ip6_tunnel: Update skb->protocol to ETH_P_IPV6 in ip6_tnl_xmit()
[cascardo/linux.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/moduleparam.h>
35 #include <linux/gfp.h>
36 #include <net/sock.h>
37 #include <linux/in.h>
38 #include <linux/list.h>
39 #include <linux/ratelimit.h>
40 #include <linux/export.h>
41 #include <linux/sizes.h>
42
43 #include "rds.h"
44
45 /* When transmitting messages in rds_send_xmit, we need to emerge from
46  * time to time and briefly release the CPU. Otherwise the softlock watchdog
47  * will kick our shin.
48  * Also, it seems fairer to not let one busy connection stall all the
49  * others.
50  *
51  * send_batch_count is the number of times we'll loop in send_xmit. Setting
52  * it to 0 will restore the old behavior (where we looped until we had
53  * drained the queue).
54  */
55 static int send_batch_count = SZ_1K;
56 module_param(send_batch_count, int, 0444);
57 MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
58
59 static void rds_send_remove_from_sock(struct list_head *messages, int status);
60
61 /*
62  * Reset the send state.  Callers must ensure that this doesn't race with
63  * rds_send_xmit().
64  */
65 void rds_send_path_reset(struct rds_conn_path *cp)
66 {
67         struct rds_message *rm, *tmp;
68         unsigned long flags;
69
70         if (cp->cp_xmit_rm) {
71                 rm = cp->cp_xmit_rm;
72                 cp->cp_xmit_rm = NULL;
73                 /* Tell the user the RDMA op is no longer mapped by the
74                  * transport. This isn't entirely true (it's flushed out
75                  * independently) but as the connection is down, there's
76                  * no ongoing RDMA to/from that memory */
77                 rds_message_unmapped(rm);
78                 rds_message_put(rm);
79         }
80
81         cp->cp_xmit_sg = 0;
82         cp->cp_xmit_hdr_off = 0;
83         cp->cp_xmit_data_off = 0;
84         cp->cp_xmit_atomic_sent = 0;
85         cp->cp_xmit_rdma_sent = 0;
86         cp->cp_xmit_data_sent = 0;
87
88         cp->cp_conn->c_map_queued = 0;
89
90         cp->cp_unacked_packets = rds_sysctl_max_unacked_packets;
91         cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes;
92
93         /* Mark messages as retransmissions, and move them to the send q */
94         spin_lock_irqsave(&cp->cp_lock, flags);
95         list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
96                 set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
97                 set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
98         }
99         list_splice_init(&cp->cp_retrans, &cp->cp_send_queue);
100         spin_unlock_irqrestore(&cp->cp_lock, flags);
101 }
102 EXPORT_SYMBOL_GPL(rds_send_path_reset);
103
104 static int acquire_in_xmit(struct rds_conn_path *cp)
105 {
106         return test_and_set_bit(RDS_IN_XMIT, &cp->cp_flags) == 0;
107 }
108
109 static void release_in_xmit(struct rds_conn_path *cp)
110 {
111         clear_bit(RDS_IN_XMIT, &cp->cp_flags);
112         smp_mb__after_atomic();
113         /*
114          * We don't use wait_on_bit()/wake_up_bit() because our waking is in a
115          * hot path and finding waiters is very rare.  We don't want to walk
116          * the system-wide hashed waitqueue buckets in the fast path only to
117          * almost never find waiters.
118          */
119         if (waitqueue_active(&cp->cp_waitq))
120                 wake_up_all(&cp->cp_waitq);
121 }
122
123 /*
124  * We're making the conscious trade-off here to only send one message
125  * down the connection at a time.
126  *   Pro:
127  *      - tx queueing is a simple fifo list
128  *      - reassembly is optional and easily done by transports per conn
129  *      - no per flow rx lookup at all, straight to the socket
130  *      - less per-frag memory and wire overhead
131  *   Con:
132  *      - queued acks can be delayed behind large messages
133  *   Depends:
134  *      - small message latency is higher behind queued large messages
135  *      - large message latency isn't starved by intervening small sends
136  */
137 int rds_send_xmit(struct rds_conn_path *cp)
138 {
139         struct rds_connection *conn = cp->cp_conn;
140         struct rds_message *rm;
141         unsigned long flags;
142         unsigned int tmp;
143         struct scatterlist *sg;
144         int ret = 0;
145         LIST_HEAD(to_be_dropped);
146         int batch_count;
147         unsigned long send_gen = 0;
148
149 restart:
150         batch_count = 0;
151
152         /*
153          * sendmsg calls here after having queued its message on the send
154          * queue.  We only have one task feeding the connection at a time.  If
155          * another thread is already feeding the queue then we back off.  This
156          * avoids blocking the caller and trading per-connection data between
157          * caches per message.
158          */
159         if (!acquire_in_xmit(cp)) {
160                 rds_stats_inc(s_send_lock_contention);
161                 ret = -ENOMEM;
162                 goto out;
163         }
164
165         /*
166          * we record the send generation after doing the xmit acquire.
167          * if someone else manages to jump in and do some work, we'll use
168          * this to avoid a goto restart farther down.
169          *
170          * The acquire_in_xmit() check above ensures that only one
171          * caller can increment c_send_gen at any time.
172          */
173         cp->cp_send_gen++;
174         send_gen = cp->cp_send_gen;
175
176         /*
177          * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
178          * we do the opposite to avoid races.
179          */
180         if (!rds_conn_path_up(cp)) {
181                 release_in_xmit(cp);
182                 ret = 0;
183                 goto out;
184         }
185
186         if (conn->c_trans->xmit_path_prepare)
187                 conn->c_trans->xmit_path_prepare(cp);
188
189         /*
190          * spin trying to push headers and data down the connection until
191          * the connection doesn't make forward progress.
192          */
193         while (1) {
194
195                 rm = cp->cp_xmit_rm;
196
197                 /*
198                  * If between sending messages, we can send a pending congestion
199                  * map update.
200                  */
201                 if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
202                         rm = rds_cong_update_alloc(conn);
203                         if (IS_ERR(rm)) {
204                                 ret = PTR_ERR(rm);
205                                 break;
206                         }
207                         rm->data.op_active = 1;
208                         rm->m_inc.i_conn_path = cp;
209                         rm->m_inc.i_conn = cp->cp_conn;
210
211                         cp->cp_xmit_rm = rm;
212                 }
213
214                 /*
215                  * If not already working on one, grab the next message.
216                  *
217                  * cp_xmit_rm holds a ref while we're sending this message down
218                  * the connction.  We can use this ref while holding the
219                  * send_sem.. rds_send_reset() is serialized with it.
220                  */
221                 if (!rm) {
222                         unsigned int len;
223
224                         batch_count++;
225
226                         /* we want to process as big a batch as we can, but
227                          * we also want to avoid softlockups.  If we've been
228                          * through a lot of messages, lets back off and see
229                          * if anyone else jumps in
230                          */
231                         if (batch_count >= send_batch_count)
232                                 goto over_batch;
233
234                         spin_lock_irqsave(&cp->cp_lock, flags);
235
236                         if (!list_empty(&cp->cp_send_queue)) {
237                                 rm = list_entry(cp->cp_send_queue.next,
238                                                 struct rds_message,
239                                                 m_conn_item);
240                                 rds_message_addref(rm);
241
242                                 /*
243                                  * Move the message from the send queue to the retransmit
244                                  * list right away.
245                                  */
246                                 list_move_tail(&rm->m_conn_item,
247                                                &cp->cp_retrans);
248                         }
249
250                         spin_unlock_irqrestore(&cp->cp_lock, flags);
251
252                         if (!rm)
253                                 break;
254
255                         /* Unfortunately, the way Infiniband deals with
256                          * RDMA to a bad MR key is by moving the entire
257                          * queue pair to error state. We cold possibly
258                          * recover from that, but right now we drop the
259                          * connection.
260                          * Therefore, we never retransmit messages with RDMA ops.
261                          */
262                         if (rm->rdma.op_active &&
263                             test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) {
264                                 spin_lock_irqsave(&cp->cp_lock, flags);
265                                 if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
266                                         list_move(&rm->m_conn_item, &to_be_dropped);
267                                 spin_unlock_irqrestore(&cp->cp_lock, flags);
268                                 continue;
269                         }
270
271                         /* Require an ACK every once in a while */
272                         len = ntohl(rm->m_inc.i_hdr.h_len);
273                         if (cp->cp_unacked_packets == 0 ||
274                             cp->cp_unacked_bytes < len) {
275                                 __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
276
277                                 cp->cp_unacked_packets =
278                                         rds_sysctl_max_unacked_packets;
279                                 cp->cp_unacked_bytes =
280                                         rds_sysctl_max_unacked_bytes;
281                                 rds_stats_inc(s_send_ack_required);
282                         } else {
283                                 cp->cp_unacked_bytes -= len;
284                                 cp->cp_unacked_packets--;
285                         }
286
287                         cp->cp_xmit_rm = rm;
288                 }
289
290                 /* The transport either sends the whole rdma or none of it */
291                 if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
292                         rm->m_final_op = &rm->rdma;
293                         /* The transport owns the mapped memory for now.
294                          * You can't unmap it while it's on the send queue
295                          */
296                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
297                         ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
298                         if (ret) {
299                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
300                                 wake_up_interruptible(&rm->m_flush_wait);
301                                 break;
302                         }
303                         cp->cp_xmit_rdma_sent = 1;
304
305                 }
306
307                 if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
308                         rm->m_final_op = &rm->atomic;
309                         /* The transport owns the mapped memory for now.
310                          * You can't unmap it while it's on the send queue
311                          */
312                         set_bit(RDS_MSG_MAPPED, &rm->m_flags);
313                         ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
314                         if (ret) {
315                                 clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
316                                 wake_up_interruptible(&rm->m_flush_wait);
317                                 break;
318                         }
319                         cp->cp_xmit_atomic_sent = 1;
320
321                 }
322
323                 /*
324                  * A number of cases require an RDS header to be sent
325                  * even if there is no data.
326                  * We permit 0-byte sends; rds-ping depends on this.
327                  * However, if there are exclusively attached silent ops,
328                  * we skip the hdr/data send, to enable silent operation.
329                  */
330                 if (rm->data.op_nents == 0) {
331                         int ops_present;
332                         int all_ops_are_silent = 1;
333
334                         ops_present = (rm->atomic.op_active || rm->rdma.op_active);
335                         if (rm->atomic.op_active && !rm->atomic.op_silent)
336                                 all_ops_are_silent = 0;
337                         if (rm->rdma.op_active && !rm->rdma.op_silent)
338                                 all_ops_are_silent = 0;
339
340                         if (ops_present && all_ops_are_silent
341                             && !rm->m_rdma_cookie)
342                                 rm->data.op_active = 0;
343                 }
344
345                 if (rm->data.op_active && !cp->cp_xmit_data_sent) {
346                         rm->m_final_op = &rm->data;
347
348                         ret = conn->c_trans->xmit(conn, rm,
349                                                   cp->cp_xmit_hdr_off,
350                                                   cp->cp_xmit_sg,
351                                                   cp->cp_xmit_data_off);
352                         if (ret <= 0)
353                                 break;
354
355                         if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
356                                 tmp = min_t(int, ret,
357                                             sizeof(struct rds_header) -
358                                             cp->cp_xmit_hdr_off);
359                                 cp->cp_xmit_hdr_off += tmp;
360                                 ret -= tmp;
361                         }
362
363                         sg = &rm->data.op_sg[cp->cp_xmit_sg];
364                         while (ret) {
365                                 tmp = min_t(int, ret, sg->length -
366                                                       cp->cp_xmit_data_off);
367                                 cp->cp_xmit_data_off += tmp;
368                                 ret -= tmp;
369                                 if (cp->cp_xmit_data_off == sg->length) {
370                                         cp->cp_xmit_data_off = 0;
371                                         sg++;
372                                         cp->cp_xmit_sg++;
373                                         BUG_ON(ret != 0 && cp->cp_xmit_sg ==
374                                                rm->data.op_nents);
375                                 }
376                         }
377
378                         if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
379                             (cp->cp_xmit_sg == rm->data.op_nents))
380                                 cp->cp_xmit_data_sent = 1;
381                 }
382
383                 /*
384                  * A rm will only take multiple times through this loop
385                  * if there is a data op. Thus, if the data is sent (or there was
386                  * none), then we're done with the rm.
387                  */
388                 if (!rm->data.op_active || cp->cp_xmit_data_sent) {
389                         cp->cp_xmit_rm = NULL;
390                         cp->cp_xmit_sg = 0;
391                         cp->cp_xmit_hdr_off = 0;
392                         cp->cp_xmit_data_off = 0;
393                         cp->cp_xmit_rdma_sent = 0;
394                         cp->cp_xmit_atomic_sent = 0;
395                         cp->cp_xmit_data_sent = 0;
396
397                         rds_message_put(rm);
398                 }
399         }
400
401 over_batch:
402         if (conn->c_trans->xmit_path_complete)
403                 conn->c_trans->xmit_path_complete(cp);
404         release_in_xmit(cp);
405
406         /* Nuke any messages we decided not to retransmit. */
407         if (!list_empty(&to_be_dropped)) {
408                 /* irqs on here, so we can put(), unlike above */
409                 list_for_each_entry(rm, &to_be_dropped, m_conn_item)
410                         rds_message_put(rm);
411                 rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
412         }
413
414         /*
415          * Other senders can queue a message after we last test the send queue
416          * but before we clear RDS_IN_XMIT.  In that case they'd back off and
417          * not try and send their newly queued message.  We need to check the
418          * send queue after having cleared RDS_IN_XMIT so that their message
419          * doesn't get stuck on the send queue.
420          *
421          * If the transport cannot continue (i.e ret != 0), then it must
422          * call us when more room is available, such as from the tx
423          * completion handler.
424          *
425          * We have an extra generation check here so that if someone manages
426          * to jump in after our release_in_xmit, we'll see that they have done
427          * some work and we will skip our goto
428          */
429         if (ret == 0) {
430                 smp_mb();
431                 if ((test_bit(0, &conn->c_map_queued) ||
432                      !list_empty(&cp->cp_send_queue)) &&
433                     send_gen == cp->cp_send_gen) {
434                         rds_stats_inc(s_send_lock_queue_raced);
435                         if (batch_count < send_batch_count)
436                                 goto restart;
437                         queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
438                 }
439         }
440 out:
441         return ret;
442 }
443 EXPORT_SYMBOL_GPL(rds_send_xmit);
444
445 static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
446 {
447         u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
448
449         assert_spin_locked(&rs->rs_lock);
450
451         BUG_ON(rs->rs_snd_bytes < len);
452         rs->rs_snd_bytes -= len;
453
454         if (rs->rs_snd_bytes == 0)
455                 rds_stats_inc(s_send_queue_empty);
456 }
457
458 static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
459                                     is_acked_func is_acked)
460 {
461         if (is_acked)
462                 return is_acked(rm, ack);
463         return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
464 }
465
466 /*
467  * This is pretty similar to what happens below in the ACK
468  * handling code - except that we call here as soon as we get
469  * the IB send completion on the RDMA op and the accompanying
470  * message.
471  */
472 void rds_rdma_send_complete(struct rds_message *rm, int status)
473 {
474         struct rds_sock *rs = NULL;
475         struct rm_rdma_op *ro;
476         struct rds_notifier *notifier;
477         unsigned long flags;
478
479         spin_lock_irqsave(&rm->m_rs_lock, flags);
480
481         ro = &rm->rdma;
482         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
483             ro->op_active && ro->op_notify && ro->op_notifier) {
484                 notifier = ro->op_notifier;
485                 rs = rm->m_rs;
486                 sock_hold(rds_rs_to_sk(rs));
487
488                 notifier->n_status = status;
489                 spin_lock(&rs->rs_lock);
490                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
491                 spin_unlock(&rs->rs_lock);
492
493                 ro->op_notifier = NULL;
494         }
495
496         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
497
498         if (rs) {
499                 rds_wake_sk_sleep(rs);
500                 sock_put(rds_rs_to_sk(rs));
501         }
502 }
503 EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
504
505 /*
506  * Just like above, except looks at atomic op
507  */
508 void rds_atomic_send_complete(struct rds_message *rm, int status)
509 {
510         struct rds_sock *rs = NULL;
511         struct rm_atomic_op *ao;
512         struct rds_notifier *notifier;
513         unsigned long flags;
514
515         spin_lock_irqsave(&rm->m_rs_lock, flags);
516
517         ao = &rm->atomic;
518         if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
519             && ao->op_active && ao->op_notify && ao->op_notifier) {
520                 notifier = ao->op_notifier;
521                 rs = rm->m_rs;
522                 sock_hold(rds_rs_to_sk(rs));
523
524                 notifier->n_status = status;
525                 spin_lock(&rs->rs_lock);
526                 list_add_tail(&notifier->n_list, &rs->rs_notify_queue);
527                 spin_unlock(&rs->rs_lock);
528
529                 ao->op_notifier = NULL;
530         }
531
532         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
533
534         if (rs) {
535                 rds_wake_sk_sleep(rs);
536                 sock_put(rds_rs_to_sk(rs));
537         }
538 }
539 EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
540
541 /*
542  * This is the same as rds_rdma_send_complete except we
543  * don't do any locking - we have all the ingredients (message,
544  * socket, socket lock) and can just move the notifier.
545  */
546 static inline void
547 __rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
548 {
549         struct rm_rdma_op *ro;
550         struct rm_atomic_op *ao;
551
552         ro = &rm->rdma;
553         if (ro->op_active && ro->op_notify && ro->op_notifier) {
554                 ro->op_notifier->n_status = status;
555                 list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
556                 ro->op_notifier = NULL;
557         }
558
559         ao = &rm->atomic;
560         if (ao->op_active && ao->op_notify && ao->op_notifier) {
561                 ao->op_notifier->n_status = status;
562                 list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
563                 ao->op_notifier = NULL;
564         }
565
566         /* No need to wake the app - caller does this */
567 }
568
569 /*
570  * This removes messages from the socket's list if they're on it.  The list
571  * argument must be private to the caller, we must be able to modify it
572  * without locks.  The messages must have a reference held for their
573  * position on the list.  This function will drop that reference after
574  * removing the messages from the 'messages' list regardless of if it found
575  * the messages on the socket list or not.
576  */
577 static void rds_send_remove_from_sock(struct list_head *messages, int status)
578 {
579         unsigned long flags;
580         struct rds_sock *rs = NULL;
581         struct rds_message *rm;
582
583         while (!list_empty(messages)) {
584                 int was_on_sock = 0;
585
586                 rm = list_entry(messages->next, struct rds_message,
587                                 m_conn_item);
588                 list_del_init(&rm->m_conn_item);
589
590                 /*
591                  * If we see this flag cleared then we're *sure* that someone
592                  * else beat us to removing it from the sock.  If we race
593                  * with their flag update we'll get the lock and then really
594                  * see that the flag has been cleared.
595                  *
596                  * The message spinlock makes sure nobody clears rm->m_rs
597                  * while we're messing with it. It does not prevent the
598                  * message from being removed from the socket, though.
599                  */
600                 spin_lock_irqsave(&rm->m_rs_lock, flags);
601                 if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
602                         goto unlock_and_drop;
603
604                 if (rs != rm->m_rs) {
605                         if (rs) {
606                                 rds_wake_sk_sleep(rs);
607                                 sock_put(rds_rs_to_sk(rs));
608                         }
609                         rs = rm->m_rs;
610                         if (rs)
611                                 sock_hold(rds_rs_to_sk(rs));
612                 }
613                 if (!rs)
614                         goto unlock_and_drop;
615                 spin_lock(&rs->rs_lock);
616
617                 if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
618                         struct rm_rdma_op *ro = &rm->rdma;
619                         struct rds_notifier *notifier;
620
621                         list_del_init(&rm->m_sock_item);
622                         rds_send_sndbuf_remove(rs, rm);
623
624                         if (ro->op_active && ro->op_notifier &&
625                                (ro->op_notify || (ro->op_recverr && status))) {
626                                 notifier = ro->op_notifier;
627                                 list_add_tail(&notifier->n_list,
628                                                 &rs->rs_notify_queue);
629                                 if (!notifier->n_status)
630                                         notifier->n_status = status;
631                                 rm->rdma.op_notifier = NULL;
632                         }
633                         was_on_sock = 1;
634                         rm->m_rs = NULL;
635                 }
636                 spin_unlock(&rs->rs_lock);
637
638 unlock_and_drop:
639                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
640                 rds_message_put(rm);
641                 if (was_on_sock)
642                         rds_message_put(rm);
643         }
644
645         if (rs) {
646                 rds_wake_sk_sleep(rs);
647                 sock_put(rds_rs_to_sk(rs));
648         }
649 }
650
651 /*
652  * Transports call here when they've determined that the receiver queued
653  * messages up to, and including, the given sequence number.  Messages are
654  * moved to the retrans queue when rds_send_xmit picks them off the send
655  * queue. This means that in the TCP case, the message may not have been
656  * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
657  * checks the RDS_MSG_HAS_ACK_SEQ bit.
658  */
659 void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
660                               is_acked_func is_acked)
661 {
662         struct rds_message *rm, *tmp;
663         unsigned long flags;
664         LIST_HEAD(list);
665
666         spin_lock_irqsave(&cp->cp_lock, flags);
667
668         list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
669                 if (!rds_send_is_acked(rm, ack, is_acked))
670                         break;
671
672                 list_move(&rm->m_conn_item, &list);
673                 clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
674         }
675
676         /* order flag updates with spin locks */
677         if (!list_empty(&list))
678                 smp_mb__after_atomic();
679
680         spin_unlock_irqrestore(&cp->cp_lock, flags);
681
682         /* now remove the messages from the sock list as needed */
683         rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
684 }
685 EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
686
687 void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
688                          is_acked_func is_acked)
689 {
690         WARN_ON(conn->c_trans->t_mp_capable);
691         rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
692 }
693 EXPORT_SYMBOL_GPL(rds_send_drop_acked);
694
695 void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
696 {
697         struct rds_message *rm, *tmp;
698         struct rds_connection *conn;
699         struct rds_conn_path *cp;
700         unsigned long flags;
701         LIST_HEAD(list);
702
703         /* get all the messages we're dropping under the rs lock */
704         spin_lock_irqsave(&rs->rs_lock, flags);
705
706         list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
707                 if (dest && (dest->sin_addr.s_addr != rm->m_daddr ||
708                              dest->sin_port != rm->m_inc.i_hdr.h_dport))
709                         continue;
710
711                 list_move(&rm->m_sock_item, &list);
712                 rds_send_sndbuf_remove(rs, rm);
713                 clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
714         }
715
716         /* order flag updates with the rs lock */
717         smp_mb__after_atomic();
718
719         spin_unlock_irqrestore(&rs->rs_lock, flags);
720
721         if (list_empty(&list))
722                 return;
723
724         /* Remove the messages from the conn */
725         list_for_each_entry(rm, &list, m_sock_item) {
726
727                 conn = rm->m_inc.i_conn;
728                 if (conn->c_trans->t_mp_capable)
729                         cp = rm->m_inc.i_conn_path;
730                 else
731                         cp = &conn->c_path[0];
732
733                 spin_lock_irqsave(&cp->cp_lock, flags);
734                 /*
735                  * Maybe someone else beat us to removing rm from the conn.
736                  * If we race with their flag update we'll get the lock and
737                  * then really see that the flag has been cleared.
738                  */
739                 if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
740                         spin_unlock_irqrestore(&cp->cp_lock, flags);
741                         spin_lock_irqsave(&rm->m_rs_lock, flags);
742                         rm->m_rs = NULL;
743                         spin_unlock_irqrestore(&rm->m_rs_lock, flags);
744                         continue;
745                 }
746                 list_del_init(&rm->m_conn_item);
747                 spin_unlock_irqrestore(&cp->cp_lock, flags);
748
749                 /*
750                  * Couldn't grab m_rs_lock in top loop (lock ordering),
751                  * but we can now.
752                  */
753                 spin_lock_irqsave(&rm->m_rs_lock, flags);
754
755                 spin_lock(&rs->rs_lock);
756                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
757                 spin_unlock(&rs->rs_lock);
758
759                 rm->m_rs = NULL;
760                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
761
762                 rds_message_put(rm);
763         }
764
765         rds_wake_sk_sleep(rs);
766
767         while (!list_empty(&list)) {
768                 rm = list_entry(list.next, struct rds_message, m_sock_item);
769                 list_del_init(&rm->m_sock_item);
770                 rds_message_wait(rm);
771
772                 /* just in case the code above skipped this message
773                  * because RDS_MSG_ON_CONN wasn't set, run it again here
774                  * taking m_rs_lock is the only thing that keeps us
775                  * from racing with ack processing.
776                  */
777                 spin_lock_irqsave(&rm->m_rs_lock, flags);
778
779                 spin_lock(&rs->rs_lock);
780                 __rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
781                 spin_unlock(&rs->rs_lock);
782
783                 rm->m_rs = NULL;
784                 spin_unlock_irqrestore(&rm->m_rs_lock, flags);
785
786                 rds_message_put(rm);
787         }
788 }
789
790 /*
791  * we only want this to fire once so we use the callers 'queued'.  It's
792  * possible that another thread can race with us and remove the
793  * message from the flow with RDS_CANCEL_SENT_TO.
794  */
795 static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
796                              struct rds_conn_path *cp,
797                              struct rds_message *rm, __be16 sport,
798                              __be16 dport, int *queued)
799 {
800         unsigned long flags;
801         u32 len;
802
803         if (*queued)
804                 goto out;
805
806         len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
807
808         /* this is the only place which holds both the socket's rs_lock
809          * and the connection's c_lock */
810         spin_lock_irqsave(&rs->rs_lock, flags);
811
812         /*
813          * If there is a little space in sndbuf, we don't queue anything,
814          * and userspace gets -EAGAIN. But poll() indicates there's send
815          * room. This can lead to bad behavior (spinning) if snd_bytes isn't
816          * freed up by incoming acks. So we check the *old* value of
817          * rs_snd_bytes here to allow the last msg to exceed the buffer,
818          * and poll() now knows no more data can be sent.
819          */
820         if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
821                 rs->rs_snd_bytes += len;
822
823                 /* let recv side know we are close to send space exhaustion.
824                  * This is probably not the optimal way to do it, as this
825                  * means we set the flag on *all* messages as soon as our
826                  * throughput hits a certain threshold.
827                  */
828                 if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
829                         __set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
830
831                 list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
832                 set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
833                 rds_message_addref(rm);
834                 rm->m_rs = rs;
835
836                 /* The code ordering is a little weird, but we're
837                    trying to minimize the time we hold c_lock */
838                 rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
839                 rm->m_inc.i_conn = conn;
840                 rm->m_inc.i_conn_path = cp;
841                 rds_message_addref(rm);
842
843                 spin_lock(&cp->cp_lock);
844                 rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
845                 list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
846                 set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
847                 spin_unlock(&cp->cp_lock);
848
849                 rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
850                          rm, len, rs, rs->rs_snd_bytes,
851                          (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
852
853                 *queued = 1;
854         }
855
856         spin_unlock_irqrestore(&rs->rs_lock, flags);
857 out:
858         return *queued;
859 }
860
861 /*
862  * rds_message is getting to be quite complicated, and we'd like to allocate
863  * it all in one go. This figures out how big it needs to be up front.
864  */
865 static int rds_rm_size(struct msghdr *msg, int data_len)
866 {
867         struct cmsghdr *cmsg;
868         int size = 0;
869         int cmsg_groups = 0;
870         int retval;
871
872         for_each_cmsghdr(cmsg, msg) {
873                 if (!CMSG_OK(msg, cmsg))
874                         return -EINVAL;
875
876                 if (cmsg->cmsg_level != SOL_RDS)
877                         continue;
878
879                 switch (cmsg->cmsg_type) {
880                 case RDS_CMSG_RDMA_ARGS:
881                         cmsg_groups |= 1;
882                         retval = rds_rdma_extra_size(CMSG_DATA(cmsg));
883                         if (retval < 0)
884                                 return retval;
885                         size += retval;
886
887                         break;
888
889                 case RDS_CMSG_RDMA_DEST:
890                 case RDS_CMSG_RDMA_MAP:
891                         cmsg_groups |= 2;
892                         /* these are valid but do no add any size */
893                         break;
894
895                 case RDS_CMSG_ATOMIC_CSWP:
896                 case RDS_CMSG_ATOMIC_FADD:
897                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
898                 case RDS_CMSG_MASKED_ATOMIC_FADD:
899                         cmsg_groups |= 1;
900                         size += sizeof(struct scatterlist);
901                         break;
902
903                 default:
904                         return -EINVAL;
905                 }
906
907         }
908
909         size += ceil(data_len, PAGE_SIZE) * sizeof(struct scatterlist);
910
911         /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
912         if (cmsg_groups == 3)
913                 return -EINVAL;
914
915         return size;
916 }
917
918 static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
919                          struct msghdr *msg, int *allocated_mr)
920 {
921         struct cmsghdr *cmsg;
922         int ret = 0;
923
924         for_each_cmsghdr(cmsg, msg) {
925                 if (!CMSG_OK(msg, cmsg))
926                         return -EINVAL;
927
928                 if (cmsg->cmsg_level != SOL_RDS)
929                         continue;
930
931                 /* As a side effect, RDMA_DEST and RDMA_MAP will set
932                  * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
933                  */
934                 switch (cmsg->cmsg_type) {
935                 case RDS_CMSG_RDMA_ARGS:
936                         ret = rds_cmsg_rdma_args(rs, rm, cmsg);
937                         break;
938
939                 case RDS_CMSG_RDMA_DEST:
940                         ret = rds_cmsg_rdma_dest(rs, rm, cmsg);
941                         break;
942
943                 case RDS_CMSG_RDMA_MAP:
944                         ret = rds_cmsg_rdma_map(rs, rm, cmsg);
945                         if (!ret)
946                                 *allocated_mr = 1;
947                         break;
948                 case RDS_CMSG_ATOMIC_CSWP:
949                 case RDS_CMSG_ATOMIC_FADD:
950                 case RDS_CMSG_MASKED_ATOMIC_CSWP:
951                 case RDS_CMSG_MASKED_ATOMIC_FADD:
952                         ret = rds_cmsg_atomic(rs, rm, cmsg);
953                         break;
954
955                 default:
956                         return -EINVAL;
957                 }
958
959                 if (ret)
960                         break;
961         }
962
963         return ret;
964 }
965
966 static void rds_send_ping(struct rds_connection *conn);
967
968 static int rds_send_mprds_hash(struct rds_sock *rs, struct rds_connection *conn)
969 {
970         int hash;
971
972         if (conn->c_npaths == 0)
973                 hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS);
974         else
975                 hash = RDS_MPATH_HASH(rs, conn->c_npaths);
976         if (conn->c_npaths == 0 && hash != 0) {
977                 rds_send_ping(conn);
978
979                 if (conn->c_npaths == 0) {
980                         wait_event_interruptible(conn->c_hs_waitq,
981                                                  (conn->c_npaths != 0));
982                 }
983                 if (conn->c_npaths == 1)
984                         hash = 0;
985         }
986         return hash;
987 }
988
989 int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len)
990 {
991         struct sock *sk = sock->sk;
992         struct rds_sock *rs = rds_sk_to_rs(sk);
993         DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name);
994         __be32 daddr;
995         __be16 dport;
996         struct rds_message *rm = NULL;
997         struct rds_connection *conn;
998         int ret = 0;
999         int queued = 0, allocated_mr = 0;
1000         int nonblock = msg->msg_flags & MSG_DONTWAIT;
1001         long timeo = sock_sndtimeo(sk, nonblock);
1002         struct rds_conn_path *cpath;
1003
1004         /* Mirror Linux UDP mirror of BSD error message compatibility */
1005         /* XXX: Perhaps MSG_MORE someday */
1006         if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT)) {
1007                 ret = -EOPNOTSUPP;
1008                 goto out;
1009         }
1010
1011         if (msg->msg_namelen) {
1012                 /* XXX fail non-unicast destination IPs? */
1013                 if (msg->msg_namelen < sizeof(*usin) || usin->sin_family != AF_INET) {
1014                         ret = -EINVAL;
1015                         goto out;
1016                 }
1017                 daddr = usin->sin_addr.s_addr;
1018                 dport = usin->sin_port;
1019         } else {
1020                 /* We only care about consistency with ->connect() */
1021                 lock_sock(sk);
1022                 daddr = rs->rs_conn_addr;
1023                 dport = rs->rs_conn_port;
1024                 release_sock(sk);
1025         }
1026
1027         lock_sock(sk);
1028         if (daddr == 0 || rs->rs_bound_addr == 0) {
1029                 release_sock(sk);
1030                 ret = -ENOTCONN; /* XXX not a great errno */
1031                 goto out;
1032         }
1033         release_sock(sk);
1034
1035         if (payload_len > rds_sk_sndbuf(rs)) {
1036                 ret = -EMSGSIZE;
1037                 goto out;
1038         }
1039
1040         /* size of rm including all sgs */
1041         ret = rds_rm_size(msg, payload_len);
1042         if (ret < 0)
1043                 goto out;
1044
1045         rm = rds_message_alloc(ret, GFP_KERNEL);
1046         if (!rm) {
1047                 ret = -ENOMEM;
1048                 goto out;
1049         }
1050
1051         /* Attach data to the rm */
1052         if (payload_len) {
1053                 rm->data.op_sg = rds_message_alloc_sgs(rm, ceil(payload_len, PAGE_SIZE));
1054                 if (!rm->data.op_sg) {
1055                         ret = -ENOMEM;
1056                         goto out;
1057                 }
1058                 ret = rds_message_copy_from_user(rm, &msg->msg_iter);
1059                 if (ret)
1060                         goto out;
1061         }
1062         rm->data.op_active = 1;
1063
1064         rm->m_daddr = daddr;
1065
1066         /* rds_conn_create has a spinlock that runs with IRQ off.
1067          * Caching the conn in the socket helps a lot. */
1068         if (rs->rs_conn && rs->rs_conn->c_faddr == daddr)
1069                 conn = rs->rs_conn;
1070         else {
1071                 conn = rds_conn_create_outgoing(sock_net(sock->sk),
1072                                                 rs->rs_bound_addr, daddr,
1073                                         rs->rs_transport,
1074                                         sock->sk->sk_allocation);
1075                 if (IS_ERR(conn)) {
1076                         ret = PTR_ERR(conn);
1077                         goto out;
1078                 }
1079                 rs->rs_conn = conn;
1080         }
1081
1082         /* Parse any control messages the user may have included. */
1083         ret = rds_cmsg_send(rs, rm, msg, &allocated_mr);
1084         if (ret)
1085                 goto out;
1086
1087         if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) {
1088                 printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n",
1089                                &rm->rdma, conn->c_trans->xmit_rdma);
1090                 ret = -EOPNOTSUPP;
1091                 goto out;
1092         }
1093
1094         if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) {
1095                 printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n",
1096                                &rm->atomic, conn->c_trans->xmit_atomic);
1097                 ret = -EOPNOTSUPP;
1098                 goto out;
1099         }
1100
1101         if (conn->c_trans->t_mp_capable)
1102                 cpath = &conn->c_path[rds_send_mprds_hash(rs, conn)];
1103         else
1104                 cpath = &conn->c_path[0];
1105
1106         rds_conn_path_connect_if_down(cpath);
1107
1108         ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs);
1109         if (ret) {
1110                 rs->rs_seen_congestion = 1;
1111                 goto out;
1112         }
1113         while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port,
1114                                   dport, &queued)) {
1115                 rds_stats_inc(s_send_queue_full);
1116
1117                 if (nonblock) {
1118                         ret = -EAGAIN;
1119                         goto out;
1120                 }
1121
1122                 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
1123                                         rds_send_queue_rm(rs, conn, cpath, rm,
1124                                                           rs->rs_bound_port,
1125                                                           dport,
1126                                                           &queued),
1127                                         timeo);
1128                 rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo);
1129                 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
1130                         continue;
1131
1132                 ret = timeo;
1133                 if (ret == 0)
1134                         ret = -ETIMEDOUT;
1135                 goto out;
1136         }
1137
1138         /*
1139          * By now we've committed to the send.  We reuse rds_send_worker()
1140          * to retry sends in the rds thread if the transport asks us to.
1141          */
1142         rds_stats_inc(s_send_queued);
1143
1144         ret = rds_send_xmit(cpath);
1145         if (ret == -ENOMEM || ret == -EAGAIN)
1146                 queue_delayed_work(rds_wq, &cpath->cp_send_w, 1);
1147
1148         rds_message_put(rm);
1149         return payload_len;
1150
1151 out:
1152         /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly.
1153          * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN
1154          * or in any other way, we need to destroy the MR again */
1155         if (allocated_mr)
1156                 rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1);
1157
1158         if (rm)
1159                 rds_message_put(rm);
1160         return ret;
1161 }
1162
1163 /*
1164  * send out a probe. Can be shared by rds_send_ping,
1165  * rds_send_pong, rds_send_hb.
1166  * rds_send_hb should use h_flags
1167  *   RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED
1168  * or
1169  *   RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED
1170  */
1171 int
1172 rds_send_probe(struct rds_conn_path *cp, __be16 sport,
1173                __be16 dport, u8 h_flags)
1174 {
1175         struct rds_message *rm;
1176         unsigned long flags;
1177         int ret = 0;
1178
1179         rm = rds_message_alloc(0, GFP_ATOMIC);
1180         if (!rm) {
1181                 ret = -ENOMEM;
1182                 goto out;
1183         }
1184
1185         rm->m_daddr = cp->cp_conn->c_faddr;
1186         rm->data.op_active = 1;
1187
1188         rds_conn_path_connect_if_down(cp);
1189
1190         ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL);
1191         if (ret)
1192                 goto out;
1193
1194         spin_lock_irqsave(&cp->cp_lock, flags);
1195         list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
1196         set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
1197         rds_message_addref(rm);
1198         rm->m_inc.i_conn = cp->cp_conn;
1199         rm->m_inc.i_conn_path = cp;
1200
1201         rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport,
1202                                     cp->cp_next_tx_seq);
1203         rm->m_inc.i_hdr.h_flags |= h_flags;
1204         cp->cp_next_tx_seq++;
1205
1206         if (RDS_HS_PROBE(sport, dport) && cp->cp_conn->c_trans->t_mp_capable) {
1207                 u16 npaths = RDS_MPATH_WORKERS;
1208
1209                 rds_message_add_extension(&rm->m_inc.i_hdr,
1210                                           RDS_EXTHDR_NPATHS, &npaths,
1211                                           sizeof(npaths));
1212         }
1213         spin_unlock_irqrestore(&cp->cp_lock, flags);
1214
1215         rds_stats_inc(s_send_queued);
1216         rds_stats_inc(s_send_pong);
1217
1218         /* schedule the send work on rds_wq */
1219         queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
1220
1221         rds_message_put(rm);
1222         return 0;
1223
1224 out:
1225         if (rm)
1226                 rds_message_put(rm);
1227         return ret;
1228 }
1229
1230 int
1231 rds_send_pong(struct rds_conn_path *cp, __be16 dport)
1232 {
1233         return rds_send_probe(cp, 0, dport, 0);
1234 }
1235
1236 void
1237 rds_send_ping(struct rds_connection *conn)
1238 {
1239         unsigned long flags;
1240         struct rds_conn_path *cp = &conn->c_path[0];
1241
1242         spin_lock_irqsave(&cp->cp_lock, flags);
1243         if (conn->c_ping_triggered) {
1244                 spin_unlock_irqrestore(&cp->cp_lock, flags);
1245                 return;
1246         }
1247         conn->c_ping_triggered = 1;
1248         spin_unlock_irqrestore(&cp->cp_lock, flags);
1249         rds_send_probe(&conn->c_path[0], RDS_FLAG_PROBE_PORT, 0, 0);
1250 }