-
Notifications
You must be signed in to change notification settings - Fork 389
/
efa_rdm_ep_fiops.c
1907 lines (1664 loc) · 55.4 KB
/
efa_rdm_ep_fiops.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
/* SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0-only */
/* SPDX-FileCopyrightText: Copyright Amazon.com, Inc. or its affiliates. All rights reserved. */
#include "efa.h"
#include "efa_av.h"
#include "efa_rdm_ep.h"
#include "efa_rdm_cq.h"
#include "efa_rdm_srx.h"
#include "efa_rdm_rma.h"
#include "efa_rdm_msg.h"
#include "efa_rdm_atomic.h"
#include "efa_rdm_rxe_map.h"
#include "efa_rdm_pkt_type.h"
#include "efa_rdm_pke_req.h"
#include "efa_cntr.h"
static
void efa_rdm_ep_construct_ibv_qp_init_attr_ex(struct efa_rdm_ep *ep,
struct ibv_qp_init_attr_ex *attr_ex,
struct ibv_cq_ex *tx_cq,
struct ibv_cq_ex *rx_cq)
{
attr_ex->cap.max_send_wr = ep->base_ep.domain->device->rdm_info->tx_attr->size;
attr_ex->cap.max_send_sge = ep->base_ep.domain->device->rdm_info->tx_attr->iov_limit;
attr_ex->cap.max_recv_wr = ep->base_ep.domain->device->rdm_info->rx_attr->size;
attr_ex->cap.max_recv_sge = ep->base_ep.domain->device->rdm_info->rx_attr->iov_limit;
attr_ex->cap.max_inline_data = ep->base_ep.domain->device->efa_attr.inline_buf_size;
attr_ex->qp_type = IBV_QPT_DRIVER;
attr_ex->pd = efa_rdm_ep_domain(ep)->ibv_pd;
attr_ex->qp_context = ep;
attr_ex->sq_sig_all = 1;
attr_ex->send_cq = ibv_cq_ex_to_cq(tx_cq);
attr_ex->recv_cq = ibv_cq_ex_to_cq(rx_cq);
}
static inline
struct efa_rdm_cq *efa_rdm_ep_get_tx_rdm_cq(struct efa_rdm_ep *ep)
{
return ep->base_ep.util_ep.tx_cq ? container_of(ep->base_ep.util_ep.tx_cq, struct efa_rdm_cq, util_cq) : NULL;
}
static inline
struct efa_rdm_cq *efa_rdm_ep_get_rx_rdm_cq(struct efa_rdm_ep *ep)
{
return ep->base_ep.util_ep.rx_cq ? container_of(ep->base_ep.util_ep.rx_cq, struct efa_rdm_cq, util_cq) : NULL;
}
/**
* @brief set the "efa_qp" field in the efa_rdm_ep->efa_base_ep
* called by efa_rdm_ep_open()
*
* @param[in,out] ep The EFA RDM endpoint to set the qp in
* @return int 0 on success, negative libfabric error code otherwise
* @todo merge this function with #efa_base_ep_construct
*/
static
int efa_rdm_ep_create_base_ep_ibv_qp(struct efa_rdm_ep *ep)
{
struct ibv_qp_init_attr_ex attr_ex = { 0 };
struct efa_rdm_cq *tx_rdm_cq, *rx_rdm_cq;
struct ibv_cq_ex *tx_ibv_cq, *rx_ibv_cq;
int ret;
tx_rdm_cq = efa_rdm_ep_get_tx_rdm_cq(ep);
rx_rdm_cq = efa_rdm_ep_get_rx_rdm_cq(ep);
if (!tx_rdm_cq && !rx_rdm_cq) {
EFA_WARN(FI_LOG_EP_CTRL,
"Endpoint is not bound to a send or receive completion queue\n");
return -FI_ENOCQ;
}
if (!tx_rdm_cq && ofi_needs_tx(ep->base_ep.info->caps)) {
EFA_WARN(FI_LOG_EP_CTRL,
"Endpoint is not bound to a send completion queue when it has transmit capabilities enabled (FI_SEND).\n");
return -FI_ENOCQ;
}
if (!rx_rdm_cq && ofi_needs_rx(ep->base_ep.info->caps)) {
EFA_WARN(FI_LOG_EP_CTRL,
"Endpoint is not bound to a receive completion queue when it has receive capabilities enabled (FI_RECV).\n");
return -FI_ENOCQ;
}
tx_ibv_cq = tx_rdm_cq ? tx_rdm_cq->ibv_cq.ibv_cq_ex : rx_rdm_cq->ibv_cq.ibv_cq_ex;
rx_ibv_cq = rx_rdm_cq ? rx_rdm_cq->ibv_cq.ibv_cq_ex : tx_rdm_cq->ibv_cq.ibv_cq_ex;
efa_rdm_ep_construct_ibv_qp_init_attr_ex(ep, &attr_ex, tx_ibv_cq, rx_ibv_cq);
ret = efa_base_ep_create_qp(&ep->base_ep, &attr_ex);
if (ret)
return ret;
/**
* Create separate user_recv_qp to receive pkts that carries user data
* without any headers.
*/
if (ep->use_zcpy_rx) {
ret = efa_qp_create(&ep->base_ep.user_recv_qp, &attr_ex, ep->base_ep.info->tx_attr->tclass);
if (ret) {
efa_base_ep_destruct_qp(&ep->base_ep);
return ret;
}
ep->base_ep.user_recv_qp->base_ep = &ep->base_ep;
}
return FI_SUCCESS;
}
static
int efa_rdm_pke_pool_mr_reg_handler(struct ofi_bufpool_region *region)
{
size_t ret;
struct fid_mr *mr;
struct efa_domain *domain = region->pool->attr.context;
ret = fi_mr_reg(&domain->util_domain.domain_fid, region->alloc_region,
region->pool->alloc_size, FI_SEND | FI_RECV, 0, 0, 0,
&mr, NULL);
region->context = mr;
return ret;
}
static
void efa_rdm_pke_pool_mr_dereg_handler(struct ofi_bufpool_region *region)
{
ssize_t ret;
ret = fi_close((struct fid *)region->context);
if (ret)
EFA_WARN(FI_LOG_EP_CTRL,
"Unable to deregister memory in a buf pool: %s\n",
fi_strerror(-ret));
}
/**
* @brief creates a packet entry pool.
*
* The pool is allowed to grow if
* max_cnt is 0 and is fixed size otherwise.
*
* @param ep efa_rdm_ep
* @param pkt_pool_type type of pkt pool
* @param chunk_cnt count of chunks in the pool
* @param max_cnt maximal count of chunks
* @param alignment memory alignment
* @param pkt_pool pkt pool
* @return int 0 on success, a negative integer on failure
*/
int efa_rdm_ep_create_pke_pool(struct efa_rdm_ep *ep,
bool need_mr,
size_t chunk_cnt,
size_t max_cnt,
size_t alignment,
struct ofi_bufpool **pke_pool)
{
/*
* use bufpool flags to make sure that no data structures can share
* the memory pages used for this buffer pool if the pool's memory
* need to be registered with EFA device.
*
* Using huge page has a small performance advantage, so we use it
* unless it is explicitly prihibitted by user.
*
* When the bufpools's memory need to be registered, we use
* either OFI_BUFPOOL_NONSHARED or OFI_BUFPOOL_HUGEPAGES, both
* would ensure that the allocated memory for bufpool does not share
* page with other memory regions. This is because memory registration
* is page based, e.g. it will always register the whole page.
*
* This is especially important when rdma-core's fork support is turned on,
* which will mark the entire pages of registered memory to be MADV_DONTFORK.
* As a result, the child process does not have the page in its memory space.
*/
uint64_t mr_flags = (efa_env.huge_page_setting == EFA_ENV_HUGE_PAGE_DISABLED)
? OFI_BUFPOOL_NONSHARED
: OFI_BUFPOOL_HUGEPAGES;
struct ofi_bufpool_attr wiredata_attr = {
.size = sizeof(struct efa_rdm_pke) + ep->mtu_size,
.alignment = alignment,
.max_cnt = max_cnt,
.chunk_cnt = chunk_cnt,
.alloc_fn = need_mr ? efa_rdm_pke_pool_mr_reg_handler : NULL,
.free_fn = need_mr ? efa_rdm_pke_pool_mr_dereg_handler : NULL,
.init_fn = NULL,
.context = efa_rdm_ep_domain(ep),
.flags = need_mr ? mr_flags : 0,
};
return ofi_bufpool_create_attr(&wiredata_attr, pke_pool);
}
/** @brief initializes the various buffer pools of EFA RDM endpoint.
*
* called by efa_rdm_ep_open()
*
* @param ep efa_rdm_ep struct to initialize.
* @return 0 on success, fi_errno on error.
* @related #efa_rdm_ep
*/
int efa_rdm_ep_create_buffer_pools(struct efa_rdm_ep *ep)
{
int ret;
ret = efa_rdm_ep_create_pke_pool(
ep,
true, /* need memory registration */
efa_rdm_ep_get_tx_pool_size(ep),
efa_rdm_ep_get_tx_pool_size(ep), /* max count==chunk_cnt means pool is not allowed to grow */
EFA_RDM_BUFPOOL_ALIGNMENT,
&ep->efa_tx_pkt_pool);
if (ret)
goto err_free;
ret = efa_rdm_ep_create_pke_pool(
ep,
true, /* need memory registration */
efa_rdm_ep_get_rx_pool_size(ep),
efa_rdm_ep_get_rx_pool_size(ep), /* max count==chunk_cnt means pool is not allowed to grow */
EFA_RDM_BUFPOOL_ALIGNMENT,
&ep->efa_rx_pkt_pool);
if (ret)
goto err_free;
ret = ofi_bufpool_create(&ep->user_rx_pkt_pool,
sizeof(struct efa_rdm_pke),
EFA_RDM_BUFPOOL_ALIGNMENT,
ep->base_ep.info->rx_attr->size,
ep->base_ep.info->rx_attr->size, /* max count==chunk_cnt means pool is not allowed to grow */
0);
if (ret)
goto err_free;
if (efa_env.rx_copy_unexp) {
ret = efa_rdm_ep_create_pke_pool(
ep,
false, /* do not need memory registration */
efa_env.unexp_pool_chunk_size,
0, /* max count = 0, so pool is allowed to grow */
EFA_RDM_BUFPOOL_ALIGNMENT,
&ep->rx_unexp_pkt_pool);
if (ret)
goto err_free;
}
if (efa_env.rx_copy_ooo) {
ret = efa_rdm_ep_create_pke_pool(
ep,
false, /* do not need memory registration */
efa_env.ooo_pool_chunk_size,
0, /* max count = 0, so pool is allowed to grow */
EFA_RDM_BUFPOOL_ALIGNMENT,
&ep->rx_ooo_pkt_pool);
if (ret)
goto err_free;
}
if ((efa_env.rx_copy_unexp || efa_env.rx_copy_ooo) &&
(efa_rdm_ep_domain(ep)->util_domain.mr_mode & FI_MR_HMEM)) {
/* this pool is only needed when application requested FI_HMEM capability */
ret = efa_rdm_ep_create_pke_pool(
ep,
true, /* need memory registration */
efa_env.readcopy_pool_size,
efa_env.readcopy_pool_size, /* max_cnt==chunk_cnt means pool is not allowed to grow */
EFA_RDM_IN_ORDER_ALIGNMENT, /* support in-order aligned send/recv */
&ep->rx_readcopy_pkt_pool);
if (ret)
goto err_free;
ep->rx_readcopy_pkt_pool_used = 0;
ep->rx_readcopy_pkt_pool_max_used = 0;
}
ret = ofi_bufpool_create(&ep->map_entry_pool,
sizeof(struct efa_rdm_rxe_map_entry),
EFA_RDM_BUFPOOL_ALIGNMENT,
0, /* no limit for max_cnt */
ep->base_ep.info->rx_attr->size, 0);
if (ret)
goto err_free;
ret = ofi_bufpool_create(&ep->rx_atomrsp_pool, ep->mtu_size,
EFA_RDM_BUFPOOL_ALIGNMENT,
0, /* no limit for max_cnt */
efa_env.atomrsp_pool_size, 0);
if (ret)
goto err_free;
ret = ofi_bufpool_create(&ep->ope_pool,
sizeof(struct efa_rdm_ope),
EFA_RDM_BUFPOOL_ALIGNMENT,
0, /* no limit for max_cnt */
ep->base_ep.info->tx_attr->size + ep->base_ep.info->rx_attr->size, 0);
if (ret)
goto err_free;
ret = ofi_bufpool_create(&ep->overflow_pke_pool,
sizeof(struct efa_rdm_peer_overflow_pke_list_entry),
EFA_RDM_BUFPOOL_ALIGNMENT,
0, /* no limit for max_cnt */
ep->base_ep.info->rx_attr->size, 0);
if (ret)
goto err_free;
efa_rdm_rxe_map_construct(&ep->rxe_map);
return 0;
err_free:
if (ep->rx_atomrsp_pool)
ofi_bufpool_destroy(ep->rx_atomrsp_pool);
if (ep->map_entry_pool)
ofi_bufpool_destroy(ep->map_entry_pool);
if (ep->ope_pool)
ofi_bufpool_destroy(ep->ope_pool);
if (ep->overflow_pke_pool)
ofi_bufpool_destroy(ep->overflow_pke_pool);
if (ep->rx_readcopy_pkt_pool)
ofi_bufpool_destroy(ep->rx_readcopy_pkt_pool);
if (efa_env.rx_copy_ooo && ep->rx_ooo_pkt_pool)
ofi_bufpool_destroy(ep->rx_ooo_pkt_pool);
if (efa_env.rx_copy_unexp && ep->rx_unexp_pkt_pool)
ofi_bufpool_destroy(ep->rx_unexp_pkt_pool);
if (ep->user_rx_pkt_pool)
ofi_bufpool_destroy(ep->user_rx_pkt_pool);
if (ep->efa_rx_pkt_pool)
ofi_bufpool_destroy(ep->efa_rx_pkt_pool);
if (ep->efa_tx_pkt_pool)
ofi_bufpool_destroy(ep->efa_tx_pkt_pool);
return ret;
}
/**
* @brief Initialize the various linked lists in an EFA RDM endpoint
* @param[in,out] ep EFA RDM endpoint
* @related #efa_rdm_ep
*/
void efa_rdm_ep_init_linked_lists(struct efa_rdm_ep *ep)
{
dlist_init(&ep->rx_posted_buf_list);
#if ENABLE_DEBUG
dlist_init(&ep->ope_recv_list);
dlist_init(&ep->rx_pkt_list);
dlist_init(&ep->tx_pkt_list);
#endif
dlist_init(&ep->rxe_list);
dlist_init(&ep->txe_list);
}
/**
* @brief function pointers to libfabric connection management API
*/
struct fi_ops_cm efa_rdm_ep_cm_ops = {
.size = sizeof(struct fi_ops_cm),
.setname = fi_no_setname,
.getname = efa_base_ep_getname,
.getpeer = fi_no_getpeer,
.connect = fi_no_connect,
.listen = fi_no_listen,
.accept = fi_no_accept,
.reject = fi_no_reject,
.shutdown = fi_no_shutdown,
.join = fi_no_join,
};
static int efa_rdm_ep_setopt(fid_t fid, int level, int optname,
const void *optval, size_t optlen);
static int efa_rdm_ep_getopt(fid_t fid, int level, int optname, void *optval,
size_t *optlen);
static ssize_t efa_rdm_ep_cancel(fid_t fid_ep, void *context);
/**
* @brief function pointers to libfabric endpoint's endpoint API
* These functions applies to an endpoint
*/
static struct fi_ops_ep efa_rdm_ep_ep_ops = {
.size = sizeof(struct fi_ops_ep),
.cancel = efa_rdm_ep_cancel,
.getopt = efa_rdm_ep_getopt,
.setopt = efa_rdm_ep_setopt,
.tx_ctx = fi_no_tx_ctx,
.rx_ctx = fi_no_rx_ctx,
.rx_size_left = fi_no_rx_size_left,
.tx_size_left = fi_no_tx_size_left,
};
static int efa_rdm_ep_close(struct fid *fid);
static int efa_rdm_ep_ctrl(struct fid *fid, int command, void *arg);
static int efa_rdm_ep_bind(struct fid *ep_fid, struct fid *bfid, uint64_t flags);
/**
* @brief function pointers to libfabric endpoint's fabric interface API
* These functions applies to a libfabric object
*/
static struct fi_ops efa_rdm_ep_base_ops = {
.size = sizeof(struct fi_ops),
.close = efa_rdm_ep_close,
.bind = efa_rdm_ep_bind,
.control = efa_rdm_ep_ctrl,
.ops_open = fi_no_ops_open,
};
/**
* @brief set the "use_zcpy_rx" flag in an EFA RDM endpoint.
* called by efa_rdm_ep_open()
*
* @param[in,out] ep EFA RDM endpoint
*/
static inline
void efa_rdm_ep_set_use_zcpy_rx(struct efa_rdm_ep *ep)
{
enum fi_hmem_iface iface;
uint64_t unsupported_caps = FI_DIRECTED_RECV | FI_TAGGED | FI_ATOMIC;
ep->use_zcpy_rx = true;
/* User requests to turn off zcpy recv */
if (!efa_env.use_zcpy_rx) {
EFA_INFO(FI_LOG_EP_CTRL, "User disables zero-copy receive protocol via environment\n");
ep->use_zcpy_rx = false;
goto out;
}
/* Unsupported capabilities */
if (ep->base_ep.util_ep.caps & unsupported_caps) {
EFA_INFO(FI_LOG_EP_CTRL, "Unsupported capabilities, zero-copy receive protocol will be disabled\n");
ep->use_zcpy_rx = false;
goto out;
}
/* Max msg size is too large, turn off zcpy recv */
if (ep->base_ep.max_msg_size > ep->mtu_size - ep->base_ep.info->ep_attr->msg_prefix_size) {
EFA_INFO(FI_LOG_EP_CTRL,
"max_msg_size (%zu) is greater than the mtu size limit: %zu. "
"Zero-copy receive protocol will be disabled.\n",
ep->base_ep.max_msg_size,
ep->mtu_size - ep->base_ep.info->ep_attr->msg_prefix_size);
ep->use_zcpy_rx = false;
goto out;
}
/* If app needs sas ordering, turn off zcpy recv */
if (efa_rdm_ep_need_sas(ep)) {
EFA_INFO(FI_LOG_EP_CTRL, "FI_ORDER_SAS is requested, zero-copy receive protocol will be disabled\n");
ep->use_zcpy_rx = false;
goto out;
}
/* FI_MR_LOCAL is not set, turn off zcpy recv */
if (!ep->base_ep.domain->mr_local) {
EFA_INFO(FI_LOG_EP_CTRL, "FI_MR_LOCAL mode bit is not set, zero-copy receive protocol will be disabled\n");
ep->use_zcpy_rx = false;
goto out;
}
if (ep->shm_ep) {
EFA_INFO(FI_LOG_EP_CTRL, "Libfabric SHM is not turned off, zero-copy receive protocol will be disabled\n");
ep->use_zcpy_rx = false;
goto out;
}
/* Zero-copy receive requires P2P support. Disable it if any initialized HMEM iface does not support P2P. */
EFA_HMEM_IFACE_FOREACH_NON_SYSTEM(iface) {
if (g_efa_hmem_info[iface].initialized &&
(ofi_hmem_p2p_disabled() ||
ep->hmem_p2p_opt == FI_HMEM_P2P_DISABLED ||
!g_efa_hmem_info[iface].p2p_supported_by_device)) {
EFA_INFO(FI_LOG_EP_CTRL,
"%s does not support P2P, zero-copy receive "
"protocol will be disabled\n",
fi_tostr(&iface, FI_TYPE_HMEM_IFACE));
ep->use_zcpy_rx = false;
goto out;
}
}
out:
EFA_INFO(FI_LOG_EP_CTRL, "efa_rdm_ep->use_zcpy_rx = %d\n",
ep->use_zcpy_rx);
return;
}
/**
* @brief progress engine for the EFA RDM endpoint
*
* This function now a no-op.
*
* @param[in] util_ep The endpoint FID to progress
*/
static
void efa_rdm_ep_progress_no_op(struct util_ep *util_ep)
{
return;
}
/**
* @brief implement the fi_endpoint() API for EFA RDM endpoint
*
* @param[in,out] domain The domain this endpoint belongs to
* @param[in] info The info struct used to create this endpoint
* @param[out] ep The endpoint to be created
* @param[in] context The context associated with this endpoint
* @return int 0 on success, negative libfabric error code otherwise
*/
int efa_rdm_ep_open(struct fid_domain *domain, struct fi_info *info,
struct fid_ep **ep, void *context)
{
struct efa_domain *efa_domain = NULL;
struct efa_rdm_ep *efa_rdm_ep = NULL;
int ret, retv, i;
enum fi_hmem_iface iface;
efa_rdm_ep = calloc(1, sizeof(*efa_rdm_ep));
if (!efa_rdm_ep)
return -FI_ENOMEM;
efa_domain = container_of(domain, struct efa_domain,
util_domain.domain_fid);
ret = efa_base_ep_construct(&efa_rdm_ep->base_ep, domain, info,
efa_rdm_ep_progress_no_op, context);
if (ret)
goto err_free_ep;
if (efa_domain->shm_domain) {
ret = fi_endpoint(efa_domain->shm_domain, efa_domain->shm_info,
&efa_rdm_ep->shm_ep, efa_rdm_ep);
if (ret)
goto err_destroy_base_ep;
} else {
efa_rdm_ep->shm_ep = NULL;
}
efa_rdm_ep->host_id = efa_get_host_id(efa_env.host_id_file);
if (efa_rdm_ep->host_id) {
EFA_INFO(FI_LOG_EP_CTRL, "efa_rdm_ep->host_id: i-%017lx\n", efa_rdm_ep->host_id);
}
efa_rdm_ep->max_tagged_size = info->ep_attr->max_msg_size;
efa_rdm_ep->max_atomic_size = info->ep_attr->max_msg_size;
efa_rdm_ep->inject_tagged_size = info->tx_attr->inject_size;
efa_rdm_ep->inject_atomic_size = info->tx_attr->inject_size;
efa_rdm_ep->base_ep.inject_rma_size = info->tx_attr->inject_size;
efa_rdm_ep->efa_max_outstanding_tx_ops = efa_domain->device->rdm_info->tx_attr->size;
efa_rdm_ep->efa_max_outstanding_rx_ops = efa_domain->device->rdm_info->rx_attr->size;
efa_rdm_ep->use_device_rdma = efa_rdm_get_use_device_rdma(info->fabric_attr->api_version);
efa_rdm_ep->shm_permitted = true;
efa_rdm_ep->msg_prefix_size = info->ep_attr->msg_prefix_size;
efa_rdm_ep->mtu_size = efa_domain->device->rdm_info->ep_attr->max_msg_size;
efa_rdm_ep->max_data_payload_size = efa_rdm_ep->mtu_size - sizeof(struct efa_rdm_ctsdata_hdr) - sizeof(struct efa_rdm_ctsdata_opt_connid_hdr);
efa_rdm_ep->min_multi_recv_size = efa_rdm_ep->mtu_size - efa_rdm_pkt_type_get_max_hdr_size();
if (efa_env.tx_queue_size > 0 &&
efa_env.tx_queue_size < efa_rdm_ep->efa_max_outstanding_tx_ops)
efa_rdm_ep->efa_max_outstanding_tx_ops = efa_env.tx_queue_size;
efa_rdm_ep->handle_resource_management = info->domain_attr->resource_mgmt;
EFA_INFO(FI_LOG_EP_CTRL,
"efa_rdm_ep->handle_resource_management = %d\n",
efa_rdm_ep->handle_resource_management);
#if ENABLE_DEBUG
efa_rdm_ep->efa_total_posted_tx_ops = 0;
efa_rdm_ep->send_comps = 0;
efa_rdm_ep->recv_comps = 0;
#endif
efa_rdm_ep->efa_rx_pkts_posted = 0;
efa_rdm_ep->efa_rx_pkts_to_post = 0;
efa_rdm_ep->efa_rx_pkts_held = 0;
efa_rdm_ep->efa_outstanding_tx_ops = 0;
if (ret) {
EFA_WARN(FI_LOG_CQ, "Unable to create extended CQ: %s\n", strerror(errno));
goto err_close_shm_ep;
}
ret = efa_rdm_ep_create_buffer_pools(efa_rdm_ep);
if (ret)
goto err_close_shm_ep;
efa_rdm_ep_init_linked_lists(efa_rdm_ep);
efa_rdm_ep->cuda_api_permitted = (FI_VERSION_GE(info->fabric_attr->api_version, FI_VERSION(1, 18)));
/* Set hmem_p2p_opt */
efa_rdm_ep->hmem_p2p_opt = FI_HMEM_P2P_DISABLED;
/*
* TODO this assumes only one non-stantard interface is initialized at a
* time. Refactor to handle multiple initialized interfaces to impose
* tighter requirements for the default p2p opt
*/
EFA_HMEM_IFACE_FOREACH_NON_SYSTEM(i) {
iface = efa_hmem_ifaces[i];
if (g_efa_hmem_info[iface].initialized &&
g_efa_hmem_info[iface].p2p_supported_by_device) {
/* If user is using libfabric API 1.18 or later, by default EFA
* provider is permitted to use CUDA library to support CUDA
* memory, therefore p2p is not required.
*/
efa_rdm_ep->hmem_p2p_opt =
(iface == FI_HMEM_CUDA && efa_rdm_ep->cuda_api_permitted) ?
FI_HMEM_P2P_PREFERRED :
FI_HMEM_P2P_REQUIRED;
break;
}
}
efa_rdm_ep->sendrecv_in_order_aligned_128_bytes = false;
efa_rdm_ep->write_in_order_aligned_128_bytes = false;
efa_rdm_ep->pke_vec = calloc(sizeof(struct efa_rdm_pke *), EFA_RDM_EP_MAX_WR_PER_IBV_POST_RECV);
if (!efa_rdm_ep->pke_vec) {
EFA_WARN(FI_LOG_EP_CTRL, "cannot alloc memory for efa_rdm_ep->pke_vec!\n");
ret = -FI_ENOMEM;
goto err_close_shm_ep;
}
*ep = &efa_rdm_ep->base_ep.util_ep.ep_fid;
(*ep)->msg = &efa_rdm_msg_ops;
(*ep)->rma = &efa_rdm_rma_ops;
(*ep)->atomic = &efa_rdm_atomic_ops;
(*ep)->tagged = &efa_rdm_msg_tagged_ops;
(*ep)->fid.ops = &efa_rdm_ep_base_ops;
(*ep)->ops = &efa_rdm_ep_ep_ops;
(*ep)->cm = &efa_rdm_ep_cm_ops;
return 0;
err_close_shm_ep:
if (efa_rdm_ep->shm_ep) {
retv = fi_close(&efa_rdm_ep->shm_ep->fid);
if (retv)
EFA_WARN(FI_LOG_EP_CTRL, "Unable to close shm EP: %s\n",
fi_strerror(-retv));
}
err_destroy_base_ep:
efa_base_ep_destruct(&efa_rdm_ep->base_ep);
err_free_ep:
if (efa_rdm_ep)
free(efa_rdm_ep);
return ret;
}
/**
* @brief implement the fi_ep_bind API for EFA RDM endpoint
* Currently supported objects are: AV, CQ, CNTR, EQ
* @param ep_fid - endpoint fid
* @param bfid - fid of the object to be binded with the endpoint
* @param flags - bind flags
* @return 0 on success, negative libfabric error code otherwise
*/
static int efa_rdm_ep_bind(struct fid *ep_fid, struct fid *bfid, uint64_t flags)
{
struct efa_rdm_ep *efa_rdm_ep =
container_of(ep_fid, struct efa_rdm_ep, base_ep.util_ep.ep_fid.fid);
struct efa_rdm_cq *cq;
struct efa_av *av;
struct efa_cntr *cntr;
struct util_eq *eq;
int ret = 0;
switch (bfid->fclass) {
case FI_CLASS_AV:
av = container_of(bfid, struct efa_av, util_av.av_fid.fid);
/* Bind util provider endpoint and av */
ret = ofi_ep_bind_av(&efa_rdm_ep->base_ep.util_ep, &av->util_av);
if (ret)
return ret;
ret = efa_base_ep_bind_av(&efa_rdm_ep->base_ep, av);
if (ret)
return ret;
/* Bind shm provider endpoint & shm av */
if (efa_rdm_ep->shm_ep) {
assert(av->shm_rdm_av);
ret = fi_ep_bind(efa_rdm_ep->shm_ep, &av->shm_rdm_av->fid, flags);
if (ret)
return ret;
}
break;
case FI_CLASS_CQ:
cq = container_of(bfid, struct efa_rdm_cq, util_cq.cq_fid.fid);
ret = ofi_ep_bind_cq(&efa_rdm_ep->base_ep.util_ep, &cq->util_cq, flags);
if (ret)
return ret;
if (cq->shm_cq) {
/* Bind ep with shm provider's cq */
ret = fi_ep_bind(efa_rdm_ep->shm_ep, &cq->shm_cq->fid, flags);
if (ret)
return ret;
}
break;
case FI_CLASS_CNTR:
cntr = container_of(bfid, struct efa_cntr, util_cntr.cntr_fid.fid);
ret = ofi_ep_bind_cntr(&efa_rdm_ep->base_ep.util_ep, &cntr->util_cntr, flags);
if (ret)
return ret;
if (cntr->shm_cntr) {
/* Bind shm ep with shm provider's cntr */
ret = fi_ep_bind(efa_rdm_ep->shm_ep, &cntr->shm_cntr->fid, flags);
if (ret)
return ret;
}
break;
case FI_CLASS_EQ:
eq = container_of(bfid, struct util_eq, eq_fid.fid);
ret = ofi_ep_bind_eq(&efa_rdm_ep->base_ep.util_ep, eq);
if (ret)
return ret;
break;
default:
EFA_WARN(FI_LOG_EP_CTRL, "invalid fid class\n");
ret = -FI_EINVAL;
break;
}
return ret;
}
/**
* @brief destroy the various buffer pools of an EFA RDM endpoint
* @param[in,out] efa_rdm_ep EFA RDM endpoint
* @related #efa_rdm_ep
*/
static void efa_rdm_ep_destroy_buffer_pools(struct efa_rdm_ep *efa_rdm_ep)
{
struct dlist_entry *entry, *tmp;
struct efa_rdm_ope *rxe;
struct efa_rdm_ope *txe;
#if ENABLE_DEBUG
struct efa_rdm_pke *pkt_entry;
dlist_foreach_safe(&efa_rdm_ep->rx_posted_buf_list, entry, tmp) {
pkt_entry = container_of(entry, struct efa_rdm_pke, dbg_entry);
efa_rdm_pke_release_rx(pkt_entry);
}
dlist_foreach_safe(&efa_rdm_ep->rx_pkt_list, entry, tmp) {
pkt_entry = container_of(entry, struct efa_rdm_pke, dbg_entry);
EFA_WARN(FI_LOG_EP_CTRL,
"Closing ep with unreleased RX pkt_entry: %p\n",
pkt_entry);
efa_rdm_pke_release_rx(pkt_entry);
}
dlist_foreach_safe(&efa_rdm_ep->tx_pkt_list, entry, tmp) {
pkt_entry = container_of(entry, struct efa_rdm_pke, dbg_entry);
EFA_WARN(FI_LOG_EP_CTRL,
"Closing ep with unreleased TX pkt_entry: %p\n",
pkt_entry);
efa_rdm_pke_release_tx(pkt_entry);
}
#endif
dlist_foreach_safe(&efa_rdm_ep->rxe_list, entry, tmp) {
rxe = container_of(entry, struct efa_rdm_ope,
ep_entry);
EFA_WARN(FI_LOG_EP_CTRL,
"Closing ep with unreleased rxe\n");
efa_rdm_rxe_release(rxe);
}
dlist_foreach_safe(&efa_rdm_ep->txe_list, entry, tmp) {
txe = container_of(entry, struct efa_rdm_ope,
ep_entry);
EFA_WARN(FI_LOG_EP_CTRL,
"Closing ep with unreleased txe: %p\n",
txe);
efa_rdm_txe_release(txe);
}
if (efa_rdm_ep->ope_pool)
ofi_bufpool_destroy(efa_rdm_ep->ope_pool);
if (efa_rdm_ep->overflow_pke_pool)
ofi_bufpool_destroy(efa_rdm_ep->overflow_pke_pool);
if (efa_rdm_ep->map_entry_pool)
ofi_bufpool_destroy(efa_rdm_ep->map_entry_pool);
if (efa_rdm_ep->rx_readcopy_pkt_pool) {
EFA_INFO(FI_LOG_EP_CTRL, "current usage of read copy packet pool is %d\n",
efa_rdm_ep->rx_readcopy_pkt_pool_used);
EFA_INFO(FI_LOG_EP_CTRL, "maximum usage of read copy packet pool is %d\n",
efa_rdm_ep->rx_readcopy_pkt_pool_max_used);
assert(!efa_rdm_ep->rx_readcopy_pkt_pool_used);
ofi_bufpool_destroy(efa_rdm_ep->rx_readcopy_pkt_pool);
}
if (efa_rdm_ep->rx_ooo_pkt_pool)
ofi_bufpool_destroy(efa_rdm_ep->rx_ooo_pkt_pool);
if (efa_rdm_ep->rx_unexp_pkt_pool)
ofi_bufpool_destroy(efa_rdm_ep->rx_unexp_pkt_pool);
if (efa_rdm_ep->user_rx_pkt_pool)
ofi_bufpool_destroy(efa_rdm_ep->user_rx_pkt_pool);
if (efa_rdm_ep->efa_rx_pkt_pool)
ofi_bufpool_destroy(efa_rdm_ep->efa_rx_pkt_pool);
if (efa_rdm_ep->efa_tx_pkt_pool)
ofi_bufpool_destroy(efa_rdm_ep->efa_tx_pkt_pool);
if (efa_rdm_ep->rx_atomrsp_pool)
ofi_bufpool_destroy(efa_rdm_ep->rx_atomrsp_pool);
}
/*
* @brief determine whether an endpoint has unfinished send
*
* Unfinished send includes queued ctrl packets, queued
* RNR packets and inflight TX packets.
*
* @param[in] efa_rdm_ep endpoint
* @return a boolean
*/
static
bool efa_rdm_ep_has_unfinished_send(struct efa_rdm_ep *efa_rdm_ep)
{
struct dlist_entry *entry, *tmp;
struct efa_rdm_ope *ope;
/* Only flush the opes queued due to rnr and ctrl */
uint64_t queued_ope_flags = EFA_RDM_OPE_QUEUED_CTRL | EFA_RDM_OPE_QUEUED_RNR;
if (efa_rdm_ep->efa_outstanding_tx_ops > 0)
return true;
dlist_foreach_safe(&efa_rdm_ep_domain(efa_rdm_ep)->ope_queued_list, entry, tmp) {
ope = container_of(entry, struct efa_rdm_ope,
queued_entry);
if (ope->ep == efa_rdm_ep && (ope->internal_flags & queued_ope_flags)) {
return true;
}
}
return false;
}
/*
* @brief wait for send to finish
*
* Wait for queued packet to be sent, and inflight send to
* complete.
*
* @param[in] efa_rdm_ep endpoint
* @return no return
*/
static inline
void efa_rdm_ep_wait_send(struct efa_rdm_ep *efa_rdm_ep)
{
struct efa_rdm_cq *tx_cq, *rx_cq;
ofi_genlock_lock(&efa_rdm_ep_domain(efa_rdm_ep)->srx_lock);
tx_cq = efa_rdm_ep_get_tx_rdm_cq(efa_rdm_ep);
rx_cq = efa_rdm_ep_get_rx_rdm_cq(efa_rdm_ep);
while (efa_rdm_ep_has_unfinished_send(efa_rdm_ep)) {
/* poll cq until empty */
if (tx_cq)
efa_rdm_cq_poll_ibv_cq(-1, &tx_cq->ibv_cq);
if (rx_cq)
efa_rdm_cq_poll_ibv_cq(-1, &rx_cq->ibv_cq);
efa_domain_progress_rdm_peers_and_queues(efa_rdm_ep_domain(efa_rdm_ep));
}
ofi_genlock_unlock(&efa_rdm_ep_domain(efa_rdm_ep)->srx_lock);
}
static inline
void efa_rdm_ep_remove_cntr_ibv_cq_poll_list(struct efa_rdm_ep *ep)
{
int i;
struct efa_cntr *efa_cntr;
struct util_cntr *util_cntr;
struct efa_rdm_cq *tx_cq, *rx_cq;
tx_cq = efa_rdm_ep_get_tx_rdm_cq(ep);
rx_cq = efa_rdm_ep_get_rx_rdm_cq(ep);
for (i = 0; i< CNTR_CNT; i++) {
util_cntr = ep->base_ep.util_ep.cntrs[i];
if (util_cntr) {
efa_cntr = container_of(util_cntr, struct efa_cntr, util_cntr);
if (tx_cq && !ofi_atomic_get32(&tx_cq->util_cq.ref))
efa_ibv_cq_poll_list_remove(&efa_cntr->ibv_cq_poll_list, &efa_cntr->util_cntr.ep_list_lock, &tx_cq->ibv_cq);
if (rx_cq && !ofi_atomic_get32(&rx_cq->util_cq.ref))
efa_ibv_cq_poll_list_remove(&efa_cntr->ibv_cq_poll_list, &efa_cntr->util_cntr.ep_list_lock, &rx_cq->ibv_cq);
}
}
}
static inline
void efa_rdm_ep_remove_cq_ibv_cq_poll_list(struct efa_rdm_ep *ep)
{
struct efa_rdm_cq *tx_cq, *rx_cq;
tx_cq = efa_rdm_ep_get_tx_rdm_cq(ep);
rx_cq = efa_rdm_ep_get_rx_rdm_cq(ep);
/* Remove the cross referencing of the CQs.
* It must happen after ofi_endpoint_close
* so we have cq's reference counters updated.
*/
if (tx_cq && !ofi_atomic_get32(&tx_cq->util_cq.ref)) {
efa_ibv_cq_poll_list_remove(&tx_cq->ibv_cq_poll_list, &tx_cq->util_cq.ep_list_lock, &tx_cq->ibv_cq);
if (rx_cq)
efa_ibv_cq_poll_list_remove(&rx_cq->ibv_cq_poll_list, &rx_cq->util_cq.ep_list_lock, &tx_cq->ibv_cq);
}
if (rx_cq && !ofi_atomic_get32(&rx_cq->util_cq.ref)) {
efa_ibv_cq_poll_list_remove(&rx_cq->ibv_cq_poll_list, &rx_cq->util_cq.ep_list_lock, &rx_cq->ibv_cq);
if (tx_cq)
efa_ibv_cq_poll_list_remove(&tx_cq->ibv_cq_poll_list, &tx_cq->util_cq.ep_list_lock, &rx_cq->ibv_cq);
}
}
/**
* @brief Clean efa_rdm_ep's shm ep level resources as the best effort
*
* @param efa_rdm_ep pointer to efa rdm ep
* @return int FI_SUCCESS on success, negative integer on failure
*/
static int efa_rdm_ep_close_shm_ep_resources(struct efa_rdm_ep *efa_rdm_ep)
{
int ret, retv = 0;
if (efa_rdm_ep->shm_srx) {
ret = fi_close(&efa_rdm_ep->shm_srx->fid);
if (ret) {
EFA_WARN(FI_LOG_EP_CTRL, "Unable to close shm srx\n");
retv = ret;
}
efa_rdm_ep->shm_srx = NULL;
}
if (efa_rdm_ep->shm_peer_srx) {
free(efa_rdm_ep->shm_peer_srx);
efa_rdm_ep->shm_peer_srx = NULL;
}
if (efa_rdm_ep->shm_ep) {
ret = fi_close(&efa_rdm_ep->shm_ep->fid);
if (ret) {
EFA_WARN(FI_LOG_EP_CTRL, "Unable to close shm ep\n");
retv = ret;
}
efa_rdm_ep->shm_ep = NULL;
}
return retv;
}
/**
* @brief implement the fi_close() API for the EFA RDM endpoint
* @param[in,out] fid Endpoint to close
*/
static int efa_rdm_ep_close(struct fid *fid)
{
int ret, retv = 0;
struct efa_rdm_ep *efa_rdm_ep;
efa_rdm_ep = container_of(fid, struct efa_rdm_ep, base_ep.util_ep.ep_fid.fid);
if (efa_rdm_ep->base_ep.efa_qp_enabled)
efa_rdm_ep_wait_send(efa_rdm_ep);
/*
* util_srx_close will clean all efa_rdm_rxes that are
* associated with peer_rx_entries in unexp msg/tag lists.
* It also decrements the ref count of rx cq. So it must
* be called before we clean up the ibv cq poll list which
* relies on the correct ref count of tx/rx cq.
*/