summaryrefslogtreecommitdiffstats
path: root/src/plugins/rdma/output.c
blob: 015208c023e8ee12a9f5abd58d6d309b218b6d24 (plain)
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
/*
 *------------------------------------------------------------------
 * Copyright (c) 2018 Cisco and/or its affiliates.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at:
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *------------------------------------------------------------------
 */

#include <vlib/vlib.h>
#include <vlib/unix/unix.h>
#include <vlib/pci/pci.h>
#include <vppinfra/ring.h>
#include <vnet/ethernet/ethernet.h>
#include <vnet/devices/devices.h>
#include <rdma/rdma.h>

#ifndef MLX5_ETH_L2_INLINE_HEADER_SIZE
#define MLX5_ETH_L2_INLINE_HEADER_SIZE  18
#endif

#define RDMA_TX_RETRIES 5

#define RDMA_TXQ_DV_DSEG_SZ(txq)        (RDMA_MLX5_WQE_DS * RDMA_TXQ_DV_SQ_SZ(txq))
#define RDMA_TXQ_DV_DSEG2WQE(d)         (((d) + RDMA_MLX5_WQE_DS - 1) / RDMA_MLX5_WQE_DS)

/*
 * MLX5 direct verbs tx/free functions
 */

static_always_inline void
rdma_device_output_free_mlx5 (vlib_main_t * vm,
			      const vlib_node_runtime_t * node,
			      rdma_txq_t * txq)
{
  u16 idx = txq->dv_cq_idx;
  u32 cq_mask = pow2_mask (txq->dv_cq_log2sz);
  u32 sq_mask = pow2_mask (txq->dv_sq_log2sz);
  u32 mask = pow2_mask (txq->bufs_log2sz);
  u32 buf_sz = RDMA_TXQ_BUF_SZ (txq);
  u32 log2_cq_sz = txq->dv_cq_log2sz;
  struct mlx5_cqe64 *cqes = txq->dv_cq_cqes, *cur = cqes + (idx & cq_mask);
  u8 op_own, saved;
  const rdma_mlx5_wqe_t *wqe;

  for (;;)
    {
      op_own = *(volatile u8 *) &cur->op_own;
      if (((idx >> log2_cq_sz) & MLX5_CQE_OWNER_MASK) !=
	  (op_own & MLX5_CQE_OWNER_MASK) || (op_own >> 4) == MLX5_CQE_INVALID)
	break;
      if (PREDICT_FALSE ((op_own >> 4)) != MLX5_CQE_REQ)
	vlib_error_count (vm, node->node_index, RDMA_TX_ERROR_COMPLETION, 1);
      idx++;
      cur = cqes + (idx & cq_mask);
    }

  if (idx == txq->dv_cq_idx)
    return;			/* nothing to do */

  cur = cqes + ((idx - 1) & cq_mask);
  saved = cur->op_own;
  (void) saved;
  cur->op_own = 0xf0;
  txq->dv_cq_idx = idx;

  /* retrieve original WQE and get new tail counter */
  wqe = txq->dv_sq_wqes + (be16toh (cur->wqe_counter) & sq_mask);
  if (PREDICT_FALSE (wqe->ctrl.imm == RDMA_TXQ_DV_INVALID_ID))
    return;			/* can happen if CQE reports error for an intermediate WQE */

  ASSERT (RDMA_TXQ_USED_SZ (txq->head, wqe->ctrl.imm) <= buf_sz &&
	  RDMA_TXQ_USED_SZ (wqe->ctrl.imm, txq->tail) < buf_sz);

  /* free sent buffers and update txq head */
  vlib_buffer_free_from_ring (vm, txq->bufs, txq->head & mask, buf_sz,
			      RDMA_TXQ_USED_SZ (txq->head, wqe->ctrl.imm));
  txq->head = wqe->ctrl.imm;

  /* ring doorbell */
  CLIB_MEMORY_STORE_BARRIER ();
  txq->dv_cq_dbrec[0] = htobe32 (idx);
}

static_always_inline void
rdma_device_output_tx_mlx5_doorbell (rdma_txq_t * txq, rdma_mlx5_wqe_t * last,
				     const u16 tail, u32 sq_mask)
{
  last->ctrl.imm = tail;	/* register item to free */
  last->ctrl.fm_ce_se = MLX5_WQE_CTRL_CQ_UPDATE;	/* generate a CQE so we can free buffers */

  ASSERT (tail != txq->tail &&
	  RDMA_TXQ_AVAIL_SZ (txq, txq->head, txq->tail) >=
	  RDMA_TXQ_USED_SZ (txq->tail, tail));

  CLIB_MEMORY_STORE_BARRIER ();
  txq->dv_sq_dbrec[MLX5_SND_DBR] = htobe32 (tail);
  CLIB_COMPILER_BARRIER ();
  txq->dv_sq_db[0] = *(u64 *) (txq->dv_sq_wqes + (txq->tail & sq_mask));
  txq->tail = tail;
}

static_always_inline void
rdma_mlx5_wqe_init (rdma_mlx5_wqe_t * wqe, const void *tmpl,
		    vlib_buffer_t * b, const u16 tail)
{
  u16 sz = b->current_length;
  u16 inline_sz = clib_min (sz, MLX5_ETH_L2_INLINE_HEADER_SIZE);

  clib_memcpy_fast (wqe, tmpl, RDMA_MLX5_WQE_SZ);

  wqe->ctrl.opmod_idx_opcode |= ((u32) htobe16 (tail)) << 8;
  /* speculatively copy at least MLX5_ETH_L2_INLINE_HEADER_SIZE (18-bytes) */
  const void *cur = vlib_buffer_get_current (b);
  clib_memcpy_fast (wqe->eseg.inline_hdr_start,
		    cur, MLX5_ETH_L2_INLINE_HEADER_SIZE);
  wqe->eseg.inline_hdr_sz = htobe16 (inline_sz);
  wqe->dseg.byte_count = htobe32 (sz - inline_sz);
  wqe->dseg.addr = htobe64 (pointer_to_uword (cur) + inline_sz);
}

/*
 * specific data path for chained buffers, supporting ring wrap-around
 * contrary to the normal path - otherwise we may fail to enqueue chained
 * buffers because we are close to the end of the ring while we still have
 * plenty of descriptors available
 */
static_always_inline u32
rdma_device_output_tx_mlx5_chained (vlib_main_t * vm,
				    const vlib_node_runtime_t * node,
				    const rdma_device_t * rd,
				    rdma_txq_t * txq, u32 n_left_from, u32 n,
				    u32 * bi, vlib_buffer_t ** b,
				    rdma_mlx5_wqe_t * wqe, u16 tail)
{
  rdma_mlx5_wqe_t *last = wqe;
  u32 wqe_n = RDMA_TXQ_AVAIL_SZ (txq, txq->head, tail);
  u32 sq_mask = pow2_mask (txq->dv_sq_log2sz);
  u32 mask = pow2_mask (txq->bufs_log2sz);
  u32 dseg_mask = RDMA_TXQ_DV_DSEG_SZ (txq) - 1;
  const u32 lkey = wqe[0].dseg.lkey;

  vlib_buffer_copy_indices (txq->bufs + (txq->tail & mask), bi,
			    n_left_from - n);

  while (n >= 1 && wqe_n >= 1)
    {
      u32 *bufs = txq->bufs + (tail & mask);
      rdma_mlx5_wqe_t *wqe = txq->dv_sq_wqes + (tail & sq_mask);

      /* setup the head WQE */
      rdma_mlx5_wqe_init (wqe, txq->dv_wqe_tmpl, b[0], tail);

      bufs[0] = bi[0];

      if (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT)
	{
	  /*
	   * max number of available dseg:
	   *  - 4 dseg per WQEBB available
	   *  - max 32 dseg per WQE (5-bits length field in WQE ctrl)
	   */
#define RDMA_MLX5_WQE_DS_MAX    (1 << 5)
	  const u32 dseg_max =
	    clib_min (RDMA_MLX5_WQE_DS * (wqe_n - 1), RDMA_MLX5_WQE_DS_MAX);
	  vlib_buffer_t *chained_b = b[0];
	  u32 chained_n = 0;

	  /* there are exactly 4 dseg per WQEBB and we rely on that */
	  STATIC_ASSERT (RDMA_MLX5_WQE_DS *
			 sizeof (struct mlx5_wqe_data_seg) ==
			 MLX5_SEND_WQE_BB, "wrong size");

	  /*
	   * iterate over fragments, supporting ring wrap-around contrary to
	   * the normal path - otherwise we may fail to enqueue chained
	   * buffers because we are close to the end of the ring while we
	   * still have plenty of descriptors available
	   */
	  while (chained_n < dseg_max
		 && chained_b->flags & VLIB_BUFFER_NEXT_PRESENT)
	    {
	      struct mlx5_wqe_data_seg *dseg = (void *) txq->dv_sq_wqes;
	      dseg += ((tail + 1) * RDMA_MLX5_WQE_DS + chained_n) & dseg_mask;
	      if (((clib_address_t) dseg & (MLX5_SEND_WQE_BB - 1)) == 0)
		{
		  /*
		   * start of new WQEBB
		   * head/tail are shared between buffers and descriptor
		   * In order to maintain 1:1 correspondance between
		   * buffer index and descriptor index, we build
		   * 4-fragments chains and save the head
		   */
		  chained_b->flags &= ~(VLIB_BUFFER_NEXT_PRESENT |
					VLIB_BUFFER_TOTAL_LENGTH_VALID);
		  u32 idx = tail + 1 + RDMA_TXQ_DV_DSEG2WQE (chained_n);
		  idx &= mask;
		  txq->bufs[idx] = chained_b->next_buffer;
		}

	      chained_b = vlib_get_buffer (vm, chained_b->next_buffer);
	      dseg->byte_count = htobe32 (chained_b->current_length);
	      dseg->lkey = lkey;
	      dseg->addr = htobe64 (vlib_buffer_get_current_va (chained_b));

	      chained_n += 1;
	    }

	  if (chained_b->flags & VLIB_BUFFER_NEXT_PRESENT)
	    {
	      /*
	       * no descriptors left: drop the chain including 1st WQE
	       * skip the problematic packet and continue
	       */
	      vlib_buffer_free_from_ring (vm, txq->bufs, tail & mask,
					  RDMA_TXQ_BUF_SZ (txq), 1 +
					  RDMA_TXQ_DV_DSEG2WQE (chained_n));
	      vlib_error_count (vm, node->node_index,
				dseg_max == chained_n ?
				RDMA_TX_ERROR_SEGMENT_SIZE_EXCEEDED :
				RDMA_TX_ERROR_NO_FREE_SLOTS, 1);

	      /* fixup tail to overwrite wqe head with next packet */
	      tail -= 1;
	    }
	  else
	    {
	      /* update WQE descriptor with new dseg number */
	      ((u8 *) & wqe[0].ctrl.qpn_ds)[3] = RDMA_MLX5_WQE_DS + chained_n;

	      tail += RDMA_TXQ_DV_DSEG2WQE (chained_n);
	      wqe_n -= RDMA_TXQ_DV_DSEG2WQE (chained_n);

	      last = wqe;
	    }
	}
      else
	{
	  /* not chained */
	  last = wqe;
	}

      tail += 1;
      bi += 1;
      b += 1;
      wqe_n -= 1;
      n -= 1;
    }

  if (n == n_left_from)
    return 0;			/* we fail to enqueue even a single packet */

  rdma_device_output_tx_mlx5_doorbell (txq, last, tail, sq_mask);
  return n_left_from - n;
}

static_always_inline u32
rdma_device_output_tx_mlx5 (vlib_main_t * vm,
			    const vlib_node_runtime_t * node,
			    const rdma_device_t * rd, rdma_txq_t * txq,
			    const u32 n_left_from, u32 * bi,
			    vlib_buffer_t ** b)
{
  u32 sq_mask = pow2_mask (txq->dv_sq_log2sz);
  u32 mask = pow2_mask (txq->bufs_log2sz);
  rdma_mlx5_wqe_t *wqe = txq->dv_sq_wqes + (txq->tail & sq_mask);
  u32 n = n_left_from;
  u16 tail = txq->tail;

  ASSERT (RDMA_TXQ_BUF_SZ (txq) <= RDMA_TXQ_DV_SQ_SZ (txq));

  while (n >= 4)
    {
      u32 flags = b[0]->flags | b[1]->flags | b[2]->flags | b[3]->flags;
      if (PREDICT_FALSE (flags & VLIB_BUFFER_NEXT_PRESENT))
	return rdma_device_output_tx_mlx5_chained (vm, node, rd, txq,
						   n_left_from, n, bi, b, wqe,
						   tail);

      if (PREDICT_TRUE (n >= 8))
	{
	  vlib_prefetch_buffer_header (b + 4, LOAD);
	  vlib_prefetch_buffer_header (b + 5, LOAD);
	  vlib_prefetch_buffer_header (b + 6, LOAD);
	  vlib_prefetch_buffer_header (b + 7, LOAD);
	  clib_prefetch_load (wqe + 4);
	}

      rdma_mlx5_wqe_init (wqe + 0, txq->dv_wqe_tmpl, b[0], tail + 0);
      rdma_mlx5_wqe_init (wqe + 1, txq->dv_wqe_tmpl, b[1], tail + 1);
      rdma_mlx5_wqe_init (wqe + 2, txq->dv_wqe_tmpl, b[2], tail + 2);
      rdma_mlx5_wqe_init (wqe + 3, txq->dv_wqe_tmpl, b[3], tail + 3);

      b += 4;
      tail += 4;
      wqe += 4;
      n -= 4;
    }

  while (n >= 1)
    {
      if (PREDICT_FALSE (b[0]->flags & VLIB_BUFFER_NEXT_PRESENT))
	return rdma_device_output_tx_mlx5_chained (vm, node, rd, txq,
						   n_left_from, n, bi, b, wqe,
						   tail);

      rdma_mlx5_wqe_init (wqe, txq->dv_wqe_tmpl, b[0], tail);

      b += 1;
      tail += 1;
      wqe += 1;
      n -= 1;
    }

  vlib_buffer_copy_indices (txq->bufs + (txq->tail & mask), bi, n_left_from);

  rdma_device_output_tx_mlx5_doorbell (txq, &wqe[-1], tail, sq_mask);
  return n_left_from;
}

/*
 * standard ibverb tx/free functions
 */

static_always_inline void
rdma_device_output_free_ibverb (vlib_main_t * vm,
				const vlib_node_runtime_t * node,
				rdma_txq_t * txq)
{
  struct ibv_wc wc[VLIB_FRAME_SIZE];
  u32 mask = pow2_mask (txq->bufs_log2sz);
  u16 tail;
  int n;

  n = ibv_poll_cq (txq->ibv_cq, VLIB_FRAME_SIZE, wc);
  if (n <= 0)
    {
      if (PREDICT_FALSE (n < 0))
	vlib_error_count (vm, node->node_index, RDMA_TX_ERROR_COMPLETION, 1);
      return;
    }

  while (PREDICT_FALSE (IBV_WC_SUCCESS != wc[n - 1].status))
    {
      vlib_error_count (vm, node->node_index, RDMA_TX_ERROR_COMPLETION, 1);
      n--;
      if (0 == n)
	return;
    }

  tail = wc[n - 1].wr_id;
  vlib_buffer_free_from_ring (vm, txq->bufs, txq->head & mask,
			      RDMA_TXQ_BUF_SZ (txq),
			      RDMA_TXQ_USED_SZ (txq->head, tail));
  txq->head = tail;
}

static_always_inline u32
rdma_device_output_tx_ibverb (vlib_main_t * vm,
			      const vlib_node_runtime_t * node,
			      const rdma_device_t * rd, rdma_txq_t * txq,
			      u32 n_left_from, u32 * bi, vlib_buffer_t ** b)
{
  struct ibv_send_wr wr[VLIB_FRAME_SIZE], *w = wr;
  struct ibv_sge sge[VLIB_FRAME_SIZE], *s = sge;
  u32 mask = txq->bufs_log2sz;
  u32 n = n_left_from;

  memset (w, 0, n_left_from * sizeof (w[0]));

  while (n >= 4)
    {
      if (PREDICT_TRUE (n >= 8))
	{
	  vlib_prefetch_buffer_header (b[4 + 0], LOAD);
	  vlib_prefetch_buffer_header (b[4 + 1], LOAD);
	  vlib_prefetch_buffer_header (b[4 + 2], LOAD);
	  vlib_prefetch_buffer_header (b[4 + 3], LOAD);

	  CLIB_PREFETCH (&s[4 + 0], 4 * sizeof (s[0]), STORE);

	  CLIB_PREFETCH (&w[4 + 0], CLIB_CACHE_LINE_BYTES, STORE);
	  CLIB_PREFETCH (&w[4 + 1], CLIB_CACHE_LINE_BYTES, STORE);
	  CLIB_PREFETCH (&w[4 + 2], CLIB_CACHE_LINE_BYTES, STORE);
	  CLIB_PREFETCH (&w[4 + 3], CLIB_CACHE_LINE_BYTES, STORE);
	}

      s[0].addr = vlib_buffer_get_current_va (b[0]);
      s[0].length = b[0]->current_length;
      s[0].lkey = rd->lkey;

      s[1].addr = vlib_buffer_get_current_va (b[1]);
      s[1].length = b[1]->current_length;
      s[1].lkey = rd->lkey;

      s[2].addr = vlib_buffer_get_current_va (b[2]);
      s[2].length = b[2]->current_length;
      s[2].lkey = rd->lkey;

      s[3].addr = vlib_buffer_get_current_va (b[3]);
      s[3].length = b[3]->current_length;
      s[3].lkey = rd->lkey;

      w[0].next = &w[0] + 1;
      w[0].sg_list = &s[0];
      w[0].num_sge = 1;
      w[0].opcode = IBV_WR_SEND;

      w[1].next = &w[1] + 1;
      w[1].sg_list = &s[1];
      w[1].num_sge = 1;
      w[1].opcode = IBV_WR_SEND;

      w[2].next = &w[2] + 1;
      w[2].sg_list = &s[2];
      w[2].num_sge = 1;
      w[2].opcode = IBV_WR_SEND;

      w[3].next = &w[3] + 1;
      w[3].sg_list = &s[3];
      w[3].num_sge = 1;
      w[3].opcode = IBV_WR_SEND;

      s += 4;
      w += 4;
      b += 4;
      n -= 4;
    }

  while (n >= 1)
    {
      s[0].addr = vlib_buffer_get_current_va (b[0]);
      s[0].length = b[0]->current_length;
      s[0].lkey = rd->lkey;

      w[0].next = &w[0] + 1;
      w[0].sg_list = &s[0];
      w[0].num_sge = 1;
      w[0].opcode = IBV_WR_SEND;

      s += 1;
      w += 1;
      b += 1;
      n -= 1;
    }

  w[-1].wr_id = txq->tail;	/* register item to free */
  w[-1].next = 0;		/* fix next pointer in WR linked-list */
  w[-1].send_flags = IBV_SEND_SIGNALED;	/* generate a CQE so we can free buffers */

  w = wr;
  if (PREDICT_FALSE (0 != ibv_post_send (txq->ibv_qp, w, &w)))
    {
      vlib_error_count (vm, node->node_index, RDMA_TX_ERROR_SUBMISSION,
			n_left_from - (w - wr));
      n_left_from = w - wr;
    }

  vlib_buffer_copy_indices (txq->bufs + (txq->tail & mask), bi, n_left_from);
  txq->tail += n_left_from;
  return n_left_from;
}

/*
 * common tx/free functions
 */

static_always_inline void
rdma_device_output_free (vlib_main_t * vm, const vlib_node_runtime_t * node,
			 rdma_txq_t * txq, int is_mlx5dv)
{
  if (is_mlx5dv)
    rdma_device_output_free_mlx5 (vm, node, txq);
  else
    rdma_device_output_free_ibverb (vm, node, txq);
}

static_always_inline u32
rdma_device_output_tx_try (vlib_main_t * vm, const vlib_node_runtime_t * node,
			   const rdma_device_t * rd, rdma_txq_t * txq,
			   u32 n_left_from, u32 * bi, int is_mlx5dv)
{
  vlib_buffer_t *b[VLIB_FRAME_SIZE];
  u32 mask = pow2_mask (txq->bufs_log2sz);

  /* do not enqueue more packet than ring space */
  n_left_from = clib_min (n_left_from, RDMA_TXQ_AVAIL_SZ (txq, txq->head,
							  txq->tail));
  /* avoid wrap-around logic in core loop */
  n_left_from = clib_min (n_left_from, RDMA_TXQ_BUF_SZ (txq) -
			  (txq->tail & mask));

  /* if ring is full, do nothing */
  if (PREDICT_FALSE (n_left_from == 0))
    return 0;

  vlib_get_buffers (vm, bi, b, n_left_from);

  return is_mlx5dv ?
    rdma_device_output_tx_mlx5 (vm, node, rd, txq, n_left_from, bi, b) :
    rdma_device_output_tx_ibverb (vm, node, rd, txq, n_left_from, bi, b);
}

static_always_inline uword
rdma_device_output_tx (vlib_main_t * vm, vlib_node_runtime_t * node,
		       vlib_frame_t * frame, rdma_device_t * rd,
		       int is_mlx5dv)
{
  u32 thread_index = vm->thread_index;
  rdma_txq_t *txq =
    vec_elt_at_index (rd->txqs, thread_index % vec_len (rd->txqs));
  u32 *from;
  u32 n_left_from;
  int i;

  ASSERT (RDMA_TXQ_BUF_SZ (txq) >= VLIB_FRAME_SIZE);

  from = vlib_frame_vector_args (frame);
  n_left_from = frame->n_vectors;

  clib_spinlock_lock_if_init (&txq->lock);

  for (i = 0; i < RDMA_TX_RETRIES && n_left_from > 0; i++)
    {
      u32 n_enq;
      rdma_device_output_free (vm, node, txq, is_mlx5dv);
      n_enq = rdma_device_output_tx_try (vm, node, rd, txq, n_left_from, from,
					 is_mlx5dv);

      n_left_from -= n_enq;
      from += n_enq;
    }

  clib_spinlock_unlock_if_init (&txq->lock);

  if (PREDICT_FALSE (n_left_from))
    {
      vlib_buffer_free (vm, from, n_left_from);
      vlib_error_count (vm, node->node_index,
			RDMA_TX_ERROR_NO_FREE_SLOTS, n_left_from);
    }

  return frame->n_vectors - n_left_from;
}

VNET_DEVICE_CLASS_TX_FN (rdma_device_class) (vlib_main_t * vm,
					     vlib_node_runtime_t * node,
					     vlib_frame_t * frame)
{
  rdma_main_t *rm = &rdma_main;
  vnet_interface_output_runtime_t *ord = (void *) node->runtime_data;
  rdma_device_t *rd = pool_elt_at_index (rm->devices, ord->dev_instance);

  if (PREDICT_TRUE (rd->flags & RDMA_DEVICE_F_MLX5DV))
    return rdma_device_output_tx (vm, node, frame, rd, 1 /* is_mlx5dv */ );

  return rdma_device_output_tx (vm, node, frame, rd, 0 /* is_mlx5dv */ );
}

/*
 * fd.io coding-style-patch-verification: ON
 *
 * Local Variables:
 * eval: (c-set-style "gnu")
 * End:
 */