forked from torvalds/linux
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathsend.c
1502 lines (1288 loc) · 39.1 KB
/
send.c
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*
* Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
*
* This software is available to you under a choice of one of two
* licenses. You may choose to be licensed under the terms of the GNU
* General Public License (GPL) Version 2, available from the file
* COPYING in the main directory of this source tree, or the
* OpenIB.org BSD license below:
*
* Redistribution and use in source and binary forms, with or
* without modification, are permitted provided that the following
* conditions are met:
*
* - Redistributions of source code must retain the above
* copyright notice, this list of conditions and the following
* disclaimer.
*
* - Redistributions in binary form must reproduce the above
* copyright notice, this list of conditions and the following
* disclaimer in the documentation and/or other materials
* provided with the distribution.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*/
#include <linux/kernel.h>
#include <linux/moduleparam.h>
#include <linux/gfp.h>
#include <net/sock.h>
#include <linux/in.h>
#include <linux/list.h>
#include <linux/ratelimit.h>
#include <linux/export.h>
#include <linux/sizes.h>
#include "rds.h"
/* When transmitting messages in rds_send_xmit, we need to emerge from
* time to time and briefly release the CPU. Otherwise the softlock watchdog
* will kick our shin.
* Also, it seems fairer to not let one busy connection stall all the
* others.
*
* send_batch_count is the number of times we'll loop in send_xmit. Setting
* it to 0 will restore the old behavior (where we looped until we had
* drained the queue).
*/
static int send_batch_count = SZ_1K;
module_param(send_batch_count, int, 0444);
MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue");
static void rds_send_remove_from_sock(struct list_head *messages, int status);
/*
* Reset the send state. Callers must ensure that this doesn't race with
* rds_send_xmit().
*/
void rds_send_path_reset(struct rds_conn_path *cp)
{
struct rds_message *rm, *tmp;
unsigned long flags;
if (cp->cp_xmit_rm) {
rm = cp->cp_xmit_rm;
cp->cp_xmit_rm = NULL;
/* Tell the user the RDMA op is no longer mapped by the
* transport. This isn't entirely true (it's flushed out
* independently) but as the connection is down, there's
* no ongoing RDMA to/from that memory */
rds_message_unmapped(rm);
rds_message_put(rm);
}
cp->cp_xmit_sg = 0;
cp->cp_xmit_hdr_off = 0;
cp->cp_xmit_data_off = 0;
cp->cp_xmit_atomic_sent = 0;
cp->cp_xmit_rdma_sent = 0;
cp->cp_xmit_data_sent = 0;
cp->cp_conn->c_map_queued = 0;
cp->cp_unacked_packets = rds_sysctl_max_unacked_packets;
cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes;
/* Mark messages as retransmissions, and move them to the send q */
spin_lock_irqsave(&cp->cp_lock, flags);
list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags);
}
list_splice_init(&cp->cp_retrans, &cp->cp_send_queue);
spin_unlock_irqrestore(&cp->cp_lock, flags);
}
EXPORT_SYMBOL_GPL(rds_send_path_reset);
static int acquire_in_xmit(struct rds_conn_path *cp)
{
return test_and_set_bit(RDS_IN_XMIT, &cp->cp_flags) == 0;
}
static void release_in_xmit(struct rds_conn_path *cp)
{
clear_bit(RDS_IN_XMIT, &cp->cp_flags);
smp_mb__after_atomic();
/*
* We don't use wait_on_bit()/wake_up_bit() because our waking is in a
* hot path and finding waiters is very rare. We don't want to walk
* the system-wide hashed waitqueue buckets in the fast path only to
* almost never find waiters.
*/
if (waitqueue_active(&cp->cp_waitq))
wake_up_all(&cp->cp_waitq);
}
/*
* We're making the conscious trade-off here to only send one message
* down the connection at a time.
* Pro:
* - tx queueing is a simple fifo list
* - reassembly is optional and easily done by transports per conn
* - no per flow rx lookup at all, straight to the socket
* - less per-frag memory and wire overhead
* Con:
* - queued acks can be delayed behind large messages
* Depends:
* - small message latency is higher behind queued large messages
* - large message latency isn't starved by intervening small sends
*/
int rds_send_xmit(struct rds_conn_path *cp)
{
struct rds_connection *conn = cp->cp_conn;
struct rds_message *rm;
unsigned long flags;
unsigned int tmp;
struct scatterlist *sg;
int ret = 0;
LIST_HEAD(to_be_dropped);
int batch_count;
unsigned long send_gen = 0;
restart:
batch_count = 0;
/*
* sendmsg calls here after having queued its message on the send
* queue. We only have one task feeding the connection at a time. If
* another thread is already feeding the queue then we back off. This
* avoids blocking the caller and trading per-connection data between
* caches per message.
*/
if (!acquire_in_xmit(cp)) {
rds_stats_inc(s_send_lock_contention);
ret = -ENOMEM;
goto out;
}
if (rds_destroy_pending(cp->cp_conn)) {
release_in_xmit(cp);
ret = -ENETUNREACH; /* dont requeue send work */
goto out;
}
/*
* we record the send generation after doing the xmit acquire.
* if someone else manages to jump in and do some work, we'll use
* this to avoid a goto restart farther down.
*
* The acquire_in_xmit() check above ensures that only one
* caller can increment c_send_gen at any time.
*/
send_gen = READ_ONCE(cp->cp_send_gen) + 1;
WRITE_ONCE(cp->cp_send_gen, send_gen);
/*
* rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT,
* we do the opposite to avoid races.
*/
if (!rds_conn_path_up(cp)) {
release_in_xmit(cp);
ret = 0;
goto out;
}
if (conn->c_trans->xmit_path_prepare)
conn->c_trans->xmit_path_prepare(cp);
/*
* spin trying to push headers and data down the connection until
* the connection doesn't make forward progress.
*/
while (1) {
rm = cp->cp_xmit_rm;
/*
* If between sending messages, we can send a pending congestion
* map update.
*/
if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) {
rm = rds_cong_update_alloc(conn);
if (IS_ERR(rm)) {
ret = PTR_ERR(rm);
break;
}
rm->data.op_active = 1;
rm->m_inc.i_conn_path = cp;
rm->m_inc.i_conn = cp->cp_conn;
cp->cp_xmit_rm = rm;
}
/*
* If not already working on one, grab the next message.
*
* cp_xmit_rm holds a ref while we're sending this message down
* the connction. We can use this ref while holding the
* send_sem.. rds_send_reset() is serialized with it.
*/
if (!rm) {
unsigned int len;
batch_count++;
/* we want to process as big a batch as we can, but
* we also want to avoid softlockups. If we've been
* through a lot of messages, lets back off and see
* if anyone else jumps in
*/
if (batch_count >= send_batch_count)
goto over_batch;
spin_lock_irqsave(&cp->cp_lock, flags);
if (!list_empty(&cp->cp_send_queue)) {
rm = list_entry(cp->cp_send_queue.next,
struct rds_message,
m_conn_item);
rds_message_addref(rm);
/*
* Move the message from the send queue to the retransmit
* list right away.
*/
list_move_tail(&rm->m_conn_item,
&cp->cp_retrans);
}
spin_unlock_irqrestore(&cp->cp_lock, flags);
if (!rm)
break;
/* Unfortunately, the way Infiniband deals with
* RDMA to a bad MR key is by moving the entire
* queue pair to error state. We cold possibly
* recover from that, but right now we drop the
* connection.
* Therefore, we never retransmit messages with RDMA ops.
*/
if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) ||
(rm->rdma.op_active &&
test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) {
spin_lock_irqsave(&cp->cp_lock, flags);
if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags))
list_move(&rm->m_conn_item, &to_be_dropped);
spin_unlock_irqrestore(&cp->cp_lock, flags);
continue;
}
/* Require an ACK every once in a while */
len = ntohl(rm->m_inc.i_hdr.h_len);
if (cp->cp_unacked_packets == 0 ||
cp->cp_unacked_bytes < len) {
set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
cp->cp_unacked_packets =
rds_sysctl_max_unacked_packets;
cp->cp_unacked_bytes =
rds_sysctl_max_unacked_bytes;
rds_stats_inc(s_send_ack_required);
} else {
cp->cp_unacked_bytes -= len;
cp->cp_unacked_packets--;
}
cp->cp_xmit_rm = rm;
}
/* The transport either sends the whole rdma or none of it */
if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) {
rm->m_final_op = &rm->rdma;
/* The transport owns the mapped memory for now.
* You can't unmap it while it's on the send queue
*/
set_bit(RDS_MSG_MAPPED, &rm->m_flags);
ret = conn->c_trans->xmit_rdma(conn, &rm->rdma);
if (ret) {
clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
wake_up_interruptible(&rm->m_flush_wait);
break;
}
cp->cp_xmit_rdma_sent = 1;
}
if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) {
rm->m_final_op = &rm->atomic;
/* The transport owns the mapped memory for now.
* You can't unmap it while it's on the send queue
*/
set_bit(RDS_MSG_MAPPED, &rm->m_flags);
ret = conn->c_trans->xmit_atomic(conn, &rm->atomic);
if (ret) {
clear_bit(RDS_MSG_MAPPED, &rm->m_flags);
wake_up_interruptible(&rm->m_flush_wait);
break;
}
cp->cp_xmit_atomic_sent = 1;
}
/*
* A number of cases require an RDS header to be sent
* even if there is no data.
* We permit 0-byte sends; rds-ping depends on this.
* However, if there are exclusively attached silent ops,
* we skip the hdr/data send, to enable silent operation.
*/
if (rm->data.op_nents == 0) {
int ops_present;
int all_ops_are_silent = 1;
ops_present = (rm->atomic.op_active || rm->rdma.op_active);
if (rm->atomic.op_active && !rm->atomic.op_silent)
all_ops_are_silent = 0;
if (rm->rdma.op_active && !rm->rdma.op_silent)
all_ops_are_silent = 0;
if (ops_present && all_ops_are_silent
&& !rm->m_rdma_cookie)
rm->data.op_active = 0;
}
if (rm->data.op_active && !cp->cp_xmit_data_sent) {
rm->m_final_op = &rm->data;
ret = conn->c_trans->xmit(conn, rm,
cp->cp_xmit_hdr_off,
cp->cp_xmit_sg,
cp->cp_xmit_data_off);
if (ret <= 0)
break;
if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) {
tmp = min_t(int, ret,
sizeof(struct rds_header) -
cp->cp_xmit_hdr_off);
cp->cp_xmit_hdr_off += tmp;
ret -= tmp;
}
sg = &rm->data.op_sg[cp->cp_xmit_sg];
while (ret) {
tmp = min_t(int, ret, sg->length -
cp->cp_xmit_data_off);
cp->cp_xmit_data_off += tmp;
ret -= tmp;
if (cp->cp_xmit_data_off == sg->length) {
cp->cp_xmit_data_off = 0;
sg++;
cp->cp_xmit_sg++;
BUG_ON(ret != 0 && cp->cp_xmit_sg ==
rm->data.op_nents);
}
}
if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) &&
(cp->cp_xmit_sg == rm->data.op_nents))
cp->cp_xmit_data_sent = 1;
}
/*
* A rm will only take multiple times through this loop
* if there is a data op. Thus, if the data is sent (or there was
* none), then we're done with the rm.
*/
if (!rm->data.op_active || cp->cp_xmit_data_sent) {
cp->cp_xmit_rm = NULL;
cp->cp_xmit_sg = 0;
cp->cp_xmit_hdr_off = 0;
cp->cp_xmit_data_off = 0;
cp->cp_xmit_rdma_sent = 0;
cp->cp_xmit_atomic_sent = 0;
cp->cp_xmit_data_sent = 0;
rds_message_put(rm);
}
}
over_batch:
if (conn->c_trans->xmit_path_complete)
conn->c_trans->xmit_path_complete(cp);
release_in_xmit(cp);
/* Nuke any messages we decided not to retransmit. */
if (!list_empty(&to_be_dropped)) {
/* irqs on here, so we can put(), unlike above */
list_for_each_entry(rm, &to_be_dropped, m_conn_item)
rds_message_put(rm);
rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED);
}
/*
* Other senders can queue a message after we last test the send queue
* but before we clear RDS_IN_XMIT. In that case they'd back off and
* not try and send their newly queued message. We need to check the
* send queue after having cleared RDS_IN_XMIT so that their message
* doesn't get stuck on the send queue.
*
* If the transport cannot continue (i.e ret != 0), then it must
* call us when more room is available, such as from the tx
* completion handler.
*
* We have an extra generation check here so that if someone manages
* to jump in after our release_in_xmit, we'll see that they have done
* some work and we will skip our goto
*/
if (ret == 0) {
bool raced;
smp_mb();
raced = send_gen != READ_ONCE(cp->cp_send_gen);
if ((test_bit(0, &conn->c_map_queued) ||
!list_empty(&cp->cp_send_queue)) && !raced) {
if (batch_count < send_batch_count)
goto restart;
rcu_read_lock();
if (rds_destroy_pending(cp->cp_conn))
ret = -ENETUNREACH;
else
queue_delayed_work(rds_wq, &cp->cp_send_w, 1);
rcu_read_unlock();
} else if (raced) {
rds_stats_inc(s_send_lock_queue_raced);
}
}
out:
return ret;
}
EXPORT_SYMBOL_GPL(rds_send_xmit);
static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm)
{
u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
assert_spin_locked(&rs->rs_lock);
BUG_ON(rs->rs_snd_bytes < len);
rs->rs_snd_bytes -= len;
if (rs->rs_snd_bytes == 0)
rds_stats_inc(s_send_queue_empty);
}
static inline int rds_send_is_acked(struct rds_message *rm, u64 ack,
is_acked_func is_acked)
{
if (is_acked)
return is_acked(rm, ack);
return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack;
}
/*
* This is pretty similar to what happens below in the ACK
* handling code - except that we call here as soon as we get
* the IB send completion on the RDMA op and the accompanying
* message.
*/
void rds_rdma_send_complete(struct rds_message *rm, int status)
{
struct rds_sock *rs = NULL;
struct rm_rdma_op *ro;
struct rds_notifier *notifier;
unsigned long flags;
unsigned int notify = 0;
spin_lock_irqsave(&rm->m_rs_lock, flags);
notify = rm->rdma.op_notify | rm->data.op_notify;
ro = &rm->rdma;
if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) &&
ro->op_active && notify && ro->op_notifier) {
notifier = ro->op_notifier;
rs = rm->m_rs;
sock_hold(rds_rs_to_sk(rs));
notifier->n_status = status;
spin_lock(&rs->rs_lock);
list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
spin_unlock(&rs->rs_lock);
ro->op_notifier = NULL;
}
spin_unlock_irqrestore(&rm->m_rs_lock, flags);
if (rs) {
rds_wake_sk_sleep(rs);
sock_put(rds_rs_to_sk(rs));
}
}
EXPORT_SYMBOL_GPL(rds_rdma_send_complete);
/*
* Just like above, except looks at atomic op
*/
void rds_atomic_send_complete(struct rds_message *rm, int status)
{
struct rds_sock *rs = NULL;
struct rm_atomic_op *ao;
struct rds_notifier *notifier;
unsigned long flags;
spin_lock_irqsave(&rm->m_rs_lock, flags);
ao = &rm->atomic;
if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)
&& ao->op_active && ao->op_notify && ao->op_notifier) {
notifier = ao->op_notifier;
rs = rm->m_rs;
sock_hold(rds_rs_to_sk(rs));
notifier->n_status = status;
spin_lock(&rs->rs_lock);
list_add_tail(¬ifier->n_list, &rs->rs_notify_queue);
spin_unlock(&rs->rs_lock);
ao->op_notifier = NULL;
}
spin_unlock_irqrestore(&rm->m_rs_lock, flags);
if (rs) {
rds_wake_sk_sleep(rs);
sock_put(rds_rs_to_sk(rs));
}
}
EXPORT_SYMBOL_GPL(rds_atomic_send_complete);
/*
* This is the same as rds_rdma_send_complete except we
* don't do any locking - we have all the ingredients (message,
* socket, socket lock) and can just move the notifier.
*/
static inline void
__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status)
{
struct rm_rdma_op *ro;
struct rm_atomic_op *ao;
ro = &rm->rdma;
if (ro->op_active && ro->op_notify && ro->op_notifier) {
ro->op_notifier->n_status = status;
list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue);
ro->op_notifier = NULL;
}
ao = &rm->atomic;
if (ao->op_active && ao->op_notify && ao->op_notifier) {
ao->op_notifier->n_status = status;
list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue);
ao->op_notifier = NULL;
}
/* No need to wake the app - caller does this */
}
/*
* This removes messages from the socket's list if they're on it. The list
* argument must be private to the caller, we must be able to modify it
* without locks. The messages must have a reference held for their
* position on the list. This function will drop that reference after
* removing the messages from the 'messages' list regardless of if it found
* the messages on the socket list or not.
*/
static void rds_send_remove_from_sock(struct list_head *messages, int status)
{
unsigned long flags;
struct rds_sock *rs = NULL;
struct rds_message *rm;
while (!list_empty(messages)) {
int was_on_sock = 0;
rm = list_entry(messages->next, struct rds_message,
m_conn_item);
list_del_init(&rm->m_conn_item);
/*
* If we see this flag cleared then we're *sure* that someone
* else beat us to removing it from the sock. If we race
* with their flag update we'll get the lock and then really
* see that the flag has been cleared.
*
* The message spinlock makes sure nobody clears rm->m_rs
* while we're messing with it. It does not prevent the
* message from being removed from the socket, though.
*/
spin_lock_irqsave(&rm->m_rs_lock, flags);
if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags))
goto unlock_and_drop;
if (rs != rm->m_rs) {
if (rs) {
rds_wake_sk_sleep(rs);
sock_put(rds_rs_to_sk(rs));
}
rs = rm->m_rs;
if (rs)
sock_hold(rds_rs_to_sk(rs));
}
if (!rs)
goto unlock_and_drop;
spin_lock(&rs->rs_lock);
if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
struct rm_rdma_op *ro = &rm->rdma;
struct rds_notifier *notifier;
list_del_init(&rm->m_sock_item);
rds_send_sndbuf_remove(rs, rm);
if (ro->op_active && ro->op_notifier &&
(ro->op_notify || (ro->op_recverr && status))) {
notifier = ro->op_notifier;
list_add_tail(¬ifier->n_list,
&rs->rs_notify_queue);
if (!notifier->n_status)
notifier->n_status = status;
rm->rdma.op_notifier = NULL;
}
was_on_sock = 1;
}
spin_unlock(&rs->rs_lock);
unlock_and_drop:
spin_unlock_irqrestore(&rm->m_rs_lock, flags);
rds_message_put(rm);
if (was_on_sock)
rds_message_put(rm);
}
if (rs) {
rds_wake_sk_sleep(rs);
sock_put(rds_rs_to_sk(rs));
}
}
/*
* Transports call here when they've determined that the receiver queued
* messages up to, and including, the given sequence number. Messages are
* moved to the retrans queue when rds_send_xmit picks them off the send
* queue. This means that in the TCP case, the message may not have been
* assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
* checks the RDS_MSG_HAS_ACK_SEQ bit.
*/
void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack,
is_acked_func is_acked)
{
struct rds_message *rm, *tmp;
unsigned long flags;
LIST_HEAD(list);
spin_lock_irqsave(&cp->cp_lock, flags);
list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) {
if (!rds_send_is_acked(rm, ack, is_acked))
break;
list_move(&rm->m_conn_item, &list);
clear_bit(RDS_MSG_ON_CONN, &rm->m_flags);
}
/* order flag updates with spin locks */
if (!list_empty(&list))
smp_mb__after_atomic();
spin_unlock_irqrestore(&cp->cp_lock, flags);
/* now remove the messages from the sock list as needed */
rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS);
}
EXPORT_SYMBOL_GPL(rds_send_path_drop_acked);
void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
is_acked_func is_acked)
{
WARN_ON(conn->c_trans->t_mp_capable);
rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked);
}
EXPORT_SYMBOL_GPL(rds_send_drop_acked);
void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest)
{
struct rds_message *rm, *tmp;
struct rds_connection *conn;
struct rds_conn_path *cp;
unsigned long flags;
LIST_HEAD(list);
/* get all the messages we're dropping under the rs lock */
spin_lock_irqsave(&rs->rs_lock, flags);
list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) {
if (dest &&
(!ipv6_addr_equal(&dest->sin6_addr, &rm->m_daddr) ||
dest->sin6_port != rm->m_inc.i_hdr.h_dport))
continue;
list_move(&rm->m_sock_item, &list);
rds_send_sndbuf_remove(rs, rm);
clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
}
/* order flag updates with the rs lock */
smp_mb__after_atomic();
spin_unlock_irqrestore(&rs->rs_lock, flags);
if (list_empty(&list))
return;
/* Remove the messages from the conn */
list_for_each_entry(rm, &list, m_sock_item) {
conn = rm->m_inc.i_conn;
if (conn->c_trans->t_mp_capable)
cp = rm->m_inc.i_conn_path;
else
cp = &conn->c_path[0];
spin_lock_irqsave(&cp->cp_lock, flags);
/*
* Maybe someone else beat us to removing rm from the conn.
* If we race with their flag update we'll get the lock and
* then really see that the flag has been cleared.
*/
if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
spin_unlock_irqrestore(&cp->cp_lock, flags);
continue;
}
list_del_init(&rm->m_conn_item);
spin_unlock_irqrestore(&cp->cp_lock, flags);
/*
* Couldn't grab m_rs_lock in top loop (lock ordering),
* but we can now.
*/
spin_lock_irqsave(&rm->m_rs_lock, flags);
spin_lock(&rs->rs_lock);
__rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
spin_unlock(&rs->rs_lock);
spin_unlock_irqrestore(&rm->m_rs_lock, flags);
rds_message_put(rm);
}
rds_wake_sk_sleep(rs);
while (!list_empty(&list)) {
rm = list_entry(list.next, struct rds_message, m_sock_item);
list_del_init(&rm->m_sock_item);
rds_message_wait(rm);
/* just in case the code above skipped this message
* because RDS_MSG_ON_CONN wasn't set, run it again here
* taking m_rs_lock is the only thing that keeps us
* from racing with ack processing.
*/
spin_lock_irqsave(&rm->m_rs_lock, flags);
spin_lock(&rs->rs_lock);
__rds_send_complete(rs, rm, RDS_RDMA_CANCELED);
spin_unlock(&rs->rs_lock);
spin_unlock_irqrestore(&rm->m_rs_lock, flags);
rds_message_put(rm);
}
}
/*
* we only want this to fire once so we use the callers 'queued'. It's
* possible that another thread can race with us and remove the
* message from the flow with RDS_CANCEL_SENT_TO.
*/
static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn,
struct rds_conn_path *cp,
struct rds_message *rm, __be16 sport,
__be16 dport, int *queued)
{
unsigned long flags;
u32 len;
if (*queued)
goto out;
len = be32_to_cpu(rm->m_inc.i_hdr.h_len);
/* this is the only place which holds both the socket's rs_lock
* and the connection's c_lock */
spin_lock_irqsave(&rs->rs_lock, flags);
/*
* If there is a little space in sndbuf, we don't queue anything,
* and userspace gets -EAGAIN. But poll() indicates there's send
* room. This can lead to bad behavior (spinning) if snd_bytes isn't
* freed up by incoming acks. So we check the *old* value of
* rs_snd_bytes here to allow the last msg to exceed the buffer,
* and poll() now knows no more data can be sent.
*/
if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) {
rs->rs_snd_bytes += len;
/* let recv side know we are close to send space exhaustion.
* This is probably not the optimal way to do it, as this
* means we set the flag on *all* messages as soon as our
* throughput hits a certain threshold.
*/
if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2)
set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags);
list_add_tail(&rm->m_sock_item, &rs->rs_send_queue);
set_bit(RDS_MSG_ON_SOCK, &rm->m_flags);
rds_message_addref(rm);
sock_hold(rds_rs_to_sk(rs));
rm->m_rs = rs;
/* The code ordering is a little weird, but we're
trying to minimize the time we hold c_lock */
rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0);
rm->m_inc.i_conn = conn;
rm->m_inc.i_conn_path = cp;
rds_message_addref(rm);
spin_lock(&cp->cp_lock);
rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++);
list_add_tail(&rm->m_conn_item, &cp->cp_send_queue);
set_bit(RDS_MSG_ON_CONN, &rm->m_flags);
spin_unlock(&cp->cp_lock);
rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n",
rm, len, rs, rs->rs_snd_bytes,
(unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence));
*queued = 1;
}
spin_unlock_irqrestore(&rs->rs_lock, flags);
out:
return *queued;
}
/*
* rds_message is getting to be quite complicated, and we'd like to allocate
* it all in one go. This figures out how big it needs to be up front.
*/
static int rds_rm_size(struct msghdr *msg, int num_sgs,
struct rds_iov_vector_arr *vct)
{
struct cmsghdr *cmsg;
int size = 0;
int cmsg_groups = 0;
int retval;
bool zcopy_cookie = false;
struct rds_iov_vector *iov, *tmp_iov;
if (num_sgs < 0)
return -EINVAL;
for_each_cmsghdr(cmsg, msg) {
if (!CMSG_OK(msg, cmsg))
return -EINVAL;
if (cmsg->cmsg_level != SOL_RDS)
continue;
switch (cmsg->cmsg_type) {
case RDS_CMSG_RDMA_ARGS:
if (vct->indx >= vct->len) {
vct->len += vct->incr;
tmp_iov =
krealloc(vct->vec,
vct->len *
sizeof(struct rds_iov_vector),
GFP_KERNEL);
if (!tmp_iov) {
vct->len -= vct->incr;
return -ENOMEM;
}
vct->vec = tmp_iov;
}
iov = &vct->vec[vct->indx];
memset(iov, 0, sizeof(struct rds_iov_vector));
vct->indx++;
cmsg_groups |= 1;
retval = rds_rdma_extra_size(CMSG_DATA(cmsg), iov);
if (retval < 0)
return retval;
size += retval;
break;
case RDS_CMSG_ZCOPY_COOKIE:
zcopy_cookie = true;
/* fall through */
case RDS_CMSG_RDMA_DEST:
case RDS_CMSG_RDMA_MAP:
cmsg_groups |= 2;
/* these are valid but do no add any size */
break;
case RDS_CMSG_ATOMIC_CSWP:
case RDS_CMSG_ATOMIC_FADD:
case RDS_CMSG_MASKED_ATOMIC_CSWP:
case RDS_CMSG_MASKED_ATOMIC_FADD:
cmsg_groups |= 1;
size += sizeof(struct scatterlist);
break;
default:
return -EINVAL;
}
}
if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie)
return -EINVAL;
size += num_sgs * sizeof(struct scatterlist);
/* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */
if (cmsg_groups == 3)
return -EINVAL;
return size;
}
static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm,
struct cmsghdr *cmsg)
{
u32 *cookie;
if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) ||
!rm->data.op_mmp_znotifier)
return -EINVAL;
cookie = CMSG_DATA(cmsg);
rm->data.op_mmp_znotifier->z_cookie = *cookie;
return 0;
}
static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm,
struct msghdr *msg, int *allocated_mr,
struct rds_iov_vector_arr *vct)
{
struct cmsghdr *cmsg;
int ret = 0, ind = 0;
for_each_cmsghdr(cmsg, msg) {
if (!CMSG_OK(msg, cmsg))
return -EINVAL;
if (cmsg->cmsg_level != SOL_RDS)
continue;
/* As a side effect, RDMA_DEST and RDMA_MAP will set
* rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr.
*/
switch (cmsg->cmsg_type) {
case RDS_CMSG_RDMA_ARGS:
if (ind >= vct->indx)
return -ENOMEM;
ret = rds_cmsg_rdma_args(rs, rm, cmsg, &vct->vec[ind]);
ind++;
break;
case RDS_CMSG_RDMA_DEST:
ret = rds_cmsg_rdma_dest(rs, rm, cmsg);