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
|
// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (c) 2015, 2017 Oracle. All rights reserved.
* Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
*/
/* Lightweight memory registration using Fast Registration Work
* Requests (FRWR).
*
* FRWR features ordered asynchronous registration and deregistration
* of arbitrarily sized memory regions. This is the fastest and safest
* but most complex memory registration mode.
*/
/* Normal operation
*
* A Memory Region is prepared for RDMA READ or WRITE using a FAST_REG
* Work Request (frwr_map). When the RDMA operation is finished, this
* Memory Region is invalidated using a LOCAL_INV Work Request
* (frwr_unmap_sync).
*
* Typically these Work Requests are not signaled, and neither are RDMA
* SEND Work Requests (with the exception of signaling occasionally to
* prevent provider work queue overflows). This greatly reduces HCA
* interrupt workload.
*
* As an optimization, frwr_unmap marks MRs INVALID before the
* LOCAL_INV WR is posted. If posting succeeds, the MR is placed on
* rb_mrs immediately so that no work (like managing a linked list
* under a spinlock) is needed in the completion upcall.
*
* But this means that frwr_map() can occasionally encounter an MR
* that is INVALID but the LOCAL_INV WR has not completed. Work Queue
* ordering prevents a subsequent FAST_REG WR from executing against
* that MR while it is still being invalidated.
*/
/* Transport recovery
*
* ->op_map and the transport connect worker cannot run at the same
* time, but ->op_unmap can fire while the transport connect worker
* is running. Thus MR recovery is handled in ->op_map, to guarantee
* that recovered MRs are owned by a sending RPC, and not one where
* ->op_unmap could fire at the same time transport reconnect is
* being done.
*
* When the underlying transport disconnects, MRs are left in one of
* four states:
*
* INVALID: The MR was not in use before the QP entered ERROR state.
*
* VALID: The MR was registered before the QP entered ERROR state.
*
* FLUSHED_FR: The MR was being registered when the QP entered ERROR
* state, and the pending WR was flushed.
*
* FLUSHED_LI: The MR was being invalidated when the QP entered ERROR
* state, and the pending WR was flushed.
*
* When frwr_map encounters FLUSHED and VALID MRs, they are recovered
* with ib_dereg_mr and then are re-initialized. Because MR recovery
* allocates fresh resources, it is deferred to a workqueue, and the
* recovered MRs are placed back on the rb_mrs list when recovery is
* complete. frwr_map allocates another MR for the current RPC while
* the broken MR is reset.
*
* To ensure that frwr_map doesn't encounter an MR that is marked
* INVALID but that is about to be flushed due to a previous transport
* disconnect, the transport connect worker attempts to drain all
* pending send queue WRs before the transport is reconnected.
*/
#include <linux/sunrpc/rpc_rdma.h>
#include <linux/sunrpc/svc_rdma.h>
#include "xprt_rdma.h"
#include <trace/events/rpcrdma.h>
#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
# define RPCDBG_FACILITY RPCDBG_TRANS
#endif
/**
* frwr_is_supported - Check if device supports FRWR
* @device: interface adapter to check
*
* Returns true if device supports FRWR, otherwise false
*/
bool frwr_is_supported(struct ib_device *device)
{
struct ib_device_attr *attrs = &device->attrs;
if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
goto out_not_supported;
if (attrs->max_fast_reg_page_list_len == 0)
goto out_not_supported;
return true;
out_not_supported:
pr_info("rpcrdma: 'frwr' mode is not supported by device %s\n",
device->name);
return false;
}
/**
* frwr_release_mr - Destroy one MR
* @mr: MR allocated by frwr_init_mr
*
*/
void frwr_release_mr(struct rpcrdma_mr *mr)
{
int rc;
rc = ib_dereg_mr(mr->frwr.fr_mr);
if (rc)
trace_xprtrdma_frwr_dereg(mr, rc);
kfree(mr->mr_sg);
kfree(mr);
}
/* MRs are dynamically allocated, so simply clean up and release the MR.
* A replacement MR will subsequently be allocated on demand.
*/
static void
frwr_mr_recycle_worker(struct work_struct *work)
{
struct rpcrdma_mr *mr = container_of(work, struct rpcrdma_mr, mr_recycle);
struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
trace_xprtrdma_mr_recycle(mr);
if (mr->mr_dir != DMA_NONE) {
trace_xprtrdma_mr_unmap(mr);
ib_dma_unmap_sg(r_xprt->rx_ia.ri_id->device,
mr->mr_sg, mr->mr_nents, mr->mr_dir);
mr->mr_dir = DMA_NONE;
}
spin_lock(&r_xprt->rx_buf.rb_mrlock);
list_del(&mr->mr_all);
r_xprt->rx_stats.mrs_recycled++;
spin_unlock(&r_xprt->rx_buf.rb_mrlock);
frwr_release_mr(mr);
}
/**
* frwr_init_mr - Initialize one MR
* @ia: interface adapter
* @mr: generic MR to prepare for FRWR
*
* Returns zero if successful. Otherwise a negative errno
* is returned.
*/
int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr)
{
unsigned int depth = ia->ri_max_frwr_depth;
struct scatterlist *sg;
struct ib_mr *frmr;
int rc;
frmr = ib_alloc_mr(ia->ri_pd, ia->ri_mrtype, depth);
if (IS_ERR(frmr))
goto out_mr_err;
sg = kcalloc(depth, sizeof(*sg), GFP_KERNEL);
if (!sg)
goto out_list_err;
mr->frwr.fr_mr = frmr;
mr->frwr.fr_state = FRWR_IS_INVALID;
mr->mr_dir = DMA_NONE;
INIT_LIST_HEAD(&mr->mr_list);
INIT_WORK(&mr->mr_recycle, frwr_mr_recycle_worker);
init_completion(&mr->frwr.fr_linv_done);
sg_init_table(sg, depth);
mr->mr_sg = sg;
return 0;
out_mr_err:
rc = PTR_ERR(frmr);
trace_xprtrdma_frwr_alloc(mr, rc);
return rc;
out_list_err:
dprintk("RPC: %s: sg allocation failure\n",
__func__);
ib_dereg_mr(frmr);
return -ENOMEM;
}
/**
* frwr_open - Prepare an endpoint for use with FRWR
* @ia: interface adapter this endpoint will use
* @ep: endpoint to prepare
*
* On success, sets:
* ep->rep_attr.cap.max_send_wr
* ep->rep_attr.cap.max_recv_wr
* ep->rep_max_requests
* ia->ri_max_segs
*
* And these FRWR-related fields:
* ia->ri_max_frwr_depth
* ia->ri_mrtype
*
* On failure, a negative errno is returned.
*/
int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep)
{
struct ib_device_attr *attrs = &ia->ri_id->device->attrs;
int max_qp_wr, depth, delta;
ia->ri_mrtype = IB_MR_TYPE_MEM_REG;
if (attrs->device_cap_flags & IB_DEVICE_SG_GAPS_REG)
ia->ri_mrtype = IB_MR_TYPE_SG_GAPS;
/* Quirk: Some devices advertise a large max_fast_reg_page_list_len
* capability, but perform optimally when the MRs are not larger
* than a page.
*/
if (attrs->max_sge_rd > 1)
ia->ri_max_frwr_depth = attrs->max_sge_rd;
else
ia->ri_max_frwr_depth = attrs->max_fast_reg_page_list_len;
if (ia->ri_max_frwr_depth > RPCRDMA_MAX_DATA_SEGS)
ia->ri_max_frwr_depth = RPCRDMA_MAX_DATA_SEGS;
dprintk("RPC: %s: max FR page list depth = %u\n",
__func__, ia->ri_max_frwr_depth);
/* Add room for frwr register and invalidate WRs.
* 1. FRWR reg WR for head
* 2. FRWR invalidate WR for head
* 3. N FRWR reg WRs for pagelist
* 4. N FRWR invalidate WRs for pagelist
* 5. FRWR reg WR for tail
* 6. FRWR invalidate WR for tail
* 7. The RDMA_SEND WR
*/
depth = 7;
/* Calculate N if the device max FRWR depth is smaller than
* RPCRDMA_MAX_DATA_SEGS.
*/
if (ia->ri_max_frwr_depth < RPCRDMA_MAX_DATA_SEGS) {
delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frwr_depth;
do {
depth += 2; /* FRWR reg + invalidate */
delta -= ia->ri_max_frwr_depth;
} while (delta > 0);
}
max_qp_wr = ia->ri_id->device->attrs.max_qp_wr;
max_qp_wr -= RPCRDMA_BACKWARD_WRS;
max_qp_wr -= 1;
if (max_qp_wr < RPCRDMA_MIN_SLOT_TABLE)
return -ENOMEM;
if (ep->rep_max_requests > max_qp_wr)
ep->rep_max_requests = max_qp_wr;
ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
if (ep->rep_attr.cap.max_send_wr > max_qp_wr) {
ep->rep_max_requests = max_qp_wr / depth;
if (!ep->rep_max_requests)
return -EINVAL;
ep->rep_attr.cap.max_send_wr = ep->rep_max_requests * depth;
}
ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
ep->rep_attr.cap.max_send_wr += 1; /* for ib_drain_sq */
ep->rep_attr.cap.max_recv_wr = ep->rep_max_requests;
ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
ep->rep_attr.cap.max_recv_wr += 1; /* for ib_drain_rq */
ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
ia->ri_max_frwr_depth);
/* Reply chunks require segments for head and tail buffers */
ia->ri_max_segs +=
|