rs_to_sk(rs));
440         }
441 }
442
443 /*
444  * This is the same as rds_rdma_send_complete except we
445  * don't do any locking - we have all the ingredients (message,
446  * socket, socket lock) and can just move the notifier.
447  */
448 static inline void
449 __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
450 {
451         struct rds_rdma_op *ro;
452
453         ro = rm->m_rdma_op;
454         if (ro && ro->r_notify && ro->r_notifier) {
455                 ro->r_notifier->n_status = status;
456                 list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue);
457                 ro->r_notifier = NULL;
458         }
459
460         /* No need to wake the app - caller does this */
461 }
462
463 /*
464  * This is called from the IB send completion when we detect
465  * a RDMA operation that failed with remote access error.
466  * So speed is not an issue here.
467  */
468 struct rds_message *rds_send_get_message(struct rds_connection *conn,
469                                          struct rds_rdma_op *op)
470 {
471         struct rds_message *rm, *tmp, *found = NULL;
472         unsigned long flags;
473
474         spin_lock_irqsave(&conn->c_lock, flags);
475
476         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
477                 if (rm->m_rdma_op == op) {
478                         atomic_inc(&rm->m_refcount);
479                         found = rm;
480                         goto out;
481                 }
482         }
483
484         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
485                 if (rm->m_rdma_op == op) {
486                         atomic_inc(&rm->m_refcount);
487                         found = rm;
488                         break;
489                 }
490         }
491
492 out:
493         spin_unlock_irqrestore(&conn->c_lock, flags);
494
495         return found;
496 }
497
498 /*
499  * This removes messages from the socket's list if they're on it.  The list
500  * argument must be private to the caller, we must be able to modify it
501  * without locks.  The messages must have a reference held for their
502  * position on the list.  This function will drop that reference after
503  * removing the messages from the 'messages' list regardless of if it found
504  * the messages on the socket list or not.
505  */
506 void rds_send_remove_from_sock(struct list_head *messages, int status)
507 {
508         unsigned long flags = 0; /* silence gcc :P */
509         struct rds_sock *rs = NULL;
510         struct rds_message *rm;
511
512         local_irq_save(flags);
513         while (!list_empty(messages)) {
514                 rm = list_entry(messages->next, struct rds_message,
515                                 m_conn_item);
516                 list_del_init(&rm->m_conn_item);
517
518                 /*
519                  * If we see this flag cleared then we're *sure* that someone
520                  * else beat us to removing it from the sock.  If we race
521                  * with their flag update we'll get the lock and then really
522                  * see that the flag has been cleared.
523                  *
524                  * The message spinlock makes sure nobody clears rm->m_rs
525                  * while we're messing with it. It does not prevent the
526                  * message from being removed from the socket, though.
527                  */
528                 spin_lock(&rm->m_rs_lock);
529                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
530                         goto unlock_and_drop;
531
532                 if (rs != rm->m_rs) {
533                         if (rs) {
534                                 spin_unlock(&rs->rs_lock);
535                                 rds_wake_sk_sleep(rs);
536                                 sock_put(rds_rs_to_sk(rs));
537                         }
538                         rs = rm->m_rs;
539                         spin_lock(&rs->rs_lock);
540                         sock_hold(rds_rs_to_sk(rs));
541                 }
542
543                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
544                         struct rds_rdma_op *ro = rm->m_rdma_op;
545                         struct rds_notifier *notifier;
546
547                         list_del_init(&rm->m_sock_item);
548                         rds_send_sndbuf_remove(rs, rm);
549
550                         if (ro && ro->r_notifier
551                            && (status || ro->r_notify)) {
552                                 notifier = ro->r_notifier;
553                                 list_add_tail(&notifier->n_list,
554                                                 &rs->rs_notify_queue);
555                                 if (!notifier->n_status)
556                                         notifier->n_status = status;
557                                 rm->m_rdma_op->r_notifier = NULL;
558                         }
559                         rds_message_put(rm);
560                         rm->m_rs = NULL;
561                 }
562
563 unlock_and_drop:
564                 spin_unlock(&rm->m_rs_lock);
565                 rds_message_put(rm);
566         }
567
568         if (rs) {
569                 spin_unlock(&rs->rs_lock);
570                 rds_wake_sk_sleep(rs);
571                 sock_put(rds_rs_to_sk(rs));
572         }
573         local_irq_restore(flags);
574 }
575
576 /*
577  * Transports call here when they've determined that the receiver queued
578  * messages up to, and including, the given sequence number.  Messages are
579  * moved to the retrans queue when rds_send_xmit picks them off the send
580  * queue. This means that in the TCP case, the message may not have been
581  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
582  * checks the RDS_MSG_HAS_ACK_SEQ bit.
583  *
584  * XXX It's not clear to me how this is safely serialized with socket
585  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
586  */
587 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
588                          is_acked_func is_acked)
589 {
590         struct rds_message *rm, *tmp;
591         unsigned long flags;
592         LIST_HEAD(list);
593
594         spin_lock_irqsave(&conn->c_lock, flags);
595
596         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
597                 if (!rds_send_is_acked(rm, ack, is_acked))
598                         break;
599
600                 list_move(&rm->m_conn_item, &list);
601                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
602         }
603
604         /* order flag updates with spin locks */
605         if (!list_empty(&list))
606                 smp_mb__after_clear_bit();
607
608         spin_unlock_irqrestore(&conn->c_lock, flags);
609
610         /* now remove the messages from the sock list as needed */
611         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
612 }
613
614 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
615 {
616         struct rds_message *rm, *tmp;
617         struct rds_connection *conn;
618         unsigned long flags, flags2;
619         LIST_HEAD(list);
620         int wake = 0;
621
622         /* get all the messages we're dropping under the rs lock */
623         spin_lock_irqsave(&rs->rs_lock, flags);
624
625         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
626                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
627                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
628                         continue;
629
630                 wake = 1;
631                 list_move(&rm->m_sock_item, &list);
632                 rds_send_sndbuf_remove(rs, rm);
633                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
634
635                 /* If this is a RDMA operation, notify the app. */
636                 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
637         }
638
639         /* order flag updates with the rs lock */
640         if (wake)
641                 smp_mb__after_clear_bit();
642
643         spin_unlock_irqrestore(&rs->rs_lock, flags);
644
645         if (wake)
646                 rds_wake_sk_sleep(rs);
647
648         conn = NULL;
649
650         /* now remove the messages from the conn list as needed */
651         list_for_each_entry(rm, &list, m_sock_item) {
652                 /* We do this here rather than in the loop above, so that
653                  * we don't have to nest m_rs_lock under rs->rs_lock */
654                 spin_lock_irqsave(&rm->m_rs_lock, flags2);
655                 rm->m_rs = NULL;
656                 spin_unlock_irqrestore(&rm->m_rs_lock, flags2);
657
658                 /*
659                  * If we see this flag cleared then we're *sure* that someone
660                  * else beat us to removing it from the conn.  If we race
661                  * with their flag update we'll get the lock and then really
662                  * see that the flag has been cleared.
663                  */
664                 if (!test_bit(RDS_MSG_ON_CONN, &rm->m_flags))
665                         continue;
666
667                 if (conn != rm->m_inc.i_conn) {
668                         if (conn)
669                                 spin_unlock_irqrestore(&conn->c_lock, flags);
670                         conn = rm->m_inc.i_conn;
671                         spin_lock_irqsave(&conn->c_lock, flags);
672                 }
673
674                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
675                         list_del_init(&rm->m_conn_item);
676                         rds_message_put(rm);
677                 }
678         }
679
680         if (conn)
681                 spin_unlock_irqrestore(&conn->c_lock, flags);
682
683         while (!list_empty(&list)) {
684                 rm = list_entry(list.next, struct rds_message, m_sock_item);
685                 list_del_init(&rm->m_sock_item);
686
687                 rds_message_wait(rm);
688                 rds_message_put(rm);
689         }
690 }
691
692 /*
693  * we only want this to fire once so we use the callers 'queued'.  It's
694  * possible that another thread can race with us and remove the
695  * message from the flow with RDS_CANCEL_SENT_TO.
696  */
697 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
698                              struct rds_message *rm, __be16 sport,
699                              __be16 dport, int *queued)
700 {
701         unsigned long flags;
702         u32 len;
703
704         if (*queued)
705                 goto out;
706
707         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
708
709         /* this is the only place which holds both the socket's rs_lock
710          * and the connection's c_lock */
711         spin_lock_irqsave(&rs->rs_lock, flags);
712
713         /*
714          * If there is a little space in sndbuf, we don't queue anything,
715          * and userspace gets -EAGAIN. But poll() indicates there's send
716          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
717          * freed up by incoming acks. So we check the *old* value of
718          * rs_snd_bytes here to allow the last msg to exceed the buffer,
719          * and poll() now knows no more data can be sent.
720          */
721         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
722                 rs->rs_snd_bytes += len;
723
724                 /* let recv side know we are close to send space exhaustion.
725                  * This is probably not the optimal way to do it, as this
726                  * means we set the flag on *all* messages as soon as our
727                  * throughput hits a certain threshold.
728                  */
729                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
730                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
731
732                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
733                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
734                 rds_message_addref(rm);
735                 rm->m_rs = rs;
736
737                 /* The code ordering is a little weird, but we're
738                    trying to minimize the time we hold c_lock */
739                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
740                 rm->m_inc.i_conn = conn;
741                 rds_message_addref(rm);
742
743                 spin_lock(&conn->c_lock);
744                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
745                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
746                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
747                 spin_unlock(&conn->c_lock);
748
749                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
750                          rm, len, rs, rs->rs_snd_bytes,
751                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
752
753                 *queued = 1;
754         }
755
756         spin_unlock_irqrestore(&rs->rs_lock, flags);
757 out:
758         return *queued;
759 }
760
761 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
762                          struct msghdr *msg, int *allocated_mr)
763 {
764         struct cmsghdr *cmsg;
765         int ret = 0;
766
767         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
768                 if (!CMSG_OK(msg, cmsg))
769                         return -EINVAL;
770
771                 if (cmsg->cmsg_level != SOL_RDS)
772                         continue;
773
774                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
775                  * rm->m_rdma_cookie and rm->m_rdma_mr.
776                  */
777                 switch (cmsg->cmsg_type) {
778                 case RDS_CMSG_RDMA_ARGS:
779                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
780                         break;
781
782                 case RDS_CMSG_RDMA_DEST:
783                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
784                         break;
785
786                 case RDS_CMSG_RDMA_MAP:
787                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
788                         if (!ret)
789                                 *allocated_mr = 1;
790                         break;
791
792                 default:
793                         return -EINVAL;
794                 }
795
796                 if (ret)
797                         break;
798         }
799
800         return ret;
801 }
802
803 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
804                 size_t payload_len)
805 {
806         struct sock *sk = sock->sk;
807         struct rds_sock *rs = rds_sk_to_rs(sk);
808         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
809         __be32 daddr;
810         __be16 dport;
811         struct rds_message *rm = NULL;
812         struct rds_connection *conn;
813         int ret = 0;
814         int queued = 0, allocated_mr = 0;
815         int nonblock = msg->msg_flags & MSG_DONTWAIT;
816         long timeo = sock_rcvtimeo(sk, nonblock);
817
818         /* Mirror Linux UDP mirror of BSD error message compatibility */
819         /* XXX: Perhaps MSG_MORE someday */
820         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
821                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
822                 ret = -EOPNOTSUPP;
823                 goto out;
824         }
825
826         if (msg->msg_namelen) {
827                 /* XXX fail non-unicast destination IPs? */
828                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
829                         ret = -EINVAL;
830                         goto out;
831                 }
832                 daddr = usin->sin_addr.s_addr;
833                 dport = usin->sin_port;
834         } else {
835                 /* We only care about consistency with ->connect() */
836                 lock_sock(sk);
837                 daddr = rs->rs_conn_addr;
838                 dport = rs->rs_conn_port;
839                 release_sock(sk);
840         }
841
842         /* racing with another thread binding seems ok here */
843         if (daddr == 0 || rs->rs_bound_addr == 0) {
844                 ret = -ENOTCONN; /* XXX not a great errno */
845                 goto out;
846         }
847
848         rm = rds_message_copy_from_user(msg->msg_iov, payload_len);
849         if (IS_ERR(rm)) {
850                 ret = PTR_ERR(rm);
851                 rm = NULL;
852                 goto out;
853         }
854
855         rm->m_daddr = daddr;
856
857         /* Parse any control messages the user may have included. */
858         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
859         if (ret)
860                 goto out;
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         if ((rm->m_rdma_cookie || rm->m_rdma_op)
878          && conn->c_trans->xmit_rdma == NULL) {
879                 if (printk_ratelimit())
880                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
881                                 rm->m_rdma_op, conn->c_trans->xmit_rdma);
882                 ret = -EOPNOTSUPP;
883                 goto out;
884         }
885
886         /* If the connection is down, trigger a connect. We may
887          * have scheduled a delayed reconnect however - in this case
888          * we should not interfere.
889          */
890         if (rds_conn_state(conn) == RDS_CONN_DOWN
891          && !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
892                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
893
894         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
895         if (ret)
896                 goto out;
897
898         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
899                                   dport, &queued)) {
900                 rds_stats_inc(s_send_queue_full);
901                 /* XXX make sure this is reasonable */
902                 if (payload_len > rds_sk_sndbuf(rs)) {
903                         ret = -EMSGSIZE;
904                         goto out;
905                 }
906                 if (nonblock) {
907                         ret = -EAGAIN;
908                         goto out;
909                 }
910
911                 timeo = wait_event_interruptible_timeout(*sk->sk_sleep,
912                                         rds_send_queue_rm(rs, conn, rm,
913                                                           rs->rs_bound_port,
914                                                           dport,
915                                                           &queued),
916                                         timeo);
917                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
918                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
919                         continue;
920
921                 ret = timeo;
922                 if (ret == 0)
923                         ret = -ETIMEDOUT;
924                 goto out;
925         }
926
927         /*
928          * By now we've committed to the send.  We reuse rds_send_worker()
929          * to retry sends in the rds thread if the transport asks us to.
930          */
931         rds_stats_inc(s_send_queued);
932
933         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
934                 rds_send_worker(&conn->c_send_w.work);
935
936         rds_message_put(rm);
937         return payload_len;
938
939 out:
940         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
941          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
942          * or in any other way, we need to destroy the MR again */
943         if (allocated_mr)
944                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
945
946         if (rm)
947                 rds_message_put(rm);
948         return ret;
949 }
950
951 /*
952  * Reply to a ping packet.
953  */
954 int
955 rds_send_pong(struct rds_connection *conn, __be16 dport)
956 {
957         struct rds_message *rm;
958         unsigned long flags;
959         int ret = 0;
960
961         rm = rds_message_alloc(0, GFP_ATOMIC);
962         if (rm == NULL) {
963                 ret = -ENOMEM;
964                 goto out;
965         }
966
967         rm->m_daddr = conn->c_faddr;
968
969         /* If the connection is down, trigger a connect. We may
970          * have scheduled a delayed reconnect however - in this case
971          * we should not interfere.
972          */
973         if (rds_conn_state(conn) == RDS_CONN_DOWN
974          && !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
975                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
976
977         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
978         if (ret)
979                 goto out;
980
981         spin_lock_irqsave(&conn->c_lock, flags);
982         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
983         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
984         rds_message_addref(rm);
985         rm->m_inc.i_conn = conn;
986
987         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
988                                     conn->c_next_tx_seq);
989         conn->c_next_tx_seq++;
990         spin_unlock_irqrestore(&conn->c_lock, flags);
991
992         rds_stats_inc(s_send_queued);
993         rds_stats_inc(s_send_pong);
994
995         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
996         rds_message_put(rm);
997         return 0;
998
999 out:
1000         if (rm)
1001                 rds_message_put(rm);
1002         return ret;
1003 }