d35c43ff792e70ed56dfa1bd372f791638bf921b
[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 printk(KERN_CRIT "send reset unmapping %p\n", rm);
72                 rds_message_unmapped(rm);
73                 spin_unlock_irqrestore(&conn->c_send_lock, flags);
74
75                 rds_message_put(rm);
76         } else {
77                 spin_unlock_irqrestore(&conn->c_send_lock, flags);
78         }
79
80         conn->c_xmit_sg = 0;
81         conn->c_xmit_hdr_off = 0;
82         conn->c_xmit_data_off = 0;
83         conn->c_xmit_atomic_sent = 0;
84         conn->c_xmit_rdma_sent = 0;
85         conn->c_xmit_data_sent = 0;
86
87         conn->c_map_queued = 0;
88
89         conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
90         conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
91
92         /* Mark messages as retransmissions, and move them to the send q */
93         spin_lock_irqsave(&conn->c_lock, flags);
94         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
95                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
96                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
97         }
98         list_splice_init(&conn->c_retrans, &conn->c_send_queue);
99         spin_unlock_irqrestore(&conn->c_lock, flags);
100 }
101
102 /*
103  * We're making the concious trade-off here to only send one message
104  * down the connection at a time.
105  *   Pro:
106  *      - tx queueing is a simple fifo list
107  *      - reassembly is optional and easily done by transports per conn
108  *      - no per flow rx lookup at all, straight to the socket
109  *      - less per-frag memory and wire overhead
110  *   Con:
111  *      - queued acks can be delayed behind large messages
112  *   Depends:
113  *      - small message latency is higher behind queued large messages
114  *      - large message latency isn't starved by intervening small sends
115  */
116 int rds_send_xmit(struct rds_connection *conn)
117 {
118         struct rds_message *rm;
119         unsigned long flags;
120         unsigned int tmp;
121         struct scatterlist *sg;
122         int ret = 0;
123         int gen = 0;
124         LIST_HEAD(to_be_dropped);
125
126 restart:
127         if (!rds_conn_up(conn))
128                 goto out;
129
130         /*
131          * sendmsg calls here after having queued its message on the send
132          * queue.  We only have one task feeding the connection at a time.  If
133          * another thread is already feeding the queue then we back off.  This
134          * avoids blocking the caller and trading per-connection data between
135          * caches per message.
136          */
137         if (!spin_trylock_irqsave(&conn->c_send_lock, flags)) {
138                 rds_stats_inc(s_send_lock_contention);
139                 ret = -ENOMEM;
140                 goto out;
141         }
142         atomic_inc(&conn->c_senders);
143
144         if (conn->c_trans->xmit_prepare)
145                 conn->c_trans->xmit_prepare(conn);
146
147         gen = atomic_inc_return(&conn->c_send_generation);
148
149         /*
150          * spin trying to push headers and data down the connection until
151          * the connection doesn't make forward progress.
152          */
153         while (1) {
154
155                 rm = conn->c_xmit_rm;
156
157                 /*
158                  * If between sending messages, we can send a pending congestion
159                  * map update.
160                  */
161                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
162                         rm = rds_cong_update_alloc(conn);
163                         if (IS_ERR(rm)) {
164                                 ret = PTR_ERR(rm);
165                                 break;
166                         }
167                         rm->data.op_active = 1;
168
169                         conn->c_xmit_rm = rm;
170                 }
171
172                 /*
173                  * If not already working on one, grab the next message.
174                  *
175                  * c_xmit_rm holds a ref while we're sending this message down
176                  * the connction.  We can use this ref while holding the
177                  * send_sem.. rds_send_reset() is serialized with it.
178                  */
179                 if (!rm) {
180                         unsigned int len;
181
182                         spin_lock(&conn->c_lock);
183
184                         if (!list_empty(&conn->c_send_queue)) {
185                                 rm = list_entry(conn->c_send_queue.next,
186                                                 struct rds_message,
187                                                 m_conn_item);
188                                 rds_message_addref(rm);
189
190                                 /*
191                                  * Move the message from the send queue to the retransmit
192                                  * list right away.
193                                  */
194                                 list_move_tail(&rm->m_conn_item, &conn->c_retrans);
195                         }
196
197                         spin_unlock(&conn->c_lock);
198
199                         if (!rm)
200                                 break;
201
202                         /* Unfortunately, the way Infiniband deals with
203                          * RDMA to a bad MR key is by moving the entire
204                          * queue pair to error state. We cold possibly
205                          * recover from that, but right now we drop the
206                          * connection.
207                          * Therefore, we never retransmit messages with RDMA ops.
208                          */
209                         if (rm->rdma.op_active &&
210                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
211                                 spin_lock(&conn->c_lock);
212                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
213                                         list_move(&rm->m_conn_item, &to_be_dropped);
214                                 spin_unlock(&conn->c_lock);
215                                 continue;
216                         }
217
218                         /* Require an ACK every once in a while */
219                         len = ntohl(rm->m_inc.i_hdr.h_len);
220                         if (conn->c_unacked_packets == 0 ||
221                             conn->c_unacked_bytes < len) {
222                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
223
224                                 conn->c_unacked_packets = rds_sysctl_max_unacked_packets;
225                                 conn->c_unacked_bytes = rds_sysctl_max_unacked_bytes;
226                                 rds_stats_inc(s_send_ack_required);
227                         } else {
228                                 conn->c_unacked_bytes -= len;
229                                 conn->c_unacked_packets--;
230                         }
231
232                         conn->c_xmit_rm = rm;
233                 }
234
235                 /* The transport either sends the whole rdma or none of it */
236                 if (rm->rdma.op_active && !conn->c_xmit_rdma_sent) {
237                         rm->m_final_op = &rm->rdma;
238                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
239                         if (ret)
240                                 break;
241                         conn->c_xmit_rdma_sent = 1;
242
243                         /* The transport owns the mapped memory for now.
244                          * You can't unmap it while it's on the send queue */
245                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
246                 }
247
248                 if (rm->atomic.op_active && !conn->c_xmit_atomic_sent) {
249                         rm->m_final_op = &rm->atomic;
250                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
251                         if (ret)
252                                 break;
253                         conn->c_xmit_atomic_sent = 1;
254
255                         /* The transport owns the mapped memory for now.
256                          * You can't unmap it while it's on the send queue */
257                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
258                 }
259
260                 /*
261                  * A number of cases require an RDS header to be sent
262                  * even if there is no data.
263                  * We permit 0-byte sends; rds-ping depends on this.
264                  * However, if there are exclusively attached silent ops,
265                  * we skip the hdr/data send, to enable silent operation.
266                  */
267                 if (rm->data.op_nents == 0) {
268                         int ops_present;
269                         int all_ops_are_silent = 1;
270
271                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
272                         if (rm->atomic.op_active && !rm->atomic.op_silent)
273                                 all_ops_are_silent = 0;
274                         if (rm->rdma.op_active && !rm->rdma.op_silent)
275                                 all_ops_are_silent = 0;
276
277                         if (ops_present && all_ops_are_silent
278                             && !rm->m_rdma_cookie)
279                                 rm->data.op_active = 0;
280                 }
281
282                 if (rm->data.op_active && !conn->c_xmit_data_sent) {
283                         rm->m_final_op = &rm->data;
284                         ret = conn->c_trans->xmit(conn, rm,
285                                                   conn->c_xmit_hdr_off,
286                                                   conn->c_xmit_sg,
287                                                   conn->c_xmit_data_off);
288                         if (ret <= 0)
289                                 break;
290
291                         if (conn->c_xmit_hdr_off < sizeof(struct rds_header)) {
292                                 tmp = min_t(int, ret,
293                                             sizeof(struct rds_header) -
294                                             conn->c_xmit_hdr_off);
295                                 conn->c_xmit_hdr_off += tmp;
296                                 ret -= tmp;
297                         }
298
299                         sg = &rm->data.op_sg[conn->c_xmit_sg];
300                         while (ret) {
301                                 tmp = min_t(int, ret, sg->length -
302                                                       conn->c_xmit_data_off);
303                                 conn->c_xmit_data_off += tmp;
304                                 ret -= tmp;
305                                 if (conn->c_xmit_data_off == sg->length) {
306                                         conn->c_xmit_data_off = 0;
307                                         sg++;
308                                         conn->c_xmit_sg++;
309                                         BUG_ON(ret != 0 &&
310                                                conn->c_xmit_sg == rm->data.op_nents);
311                                 }
312                         }
313
314                         if (conn->c_xmit_hdr_off == sizeof(struct rds_header) &&
315                             (conn->c_xmit_sg == rm->data.op_nents))
316                                 conn->c_xmit_data_sent = 1;
317                 }
318
319                 /*
320                  * A rm will only take multiple times through this loop
321                  * if there is a data op. Thus, if the data is sent (or there was
322                  * none), then we're done with the rm.
323                  */
324                 if (!rm->data.op_active || conn->c_xmit_data_sent) {
325                         conn->c_xmit_rm = NULL;
326                         conn->c_xmit_sg = 0;
327                         conn->c_xmit_hdr_off = 0;
328                         conn->c_xmit_data_off = 0;
329                         conn->c_xmit_rdma_sent = 0;
330                         conn->c_xmit_atomic_sent = 0;
331                         conn->c_xmit_data_sent = 0;
332
333                         rds_message_put(rm);
334                 }
335         }
336
337         if (conn->c_trans->xmit_complete)
338                 conn->c_trans->xmit_complete(conn);
339
340         /*
341          * We might be racing with another sender who queued a message but
342          * backed off on noticing that we held the c_send_lock.  If we check
343          * for queued messages after dropping the sem then either we'll
344          * see the queued message or the queuer will get the sem.  If we
345          * notice the queued message then we trigger an immediate retry.
346          *
347          * We need to be careful only to do this when we stopped processing
348          * the send queue because it was empty.  It's the only way we
349          * stop processing the loop when the transport hasn't taken
350          * responsibility for forward progress.
351          */
352         spin_unlock_irqrestore(&conn->c_send_lock, flags);
353
354         /* Nuke any messages we decided not to retransmit. */
355         if (!list_empty(&to_be_dropped)) {
356                 /* irqs on here, so we can put(), unlike above */
357                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
358                         rds_message_put(rm);
359                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
360         }
361
362         atomic_dec(&conn->c_senders);
363
364         /*
365          * Other senders will see we have c_send_lock and exit. We
366          * need to recheck the send queue and race again for c_send_lock
367          * to make sure messages don't just sit on the send queue, if
368          * somebody hasn't already beat us into the loop.
369          *
370          * If the transport cannot continue (i.e ret != 0), then it must
371          * call us when more room is available, such as from the tx
372          * completion handler.
373          */
374         if (ret == 0) {
375                 smp_mb();
376                 if (!list_empty(&conn->c_send_queue)) {
377                         rds_stats_inc(s_send_lock_queue_raced);
378                         if (gen == atomic_read(&conn->c_send_generation)) {
379                                 goto restart;
380                         }
381                 }
382         }
383 out:
384         return ret;
385 }
386
387 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
388 {
389         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
390
391         assert_spin_locked(&rs->rs_lock);
392
393         BUG_ON(rs->rs_snd_bytes < len);
394         rs->rs_snd_bytes -= len;
395
396         if (rs->rs_snd_bytes == 0)
397                 rds_stats_inc(s_send_queue_empty);
398 }
399
400 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
401                                     is_acked_func is_acked)
402 {
403         if (is_acked)
404                 return is_acked(rm, ack);
405         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
406 }
407
408 /*
409  * Returns true if there are no messages on the send and retransmit queues
410  * which have a sequence number greater than or equal to the given sequence
411  * number.
412  */
413 int rds_send_acked_before(struct rds_connection *conn, u64 seq)
414 {
415         struct rds_message *rm, *tmp;
416         int ret = 1;
417
418         spin_lock(&conn->c_lock);
419
420         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
421                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
422                         ret = 0;
423                 break;
424         }
425
426         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
427                 if (be64_to_cpu(rm->m_inc.i_hdr.h_sequence) < seq)
428                         ret = 0;
429                 break;
430         }
431
432         spin_unlock(&conn->c_lock);
433
434         return ret;
435 }
436
437 /*
438  * This is pretty similar to what happens below in the ACK
439  * handling code - except that we call here as soon as we get
440  * the IB send completion on the RDMA op and the accompanying
441  * message.
442  */
443 void rds_rdma_send_complete(struct rds_message *rm, int status)
444 {
445         struct rds_sock *rs = NULL;
446         struct rm_rdma_op *ro;
447         struct rds_notifier *notifier;
448         unsigned long flags;
449
450         spin_lock_irqsave(&rm->m_rs_lock, flags);
451
452         ro = &rm->rdma;
453         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
454             ro->op_active && ro->op_notify && ro->op_notifier) {
455                 notifier = ro->op_notifier;
456                 rs = rm->m_rs;
457                 sock_hold(rds_rs_to_sk(rs));
458
459                 notifier->n_status = status;
460                 spin_lock(&rs->rs_lock);
461                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
462                 spin_unlock(&rs->rs_lock);
463
464                 ro->op_notifier = NULL;
465         }
466
467         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
468
469         if (rs) {
470                 rds_wake_sk_sleep(rs);
471                 sock_put(rds_rs_to_sk(rs));
472         }
473 }
474 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
475
476 /*
477  * Just like above, except looks at atomic op
478  */
479 void rds_atomic_send_complete(struct rds_message *rm, int status)
480 {
481         struct rds_sock *rs = NULL;
482         struct rm_atomic_op *ao;
483         struct rds_notifier *notifier;
484         unsigned long flags;
485
486         spin_lock_irqsave(&rm->m_rs_lock, flags);
487
488         ao = &rm->atomic;
489         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
490             && ao->op_active && ao->op_notify && ao->op_notifier) {
491                 notifier = ao->op_notifier;
492                 rs = rm->m_rs;
493                 sock_hold(rds_rs_to_sk(rs));
494
495                 notifier->n_status = status;
496                 spin_lock(&rs->rs_lock);
497                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
498                 spin_unlock(&rs->rs_lock);
499
500                 ao->op_notifier = NULL;
501         }
502
503         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
504
505         if (rs) {
506                 rds_wake_sk_sleep(rs);
507                 sock_put(rds_rs_to_sk(rs));
508         }
509 }
510 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
511
512 /*
513  * This is the same as rds_rdma_send_complete except we
514  * don't do any locking - we have all the ingredients (message,
515  * socket, socket lock) and can just move the notifier.
516  */
517 static inline void
518 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
519 {
520         struct rm_rdma_op *ro;
521         struct rm_atomic_op *ao;
522
523         ro = &rm->rdma;
524         if (ro->op_active && ro->op_notify && ro->op_notifier) {
525                 ro->op_notifier->n_status = status;
526                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
527                 ro->op_notifier = NULL;
528         }
529
530         ao = &rm->atomic;
531         if (ao->op_active && ao->op_notify && ao->op_notifier) {
532                 ao->op_notifier->n_status = status;
533                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
534                 ao->op_notifier = NULL;
535         }
536
537         /* No need to wake the app - caller does this */
538 }
539
540 /*
541  * This is called from the IB send completion when we detect
542  * a RDMA operation that failed with remote access error.
543  * So speed is not an issue here.
544  */
545 struct rds_message *rds_send_get_message(struct rds_connection *conn,
546                                          struct rm_rdma_op *op)
547 {
548         struct rds_message *rm, *tmp, *found = NULL;
549         unsigned long flags;
550
551         spin_lock_irqsave(&conn->c_lock, flags);
552
553         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
554                 if (&rm->rdma == op) {
555                         atomic_inc(&rm->m_refcount);
556                         found = rm;
557                         goto out;
558                 }
559         }
560
561         list_for_each_entry_safe(rm, tmp, &conn->c_send_queue, m_conn_item) {
562                 if (&rm->rdma == op) {
563                         atomic_inc(&rm->m_refcount);
564                         found = rm;
565                         break;
566                 }
567         }
568
569 out:
570         spin_unlock_irqrestore(&conn->c_lock, flags);
571
572         return found;
573 }
574 EXPORT_SYMBOL_GPL(rds_send_get_message);
575
576 /*
577  * This removes messages from the socket's list if they're on it.  The list
578  * argument must be private to the caller, we must be able to modify it
579  * without locks.  The messages must have a reference held for their
580  * position on the list.  This function will drop that reference after
581  * removing the messages from the 'messages' list regardless of if it found
582  * the messages on the socket list or not.
583  */
584 void rds_send_remove_from_sock(struct list_head *messages, int status)
585 {
586         unsigned long flags;
587         struct rds_sock *rs = NULL;
588         struct rds_message *rm;
589
590         while (!list_empty(messages)) {
591                 int was_on_sock = 0;
592
593                 rm = list_entry(messages->next, struct rds_message,
594                                 m_conn_item);
595                 list_del_init(&rm->m_conn_item);
596
597                 /*
598                  * If we see this flag cleared then we're *sure* that someone
599                  * else beat us to removing it from the sock.  If we race
600                  * with their flag update we'll get the lock and then really
601                  * see that the flag has been cleared.
602                  *
603                  * The message spinlock makes sure nobody clears rm->m_rs
604                  * while we're messing with it. It does not prevent the
605                  * message from being removed from the socket, though.
606                  */
607                 spin_lock_irqsave(&rm->m_rs_lock, flags);
608                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
609                         goto unlock_and_drop;
610
611                 if (rs != rm->m_rs) {
612                         if (rs) {
613                                 rds_wake_sk_sleep(rs);
614                                 sock_put(rds_rs_to_sk(rs));
615                         }
616                         rs = rm->m_rs;
617                         sock_hold(rds_rs_to_sk(rs));
618                 }
619                 spin_lock(&rs->rs_lock);
620
621                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
622                         struct rm_rdma_op *ro = &rm->rdma;
623                         struct rds_notifier *notifier;
624
625                         list_del_init(&rm->m_sock_item);
626                         rds_send_sndbuf_remove(rs, rm);
627
628                         if (ro->op_active && ro->op_notifier &&
629                                (ro->op_notify || (ro->op_recverr && status))) {
630                                 notifier = ro->op_notifier;
631                                 list_add_tail(&notifier->n_list,
632                                                 &rs->rs_notify_queue);
633                                 if (!notifier->n_status)
634                                         notifier->n_status = status;
635                                 rm->rdma.op_notifier = NULL;
636                         }
637                         was_on_sock = 1;
638                         rm->m_rs = NULL;
639                 }
640                 spin_unlock(&rs->rs_lock);
641
642 unlock_and_drop:
643                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
644                 rds_message_put(rm);
645                 if (was_on_sock)
646                         rds_message_put(rm);
647         }
648
649         if (rs) {
650                 rds_wake_sk_sleep(rs);
651                 sock_put(rds_rs_to_sk(rs));
652         }
653 }
654
655 /*
656  * Transports call here when they've determined that the receiver queued
657  * messages up to, and including, the given sequence number.  Messages are
658  * moved to the retrans queue when rds_send_xmit picks them off the send
659  * queue. This means that in the TCP case, the message may not have been
660  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
661  * checks the RDS_MSG_HAS_ACK_SEQ bit.
662  *
663  * XXX It's not clear to me how this is safely serialized with socket
664  * destruction.  Maybe it should bail if it sees SOCK_DEAD.
665  */
666 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
667                          is_acked_func is_acked)
668 {
669         struct rds_message *rm, *tmp;
670         unsigned long flags;
671         LIST_HEAD(list);
672
673         spin_lock_irqsave(&conn->c_lock, flags);
674
675         list_for_each_entry_safe(rm, tmp, &conn->c_retrans, m_conn_item) {
676                 if (!rds_send_is_acked(rm, ack, is_acked))
677                         break;
678
679                 list_move(&rm->m_conn_item, &list);
680                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
681         }
682
683         /* order flag updates with spin locks */
684         if (!list_empty(&list))
685                 smp_mb__after_clear_bit();
686
687         spin_unlock_irqrestore(&conn->c_lock, flags);
688
689         /* now remove the messages from the sock list as needed */
690         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
691 }
692 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
693
694 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
695 {
696         struct rds_message *rm, *tmp;
697         struct rds_connection *conn;
698         unsigned long flags;
699         LIST_HEAD(list);
700
701         /* get all the messages we're dropping under the rs lock */
702         spin_lock_irqsave(&rs->rs_lock, flags);
703
704         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
705                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
706                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
707                         continue;
708
709                 list_move(&rm->m_sock_item, &list);
710                 rds_send_sndbuf_remove(rs, rm);
711                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
712         }
713
714         /* order flag updates with the rs lock */
715         smp_mb__after_clear_bit();
716
717         spin_unlock_irqrestore(&rs->rs_lock, flags);
718
719         if (list_empty(&list))
720                 return;
721
722         /* Remove the messages from the conn */
723         list_for_each_entry(rm, &list, m_sock_item) {
724
725                 conn = rm->m_inc.i_conn;
726
727                 spin_lock_irqsave(&conn->c_lock, flags);
728                 /*
729                  * Maybe someone else beat us to removing rm from the conn.
730                  * If we race with their flag update we'll get the lock and
731                  * then really see that the flag has been cleared.
732                  */
733                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
734                         spin_unlock_irqrestore(&conn->c_lock, flags);
735                         continue;
736                 }
737                 list_del_init(&rm->m_conn_item);
738                 spin_unlock_irqrestore(&conn->c_lock, flags);
739
740                 /*
741                  * Couldn't grab m_rs_lock in top loop (lock ordering),
742                  * but we can now.
743                  */
744                 spin_lock_irqsave(&rm->m_rs_lock, flags);
745
746                 spin_lock(&rs->rs_lock);
747                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
748                 spin_unlock(&rs->rs_lock);
749
750                 rm->m_rs = NULL;
751                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
752
753                 rds_message_put(rm);
754         }
755
756         rds_wake_sk_sleep(rs);
757
758         while (!list_empty(&list)) {
759                 rm = list_entry(list.next, struct rds_message, m_sock_item);
760                 list_del_init(&rm->m_sock_item);
761
762                 rds_message_wait(rm);
763                 rds_message_put(rm);
764         }
765 }
766
767 /*
768  * we only want this to fire once so we use the callers 'queued'.  It's
769  * possible that another thread can race with us and remove the
770  * message from the flow with RDS_CANCEL_SENT_TO.
771  */
772 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
773                              struct rds_message *rm, __be16 sport,
774                              __be16 dport, int *queued)
775 {
776         unsigned long flags;
777         u32 len;
778
779         if (*queued)
780                 goto out;
781
782         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
783
784         /* this is the only place which holds both the socket's rs_lock
785          * and the connection's c_lock */
786         spin_lock_irqsave(&rs->rs_lock, flags);
787
788         /*
789          * If there is a little space in sndbuf, we don't queue anything,
790          * and userspace gets -EAGAIN. But poll() indicates there's send
791          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
792          * freed up by incoming acks. So we check the *old* value of
793          * rs_snd_bytes here to allow the last msg to exceed the buffer,
794          * and poll() now knows no more data can be sent.
795          */
796         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
797                 rs->rs_snd_bytes += len;
798
799                 /* let recv side know we are close to send space exhaustion.
800                  * This is probably not the optimal way to do it, as this
801                  * means we set the flag on *all* messages as soon as our
802                  * throughput hits a certain threshold.
803                  */
804                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
805                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
806
807                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
808                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
809                 rds_message_addref(rm);
810                 rm->m_rs = rs;
811
812                 /* The code ordering is a little weird, but we're
813                    trying to minimize the time we hold c_lock */
814                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
815                 rm->m_inc.i_conn = conn;
816                 rds_message_addref(rm);
817
818                 spin_lock(&conn->c_lock);
819                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(conn->c_next_tx_seq++);
820                 list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
821                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
822                 spin_unlock(&conn->c_lock);
823
824                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
825                          rm, len, rs, rs->rs_snd_bytes,
826                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
827
828                 *queued = 1;
829         }
830
831         spin_unlock_irqrestore(&rs->rs_lock, flags);
832 out:
833         return *queued;
834 }
835
836 /*
837  * rds_message is getting to be quite complicated, and we'd like to allocate
838  * it all in one go. This figures out how big it needs to be up front.
839  */
840 static int rds_rm_size(struct msghdr *msg, int data_len)
841 {
842         struct cmsghdr *cmsg;
843         int size = 0;
844         int cmsg_groups = 0;
845         int retval;
846
847         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
848                 if (!CMSG_OK(msg, cmsg))
849                         return -EINVAL;
850
851                 if (cmsg->cmsg_level != SOL_RDS)
852                         continue;
853
854                 switch (cmsg->cmsg_type) {
855                 case RDS_CMSG_RDMA_ARGS:
856                         cmsg_groups |= 1;
857                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
858                         if (retval < 0)
859                                 return retval;
860                         size += retval;
861
862                         break;
863
864                 case RDS_CMSG_RDMA_DEST:
865                 case RDS_CMSG_RDMA_MAP:
866                         cmsg_groups |= 2;
867                         /* these are valid but do no add any size */
868                         break;
869
870                 case RDS_CMSG_ATOMIC_CSWP:
871                 case RDS_CMSG_ATOMIC_FADD:
872                         cmsg_groups |= 1;
873                         size += sizeof(struct scatterlist);
874                         break;
875
876                 default:
877                         return -EINVAL;
878                 }
879
880         }
881
882         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
883
884         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
885         if (cmsg_groups == 3)
886                 return -EINVAL;
887
888         return size;
889 }
890
891 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
892                          struct msghdr *msg, int *allocated_mr)
893 {
894         struct cmsghdr *cmsg;
895         int ret = 0;
896
897         for (cmsg = CMSG_FIRSTHDR(msg); cmsg; cmsg = CMSG_NXTHDR(msg, cmsg)) {
898                 if (!CMSG_OK(msg, cmsg))
899                         return -EINVAL;
900
901                 if (cmsg->cmsg_level != SOL_RDS)
902                         continue;
903
904                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
905                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
906                  */
907                 switch (cmsg->cmsg_type) {
908                 case RDS_CMSG_RDMA_ARGS:
909                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
910                         break;
911
912                 case RDS_CMSG_RDMA_DEST:
913                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
914                         break;
915
916                 case RDS_CMSG_RDMA_MAP:
917                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
918                         if (!ret)
919                                 *allocated_mr = 1;
920                         break;
921                 case RDS_CMSG_ATOMIC_CSWP:
922                 case RDS_CMSG_ATOMIC_FADD:
923                         ret = rds_cmsg_atomic(rs, rm, cmsg);
924                         break;
925
926                 default:
927                         return -EINVAL;
928                 }
929
930                 if (ret)
931                         break;
932         }
933
934         return ret;
935 }
936
937 int rds_sendmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
938                 size_t payload_len)
939 {
940         struct sock *sk = sock->sk;
941         struct rds_sock *rs = rds_sk_to_rs(sk);
942         struct sockaddr_in *usin = (struct sockaddr_in *)msg->msg_name;
943         __be32 daddr;
944         __be16 dport;
945         struct rds_message *rm = NULL;
946         struct rds_connection *conn;
947         int ret = 0;
948         int queued = 0, allocated_mr = 0;
949         int nonblock = msg->msg_flags & MSG_DONTWAIT;
950         long timeo = sock_sndtimeo(sk, nonblock);
951
952         /* Mirror Linux UDP mirror of BSD error message compatibility */
953         /* XXX: Perhaps MSG_MORE someday */
954         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
955                 printk(KERN_INFO "msg_flags 0x%08X\n", msg->msg_flags);
956                 ret = -EOPNOTSUPP;
957                 goto out;
958         }
959
960         if (msg->msg_namelen) {
961                 /* XXX fail non-unicast destination IPs? */
962                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
963                         ret = -EINVAL;
964                         goto out;
965                 }
966                 daddr = usin->sin_addr.s_addr;
967                 dport = usin->sin_port;
968         } else {
969                 /* We only care about consistency with ->connect() */
970                 lock_sock(sk);
971                 daddr = rs->rs_conn_addr;
972                 dport = rs->rs_conn_port;
973                 release_sock(sk);
974         }
975
976         /* racing with another thread binding seems ok here */
977         if (daddr == 0 || rs->rs_bound_addr == 0) {
978                 ret = -ENOTCONN; /* XXX not a great errno */
979                 goto out;
980         }
981
982         /* size of rm including all sgs */
983         ret = rds_rm_size(msg, payload_len);
984         if (ret < 0)
985                 goto out;
986
987         rm = rds_message_alloc(ret, GFP_KERNEL);
988         if (!rm) {
989                 ret = -ENOMEM;
990                 goto out;
991         }
992
993         /* Attach data to the rm */
994         if (payload_len) {
995                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
996                 ret = rds_message_copy_from_user(rm, msg->msg_iov, payload_len);
997                 if (ret)
998                         goto out;
999         }
1000         rm->data.op_active = 1;
1001
1002         rm->m_daddr = daddr;
1003
1004         /* rds_conn_create has a spinlock that runs with IRQ off.
1005          * Caching the conn in the socket helps a lot. */
1006         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1007                 conn = rs->rs_conn;
1008         else {
1009                 conn = rds_conn_create_outgoing(rs->rs_bound_addr, daddr,
1010                                         rs->rs_transport,
1011                                         sock->sk->sk_allocation);
1012                 if (IS_ERR(conn)) {
1013                         ret = PTR_ERR(conn);
1014                         goto out;
1015                 }
1016                 rs->rs_conn = conn;
1017         }
1018
1019         /* Parse any control messages the user may have included. */
1020         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1021         if (ret)
1022                 goto out;
1023
1024         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1025                 if (printk_ratelimit())
1026                         printk(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1027                                &rm->rdma, conn->c_trans->xmit_rdma);
1028                 ret = -EOPNOTSUPP;
1029                 goto out;
1030         }
1031
1032         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1033                 if (printk_ratelimit())
1034                         printk(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1035                                &rm->atomic, conn->c_trans->xmit_atomic);
1036                 ret = -EOPNOTSUPP;
1037                 goto out;
1038         }
1039
1040         /* If the connection is down, trigger a connect. We may
1041          * have scheduled a delayed reconnect however - in this case
1042          * we should not interfere.
1043          */
1044         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1045             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1046                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1047
1048         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1049         if (ret) {
1050                 rs->rs_seen_congestion = 1;
1051                 goto out;
1052         }
1053
1054         while (!rds_send_queue_rm(rs, conn, rm, rs->rs_bound_port,
1055                                   dport, &queued)) {
1056                 rds_stats_inc(s_send_queue_full);
1057                 /* XXX make sure this is reasonable */
1058                 if (payload_len > rds_sk_sndbuf(rs)) {
1059                         ret = -EMSGSIZE;
1060                         goto out;
1061                 }
1062                 if (nonblock) {
1063                         ret = -EAGAIN;
1064                         goto out;
1065                 }
1066
1067                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1068                                         rds_send_queue_rm(rs, conn, rm,
1069                                                           rs->rs_bound_port,
1070                                                           dport,
1071                                                           &queued),
1072                                         timeo);
1073                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1074                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1075                         continue;
1076
1077                 ret = timeo;
1078                 if (ret == 0)
1079                         ret = -ETIMEDOUT;
1080                 goto out;
1081         }
1082
1083         /*
1084          * By now we've committed to the send.  We reuse rds_send_worker()
1085          * to retry sends in the rds thread if the transport asks us to.
1086          */
1087         rds_stats_inc(s_send_queued);
1088
1089         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1090                 rds_send_xmit(conn);
1091
1092         rds_message_put(rm);
1093         return payload_len;
1094
1095 out:
1096         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1097          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1098          * or in any other way, we need to destroy the MR again */
1099         if (allocated_mr)
1100                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1101
1102         if (rm)
1103                 rds_message_put(rm);
1104         return ret;
1105 }
1106
1107 /*
1108  * Reply to a ping packet.
1109  */
1110 int
1111 rds_send_pong(struct rds_connection *conn, __be16 dport)
1112 {
1113         struct rds_message *rm;
1114         unsigned long flags;
1115         int ret = 0;
1116
1117         rm = rds_message_alloc(0, GFP_ATOMIC);
1118         if (!rm) {
1119                 ret = -ENOMEM;
1120                 goto out;
1121         }
1122
1123         rm->m_daddr = conn->c_faddr;
1124         rm->data.op_active = 1;
1125
1126         /* If the connection is down, trigger a connect. We may
1127          * have scheduled a delayed reconnect however - in this case
1128          * we should not interfere.
1129          */
1130         if (rds_conn_state(conn) == RDS_CONN_DOWN &&
1131             !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags))
1132                 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
1133
1134         ret = rds_cong_wait(conn->c_fcong, dport, 1, NULL);
1135         if (ret)
1136                 goto out;
1137
1138         spin_lock_irqsave(&conn->c_lock, flags);
1139         list_add_tail(&rm->m_conn_item, &conn->c_send_queue);
1140         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1141         rds_message_addref(rm);
1142         rm->m_inc.i_conn = conn;
1143
1144         rds_message_populate_header(&rm->m_inc.i_hdr, 0, dport,
1145                                     conn->c_next_tx_seq);
1146         conn->c_next_tx_seq++;
1147         spin_unlock_irqrestore(&conn->c_lock, flags);
1148
1149         rds_stats_inc(s_send_queued);
1150         rds_stats_inc(s_send_pong);
1151
1152         if (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags))
1153                 rds_send_xmit(conn);
1154
1155         rds_message_put(rm);
1156         return 0;
1157
1158 out:
1159         if (rm)
1160                 rds_message_put(rm);
1161         return ret;
1162 }