28d09447207b8b765b3df543acbeffa544fde906
[linux-3.10.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 &&
168                     conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
169                     conn->c_xmit_sg == rm->data.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 && test_and_clear_bit(0, &conn->c_map_queued)) {
184                         if (conn->c_trans->xmit_cong_map) {
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) {
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) {
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->rdma.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->rdma.m_rdma_op && !conn->c_xmit_rdma_sent) {
272                         ret = conn->c_trans->xmit_rdma(conn, rm->rdma.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->data.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->data.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->data.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         unsigned long flags;
419
420         spin_lock_irqsave(&rm->m_rs_lock, flags);
421
422         ro = rm->rdma.m_rdma_op;
423         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
424             ro && ro->r_notify && ro->r_notifier) {
425                 notifier = ro->r_notifier;
426                 rs = rm->m_rs;
427                 sock_hold(rds_rs_to_sk(rs));
428
429                 notifier->n_status = status;
430                 spin_lock(&rs->rs_lock);
431                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
432                 spin_unlock(&rs->rs_lock);
433
434                 ro->r_notifier = NULL;
435         }
436
437         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
438
439         if (rs) {
440                 rds_wake_sk_sleep(rs);
441                 sock_put(rds_rs_to_sk(rs));
442         }
443 }
444 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
445
446 /*
447  * This is the same as rds_rdma_send_complete except we
448  * don't do any locking - we have all the ingredients (message,
449  * socket, socket lock) and can just move the notifier.
450  */
451 static inline void
452 __rds_rdma_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
453 {
454         struct rds_rdma_op *ro;
455
456         ro = rm->rdma.m_rdma_op;
457         if (ro && ro->r_notify && ro->r_notifier) {
458                 ro->r_notifier->n_status = status;
459                 list_add_tail(&ro->r_notifier->n_list, &rs->rs_notify_queue);
460                 ro->r_notifier = NULL;
461         }
462
463         /* No need to wake the app - caller does this */
464 }
465
466 /*
467  * This is called from the IB send completion when we detect
468  * a RDMA operation that failed with remote access error.
469  * So speed is not an issue here.
470  */
471 struct rds_message *rds_send_get_message(struct rds_connection *conn,
472                                          struct rds_rdma_op *op)
473 {
474         struct rds_message *rm, *tmp, *found = NULL;
475         unsigned long flags;
476
477         spin_lock_irqsave(&conn->c_lock, flags);
478
479         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
480                 if (rm->rdma.m_rdma_op == op) {
481                         atomic_inc(&rm->m_refcount);
482                         found = rm;
483                         goto out;
484                 }
485         }
486
487         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
488                 if (rm->rdma.m_rdma_op == op) {
489                         atomic_inc(&rm->m_refcount);
490                         found = rm;
491                         break;
492                 }
493         }
494
495 out:
496         spin_unlock_irqrestore(&conn->c_lock, flags);
497
498         return found;
499 }
500 EXPORT_SYMBOL_GPL(rds_send_get_message);
501
502 /*
503  * This removes messages from the socket's list if they're on it.  The list
504  * argument must be private to the caller, we must be able to modify it
505  * without locks.  The messages must have a reference held for their
506  * position on the list.  This function will drop that reference after
507  * removing the messages from the 'messages' list regardless of if it found
508  * the messages on the socket list or not.
509  */
510 void rds_send_remove_from_sock(struct list_head *messages, int status)
511 {
512         unsigned long flags;
513         struct rds_sock *rs = NULL;
514         struct rds_message *rm;
515
516         while (!list_empty(messages)) {
517                 int was_on_sock = 0;
518
519                 rm = list_entry(messages->next, struct rds_message,
520                                 m_conn_item);
521                 list_del_init(&rm->m_conn_item);
522
523                 /*
524                  * If we see this flag cleared then we're *sure* that someone
525                  * else beat us to removing it from the sock.  If we race
526                  * with their flag update we'll get the lock and then really
527                  * see that the flag has been cleared.
528                  *
529                  * The message spinlock makes sure nobody clears rm->m_rs
530                  * while we're messing with it. It does not prevent the
531                  * message from being removed from the socket, though.
532                  */
533                 spin_lock_irqsave(&rm->m_rs_lock, flags);
534                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
535                         goto unlock_and_drop;
536
537                 if (rs != rm->m_rs) {
538                         if (rs) {
539                                 rds_wake_sk_sleep(rs);
540                                 sock_put(rds_rs_to_sk(rs));
541                         }
542                         rs = rm->m_rs;
543                         sock_hold(rds_rs_to_sk(rs));
544                 }
545                 spin_lock(&rs->rs_lock);
546
547                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
548                         struct rds_rdma_op *ro = rm->rdma.m_rdma_op;
549                         struct rds_notifier *notifier;
550
551                         list_del_init(&rm->m_sock_item);
552                         rds_send_sndbuf_remove(rs, rm);
553
554                         if (ro && ro->r_notifier && (status || ro->r_notify)) {
555                                 notifier = ro->r_notifier;
556                                 list_add_tail(&notifier->n_list,
557                                                 &rs->rs_notify_queue);
558                                 if (!notifier->n_status)
559                                         notifier->n_status = status;
560                                 rm->rdma.m_rdma_op->r_notifier = NULL;
561                         }
562                         was_on_sock = 1;
563                         rm->m_rs = NULL;
564                 }
565                 spin_unlock(&rs->rs_lock);
566
567 unlock_and_drop:
568                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
569                 rds_message_put(rm);
570                 if (was_on_sock)
571                         rds_message_put(rm);
572         }
573
574         if (rs) {
575                 rds_wake_sk_sleep(rs);
576                 sock_put(rds_rs_to_sk(rs));
577         }
578 }
579
580 /*
581  * Transports call here when they've determined that the receiver queued
582  * messages up to, and including, the given sequence number.  Messages are
583  * moved to the retrans queue when rds_send_xmit picks them off the send
584  * queue. This means that in the TCP case, the message may not have been
585  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
586  * checks the RDS_MSG_HAS_ACK_SEQ bit.
587  *
588  * XXX It's not clear to me how this is safely serialized with socket
589  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
590  */
591 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
592                          is_acked_func is_acked)
593 {
594         struct rds_message *rm, *tmp;
595         unsigned long flags;
596         LIST_HEAD(list);
597
598         spin_lock_irqsave(&conn->c_lock, flags);
599
600         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
601                 if (!rds_send_is_acked(rm, ack, is_acked))
602                         break;
603
604                 list_move(&rm->m_conn_item, &list);
605                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
606         }
607
608         /* order flag updates with spin locks */
609         if (!list_empty(&list))
610                 smp_mb__after_clear_bit();
611
612         spin_unlock_irqrestore(&conn->c_lock, flags);
613
614         /* now remove the messages from the sock list as needed */
615         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
616 }
617 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
618
619 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
620 {
621         struct rds_message *rm, *tmp;
622         struct rds_connection *conn;
623         unsigned long flags;
624         LIST_HEAD(list);
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                 list_move(&rm->m_sock_item, &list);
635                 rds_send_sndbuf_remove(rs, rm);
636                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
637         }
638
639         /* order flag updates with the rs lock */
640         smp_mb__after_clear_bit();
641
642         spin_unlock_irqrestore(&rs->rs_lock, flags);
643
644         if (list_empty(&list))
645                 return;
646
647         /* Remove the messages from the conn */
648         list_for_each_entry(rm, &list, m_sock_item) {
649
650                 conn = rm->m_inc.i_conn;
651
652                 spin_lock_irqsave(&conn->c_lock, flags);
653                 /*
654                  * Maybe someone else beat us to removing rm from the conn.
655                  * If we race with their flag update we'll get the lock and
656                  * then really see that the flag has been cleared.
657                  */
658                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
659                         spin_unlock_irqrestore(&conn->c_lock, flags);
660                         continue;
661                 }
662                 list_del_init(&rm->m_conn_item);
663                 spin_unlock_irqrestore(&conn->c_lock, flags);
664
665                 /*
666                  * Couldn't grab m_rs_lock in top loop (lock ordering),
667                  * but we can now.
668                  */
669                 spin_lock_irqsave(&rm->m_rs_lock, flags);
670
671                 spin_lock(&rs->rs_lock);
672                 __rds_rdma_send_complete(rs, rm, RDS_RDMA_CANCELED);
673                 spin_unlock(&rs->rs_lock);
674
675                 rm->m_rs = NULL;
676                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
677
678                 rds_message_put(rm);
679         }
680
681         rds_wake_sk_sleep(rs);
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 /*
762  * rds_message is getting to be quite complicated, and we'd like to allocate
763  * it all in one go. This figures out how big it needs to be up front.
764  */
765 static int rds_rm_size(struct msghdr *msg, int data_len)
766 {
767         int size = 0;
768
769         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
770
771         return size;
772 }
773
774 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
775                          struct msghdr *msg, int *allocated_mr)
776 {
777         struct cmsghdr *cmsg;
778         int ret = 0;
779
780         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
781                 if (!CMSG_OK(msg, cmsg))
782                         return -EINVAL;
783
784                 if (cmsg->cmsg_level != SOL_RDS)
785                         continue;
786
787                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
788                  * rm->m_rdma_cookie and rm->m_rdma_mr.
789                  */
790                 switch (cmsg->cmsg_type) {
791                 case RDS_CMSG_RDMA_ARGS:
792                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
793                         break;
794
795                 case RDS_CMSG_RDMA_DEST:
796                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
797                         break;
798
799                 case RDS_CMSG_RDMA_MAP:
800                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
801                         if (!ret)
802                                 *allocated_mr = 1;
803                         break;
804
805                 default:
806                         return -EINVAL;
807                 }
808
809                 if (ret)
810                         break;
811         }
812
813         return ret;
814 }
815
816 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
817                 size_t payload_len)
818 {
819         struct sock *sk = sock->sk;
820         struct rds_sock *rs = rds_sk_to_rs(sk);
821         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
822         __be32 daddr;
823         __be16 dport;
824         struct rds_message *rm = NULL;
825         struct rds_connection *conn;
826         int ret = 0;
827         int queued = 0, allocated_mr = 0;
828         int nonblock = msg->msg_flags & MSG_DONTWAIT;
829         long timeo = sock_sndtimeo(sk, nonblock);
830
831         /* Mirror Linux UDP mirror of BSD error message compatibility */
832         /* XXX: Perhaps MSG_MORE someday */
833         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
834                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
835                 ret = -EOPNOTSUPP;
836                 goto out;
837         }
838
839         if (msg->msg_namelen) {
840                 /* XXX fail non-unicast destination IPs? */
841                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
842                         ret = -EINVAL;
843                         goto out;
844                 }
845                 daddr = usin->sin_addr.s_addr;
846                 dport = usin->sin_port;
847         } else {
848                 /* We only care about consistency with ->connect() */
849                 lock_sock(sk);
850                 daddr = rs->rs_conn_addr;
851                 dport = rs->rs_conn_port;
852                 release_sock(sk);
853         }
854
855         /* racing with another thread binding seems ok here */
856         if (daddr == 0 || rs->rs_bound_addr == 0) {
857                 ret = -ENOTCONN; /* XXX not a great errno */
858                 goto out;
859         }
860
861         /* size of rm including all sgs */
862         ret = rds_rm_size(msg, payload_len);
863         if (ret < 0)
864                 goto out;
865
866         rm = rds_message_alloc(ret, GFP_KERNEL);
867         if (!rm) {
868                 ret = -ENOMEM;
869                 goto out;
870         }
871
872         rm->data.m_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
873         /* XXX fix this to not allocate memory */
874         ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
875         if (ret)
876                 goto out;
877
878         rm->m_daddr = daddr;
879
880         /* rds_conn_create has a spinlock that runs with IRQ off.
881          * Caching the conn in the socket helps a lot. */
882         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
883                 conn = rs->rs_conn;
884         else {
885                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
886                                         rs->rs_transport,
887                                         sock->sk->sk_allocation);
888                 if (IS_ERR(conn)) {
889                         ret = PTR_ERR(conn);
890                         goto out;
891                 }
892                 rs->rs_conn = conn;
893         }
894
895         /* Parse any control messages the user may have included. */
896         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
897         if (ret)
898                 goto out;
899
900         if ((rm->m_rdma_cookie || rm->rdma.m_rdma_op) &&
901             !conn->c_trans->xmit_rdma) {
902                 if (printk_ratelimit())
903                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
904                                 rm->rdma.m_rdma_op, conn->c_trans->xmit_rdma);
905                 ret = -EOPNOTSUPP;
906                 goto out;
907         }
908
909         /* If the connection is down, trigger a connect. We may
910          * have scheduled a delayed reconnect however - in this case
911          * we should not interfere.
912          */
913         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
914             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
915                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
916
917         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
918         if (ret) {
919                 rs->rs_seen_congestion = 1;
920                 goto out;
921         }
922
923         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
924                                   dport, &queued)) {
925                 rds_stats_inc(s_send_queue_full);
926                 /* XXX make sure this is reasonable */
927                 if (payload_len > rds_sk_sndbuf(rs)) {
928                         ret = -EMSGSIZE;
929                         goto out;
930                 }
931                 if (nonblock) {
932                         ret = -EAGAIN;
933                         goto out;
934                 }
935
936                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
937                                         rds_send_queue_rm(rs, conn, rm,
938                                                           rs->rs_bound_port,
939                                                           dport,
940                                                           &queued),
941                                         timeo);
942                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
943                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
944                         continue;
945
946                 ret = timeo;
947                 if (ret == 0)
948                         ret = -ETIMEDOUT;
949                 goto out;
950         }
951
952         /*
953          * By now we've committed to the send.  We reuse rds_send_worker()
954          * to retry sends in the rds thread if the transport asks us to.
955          */
956         rds_stats_inc(s_send_queued);
957
958         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
959                 rds_send_worker(&conn->c_send_w.work);
960
961         rds_message_put(rm);
962         return payload_len;
963
964 out:
965         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
966          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
967          * or in any other way, we need to destroy the MR again */
968         if (allocated_mr)
969                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
970
971         if (rm)
972                 rds_message_put(rm);
973         return ret;
974 }
975
976 /*
977  * Reply to a ping packet.
978  */
979 int
980 rds_send_pong(struct rds_connection *conn, __be16 dport)
981 {
982         struct rds_message *rm;
983         unsigned long flags;
984         int ret = 0;
985
986         rm = rds_message_alloc(0, GFP_ATOMIC);
987         if (!rm) {
988                 ret = -ENOMEM;
989                 goto out;
990         }
991
992         rm->m_daddr = conn->c_faddr;
993
994         /* If the connection is down, trigger a connect. We may
995          * have scheduled a delayed reconnect however - in this case
996          * we should not interfere.
997          */
998         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
999             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1000                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1001
1002         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1003         if (ret)
1004                 goto out;
1005
1006         spin_lock_irqsave(&conn->c_lock, flags);
1007         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1008         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1009         rds_message_addref(rm);
1010         rm->m_inc.i_conn = conn;
1011
1012         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1013                                     conn->c_next_tx_seq);
1014         conn->c_next_tx_seq++;
1015         spin_unlock_irqrestore(&conn->c_lock, flags);
1016
1017         rds_stats_inc(s_send_queued);
1018         rds_stats_inc(s_send_pong);
1019
1020         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
1021         rds_message_put(rm);
1022         return 0;
1023
1024 out:
1025         if (rm)
1026                 rds_message_put(rm);
1027         return ret;
1028 }