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