xref: /dpdk/drivers/net/mlx5/mlx5_rxq.c (revision 2490bb897182f57de80fd924dd3ae48dda819b8c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <errno.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <fcntl.h>
11 #include <sys/queue.h>
12 
13 #include <rte_mbuf.h>
14 #include <rte_malloc.h>
15 #include <ethdev_driver.h>
16 #include <rte_common.h>
17 #include <rte_interrupts.h>
18 #include <rte_debug.h>
19 #include <rte_io.h>
20 #include <rte_eal_paging.h>
21 
22 #include <mlx5_glue.h>
23 #include <mlx5_malloc.h>
24 #include <mlx5_common.h>
25 #include <mlx5_common_mr.h>
26 
27 #include "mlx5_defs.h"
28 #include "mlx5.h"
29 #include "mlx5_rx.h"
30 #include "mlx5_utils.h"
31 #include "mlx5_autoconf.h"
32 #include "mlx5_devx.h"
33 
34 
35 /* Default RSS hash key also used for ConnectX-3. */
36 uint8_t rss_hash_default_key[] = {
37 	0x2c, 0xc6, 0x81, 0xd1,
38 	0x5b, 0xdb, 0xf4, 0xf7,
39 	0xfc, 0xa2, 0x83, 0x19,
40 	0xdb, 0x1a, 0x3e, 0x94,
41 	0x6b, 0x9e, 0x38, 0xd9,
42 	0x2c, 0x9c, 0x03, 0xd1,
43 	0xad, 0x99, 0x44, 0xa7,
44 	0xd9, 0x56, 0x3d, 0x59,
45 	0x06, 0x3c, 0x25, 0xf3,
46 	0xfc, 0x1f, 0xdc, 0x2a,
47 };
48 
49 /* Length of the default RSS hash key. */
50 static_assert(MLX5_RSS_HASH_KEY_LEN ==
51 	      (unsigned int)sizeof(rss_hash_default_key),
52 	      "wrong RSS default key size.");
53 
54 /**
55  * Calculate the number of CQEs in CQ for the Rx queue.
56  *
57  *  @param rxq_data
58  *     Pointer to receive queue structure.
59  *
60  * @return
61  *   Number of CQEs in CQ.
62  */
63 unsigned int
64 mlx5_rxq_cqe_num(struct mlx5_rxq_data *rxq_data)
65 {
66 	unsigned int cqe_n;
67 	unsigned int wqe_n = 1 << rxq_data->elts_n;
68 
69 	if (mlx5_rxq_mprq_enabled(rxq_data))
70 		cqe_n = wqe_n * (1 << rxq_data->strd_num_n) - 1;
71 	else
72 		cqe_n = wqe_n - 1;
73 	return cqe_n;
74 }
75 
76 /**
77  * Allocate RX queue elements for Multi-Packet RQ.
78  *
79  * @param rxq_ctrl
80  *   Pointer to RX queue structure.
81  *
82  * @return
83  *   0 on success, a negative errno value otherwise and rte_errno is set.
84  */
85 static int
86 rxq_alloc_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
87 {
88 	struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
89 	unsigned int wqe_n = 1 << rxq->elts_n;
90 	unsigned int i;
91 	int err;
92 
93 	/* Iterate on segments. */
94 	for (i = 0; i <= wqe_n; ++i) {
95 		struct mlx5_mprq_buf *buf;
96 
97 		if (rte_mempool_get(rxq->mprq_mp, (void **)&buf) < 0) {
98 			DRV_LOG(ERR, "port %u empty mbuf pool", rxq->port_id);
99 			rte_errno = ENOMEM;
100 			goto error;
101 		}
102 		if (i < wqe_n)
103 			(*rxq->mprq_bufs)[i] = buf;
104 		else
105 			rxq->mprq_repl = buf;
106 	}
107 	DRV_LOG(DEBUG,
108 		"port %u MPRQ queue %u allocated and configured %u segments",
109 		rxq->port_id, rxq->idx, wqe_n);
110 	return 0;
111 error:
112 	err = rte_errno; /* Save rte_errno before cleanup. */
113 	wqe_n = i;
114 	for (i = 0; (i != wqe_n); ++i) {
115 		if ((*rxq->mprq_bufs)[i] != NULL)
116 			rte_mempool_put(rxq->mprq_mp,
117 					(*rxq->mprq_bufs)[i]);
118 		(*rxq->mprq_bufs)[i] = NULL;
119 	}
120 	DRV_LOG(DEBUG, "port %u MPRQ queue %u failed, freed everything",
121 		rxq->port_id, rxq->idx);
122 	rte_errno = err; /* Restore rte_errno. */
123 	return -rte_errno;
124 }
125 
126 /**
127  * Allocate RX queue elements for Single-Packet RQ.
128  *
129  * @param rxq_ctrl
130  *   Pointer to RX queue structure.
131  *
132  * @return
133  *   0 on success, negative errno value on failure.
134  */
135 static int
136 rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
137 {
138 	const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
139 	unsigned int elts_n = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
140 		(1 << rxq_ctrl->rxq.elts_n) * (1 << rxq_ctrl->rxq.strd_num_n) :
141 		(1 << rxq_ctrl->rxq.elts_n);
142 	unsigned int i;
143 	int err;
144 
145 	/* Iterate on segments. */
146 	for (i = 0; (i != elts_n); ++i) {
147 		struct mlx5_eth_rxseg *seg = &rxq_ctrl->rxq.rxseg[i % sges_n];
148 		struct rte_mbuf *buf;
149 
150 		buf = rte_pktmbuf_alloc(seg->mp);
151 		if (buf == NULL) {
152 			if (rxq_ctrl->share_group == 0)
153 				DRV_LOG(ERR, "port %u queue %u empty mbuf pool",
154 					RXQ_PORT_ID(rxq_ctrl),
155 					rxq_ctrl->rxq.idx);
156 			else
157 				DRV_LOG(ERR, "share group %u queue %u empty mbuf pool",
158 					rxq_ctrl->share_group,
159 					rxq_ctrl->share_qid);
160 			rte_errno = ENOMEM;
161 			goto error;
162 		}
163 		/* Headroom is reserved by rte_pktmbuf_alloc(). */
164 		MLX5_ASSERT(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
165 		/* Buffer is supposed to be empty. */
166 		MLX5_ASSERT(rte_pktmbuf_data_len(buf) == 0);
167 		MLX5_ASSERT(rte_pktmbuf_pkt_len(buf) == 0);
168 		MLX5_ASSERT(!buf->next);
169 		SET_DATA_OFF(buf, seg->offset);
170 		PORT(buf) = rxq_ctrl->rxq.port_id;
171 		DATA_LEN(buf) = seg->length;
172 		PKT_LEN(buf) = seg->length;
173 		NB_SEGS(buf) = 1;
174 		(*rxq_ctrl->rxq.elts)[i] = buf;
175 	}
176 	/* If Rx vector is activated. */
177 	if (mlx5_rxq_check_vec_support(&rxq_ctrl->rxq) > 0) {
178 		struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
179 		struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
180 		struct rte_pktmbuf_pool_private *priv =
181 			(struct rte_pktmbuf_pool_private *)
182 				rte_mempool_get_priv(rxq_ctrl->rxq.mp);
183 		int j;
184 
185 		/* Initialize default rearm_data for vPMD. */
186 		mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
187 		rte_mbuf_refcnt_set(mbuf_init, 1);
188 		mbuf_init->nb_segs = 1;
189 		/* For shared queues port is provided in CQE */
190 		mbuf_init->port = rxq->shared ? 0 : rxq->port_id;
191 		if (priv->flags & RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF)
192 			mbuf_init->ol_flags = RTE_MBUF_F_EXTERNAL;
193 		/*
194 		 * prevent compiler reordering:
195 		 * rearm_data covers previous fields.
196 		 */
197 		rte_compiler_barrier();
198 		rxq->mbuf_initializer =
199 			*(rte_xmm_t *)&mbuf_init->rearm_data;
200 		/* Padding with a fake mbuf for vectorized Rx. */
201 		for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
202 			(*rxq->elts)[elts_n + j] = &rxq->fake_mbuf;
203 	}
204 	if (rxq_ctrl->share_group == 0)
205 		DRV_LOG(DEBUG,
206 			"port %u SPRQ queue %u allocated and configured %u segments (max %u packets)",
207 			RXQ_PORT_ID(rxq_ctrl), rxq_ctrl->rxq.idx, elts_n,
208 			elts_n / (1 << rxq_ctrl->rxq.sges_n));
209 	else
210 		DRV_LOG(DEBUG,
211 			"share group %u SPRQ queue %u allocated and configured %u segments (max %u packets)",
212 			rxq_ctrl->share_group, rxq_ctrl->share_qid, elts_n,
213 			elts_n / (1 << rxq_ctrl->rxq.sges_n));
214 	return 0;
215 error:
216 	err = rte_errno; /* Save rte_errno before cleanup. */
217 	elts_n = i;
218 	for (i = 0; (i != elts_n); ++i) {
219 		if ((*rxq_ctrl->rxq.elts)[i] != NULL)
220 			rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
221 		(*rxq_ctrl->rxq.elts)[i] = NULL;
222 	}
223 	if (rxq_ctrl->share_group == 0)
224 		DRV_LOG(DEBUG, "port %u SPRQ queue %u failed, freed everything",
225 			RXQ_PORT_ID(rxq_ctrl), rxq_ctrl->rxq.idx);
226 	else
227 		DRV_LOG(DEBUG, "share group %u SPRQ queue %u failed, freed everything",
228 			rxq_ctrl->share_group, rxq_ctrl->share_qid);
229 	rte_errno = err; /* Restore rte_errno. */
230 	return -rte_errno;
231 }
232 
233 /**
234  * Allocate RX queue elements.
235  *
236  * @param rxq_ctrl
237  *   Pointer to RX queue structure.
238  *
239  * @return
240  *   0 on success, negative errno value on failure.
241  */
242 int
243 rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
244 {
245 	int ret = 0;
246 
247 	/**
248 	 * For MPRQ we need to allocate both MPRQ buffers
249 	 * for WQEs and simple mbufs for vector processing.
250 	 */
251 	if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
252 		ret = rxq_alloc_elts_mprq(rxq_ctrl);
253 	if (ret == 0)
254 		ret = rxq_alloc_elts_sprq(rxq_ctrl);
255 	return ret;
256 }
257 
258 /**
259  * Free RX queue elements for Multi-Packet RQ.
260  *
261  * @param rxq_ctrl
262  *   Pointer to RX queue structure.
263  */
264 static void
265 rxq_free_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
266 {
267 	struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
268 	uint16_t i;
269 
270 	DRV_LOG(DEBUG, "port %u Multi-Packet Rx queue %u freeing %d WRs",
271 		rxq->port_id, rxq->idx, (1u << rxq->elts_n));
272 	if (rxq->mprq_bufs == NULL)
273 		return;
274 	for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
275 		if ((*rxq->mprq_bufs)[i] != NULL)
276 			mlx5_mprq_buf_free((*rxq->mprq_bufs)[i]);
277 		(*rxq->mprq_bufs)[i] = NULL;
278 	}
279 	if (rxq->mprq_repl != NULL) {
280 		mlx5_mprq_buf_free(rxq->mprq_repl);
281 		rxq->mprq_repl = NULL;
282 	}
283 }
284 
285 /**
286  * Free RX queue elements for Single-Packet RQ.
287  *
288  * @param rxq_ctrl
289  *   Pointer to RX queue structure.
290  */
291 static void
292 rxq_free_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
293 {
294 	struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
295 	const uint16_t q_n = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
296 		(1 << rxq->elts_n) * (1 << rxq->strd_num_n) :
297 		(1 << rxq->elts_n);
298 	const uint16_t q_mask = q_n - 1;
299 	uint16_t elts_ci = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
300 		rxq->elts_ci : rxq->rq_ci;
301 	uint16_t used = q_n - (elts_ci - rxq->rq_pi);
302 	uint16_t i;
303 
304 	if (rxq_ctrl->share_group == 0)
305 		DRV_LOG(DEBUG, "port %u Rx queue %u freeing %d WRs",
306 			RXQ_PORT_ID(rxq_ctrl), rxq->idx, q_n);
307 	else
308 		DRV_LOG(DEBUG, "share group %u Rx queue %u freeing %d WRs",
309 			rxq_ctrl->share_group, rxq_ctrl->share_qid, q_n);
310 	if (rxq->elts == NULL)
311 		return;
312 	/**
313 	 * Some mbuf in the Ring belongs to the application.
314 	 * They cannot be freed.
315 	 */
316 	if (mlx5_rxq_check_vec_support(rxq) > 0) {
317 		for (i = 0; i < used; ++i)
318 			(*rxq->elts)[(elts_ci + i) & q_mask] = NULL;
319 		rxq->rq_pi = elts_ci;
320 	}
321 	for (i = 0; i != q_n; ++i) {
322 		if ((*rxq->elts)[i] != NULL)
323 			rte_pktmbuf_free_seg((*rxq->elts)[i]);
324 		(*rxq->elts)[i] = NULL;
325 	}
326 }
327 
328 /**
329  * Free RX queue elements.
330  *
331  * @param rxq_ctrl
332  *   Pointer to RX queue structure.
333  */
334 static void
335 rxq_free_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
336 {
337 	/*
338 	 * For MPRQ we need to allocate both MPRQ buffers
339 	 * for WQEs and simple mbufs for vector processing.
340 	 */
341 	if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
342 		rxq_free_elts_mprq(rxq_ctrl);
343 	rxq_free_elts_sprq(rxq_ctrl);
344 }
345 
346 /**
347  * Returns the per-queue supported offloads.
348  *
349  * @param dev
350  *   Pointer to Ethernet device.
351  *
352  * @return
353  *   Supported Rx offloads.
354  */
355 uint64_t
356 mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
357 {
358 	struct mlx5_priv *priv = dev->data->dev_private;
359 	struct mlx5_dev_config *config = &priv->config;
360 	uint64_t offloads = (RTE_ETH_RX_OFFLOAD_SCATTER |
361 			     RTE_ETH_RX_OFFLOAD_TIMESTAMP |
362 			     RTE_ETH_RX_OFFLOAD_RSS_HASH);
363 
364 	if (!config->mprq.enabled)
365 		offloads |= RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT;
366 	if (config->hw_fcs_strip)
367 		offloads |= RTE_ETH_RX_OFFLOAD_KEEP_CRC;
368 	if (config->hw_csum)
369 		offloads |= (RTE_ETH_RX_OFFLOAD_IPV4_CKSUM |
370 			     RTE_ETH_RX_OFFLOAD_UDP_CKSUM |
371 			     RTE_ETH_RX_OFFLOAD_TCP_CKSUM);
372 	if (config->hw_vlan_strip)
373 		offloads |= RTE_ETH_RX_OFFLOAD_VLAN_STRIP;
374 	if (MLX5_LRO_SUPPORTED(dev))
375 		offloads |= RTE_ETH_RX_OFFLOAD_TCP_LRO;
376 	return offloads;
377 }
378 
379 
380 /**
381  * Returns the per-port supported offloads.
382  *
383  * @return
384  *   Supported Rx offloads.
385  */
386 uint64_t
387 mlx5_get_rx_port_offloads(void)
388 {
389 	uint64_t offloads = RTE_ETH_RX_OFFLOAD_VLAN_FILTER;
390 
391 	return offloads;
392 }
393 
394 /**
395  * Verify if the queue can be released.
396  *
397  * @param dev
398  *   Pointer to Ethernet device.
399  * @param idx
400  *   RX queue index.
401  *
402  * @return
403  *   1 if the queue can be released
404  *   0 if the queue can not be released, there are references to it.
405  *   Negative errno and rte_errno is set if queue doesn't exist.
406  */
407 static int
408 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
409 {
410 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
411 
412 	if (rxq == NULL) {
413 		rte_errno = EINVAL;
414 		return -rte_errno;
415 	}
416 	return (__atomic_load_n(&rxq->refcnt, __ATOMIC_RELAXED) == 1);
417 }
418 
419 /* Fetches and drops all SW-owned and error CQEs to synchronize CQ. */
420 static void
421 rxq_sync_cq(struct mlx5_rxq_data *rxq)
422 {
423 	const uint16_t cqe_n = 1 << rxq->cqe_n;
424 	const uint16_t cqe_mask = cqe_n - 1;
425 	volatile struct mlx5_cqe *cqe;
426 	int ret, i;
427 
428 	i = cqe_n;
429 	do {
430 		cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask];
431 		ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
432 		if (ret == MLX5_CQE_STATUS_HW_OWN)
433 			break;
434 		if (ret == MLX5_CQE_STATUS_ERR) {
435 			rxq->cq_ci++;
436 			continue;
437 		}
438 		MLX5_ASSERT(ret == MLX5_CQE_STATUS_SW_OWN);
439 		if (MLX5_CQE_FORMAT(cqe->op_own) != MLX5_COMPRESSED) {
440 			rxq->cq_ci++;
441 			continue;
442 		}
443 		/* Compute the next non compressed CQE. */
444 		rxq->cq_ci += rte_be_to_cpu_32(cqe->byte_cnt);
445 
446 	} while (--i);
447 	/* Move all CQEs to HW ownership, including possible MiniCQEs. */
448 	for (i = 0; i < cqe_n; i++) {
449 		cqe = &(*rxq->cqes)[i];
450 		cqe->op_own = MLX5_CQE_INVALIDATE;
451 	}
452 	/* Resync CQE and WQE (WQ in RESET state). */
453 	rte_io_wmb();
454 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
455 	rte_io_wmb();
456 	*rxq->rq_db = rte_cpu_to_be_32(0);
457 	rte_io_wmb();
458 }
459 
460 /**
461  * Rx queue stop. Device queue goes to the RESET state,
462  * all involved mbufs are freed from WQ.
463  *
464  * @param dev
465  *   Pointer to Ethernet device structure.
466  * @param idx
467  *   RX queue index.
468  *
469  * @return
470  *   0 on success, a negative errno value otherwise and rte_errno is set.
471  */
472 int
473 mlx5_rx_queue_stop_primary(struct rte_eth_dev *dev, uint16_t idx)
474 {
475 	struct mlx5_priv *priv = dev->data->dev_private;
476 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
477 	struct mlx5_rxq_ctrl *rxq_ctrl = rxq->ctrl;
478 	int ret;
479 
480 	MLX5_ASSERT(rxq != NULL && rxq_ctrl != NULL);
481 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
482 	ret = priv->obj_ops.rxq_obj_modify(rxq, MLX5_RXQ_MOD_RDY2RST);
483 	if (ret) {
484 		DRV_LOG(ERR, "Cannot change Rx WQ state to RESET:  %s",
485 			strerror(errno));
486 		rte_errno = errno;
487 		return ret;
488 	}
489 	/* Remove all processes CQEs. */
490 	rxq_sync_cq(&rxq_ctrl->rxq);
491 	/* Free all involved mbufs. */
492 	rxq_free_elts(rxq_ctrl);
493 	/* Set the actual queue state. */
494 	dev->data->rx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STOPPED;
495 	return 0;
496 }
497 
498 /**
499  * Rx queue stop. Device queue goes to the RESET state,
500  * all involved mbufs are freed from WQ.
501  *
502  * @param dev
503  *   Pointer to Ethernet device structure.
504  * @param idx
505  *   RX queue index.
506  *
507  * @return
508  *   0 on success, a negative errno value otherwise and rte_errno is set.
509  */
510 int
511 mlx5_rx_queue_stop(struct rte_eth_dev *dev, uint16_t idx)
512 {
513 	eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
514 	int ret;
515 
516 	if (rte_eth_dev_is_rx_hairpin_queue(dev, idx)) {
517 		DRV_LOG(ERR, "Hairpin queue can't be stopped");
518 		rte_errno = EINVAL;
519 		return -EINVAL;
520 	}
521 	if (dev->data->rx_queue_state[idx] == RTE_ETH_QUEUE_STATE_STOPPED)
522 		return 0;
523 	/*
524 	 * Vectorized Rx burst requires the CQ and RQ indices
525 	 * synchronized, that might be broken on RQ restart
526 	 * and cause Rx malfunction, so queue stopping is
527 	 * not supported if vectorized Rx burst is engaged.
528 	 * The routine pointer depends on the process
529 	 * type, should perform check there.
530 	 */
531 	if (pkt_burst == mlx5_rx_burst_vec) {
532 		DRV_LOG(ERR, "Rx queue stop is not supported "
533 			"for vectorized Rx");
534 		rte_errno = EINVAL;
535 		return -EINVAL;
536 	}
537 	if (rte_eal_process_type() ==  RTE_PROC_SECONDARY) {
538 		ret = mlx5_mp_os_req_queue_control(dev, idx,
539 						   MLX5_MP_REQ_QUEUE_RX_STOP);
540 	} else {
541 		ret = mlx5_rx_queue_stop_primary(dev, idx);
542 	}
543 	return ret;
544 }
545 
546 /**
547  * Rx queue start. Device queue goes to the ready state,
548  * all required mbufs are allocated and WQ is replenished.
549  *
550  * @param dev
551  *   Pointer to Ethernet device structure.
552  * @param idx
553  *   RX queue index.
554  *
555  * @return
556  *   0 on success, a negative errno value otherwise and rte_errno is set.
557  */
558 int
559 mlx5_rx_queue_start_primary(struct rte_eth_dev *dev, uint16_t idx)
560 {
561 	struct mlx5_priv *priv = dev->data->dev_private;
562 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
563 	struct mlx5_rxq_data *rxq_data = &rxq->ctrl->rxq;
564 	int ret;
565 
566 	MLX5_ASSERT(rxq != NULL && rxq->ctrl != NULL);
567 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
568 	/* Allocate needed buffers. */
569 	ret = rxq_alloc_elts(rxq->ctrl);
570 	if (ret) {
571 		DRV_LOG(ERR, "Cannot reallocate buffers for Rx WQ");
572 		rte_errno = errno;
573 		return ret;
574 	}
575 	rte_io_wmb();
576 	*rxq_data->cq_db = rte_cpu_to_be_32(rxq_data->cq_ci);
577 	rte_io_wmb();
578 	/* Reset RQ consumer before moving queue to READY state. */
579 	*rxq_data->rq_db = rte_cpu_to_be_32(0);
580 	rte_io_wmb();
581 	ret = priv->obj_ops.rxq_obj_modify(rxq, MLX5_RXQ_MOD_RST2RDY);
582 	if (ret) {
583 		DRV_LOG(ERR, "Cannot change Rx WQ state to READY:  %s",
584 			strerror(errno));
585 		rte_errno = errno;
586 		return ret;
587 	}
588 	/* Reinitialize RQ - set WQEs. */
589 	mlx5_rxq_initialize(rxq_data);
590 	rxq_data->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
591 	/* Set actual queue state. */
592 	dev->data->rx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STARTED;
593 	return 0;
594 }
595 
596 /**
597  * Rx queue start. Device queue goes to the ready state,
598  * all required mbufs are allocated and WQ is replenished.
599  *
600  * @param dev
601  *   Pointer to Ethernet device structure.
602  * @param idx
603  *   RX queue index.
604  *
605  * @return
606  *   0 on success, a negative errno value otherwise and rte_errno is set.
607  */
608 int
609 mlx5_rx_queue_start(struct rte_eth_dev *dev, uint16_t idx)
610 {
611 	int ret;
612 
613 	if (rte_eth_dev_is_rx_hairpin_queue(dev, idx)) {
614 		DRV_LOG(ERR, "Hairpin queue can't be started");
615 		rte_errno = EINVAL;
616 		return -EINVAL;
617 	}
618 	if (dev->data->rx_queue_state[idx] == RTE_ETH_QUEUE_STATE_STARTED)
619 		return 0;
620 	if (rte_eal_process_type() ==  RTE_PROC_SECONDARY) {
621 		ret = mlx5_mp_os_req_queue_control(dev, idx,
622 						   MLX5_MP_REQ_QUEUE_RX_START);
623 	} else {
624 		ret = mlx5_rx_queue_start_primary(dev, idx);
625 	}
626 	return ret;
627 }
628 
629 /**
630  * Rx queue presetup checks.
631  *
632  * @param dev
633  *   Pointer to Ethernet device structure.
634  * @param idx
635  *   RX queue index.
636  * @param desc
637  *   Number of descriptors to configure in queue.
638  * @param[out] rxq_ctrl
639  *   Address of pointer to shared Rx queue control.
640  *
641  * @return
642  *   0 on success, a negative errno value otherwise and rte_errno is set.
643  */
644 static int
645 mlx5_rx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t *desc,
646 			struct mlx5_rxq_ctrl **rxq_ctrl)
647 {
648 	struct mlx5_priv *priv = dev->data->dev_private;
649 	struct mlx5_rxq_priv *rxq;
650 	bool empty;
651 
652 	if (!rte_is_power_of_2(*desc)) {
653 		*desc = 1 << log2above(*desc);
654 		DRV_LOG(WARNING,
655 			"port %u increased number of descriptors in Rx queue %u"
656 			" to the next power of two (%d)",
657 			dev->data->port_id, idx, *desc);
658 	}
659 	DRV_LOG(DEBUG, "port %u configuring Rx queue %u for %u descriptors",
660 		dev->data->port_id, idx, *desc);
661 	if (idx >= priv->rxqs_n) {
662 		DRV_LOG(ERR, "port %u Rx queue index out of range (%u >= %u)",
663 			dev->data->port_id, idx, priv->rxqs_n);
664 		rte_errno = EOVERFLOW;
665 		return -rte_errno;
666 	}
667 	if (rxq_ctrl == NULL || *rxq_ctrl == NULL)
668 		return 0;
669 	if (!(*rxq_ctrl)->rxq.shared) {
670 		if (!mlx5_rxq_releasable(dev, idx)) {
671 			DRV_LOG(ERR, "port %u unable to release queue index %u",
672 				dev->data->port_id, idx);
673 			rte_errno = EBUSY;
674 			return -rte_errno;
675 		}
676 		mlx5_rxq_release(dev, idx);
677 	} else {
678 		if ((*rxq_ctrl)->obj != NULL)
679 			/* Some port using shared Rx queue has been started. */
680 			return 0;
681 		/* Release all owner RxQ to reconfigure Shared RxQ. */
682 		do {
683 			rxq = LIST_FIRST(&(*rxq_ctrl)->owners);
684 			LIST_REMOVE(rxq, owner_entry);
685 			empty = LIST_EMPTY(&(*rxq_ctrl)->owners);
686 			mlx5_rxq_release(ETH_DEV(rxq->priv), rxq->idx);
687 		} while (!empty);
688 		*rxq_ctrl = NULL;
689 	}
690 	return 0;
691 }
692 
693 /**
694  * Get the shared Rx queue object that matches group and queue index.
695  *
696  * @param dev
697  *   Pointer to Ethernet device structure.
698  * @param group
699  *   Shared RXQ group.
700  * @param share_qid
701  *   Shared RX queue index.
702  *
703  * @return
704  *   Shared RXQ object that matching, or NULL if not found.
705  */
706 static struct mlx5_rxq_ctrl *
707 mlx5_shared_rxq_get(struct rte_eth_dev *dev, uint32_t group, uint16_t share_qid)
708 {
709 	struct mlx5_rxq_ctrl *rxq_ctrl;
710 	struct mlx5_priv *priv = dev->data->dev_private;
711 
712 	LIST_FOREACH(rxq_ctrl, &priv->sh->shared_rxqs, share_entry) {
713 		if (rxq_ctrl->share_group == group &&
714 		    rxq_ctrl->share_qid == share_qid)
715 			return rxq_ctrl;
716 	}
717 	return NULL;
718 }
719 
720 /**
721  * Check whether requested Rx queue configuration matches shared RXQ.
722  *
723  * @param rxq_ctrl
724  *   Pointer to shared RXQ.
725  * @param dev
726  *   Pointer to Ethernet device structure.
727  * @param idx
728  *   Queue index.
729  * @param desc
730  *   Number of descriptors to configure in queue.
731  * @param socket
732  *   NUMA socket on which memory must be allocated.
733  * @param[in] conf
734  *   Thresholds parameters.
735  * @param mp
736  *   Memory pool for buffer allocations.
737  *
738  * @return
739  *   0 on success, a negative errno value otherwise and rte_errno is set.
740  */
741 static bool
742 mlx5_shared_rxq_match(struct mlx5_rxq_ctrl *rxq_ctrl, struct rte_eth_dev *dev,
743 		      uint16_t idx, uint16_t desc, unsigned int socket,
744 		      const struct rte_eth_rxconf *conf,
745 		      struct rte_mempool *mp)
746 {
747 	struct mlx5_priv *spriv = LIST_FIRST(&rxq_ctrl->owners)->priv;
748 	struct mlx5_priv *priv = dev->data->dev_private;
749 	unsigned int i;
750 
751 	RTE_SET_USED(conf);
752 	if (rxq_ctrl->socket != socket) {
753 		DRV_LOG(ERR, "port %u queue index %u failed to join shared group: socket mismatch",
754 			dev->data->port_id, idx);
755 		return false;
756 	}
757 	if (rxq_ctrl->rxq.elts_n != log2above(desc)) {
758 		DRV_LOG(ERR, "port %u queue index %u failed to join shared group: descriptor number mismatch",
759 			dev->data->port_id, idx);
760 		return false;
761 	}
762 	if (priv->mtu != spriv->mtu) {
763 		DRV_LOG(ERR, "port %u queue index %u failed to join shared group: mtu mismatch",
764 			dev->data->port_id, idx);
765 		return false;
766 	}
767 	if (priv->dev_data->dev_conf.intr_conf.rxq !=
768 	    spriv->dev_data->dev_conf.intr_conf.rxq) {
769 		DRV_LOG(ERR, "port %u queue index %u failed to join shared group: interrupt mismatch",
770 			dev->data->port_id, idx);
771 		return false;
772 	}
773 	if (mp != NULL && rxq_ctrl->rxq.mp != mp) {
774 		DRV_LOG(ERR, "port %u queue index %u failed to join shared group: mempool mismatch",
775 			dev->data->port_id, idx);
776 		return false;
777 	} else if (mp == NULL) {
778 		for (i = 0; i < conf->rx_nseg; i++) {
779 			if (conf->rx_seg[i].split.mp !=
780 			    rxq_ctrl->rxq.rxseg[i].mp ||
781 			    conf->rx_seg[i].split.length !=
782 			    rxq_ctrl->rxq.rxseg[i].length) {
783 				DRV_LOG(ERR, "port %u queue index %u failed to join shared group: segment %u configuration mismatch",
784 					dev->data->port_id, idx, i);
785 				return false;
786 			}
787 		}
788 	}
789 	if (priv->config.hw_padding != spriv->config.hw_padding) {
790 		DRV_LOG(ERR, "port %u queue index %u failed to join shared group: padding mismatch",
791 			dev->data->port_id, idx);
792 		return false;
793 	}
794 	if (priv->config.cqe_comp != spriv->config.cqe_comp ||
795 	    (priv->config.cqe_comp &&
796 	     priv->config.cqe_comp_fmt != spriv->config.cqe_comp_fmt)) {
797 		DRV_LOG(ERR, "port %u queue index %u failed to join shared group: CQE compression mismatch",
798 			dev->data->port_id, idx);
799 		return false;
800 	}
801 	return true;
802 }
803 
804 /**
805  *
806  * @param dev
807  *   Pointer to Ethernet device structure.
808  * @param idx
809  *   RX queue index.
810  * @param desc
811  *   Number of descriptors to configure in queue.
812  * @param socket
813  *   NUMA socket on which memory must be allocated.
814  * @param[in] conf
815  *   Thresholds parameters.
816  * @param mp
817  *   Memory pool for buffer allocations.
818  *
819  * @return
820  *   0 on success, a negative errno value otherwise and rte_errno is set.
821  */
822 int
823 mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
824 		    unsigned int socket, const struct rte_eth_rxconf *conf,
825 		    struct rte_mempool *mp)
826 {
827 	struct mlx5_priv *priv = dev->data->dev_private;
828 	struct mlx5_rxq_priv *rxq;
829 	struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
830 	struct rte_eth_rxseg_split *rx_seg =
831 				(struct rte_eth_rxseg_split *)conf->rx_seg;
832 	struct rte_eth_rxseg_split rx_single = {.mp = mp};
833 	uint16_t n_seg = conf->rx_nseg;
834 	int res;
835 	uint64_t offloads = conf->offloads |
836 			    dev->data->dev_conf.rxmode.offloads;
837 
838 	if (mp) {
839 		/*
840 		 * The parameters should be checked on rte_eth_dev layer.
841 		 * If mp is specified it means the compatible configuration
842 		 * without buffer split feature tuning.
843 		 */
844 		rx_seg = &rx_single;
845 		n_seg = 1;
846 	}
847 	if (n_seg > 1) {
848 		/* The offloads should be checked on rte_eth_dev layer. */
849 		MLX5_ASSERT(offloads & RTE_ETH_RX_OFFLOAD_SCATTER);
850 		if (!(offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT)) {
851 			DRV_LOG(ERR, "port %u queue index %u split "
852 				     "offload not configured",
853 				     dev->data->port_id, idx);
854 			rte_errno = ENOSPC;
855 			return -rte_errno;
856 		}
857 		MLX5_ASSERT(n_seg < MLX5_MAX_RXQ_NSEG);
858 	}
859 	if (conf->share_group > 0) {
860 		if (!priv->config.hca_attr.mem_rq_rmp) {
861 			DRV_LOG(ERR, "port %u queue index %u shared Rx queue not supported by fw",
862 				     dev->data->port_id, idx);
863 			rte_errno = EINVAL;
864 			return -rte_errno;
865 		}
866 		if (priv->obj_ops.rxq_obj_new != devx_obj_ops.rxq_obj_new) {
867 			DRV_LOG(ERR, "port %u queue index %u shared Rx queue needs DevX api",
868 				     dev->data->port_id, idx);
869 			rte_errno = EINVAL;
870 			return -rte_errno;
871 		}
872 		if (conf->share_qid >= priv->rxqs_n) {
873 			DRV_LOG(ERR, "port %u shared Rx queue index %u > number of Rx queues %u",
874 				dev->data->port_id, conf->share_qid,
875 				priv->rxqs_n);
876 			rte_errno = EINVAL;
877 			return -rte_errno;
878 		}
879 		if (priv->config.mprq.enabled) {
880 			DRV_LOG(ERR, "port %u shared Rx queue index %u: not supported when MPRQ enabled",
881 				dev->data->port_id, conf->share_qid);
882 			rte_errno = EINVAL;
883 			return -rte_errno;
884 		}
885 		/* Try to reuse shared RXQ. */
886 		rxq_ctrl = mlx5_shared_rxq_get(dev, conf->share_group,
887 					       conf->share_qid);
888 		if (rxq_ctrl != NULL &&
889 		    !mlx5_shared_rxq_match(rxq_ctrl, dev, idx, desc, socket,
890 					   conf, mp)) {
891 			rte_errno = EINVAL;
892 			return -rte_errno;
893 		}
894 	}
895 	res = mlx5_rx_queue_pre_setup(dev, idx, &desc, &rxq_ctrl);
896 	if (res)
897 		return res;
898 	/* Allocate RXQ. */
899 	rxq = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*rxq), 0,
900 			  SOCKET_ID_ANY);
901 	if (!rxq) {
902 		DRV_LOG(ERR, "port %u unable to allocate rx queue index %u private data",
903 			dev->data->port_id, idx);
904 		rte_errno = ENOMEM;
905 		return -rte_errno;
906 	}
907 	rxq->priv = priv;
908 	rxq->idx = idx;
909 	(*priv->rxq_privs)[idx] = rxq;
910 	if (rxq_ctrl != NULL) {
911 		/* Join owner list. */
912 		LIST_INSERT_HEAD(&rxq_ctrl->owners, rxq, owner_entry);
913 		rxq->ctrl = rxq_ctrl;
914 	} else {
915 		rxq_ctrl = mlx5_rxq_new(dev, rxq, desc, socket, conf, rx_seg,
916 					n_seg);
917 		if (rxq_ctrl == NULL) {
918 			DRV_LOG(ERR, "port %u unable to allocate rx queue index %u",
919 				dev->data->port_id, idx);
920 			mlx5_free(rxq);
921 			(*priv->rxq_privs)[idx] = NULL;
922 			rte_errno = ENOMEM;
923 			return -rte_errno;
924 		}
925 	}
926 	mlx5_rxq_ref(dev, idx);
927 	DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
928 		dev->data->port_id, idx);
929 	dev->data->rx_queues[idx] = &rxq_ctrl->rxq;
930 	return 0;
931 }
932 
933 /**
934  *
935  * @param dev
936  *   Pointer to Ethernet device structure.
937  * @param idx
938  *   RX queue index.
939  * @param desc
940  *   Number of descriptors to configure in queue.
941  * @param hairpin_conf
942  *   Hairpin configuration parameters.
943  *
944  * @return
945  *   0 on success, a negative errno value otherwise and rte_errno is set.
946  */
947 int
948 mlx5_rx_hairpin_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
949 			    uint16_t desc,
950 			    const struct rte_eth_hairpin_conf *hairpin_conf)
951 {
952 	struct mlx5_priv *priv = dev->data->dev_private;
953 	struct mlx5_rxq_priv *rxq;
954 	struct mlx5_rxq_ctrl *rxq_ctrl;
955 	int res;
956 
957 	res = mlx5_rx_queue_pre_setup(dev, idx, &desc, NULL);
958 	if (res)
959 		return res;
960 	if (hairpin_conf->peer_count != 1) {
961 		rte_errno = EINVAL;
962 		DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue index %u"
963 			" peer count is %u", dev->data->port_id,
964 			idx, hairpin_conf->peer_count);
965 		return -rte_errno;
966 	}
967 	if (hairpin_conf->peers[0].port == dev->data->port_id) {
968 		if (hairpin_conf->peers[0].queue >= priv->txqs_n) {
969 			rte_errno = EINVAL;
970 			DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue"
971 				" index %u, Tx %u is larger than %u",
972 				dev->data->port_id, idx,
973 				hairpin_conf->peers[0].queue, priv->txqs_n);
974 			return -rte_errno;
975 		}
976 	} else {
977 		if (hairpin_conf->manual_bind == 0 ||
978 		    hairpin_conf->tx_explicit == 0) {
979 			rte_errno = EINVAL;
980 			DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue"
981 				" index %u peer port %u with attributes %u %u",
982 				dev->data->port_id, idx,
983 				hairpin_conf->peers[0].port,
984 				hairpin_conf->manual_bind,
985 				hairpin_conf->tx_explicit);
986 			return -rte_errno;
987 		}
988 	}
989 	rxq = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*rxq), 0,
990 			  SOCKET_ID_ANY);
991 	if (!rxq) {
992 		DRV_LOG(ERR, "port %u unable to allocate hairpin rx queue index %u private data",
993 			dev->data->port_id, idx);
994 		rte_errno = ENOMEM;
995 		return -rte_errno;
996 	}
997 	rxq->priv = priv;
998 	rxq->idx = idx;
999 	(*priv->rxq_privs)[idx] = rxq;
1000 	rxq_ctrl = mlx5_rxq_hairpin_new(dev, rxq, desc, hairpin_conf);
1001 	if (!rxq_ctrl) {
1002 		DRV_LOG(ERR, "port %u unable to allocate hairpin queue index %u",
1003 			dev->data->port_id, idx);
1004 		mlx5_free(rxq);
1005 		(*priv->rxq_privs)[idx] = NULL;
1006 		rte_errno = ENOMEM;
1007 		return -rte_errno;
1008 	}
1009 	DRV_LOG(DEBUG, "port %u adding hairpin Rx queue %u to list",
1010 		dev->data->port_id, idx);
1011 	dev->data->rx_queues[idx] = &rxq_ctrl->rxq;
1012 	return 0;
1013 }
1014 
1015 /**
1016  * DPDK callback to release a RX queue.
1017  *
1018  * @param dev
1019  *   Pointer to Ethernet device structure.
1020  * @param qid
1021  *   Receive queue index.
1022  */
1023 void
1024 mlx5_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
1025 {
1026 	struct mlx5_rxq_data *rxq = dev->data->rx_queues[qid];
1027 
1028 	if (rxq == NULL)
1029 		return;
1030 	if (!mlx5_rxq_releasable(dev, qid))
1031 		rte_panic("port %u Rx queue %u is still used by a flow and"
1032 			  " cannot be removed\n", dev->data->port_id, qid);
1033 	mlx5_rxq_release(dev, qid);
1034 }
1035 
1036 /**
1037  * Allocate queue vector and fill epoll fd list for Rx interrupts.
1038  *
1039  * @param dev
1040  *   Pointer to Ethernet device.
1041  *
1042  * @return
1043  *   0 on success, a negative errno value otherwise and rte_errno is set.
1044  */
1045 int
1046 mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
1047 {
1048 	struct mlx5_priv *priv = dev->data->dev_private;
1049 	unsigned int i;
1050 	unsigned int rxqs_n = priv->rxqs_n;
1051 	unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1052 	unsigned int count = 0;
1053 	struct rte_intr_handle *intr_handle = dev->intr_handle;
1054 
1055 	if (!dev->data->dev_conf.intr_conf.rxq)
1056 		return 0;
1057 	mlx5_rx_intr_vec_disable(dev);
1058 	if (rte_intr_vec_list_alloc(intr_handle, NULL, n)) {
1059 		DRV_LOG(ERR,
1060 			"port %u failed to allocate memory for interrupt"
1061 			" vector, Rx interrupts will not be supported",
1062 			dev->data->port_id);
1063 		rte_errno = ENOMEM;
1064 		return -rte_errno;
1065 	}
1066 
1067 	if (rte_intr_type_set(intr_handle, RTE_INTR_HANDLE_EXT))
1068 		return -rte_errno;
1069 
1070 	for (i = 0; i != n; ++i) {
1071 		/* This rxq obj must not be released in this function. */
1072 		struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, i);
1073 		struct mlx5_rxq_obj *rxq_obj = rxq ? rxq->ctrl->obj : NULL;
1074 		int rc;
1075 
1076 		/* Skip queues that cannot request interrupts. */
1077 		if (!rxq_obj || (!rxq_obj->ibv_channel &&
1078 				 !rxq_obj->devx_channel)) {
1079 			/* Use invalid intr_vec[] index to disable entry. */
1080 			if (rte_intr_vec_list_index_set(intr_handle, i,
1081 			   RTE_INTR_VEC_RXTX_OFFSET + RTE_MAX_RXTX_INTR_VEC_ID))
1082 				return -rte_errno;
1083 			continue;
1084 		}
1085 		mlx5_rxq_ref(dev, i);
1086 		if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
1087 			DRV_LOG(ERR,
1088 				"port %u too many Rx queues for interrupt"
1089 				" vector size (%d), Rx interrupts cannot be"
1090 				" enabled",
1091 				dev->data->port_id, RTE_MAX_RXTX_INTR_VEC_ID);
1092 			mlx5_rx_intr_vec_disable(dev);
1093 			rte_errno = ENOMEM;
1094 			return -rte_errno;
1095 		}
1096 		rc = mlx5_os_set_nonblock_channel_fd(rxq_obj->fd);
1097 		if (rc < 0) {
1098 			rte_errno = errno;
1099 			DRV_LOG(ERR,
1100 				"port %u failed to make Rx interrupt file"
1101 				" descriptor %d non-blocking for queue index"
1102 				" %d",
1103 				dev->data->port_id, rxq_obj->fd, i);
1104 			mlx5_rx_intr_vec_disable(dev);
1105 			return -rte_errno;
1106 		}
1107 
1108 		if (rte_intr_vec_list_index_set(intr_handle, i,
1109 					RTE_INTR_VEC_RXTX_OFFSET + count))
1110 			return -rte_errno;
1111 		if (rte_intr_efds_index_set(intr_handle, count,
1112 						   rxq_obj->fd))
1113 			return -rte_errno;
1114 		count++;
1115 	}
1116 	if (!count)
1117 		mlx5_rx_intr_vec_disable(dev);
1118 	else if (rte_intr_nb_efd_set(intr_handle, count))
1119 		return -rte_errno;
1120 	return 0;
1121 }
1122 
1123 /**
1124  * Clean up Rx interrupts handler.
1125  *
1126  * @param dev
1127  *   Pointer to Ethernet device.
1128  */
1129 void
1130 mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
1131 {
1132 	struct mlx5_priv *priv = dev->data->dev_private;
1133 	struct rte_intr_handle *intr_handle = dev->intr_handle;
1134 	unsigned int i;
1135 	unsigned int rxqs_n = priv->rxqs_n;
1136 	unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1137 
1138 	if (!dev->data->dev_conf.intr_conf.rxq)
1139 		return;
1140 	if (rte_intr_vec_list_index_get(intr_handle, 0) < 0)
1141 		goto free;
1142 	for (i = 0; i != n; ++i) {
1143 		if (rte_intr_vec_list_index_get(intr_handle, i) ==
1144 		    RTE_INTR_VEC_RXTX_OFFSET + RTE_MAX_RXTX_INTR_VEC_ID)
1145 			continue;
1146 		/**
1147 		 * Need to access directly the queue to release the reference
1148 		 * kept in mlx5_rx_intr_vec_enable().
1149 		 */
1150 		mlx5_rxq_deref(dev, i);
1151 	}
1152 free:
1153 	rte_intr_free_epoll_fd(intr_handle);
1154 
1155 	rte_intr_vec_list_free(intr_handle);
1156 
1157 	rte_intr_nb_efd_set(intr_handle, 0);
1158 }
1159 
1160 /**
1161  *  MLX5 CQ notification .
1162  *
1163  *  @param rxq
1164  *     Pointer to receive queue structure.
1165  *  @param sq_n_rxq
1166  *     Sequence number per receive queue .
1167  */
1168 static inline void
1169 mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
1170 {
1171 	int sq_n = 0;
1172 	uint32_t doorbell_hi;
1173 	uint64_t doorbell;
1174 
1175 	sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
1176 	doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
1177 	doorbell = (uint64_t)doorbell_hi << 32;
1178 	doorbell |= rxq->cqn;
1179 	mlx5_doorbell_ring(&rxq->uar_data, rte_cpu_to_be_64(doorbell),
1180 			   doorbell_hi, &rxq->cq_db[MLX5_CQ_ARM_DB], 0);
1181 }
1182 
1183 /**
1184  * DPDK callback for Rx queue interrupt enable.
1185  *
1186  * @param dev
1187  *   Pointer to Ethernet device structure.
1188  * @param rx_queue_id
1189  *   Rx queue number.
1190  *
1191  * @return
1192  *   0 on success, a negative errno value otherwise and rte_errno is set.
1193  */
1194 int
1195 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1196 {
1197 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, rx_queue_id);
1198 	if (!rxq)
1199 		goto error;
1200 	if (rxq->ctrl->irq) {
1201 		if (!rxq->ctrl->obj)
1202 			goto error;
1203 		mlx5_arm_cq(&rxq->ctrl->rxq, rxq->ctrl->rxq.cq_arm_sn);
1204 	}
1205 	return 0;
1206 error:
1207 	rte_errno = EINVAL;
1208 	return -rte_errno;
1209 }
1210 
1211 /**
1212  * DPDK callback for Rx queue interrupt disable.
1213  *
1214  * @param dev
1215  *   Pointer to Ethernet device structure.
1216  * @param rx_queue_id
1217  *   Rx queue number.
1218  *
1219  * @return
1220  *   0 on success, a negative errno value otherwise and rte_errno is set.
1221  */
1222 int
1223 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1224 {
1225 	struct mlx5_priv *priv = dev->data->dev_private;
1226 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, rx_queue_id);
1227 	int ret = 0;
1228 
1229 	if (!rxq) {
1230 		rte_errno = EINVAL;
1231 		return -rte_errno;
1232 	}
1233 	if (!rxq->ctrl->obj)
1234 		goto error;
1235 	if (rxq->ctrl->irq) {
1236 		ret = priv->obj_ops.rxq_event_get(rxq->ctrl->obj);
1237 		if (ret < 0)
1238 			goto error;
1239 		rxq->ctrl->rxq.cq_arm_sn++;
1240 	}
1241 	return 0;
1242 error:
1243 	/**
1244 	 * The ret variable may be EAGAIN which means the get_event function was
1245 	 * called before receiving one.
1246 	 */
1247 	if (ret < 0)
1248 		rte_errno = errno;
1249 	else
1250 		rte_errno = EINVAL;
1251 	if (rte_errno != EAGAIN)
1252 		DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
1253 			dev->data->port_id, rx_queue_id);
1254 	return -rte_errno;
1255 }
1256 
1257 /**
1258  * Verify the Rx queue objects list is empty
1259  *
1260  * @param dev
1261  *   Pointer to Ethernet device.
1262  *
1263  * @return
1264  *   The number of objects not released.
1265  */
1266 int
1267 mlx5_rxq_obj_verify(struct rte_eth_dev *dev)
1268 {
1269 	struct mlx5_priv *priv = dev->data->dev_private;
1270 	int ret = 0;
1271 	struct mlx5_rxq_obj *rxq_obj;
1272 
1273 	LIST_FOREACH(rxq_obj, &priv->rxqsobj, next) {
1274 		if (rxq_obj->rxq_ctrl->rxq.shared &&
1275 		    !LIST_EMPTY(&rxq_obj->rxq_ctrl->owners))
1276 			continue;
1277 		DRV_LOG(DEBUG, "port %u Rx queue %u still referenced",
1278 			dev->data->port_id, rxq_obj->rxq_ctrl->rxq.idx);
1279 		++ret;
1280 	}
1281 	return ret;
1282 }
1283 
1284 /**
1285  * Callback function to initialize mbufs for Multi-Packet RQ.
1286  */
1287 static inline void
1288 mlx5_mprq_buf_init(struct rte_mempool *mp, void *opaque_arg,
1289 		    void *_m, unsigned int i __rte_unused)
1290 {
1291 	struct mlx5_mprq_buf *buf = _m;
1292 	struct rte_mbuf_ext_shared_info *shinfo;
1293 	unsigned int strd_n = (unsigned int)(uintptr_t)opaque_arg;
1294 	unsigned int j;
1295 
1296 	memset(_m, 0, sizeof(*buf));
1297 	buf->mp = mp;
1298 	__atomic_store_n(&buf->refcnt, 1, __ATOMIC_RELAXED);
1299 	for (j = 0; j != strd_n; ++j) {
1300 		shinfo = &buf->shinfos[j];
1301 		shinfo->free_cb = mlx5_mprq_buf_free_cb;
1302 		shinfo->fcb_opaque = buf;
1303 	}
1304 }
1305 
1306 /**
1307  * Free mempool of Multi-Packet RQ.
1308  *
1309  * @param dev
1310  *   Pointer to Ethernet device.
1311  *
1312  * @return
1313  *   0 on success, negative errno value on failure.
1314  */
1315 int
1316 mlx5_mprq_free_mp(struct rte_eth_dev *dev)
1317 {
1318 	struct mlx5_priv *priv = dev->data->dev_private;
1319 	struct rte_mempool *mp = priv->mprq_mp;
1320 	unsigned int i;
1321 
1322 	if (mp == NULL)
1323 		return 0;
1324 	DRV_LOG(DEBUG, "port %u freeing mempool (%s) for Multi-Packet RQ",
1325 		dev->data->port_id, mp->name);
1326 	/*
1327 	 * If a buffer in the pool has been externally attached to a mbuf and it
1328 	 * is still in use by application, destroying the Rx queue can spoil
1329 	 * the packet. It is unlikely to happen but if application dynamically
1330 	 * creates and destroys with holding Rx packets, this can happen.
1331 	 *
1332 	 * TODO: It is unavoidable for now because the mempool for Multi-Packet
1333 	 * RQ isn't provided by application but managed by PMD.
1334 	 */
1335 	if (!rte_mempool_full(mp)) {
1336 		DRV_LOG(ERR,
1337 			"port %u mempool for Multi-Packet RQ is still in use",
1338 			dev->data->port_id);
1339 		rte_errno = EBUSY;
1340 		return -rte_errno;
1341 	}
1342 	rte_mempool_free(mp);
1343 	/* Unset mempool for each Rx queue. */
1344 	for (i = 0; i != priv->rxqs_n; ++i) {
1345 		struct mlx5_rxq_data *rxq = mlx5_rxq_data_get(dev, i);
1346 
1347 		if (rxq == NULL)
1348 			continue;
1349 		rxq->mprq_mp = NULL;
1350 	}
1351 	priv->mprq_mp = NULL;
1352 	return 0;
1353 }
1354 
1355 /**
1356  * Allocate a mempool for Multi-Packet RQ. All configured Rx queues share the
1357  * mempool. If already allocated, reuse it if there're enough elements.
1358  * Otherwise, resize it.
1359  *
1360  * @param dev
1361  *   Pointer to Ethernet device.
1362  *
1363  * @return
1364  *   0 on success, negative errno value on failure.
1365  */
1366 int
1367 mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
1368 {
1369 	struct mlx5_priv *priv = dev->data->dev_private;
1370 	struct rte_mempool *mp = priv->mprq_mp;
1371 	char name[RTE_MEMPOOL_NAMESIZE];
1372 	unsigned int desc = 0;
1373 	unsigned int buf_len;
1374 	unsigned int obj_num;
1375 	unsigned int obj_size;
1376 	unsigned int strd_num_n = 0;
1377 	unsigned int strd_sz_n = 0;
1378 	unsigned int i;
1379 	unsigned int n_ibv = 0;
1380 	int ret;
1381 
1382 	if (!mlx5_mprq_enabled(dev))
1383 		return 0;
1384 	/* Count the total number of descriptors configured. */
1385 	for (i = 0; i != priv->rxqs_n; ++i) {
1386 		struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, i);
1387 		struct mlx5_rxq_data *rxq;
1388 
1389 		if (rxq_ctrl == NULL ||
1390 		    rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
1391 			continue;
1392 		rxq = &rxq_ctrl->rxq;
1393 		n_ibv++;
1394 		desc += 1 << rxq->elts_n;
1395 		/* Get the max number of strides. */
1396 		if (strd_num_n < rxq->strd_num_n)
1397 			strd_num_n = rxq->strd_num_n;
1398 		/* Get the max size of a stride. */
1399 		if (strd_sz_n < rxq->strd_sz_n)
1400 			strd_sz_n = rxq->strd_sz_n;
1401 	}
1402 	MLX5_ASSERT(strd_num_n && strd_sz_n);
1403 	buf_len = (1 << strd_num_n) * (1 << strd_sz_n);
1404 	obj_size = sizeof(struct mlx5_mprq_buf) + buf_len + (1 << strd_num_n) *
1405 		sizeof(struct rte_mbuf_ext_shared_info) + RTE_PKTMBUF_HEADROOM;
1406 	/*
1407 	 * Received packets can be either memcpy'd or externally referenced. In
1408 	 * case that the packet is attached to an mbuf as an external buffer, as
1409 	 * it isn't possible to predict how the buffers will be queued by
1410 	 * application, there's no option to exactly pre-allocate needed buffers
1411 	 * in advance but to speculatively prepares enough buffers.
1412 	 *
1413 	 * In the data path, if this Mempool is depleted, PMD will try to memcpy
1414 	 * received packets to buffers provided by application (rxq->mp) until
1415 	 * this Mempool gets available again.
1416 	 */
1417 	desc *= 4;
1418 	obj_num = desc + MLX5_MPRQ_MP_CACHE_SZ * n_ibv;
1419 	/*
1420 	 * rte_mempool_create_empty() has sanity check to refuse large cache
1421 	 * size compared to the number of elements.
1422 	 * CACHE_FLUSHTHRESH_MULTIPLIER is defined in a C file, so using a
1423 	 * constant number 2 instead.
1424 	 */
1425 	obj_num = RTE_MAX(obj_num, MLX5_MPRQ_MP_CACHE_SZ * 2);
1426 	/* Check a mempool is already allocated and if it can be resued. */
1427 	if (mp != NULL && mp->elt_size >= obj_size && mp->size >= obj_num) {
1428 		DRV_LOG(DEBUG, "port %u mempool %s is being reused",
1429 			dev->data->port_id, mp->name);
1430 		/* Reuse. */
1431 		goto exit;
1432 	} else if (mp != NULL) {
1433 		DRV_LOG(DEBUG, "port %u mempool %s should be resized, freeing it",
1434 			dev->data->port_id, mp->name);
1435 		/*
1436 		 * If failed to free, which means it may be still in use, no way
1437 		 * but to keep using the existing one. On buffer underrun,
1438 		 * packets will be memcpy'd instead of external buffer
1439 		 * attachment.
1440 		 */
1441 		if (mlx5_mprq_free_mp(dev)) {
1442 			if (mp->elt_size >= obj_size)
1443 				goto exit;
1444 			else
1445 				return -rte_errno;
1446 		}
1447 	}
1448 	snprintf(name, sizeof(name), "port-%u-mprq", dev->data->port_id);
1449 	mp = rte_mempool_create(name, obj_num, obj_size, MLX5_MPRQ_MP_CACHE_SZ,
1450 				0, NULL, NULL, mlx5_mprq_buf_init,
1451 				(void *)((uintptr_t)1 << strd_num_n),
1452 				dev->device->numa_node, 0);
1453 	if (mp == NULL) {
1454 		DRV_LOG(ERR,
1455 			"port %u failed to allocate a mempool for"
1456 			" Multi-Packet RQ, count=%u, size=%u",
1457 			dev->data->port_id, obj_num, obj_size);
1458 		rte_errno = ENOMEM;
1459 		return -rte_errno;
1460 	}
1461 	ret = mlx5_mr_mempool_register(priv->sh->cdev, mp);
1462 	if (ret < 0 && rte_errno != EEXIST) {
1463 		ret = rte_errno;
1464 		DRV_LOG(ERR, "port %u failed to register a mempool for Multi-Packet RQ",
1465 			dev->data->port_id);
1466 		rte_mempool_free(mp);
1467 		rte_errno = ret;
1468 		return -rte_errno;
1469 	}
1470 	priv->mprq_mp = mp;
1471 exit:
1472 	/* Set mempool for each Rx queue. */
1473 	for (i = 0; i != priv->rxqs_n; ++i) {
1474 		struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, i);
1475 
1476 		if (rxq_ctrl == NULL ||
1477 		    rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
1478 			continue;
1479 		rxq_ctrl->rxq.mprq_mp = mp;
1480 	}
1481 	DRV_LOG(INFO, "port %u Multi-Packet RQ is configured",
1482 		dev->data->port_id);
1483 	return 0;
1484 }
1485 
1486 #define MLX5_MAX_TCP_HDR_OFFSET ((unsigned int)(sizeof(struct rte_ether_hdr) + \
1487 					sizeof(struct rte_vlan_hdr) * 2 + \
1488 					sizeof(struct rte_ipv6_hdr)))
1489 #define MAX_TCP_OPTION_SIZE 40u
1490 #define MLX5_MAX_LRO_HEADER_FIX ((unsigned int)(MLX5_MAX_TCP_HDR_OFFSET + \
1491 				 sizeof(struct rte_tcp_hdr) + \
1492 				 MAX_TCP_OPTION_SIZE))
1493 
1494 /**
1495  * Adjust the maximum LRO massage size.
1496  *
1497  * @param dev
1498  *   Pointer to Ethernet device.
1499  * @param idx
1500  *   RX queue index.
1501  * @param max_lro_size
1502  *   The maximum size for LRO packet.
1503  */
1504 static void
1505 mlx5_max_lro_msg_size_adjust(struct rte_eth_dev *dev, uint16_t idx,
1506 			     uint32_t max_lro_size)
1507 {
1508 	struct mlx5_priv *priv = dev->data->dev_private;
1509 
1510 	if (priv->config.hca_attr.lro_max_msg_sz_mode ==
1511 	    MLX5_LRO_MAX_MSG_SIZE_START_FROM_L4 && max_lro_size >
1512 	    MLX5_MAX_TCP_HDR_OFFSET)
1513 		max_lro_size -= MLX5_MAX_TCP_HDR_OFFSET;
1514 	max_lro_size = RTE_MIN(max_lro_size, MLX5_MAX_LRO_SIZE);
1515 	MLX5_ASSERT(max_lro_size >= MLX5_LRO_SEG_CHUNK_SIZE);
1516 	max_lro_size /= MLX5_LRO_SEG_CHUNK_SIZE;
1517 	if (priv->max_lro_msg_size)
1518 		priv->max_lro_msg_size =
1519 			RTE_MIN((uint32_t)priv->max_lro_msg_size, max_lro_size);
1520 	else
1521 		priv->max_lro_msg_size = max_lro_size;
1522 	DRV_LOG(DEBUG,
1523 		"port %u Rx Queue %u max LRO message size adjusted to %u bytes",
1524 		dev->data->port_id, idx,
1525 		priv->max_lro_msg_size * MLX5_LRO_SEG_CHUNK_SIZE);
1526 }
1527 
1528 /**
1529  * Create a DPDK Rx queue.
1530  *
1531  * @param dev
1532  *   Pointer to Ethernet device.
1533  * @param rxq
1534  *   RX queue private data.
1535  * @param desc
1536  *   Number of descriptors to configure in queue.
1537  * @param socket
1538  *   NUMA socket on which memory must be allocated.
1539  *
1540  * @return
1541  *   A DPDK queue object on success, NULL otherwise and rte_errno is set.
1542  */
1543 struct mlx5_rxq_ctrl *
1544 mlx5_rxq_new(struct rte_eth_dev *dev, struct mlx5_rxq_priv *rxq,
1545 	     uint16_t desc,
1546 	     unsigned int socket, const struct rte_eth_rxconf *conf,
1547 	     const struct rte_eth_rxseg_split *rx_seg, uint16_t n_seg)
1548 {
1549 	uint16_t idx = rxq->idx;
1550 	struct mlx5_priv *priv = dev->data->dev_private;
1551 	struct mlx5_rxq_ctrl *tmpl;
1552 	unsigned int mb_len = rte_pktmbuf_data_room_size(rx_seg[0].mp);
1553 	struct mlx5_dev_config *config = &priv->config;
1554 	uint64_t offloads = conf->offloads |
1555 			   dev->data->dev_conf.rxmode.offloads;
1556 	unsigned int lro_on_queue = !!(offloads & RTE_ETH_RX_OFFLOAD_TCP_LRO);
1557 	unsigned int max_rx_pktlen = lro_on_queue ?
1558 			dev->data->dev_conf.rxmode.max_lro_pkt_size :
1559 			dev->data->mtu + (unsigned int)RTE_ETHER_HDR_LEN +
1560 				RTE_ETHER_CRC_LEN;
1561 	unsigned int non_scatter_min_mbuf_size = max_rx_pktlen +
1562 							RTE_PKTMBUF_HEADROOM;
1563 	unsigned int max_lro_size = 0;
1564 	unsigned int first_mb_free_size = mb_len - RTE_PKTMBUF_HEADROOM;
1565 	const int mprq_en = mlx5_check_mprq_support(dev) > 0 && n_seg == 1 &&
1566 			    !rx_seg[0].offset && !rx_seg[0].length;
1567 	unsigned int mprq_stride_nums = config->mprq.stride_num_n ?
1568 		config->mprq.stride_num_n : MLX5_MPRQ_STRIDE_NUM_N;
1569 	unsigned int mprq_stride_size = non_scatter_min_mbuf_size <=
1570 		(1U << config->mprq.max_stride_size_n) ?
1571 		log2above(non_scatter_min_mbuf_size) : MLX5_MPRQ_STRIDE_SIZE_N;
1572 	unsigned int mprq_stride_cap = (config->mprq.stride_num_n ?
1573 		(1U << config->mprq.stride_num_n) : (1U << mprq_stride_nums)) *
1574 		(config->mprq.stride_size_n ?
1575 		(1U << config->mprq.stride_size_n) : (1U << mprq_stride_size));
1576 	/*
1577 	 * Always allocate extra slots, even if eventually
1578 	 * the vector Rx will not be used.
1579 	 */
1580 	uint16_t desc_n = desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
1581 	const struct rte_eth_rxseg_split *qs_seg = rx_seg;
1582 	unsigned int tail_len;
1583 
1584 	tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
1585 		sizeof(*tmpl) + desc_n * sizeof(struct rte_mbuf *) +
1586 		(!!mprq_en) *
1587 		(desc >> mprq_stride_nums) * sizeof(struct mlx5_mprq_buf *),
1588 		0, socket);
1589 	if (!tmpl) {
1590 		rte_errno = ENOMEM;
1591 		return NULL;
1592 	}
1593 	LIST_INIT(&tmpl->owners);
1594 	if (conf->share_group > 0) {
1595 		tmpl->rxq.shared = 1;
1596 		tmpl->share_group = conf->share_group;
1597 		tmpl->share_qid = conf->share_qid;
1598 		LIST_INSERT_HEAD(&priv->sh->shared_rxqs, tmpl, share_entry);
1599 	}
1600 	rxq->ctrl = tmpl;
1601 	LIST_INSERT_HEAD(&tmpl->owners, rxq, owner_entry);
1602 	MLX5_ASSERT(n_seg && n_seg <= MLX5_MAX_RXQ_NSEG);
1603 	/*
1604 	 * Build the array of actual buffer offsets and lengths.
1605 	 * Pad with the buffers from the last memory pool if
1606 	 * needed to handle max size packets, replace zero length
1607 	 * with the buffer length from the pool.
1608 	 */
1609 	tail_len = max_rx_pktlen;
1610 	do {
1611 		struct mlx5_eth_rxseg *hw_seg =
1612 					&tmpl->rxq.rxseg[tmpl->rxq.rxseg_n];
1613 		uint32_t buf_len, offset, seg_len;
1614 
1615 		/*
1616 		 * For the buffers beyond descriptions offset is zero,
1617 		 * the first buffer contains head room.
1618 		 */
1619 		buf_len = rte_pktmbuf_data_room_size(qs_seg->mp);
1620 		offset = (tmpl->rxq.rxseg_n >= n_seg ? 0 : qs_seg->offset) +
1621 			 (tmpl->rxq.rxseg_n ? 0 : RTE_PKTMBUF_HEADROOM);
1622 		/*
1623 		 * For the buffers beyond descriptions the length is
1624 		 * pool buffer length, zero lengths are replaced with
1625 		 * pool buffer length either.
1626 		 */
1627 		seg_len = tmpl->rxq.rxseg_n >= n_seg ? buf_len :
1628 						       qs_seg->length ?
1629 						       qs_seg->length :
1630 						       (buf_len - offset);
1631 		/* Check is done in long int, now overflows. */
1632 		if (buf_len < seg_len + offset) {
1633 			DRV_LOG(ERR, "port %u Rx queue %u: Split offset/length "
1634 				     "%u/%u can't be satisfied",
1635 				     dev->data->port_id, idx,
1636 				     qs_seg->length, qs_seg->offset);
1637 			rte_errno = EINVAL;
1638 			goto error;
1639 		}
1640 		if (seg_len > tail_len)
1641 			seg_len = buf_len - offset;
1642 		if (++tmpl->rxq.rxseg_n > MLX5_MAX_RXQ_NSEG) {
1643 			DRV_LOG(ERR,
1644 				"port %u too many SGEs (%u) needed to handle"
1645 				" requested maximum packet size %u, the maximum"
1646 				" supported are %u", dev->data->port_id,
1647 				tmpl->rxq.rxseg_n, max_rx_pktlen,
1648 				MLX5_MAX_RXQ_NSEG);
1649 			rte_errno = ENOTSUP;
1650 			goto error;
1651 		}
1652 		/* Build the actual scattering element in the queue object. */
1653 		hw_seg->mp = qs_seg->mp;
1654 		MLX5_ASSERT(offset <= UINT16_MAX);
1655 		MLX5_ASSERT(seg_len <= UINT16_MAX);
1656 		hw_seg->offset = (uint16_t)offset;
1657 		hw_seg->length = (uint16_t)seg_len;
1658 		/*
1659 		 * Advance the segment descriptor, the padding is the based
1660 		 * on the attributes of the last descriptor.
1661 		 */
1662 		if (tmpl->rxq.rxseg_n < n_seg)
1663 			qs_seg++;
1664 		tail_len -= RTE_MIN(tail_len, seg_len);
1665 	} while (tail_len || !rte_is_power_of_2(tmpl->rxq.rxseg_n));
1666 	MLX5_ASSERT(tmpl->rxq.rxseg_n &&
1667 		    tmpl->rxq.rxseg_n <= MLX5_MAX_RXQ_NSEG);
1668 	if (tmpl->rxq.rxseg_n > 1 && !(offloads & RTE_ETH_RX_OFFLOAD_SCATTER)) {
1669 		DRV_LOG(ERR, "port %u Rx queue %u: Scatter offload is not"
1670 			" configured and no enough mbuf space(%u) to contain "
1671 			"the maximum RX packet length(%u) with head-room(%u)",
1672 			dev->data->port_id, idx, mb_len, max_rx_pktlen,
1673 			RTE_PKTMBUF_HEADROOM);
1674 		rte_errno = ENOSPC;
1675 		goto error;
1676 	}
1677 	tmpl->type = MLX5_RXQ_TYPE_STANDARD;
1678 	if (mlx5_mr_ctrl_init(&tmpl->rxq.mr_ctrl,
1679 			      &priv->sh->cdev->mr_scache.dev_gen, socket)) {
1680 		/* rte_errno is already set. */
1681 		goto error;
1682 	}
1683 	tmpl->socket = socket;
1684 	if (dev->data->dev_conf.intr_conf.rxq)
1685 		tmpl->irq = 1;
1686 	/*
1687 	 * This Rx queue can be configured as a Multi-Packet RQ if all of the
1688 	 * following conditions are met:
1689 	 *  - MPRQ is enabled.
1690 	 *  - The number of descs is more than the number of strides.
1691 	 *  - max_rx_pktlen plus overhead is less than the max size
1692 	 *    of a stride or mprq_stride_size is specified by a user.
1693 	 *    Need to make sure that there are enough strides to encap
1694 	 *    the maximum packet size in case mprq_stride_size is set.
1695 	 *  Otherwise, enable Rx scatter if necessary.
1696 	 */
1697 	if (mprq_en && desc > (1U << mprq_stride_nums) &&
1698 	    (non_scatter_min_mbuf_size <=
1699 	     (1U << config->mprq.max_stride_size_n) ||
1700 	     (config->mprq.stride_size_n &&
1701 	      non_scatter_min_mbuf_size <= mprq_stride_cap))) {
1702 		/* TODO: Rx scatter isn't supported yet. */
1703 		tmpl->rxq.sges_n = 0;
1704 		/* Trim the number of descs needed. */
1705 		desc >>= mprq_stride_nums;
1706 		tmpl->rxq.strd_num_n = config->mprq.stride_num_n ?
1707 			config->mprq.stride_num_n : mprq_stride_nums;
1708 		tmpl->rxq.strd_sz_n = config->mprq.stride_size_n ?
1709 			config->mprq.stride_size_n : mprq_stride_size;
1710 		tmpl->rxq.strd_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT;
1711 		tmpl->rxq.strd_scatter_en =
1712 				!!(offloads & RTE_ETH_RX_OFFLOAD_SCATTER);
1713 		tmpl->rxq.mprq_max_memcpy_len = RTE_MIN(first_mb_free_size,
1714 				config->mprq.max_memcpy_len);
1715 		max_lro_size = RTE_MIN(max_rx_pktlen,
1716 				       (1u << tmpl->rxq.strd_num_n) *
1717 				       (1u << tmpl->rxq.strd_sz_n));
1718 		DRV_LOG(DEBUG,
1719 			"port %u Rx queue %u: Multi-Packet RQ is enabled"
1720 			" strd_num_n = %u, strd_sz_n = %u",
1721 			dev->data->port_id, idx,
1722 			tmpl->rxq.strd_num_n, tmpl->rxq.strd_sz_n);
1723 	} else if (tmpl->rxq.rxseg_n == 1) {
1724 		MLX5_ASSERT(max_rx_pktlen <= first_mb_free_size);
1725 		tmpl->rxq.sges_n = 0;
1726 		max_lro_size = max_rx_pktlen;
1727 	} else if (offloads & RTE_ETH_RX_OFFLOAD_SCATTER) {
1728 		unsigned int sges_n;
1729 
1730 		if (lro_on_queue && first_mb_free_size <
1731 		    MLX5_MAX_LRO_HEADER_FIX) {
1732 			DRV_LOG(ERR, "Not enough space in the first segment(%u)"
1733 				" to include the max header size(%u) for LRO",
1734 				first_mb_free_size, MLX5_MAX_LRO_HEADER_FIX);
1735 			rte_errno = ENOTSUP;
1736 			goto error;
1737 		}
1738 		/*
1739 		 * Determine the number of SGEs needed for a full packet
1740 		 * and round it to the next power of two.
1741 		 */
1742 		sges_n = log2above(tmpl->rxq.rxseg_n);
1743 		if (sges_n > MLX5_MAX_LOG_RQ_SEGS) {
1744 			DRV_LOG(ERR,
1745 				"port %u too many SGEs (%u) needed to handle"
1746 				" requested maximum packet size %u, the maximum"
1747 				" supported are %u", dev->data->port_id,
1748 				1 << sges_n, max_rx_pktlen,
1749 				1u << MLX5_MAX_LOG_RQ_SEGS);
1750 			rte_errno = ENOTSUP;
1751 			goto error;
1752 		}
1753 		tmpl->rxq.sges_n = sges_n;
1754 		max_lro_size = max_rx_pktlen;
1755 	}
1756 	if (config->mprq.enabled && !mlx5_rxq_mprq_enabled(&tmpl->rxq))
1757 		DRV_LOG(WARNING,
1758 			"port %u MPRQ is requested but cannot be enabled\n"
1759 			" (requested: pkt_sz = %u, desc_num = %u,"
1760 			" rxq_num = %u, stride_sz = %u, stride_num = %u\n"
1761 			"  supported: min_rxqs_num = %u,"
1762 			" min_stride_sz = %u, max_stride_sz = %u).",
1763 			dev->data->port_id, non_scatter_min_mbuf_size,
1764 			desc, priv->rxqs_n,
1765 			config->mprq.stride_size_n ?
1766 				(1U << config->mprq.stride_size_n) :
1767 				(1U << mprq_stride_size),
1768 			config->mprq.stride_num_n ?
1769 				(1U << config->mprq.stride_num_n) :
1770 				(1U << mprq_stride_nums),
1771 			config->mprq.min_rxqs_num,
1772 			(1U << config->mprq.min_stride_size_n),
1773 			(1U << config->mprq.max_stride_size_n));
1774 	DRV_LOG(DEBUG, "port %u maximum number of segments per packet: %u",
1775 		dev->data->port_id, 1 << tmpl->rxq.sges_n);
1776 	if (desc % (1 << tmpl->rxq.sges_n)) {
1777 		DRV_LOG(ERR,
1778 			"port %u number of Rx queue descriptors (%u) is not a"
1779 			" multiple of SGEs per packet (%u)",
1780 			dev->data->port_id,
1781 			desc,
1782 			1 << tmpl->rxq.sges_n);
1783 		rte_errno = EINVAL;
1784 		goto error;
1785 	}
1786 	mlx5_max_lro_msg_size_adjust(dev, idx, max_lro_size);
1787 	/* Toggle RX checksum offload if hardware supports it. */
1788 	tmpl->rxq.csum = !!(offloads & RTE_ETH_RX_OFFLOAD_CHECKSUM);
1789 	/* Configure Rx timestamp. */
1790 	tmpl->rxq.hw_timestamp = !!(offloads & RTE_ETH_RX_OFFLOAD_TIMESTAMP);
1791 	tmpl->rxq.timestamp_rx_flag = 0;
1792 	if (tmpl->rxq.hw_timestamp && rte_mbuf_dyn_rx_timestamp_register(
1793 			&tmpl->rxq.timestamp_offset,
1794 			&tmpl->rxq.timestamp_rx_flag) != 0) {
1795 		DRV_LOG(ERR, "Cannot register Rx timestamp field/flag");
1796 		goto error;
1797 	}
1798 	/* Configure VLAN stripping. */
1799 	tmpl->rxq.vlan_strip = !!(offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP);
1800 	/* By default, FCS (CRC) is stripped by hardware. */
1801 	tmpl->rxq.crc_present = 0;
1802 	tmpl->rxq.lro = lro_on_queue;
1803 	if (offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC) {
1804 		if (config->hw_fcs_strip) {
1805 			/*
1806 			 * RQs used for LRO-enabled TIRs should not be
1807 			 * configured to scatter the FCS.
1808 			 */
1809 			if (lro_on_queue)
1810 				DRV_LOG(WARNING,
1811 					"port %u CRC stripping has been "
1812 					"disabled but will still be performed "
1813 					"by hardware, because LRO is enabled",
1814 					dev->data->port_id);
1815 			else
1816 				tmpl->rxq.crc_present = 1;
1817 		} else {
1818 			DRV_LOG(WARNING,
1819 				"port %u CRC stripping has been disabled but will"
1820 				" still be performed by hardware, make sure MLNX_OFED"
1821 				" and firmware are up to date",
1822 				dev->data->port_id);
1823 		}
1824 	}
1825 	DRV_LOG(DEBUG,
1826 		"port %u CRC stripping is %s, %u bytes will be subtracted from"
1827 		" incoming frames to hide it",
1828 		dev->data->port_id,
1829 		tmpl->rxq.crc_present ? "disabled" : "enabled",
1830 		tmpl->rxq.crc_present << 2);
1831 	/* Save port ID. */
1832 	tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
1833 		(!!(dev->data->dev_conf.rxmode.mq_mode & RTE_ETH_MQ_RX_RSS));
1834 	tmpl->rxq.port_id = dev->data->port_id;
1835 	tmpl->sh = priv->sh;
1836 	tmpl->rxq.mp = rx_seg[0].mp;
1837 	tmpl->rxq.elts_n = log2above(desc);
1838 	tmpl->rxq.rq_repl_thresh =
1839 		MLX5_VPMD_RXQ_RPLNSH_THRESH(desc_n);
1840 	tmpl->rxq.elts =
1841 		(struct rte_mbuf *(*)[desc_n])(tmpl + 1);
1842 	tmpl->rxq.mprq_bufs =
1843 		(struct mlx5_mprq_buf *(*)[desc])(*tmpl->rxq.elts + desc_n);
1844 	tmpl->rxq.idx = idx;
1845 	LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1846 	return tmpl;
1847 error:
1848 	mlx5_mr_btree_free(&tmpl->rxq.mr_ctrl.cache_bh);
1849 	mlx5_free(tmpl);
1850 	return NULL;
1851 }
1852 
1853 /**
1854  * Create a DPDK Rx hairpin queue.
1855  *
1856  * @param dev
1857  *   Pointer to Ethernet device.
1858  * @param rxq
1859  *   RX queue.
1860  * @param desc
1861  *   Number of descriptors to configure in queue.
1862  * @param hairpin_conf
1863  *   The hairpin binding configuration.
1864  *
1865  * @return
1866  *   A DPDK queue object on success, NULL otherwise and rte_errno is set.
1867  */
1868 struct mlx5_rxq_ctrl *
1869 mlx5_rxq_hairpin_new(struct rte_eth_dev *dev, struct mlx5_rxq_priv *rxq,
1870 		     uint16_t desc,
1871 		     const struct rte_eth_hairpin_conf *hairpin_conf)
1872 {
1873 	uint16_t idx = rxq->idx;
1874 	struct mlx5_priv *priv = dev->data->dev_private;
1875 	struct mlx5_rxq_ctrl *tmpl;
1876 
1877 	tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*tmpl), 0,
1878 			   SOCKET_ID_ANY);
1879 	if (!tmpl) {
1880 		rte_errno = ENOMEM;
1881 		return NULL;
1882 	}
1883 	LIST_INIT(&tmpl->owners);
1884 	rxq->ctrl = tmpl;
1885 	LIST_INSERT_HEAD(&tmpl->owners, rxq, owner_entry);
1886 	tmpl->type = MLX5_RXQ_TYPE_HAIRPIN;
1887 	tmpl->socket = SOCKET_ID_ANY;
1888 	tmpl->rxq.rss_hash = 0;
1889 	tmpl->rxq.port_id = dev->data->port_id;
1890 	tmpl->sh = priv->sh;
1891 	tmpl->rxq.mp = NULL;
1892 	tmpl->rxq.elts_n = log2above(desc);
1893 	tmpl->rxq.elts = NULL;
1894 	tmpl->rxq.mr_ctrl.cache_bh = (struct mlx5_mr_btree) { 0 };
1895 	tmpl->rxq.idx = idx;
1896 	rxq->hairpin_conf = *hairpin_conf;
1897 	mlx5_rxq_ref(dev, idx);
1898 	LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1899 	return tmpl;
1900 }
1901 
1902 /**
1903  * Increase Rx queue reference count.
1904  *
1905  * @param dev
1906  *   Pointer to Ethernet device.
1907  * @param idx
1908  *   RX queue index.
1909  *
1910  * @return
1911  *   A pointer to the queue if it exists, NULL otherwise.
1912  */
1913 struct mlx5_rxq_priv *
1914 mlx5_rxq_ref(struct rte_eth_dev *dev, uint16_t idx)
1915 {
1916 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
1917 
1918 	if (rxq != NULL)
1919 		__atomic_fetch_add(&rxq->refcnt, 1, __ATOMIC_RELAXED);
1920 	return rxq;
1921 }
1922 
1923 /**
1924  * Dereference a Rx queue.
1925  *
1926  * @param dev
1927  *   Pointer to Ethernet device.
1928  * @param idx
1929  *   RX queue index.
1930  *
1931  * @return
1932  *   Updated reference count.
1933  */
1934 uint32_t
1935 mlx5_rxq_deref(struct rte_eth_dev *dev, uint16_t idx)
1936 {
1937 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
1938 
1939 	if (rxq == NULL)
1940 		return 0;
1941 	return __atomic_sub_fetch(&rxq->refcnt, 1, __ATOMIC_RELAXED);
1942 }
1943 
1944 /**
1945  * Get a Rx queue.
1946  *
1947  * @param dev
1948  *   Pointer to Ethernet device.
1949  * @param idx
1950  *   RX queue index.
1951  *
1952  * @return
1953  *   A pointer to the queue if it exists, NULL otherwise.
1954  */
1955 struct mlx5_rxq_priv *
1956 mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
1957 {
1958 	struct mlx5_priv *priv = dev->data->dev_private;
1959 
1960 	MLX5_ASSERT(priv->rxq_privs != NULL);
1961 	return (*priv->rxq_privs)[idx];
1962 }
1963 
1964 /**
1965  * Get Rx queue shareable control.
1966  *
1967  * @param dev
1968  *   Pointer to Ethernet device.
1969  * @param idx
1970  *   RX queue index.
1971  *
1972  * @return
1973  *   A pointer to the queue control if it exists, NULL otherwise.
1974  */
1975 struct mlx5_rxq_ctrl *
1976 mlx5_rxq_ctrl_get(struct rte_eth_dev *dev, uint16_t idx)
1977 {
1978 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
1979 
1980 	return rxq == NULL ? NULL : rxq->ctrl;
1981 }
1982 
1983 /**
1984  * Get Rx queue shareable data.
1985  *
1986  * @param dev
1987  *   Pointer to Ethernet device.
1988  * @param idx
1989  *   RX queue index.
1990  *
1991  * @return
1992  *   A pointer to the queue data if it exists, NULL otherwise.
1993  */
1994 struct mlx5_rxq_data *
1995 mlx5_rxq_data_get(struct rte_eth_dev *dev, uint16_t idx)
1996 {
1997 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
1998 
1999 	return rxq == NULL ? NULL : &rxq->ctrl->rxq;
2000 }
2001 
2002 /**
2003  * Release a Rx queue.
2004  *
2005  * @param dev
2006  *   Pointer to Ethernet device.
2007  * @param idx
2008  *   RX queue index.
2009  *
2010  * @return
2011  *   1 while a reference on it exists, 0 when freed.
2012  */
2013 int
2014 mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
2015 {
2016 	struct mlx5_priv *priv = dev->data->dev_private;
2017 	struct mlx5_rxq_priv *rxq;
2018 	struct mlx5_rxq_ctrl *rxq_ctrl;
2019 	uint32_t refcnt;
2020 
2021 	if (priv->rxq_privs == NULL)
2022 		return 0;
2023 	rxq = mlx5_rxq_get(dev, idx);
2024 	if (rxq == NULL || rxq->refcnt == 0)
2025 		return 0;
2026 	rxq_ctrl = rxq->ctrl;
2027 	refcnt = mlx5_rxq_deref(dev, idx);
2028 	if (refcnt > 1) {
2029 		return 1;
2030 	} else if (refcnt == 1) { /* RxQ stopped. */
2031 		priv->obj_ops.rxq_obj_release(rxq);
2032 		if (!rxq_ctrl->started && rxq_ctrl->obj != NULL) {
2033 			LIST_REMOVE(rxq_ctrl->obj, next);
2034 			mlx5_free(rxq_ctrl->obj);
2035 			rxq_ctrl->obj = NULL;
2036 		}
2037 		if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD) {
2038 			if (!rxq_ctrl->started)
2039 				rxq_free_elts(rxq_ctrl);
2040 			dev->data->rx_queue_state[idx] =
2041 					RTE_ETH_QUEUE_STATE_STOPPED;
2042 		}
2043 	} else { /* Refcnt zero, closing device. */
2044 		LIST_REMOVE(rxq, owner_entry);
2045 		if (LIST_EMPTY(&rxq_ctrl->owners)) {
2046 			if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
2047 				mlx5_mr_btree_free
2048 					(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
2049 			if (rxq_ctrl->rxq.shared)
2050 				LIST_REMOVE(rxq_ctrl, share_entry);
2051 			LIST_REMOVE(rxq_ctrl, next);
2052 			mlx5_free(rxq_ctrl);
2053 		}
2054 		dev->data->rx_queues[idx] = NULL;
2055 		mlx5_free(rxq);
2056 		(*priv->rxq_privs)[idx] = NULL;
2057 	}
2058 	return 0;
2059 }
2060 
2061 /**
2062  * Verify the Rx Queue list is empty
2063  *
2064  * @param dev
2065  *   Pointer to Ethernet device.
2066  *
2067  * @return
2068  *   The number of object not released.
2069  */
2070 int
2071 mlx5_rxq_verify(struct rte_eth_dev *dev)
2072 {
2073 	struct mlx5_priv *priv = dev->data->dev_private;
2074 	struct mlx5_rxq_ctrl *rxq_ctrl;
2075 	int ret = 0;
2076 
2077 	LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
2078 		DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
2079 			dev->data->port_id, rxq_ctrl->rxq.idx);
2080 		++ret;
2081 	}
2082 	return ret;
2083 }
2084 
2085 /**
2086  * Get a Rx queue type.
2087  *
2088  * @param dev
2089  *   Pointer to Ethernet device.
2090  * @param idx
2091  *   Rx queue index.
2092  *
2093  * @return
2094  *   The Rx queue type.
2095  */
2096 enum mlx5_rxq_type
2097 mlx5_rxq_get_type(struct rte_eth_dev *dev, uint16_t idx)
2098 {
2099 	struct mlx5_priv *priv = dev->data->dev_private;
2100 	struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, idx);
2101 
2102 	if (idx < priv->rxqs_n && rxq_ctrl != NULL)
2103 		return rxq_ctrl->type;
2104 	return MLX5_RXQ_TYPE_UNDEFINED;
2105 }
2106 
2107 /*
2108  * Get a Rx hairpin queue configuration.
2109  *
2110  * @param dev
2111  *   Pointer to Ethernet device.
2112  * @param idx
2113  *   Rx queue index.
2114  *
2115  * @return
2116  *   Pointer to the configuration if a hairpin RX queue, otherwise NULL.
2117  */
2118 const struct rte_eth_hairpin_conf *
2119 mlx5_rxq_get_hairpin_conf(struct rte_eth_dev *dev, uint16_t idx)
2120 {
2121 	struct mlx5_priv *priv = dev->data->dev_private;
2122 	struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
2123 
2124 	if (idx < priv->rxqs_n && rxq != NULL) {
2125 		if (rxq->ctrl->type == MLX5_RXQ_TYPE_HAIRPIN)
2126 			return &rxq->hairpin_conf;
2127 	}
2128 	return NULL;
2129 }
2130 
2131 /**
2132  * Match queues listed in arguments to queues contained in indirection table
2133  * object.
2134  *
2135  * @param ind_tbl
2136  *   Pointer to indirection table to match.
2137  * @param queues
2138  *   Queues to match to ques in indirection table.
2139  * @param queues_n
2140  *   Number of queues in the array.
2141  *
2142  * @return
2143  *   1 if all queues in indirection table match 0 othrwise.
2144  */
2145 static int
2146 mlx5_ind_table_obj_match_queues(const struct mlx5_ind_table_obj *ind_tbl,
2147 		       const uint16_t *queues, uint32_t queues_n)
2148 {
2149 		return (ind_tbl->queues_n == queues_n) &&
2150 		    (!memcmp(ind_tbl->queues, queues,
2151 			    ind_tbl->queues_n * sizeof(ind_tbl->queues[0])));
2152 }
2153 
2154 /**
2155  * Get an indirection table.
2156  *
2157  * @param dev
2158  *   Pointer to Ethernet device.
2159  * @param queues
2160  *   Queues entering in the indirection table.
2161  * @param queues_n
2162  *   Number of queues in the array.
2163  *
2164  * @return
2165  *   An indirection table if found.
2166  */
2167 struct mlx5_ind_table_obj *
2168 mlx5_ind_table_obj_get(struct rte_eth_dev *dev, const uint16_t *queues,
2169 		       uint32_t queues_n)
2170 {
2171 	struct mlx5_priv *priv = dev->data->dev_private;
2172 	struct mlx5_ind_table_obj *ind_tbl;
2173 
2174 	rte_rwlock_read_lock(&priv->ind_tbls_lock);
2175 	LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
2176 		if ((ind_tbl->queues_n == queues_n) &&
2177 		    (memcmp(ind_tbl->queues, queues,
2178 			    ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
2179 		     == 0)) {
2180 			__atomic_fetch_add(&ind_tbl->refcnt, 1,
2181 					   __ATOMIC_RELAXED);
2182 			break;
2183 		}
2184 	}
2185 	rte_rwlock_read_unlock(&priv->ind_tbls_lock);
2186 	return ind_tbl;
2187 }
2188 
2189 /**
2190  * Release an indirection table.
2191  *
2192  * @param dev
2193  *   Pointer to Ethernet device.
2194  * @param ind_table
2195  *   Indirection table to release.
2196  * @param standalone
2197  *   Indirection table for Standalone queue.
2198  *
2199  * @return
2200  *   1 while a reference on it exists, 0 when freed.
2201  */
2202 int
2203 mlx5_ind_table_obj_release(struct rte_eth_dev *dev,
2204 			   struct mlx5_ind_table_obj *ind_tbl,
2205 			   bool standalone)
2206 {
2207 	struct mlx5_priv *priv = dev->data->dev_private;
2208 	unsigned int i, ret;
2209 
2210 	rte_rwlock_write_lock(&priv->ind_tbls_lock);
2211 	ret = __atomic_sub_fetch(&ind_tbl->refcnt, 1, __ATOMIC_RELAXED);
2212 	if (!ret && !standalone)
2213 		LIST_REMOVE(ind_tbl, next);
2214 	rte_rwlock_write_unlock(&priv->ind_tbls_lock);
2215 	if (ret)
2216 		return 1;
2217 	priv->obj_ops.ind_table_destroy(ind_tbl);
2218 	for (i = 0; i != ind_tbl->queues_n; ++i)
2219 		claim_nonzero(mlx5_rxq_deref(dev, ind_tbl->queues[i]));
2220 	mlx5_free(ind_tbl);
2221 	return 0;
2222 }
2223 
2224 /**
2225  * Verify the Rx Queue list is empty
2226  *
2227  * @param dev
2228  *   Pointer to Ethernet device.
2229  *
2230  * @return
2231  *   The number of object not released.
2232  */
2233 int
2234 mlx5_ind_table_obj_verify(struct rte_eth_dev *dev)
2235 {
2236 	struct mlx5_priv *priv = dev->data->dev_private;
2237 	struct mlx5_ind_table_obj *ind_tbl;
2238 	int ret = 0;
2239 
2240 	rte_rwlock_read_lock(&priv->ind_tbls_lock);
2241 	LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
2242 		DRV_LOG(DEBUG,
2243 			"port %u indirection table obj %p still referenced",
2244 			dev->data->port_id, (void *)ind_tbl);
2245 		++ret;
2246 	}
2247 	rte_rwlock_read_unlock(&priv->ind_tbls_lock);
2248 	return ret;
2249 }
2250 
2251 /**
2252  * Setup an indirection table structure fields.
2253  *
2254  * @param dev
2255  *   Pointer to Ethernet device.
2256  * @param ind_table
2257  *   Indirection table to modify.
2258  *
2259  * @return
2260  *   0 on success, a negative errno value otherwise and rte_errno is set.
2261  */
2262 int
2263 mlx5_ind_table_obj_setup(struct rte_eth_dev *dev,
2264 			 struct mlx5_ind_table_obj *ind_tbl)
2265 {
2266 	struct mlx5_priv *priv = dev->data->dev_private;
2267 	uint32_t queues_n = ind_tbl->queues_n;
2268 	uint16_t *queues = ind_tbl->queues;
2269 	unsigned int i, j;
2270 	int ret = 0, err;
2271 	const unsigned int n = rte_is_power_of_2(queues_n) ?
2272 			       log2above(queues_n) :
2273 			       log2above(priv->config.ind_table_max_size);
2274 
2275 	for (i = 0; i != queues_n; ++i) {
2276 		if (mlx5_rxq_ref(dev, queues[i]) == NULL) {
2277 			ret = -rte_errno;
2278 			goto error;
2279 		}
2280 	}
2281 	ret = priv->obj_ops.ind_table_new(dev, n, ind_tbl);
2282 	if (ret)
2283 		goto error;
2284 	__atomic_fetch_add(&ind_tbl->refcnt, 1, __ATOMIC_RELAXED);
2285 	return 0;
2286 error:
2287 	err = rte_errno;
2288 	for (j = 0; j < i; j++)
2289 		mlx5_rxq_deref(dev, ind_tbl->queues[j]);
2290 	rte_errno = err;
2291 	DRV_LOG(DEBUG, "Port %u cannot setup indirection table.",
2292 		dev->data->port_id);
2293 	return ret;
2294 }
2295 
2296 /**
2297  * Create an indirection table.
2298  *
2299  * @param dev
2300  *   Pointer to Ethernet device.
2301  * @param queues
2302  *   Queues entering in the indirection table.
2303  * @param queues_n
2304  *   Number of queues in the array.
2305  * @param standalone
2306  *   Indirection table for Standalone queue.
2307  *
2308  * @return
2309  *   The Verbs/DevX object initialized, NULL otherwise and rte_errno is set.
2310  */
2311 static struct mlx5_ind_table_obj *
2312 mlx5_ind_table_obj_new(struct rte_eth_dev *dev, const uint16_t *queues,
2313 		       uint32_t queues_n, bool standalone)
2314 {
2315 	struct mlx5_priv *priv = dev->data->dev_private;
2316 	struct mlx5_ind_table_obj *ind_tbl;
2317 	int ret;
2318 
2319 	ind_tbl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*ind_tbl) +
2320 			      queues_n * sizeof(uint16_t), 0, SOCKET_ID_ANY);
2321 	if (!ind_tbl) {
2322 		rte_errno = ENOMEM;
2323 		return NULL;
2324 	}
2325 	ind_tbl->queues_n = queues_n;
2326 	ind_tbl->queues = (uint16_t *)(ind_tbl + 1);
2327 	memcpy(ind_tbl->queues, queues, queues_n * sizeof(*queues));
2328 	ret = mlx5_ind_table_obj_setup(dev, ind_tbl);
2329 	if (ret < 0) {
2330 		mlx5_free(ind_tbl);
2331 		return NULL;
2332 	}
2333 	if (!standalone) {
2334 		rte_rwlock_write_lock(&priv->ind_tbls_lock);
2335 		LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
2336 		rte_rwlock_write_unlock(&priv->ind_tbls_lock);
2337 	}
2338 	return ind_tbl;
2339 }
2340 
2341 static int
2342 mlx5_ind_table_obj_check_standalone(struct rte_eth_dev *dev __rte_unused,
2343 				    struct mlx5_ind_table_obj *ind_tbl)
2344 {
2345 	uint32_t refcnt;
2346 
2347 	refcnt = __atomic_load_n(&ind_tbl->refcnt, __ATOMIC_RELAXED);
2348 	if (refcnt <= 1)
2349 		return 0;
2350 	/*
2351 	 * Modification of indirection tables having more than 1
2352 	 * reference is unsupported.
2353 	 */
2354 	DRV_LOG(DEBUG,
2355 		"Port %u cannot modify indirection table %p (refcnt %u > 1).",
2356 		dev->data->port_id, (void *)ind_tbl, refcnt);
2357 	rte_errno = EINVAL;
2358 	return -rte_errno;
2359 }
2360 
2361 /**
2362  * Modify an indirection table.
2363  *
2364  * @param dev
2365  *   Pointer to Ethernet device.
2366  * @param ind_table
2367  *   Indirection table to modify.
2368  * @param queues
2369  *   Queues replacement for the indirection table.
2370  * @param queues_n
2371  *   Number of queues in the array.
2372  * @param standalone
2373  *   Indirection table for Standalone queue.
2374  *
2375  * @return
2376  *   0 on success, a negative errno value otherwise and rte_errno is set.
2377  */
2378 int
2379 mlx5_ind_table_obj_modify(struct rte_eth_dev *dev,
2380 			  struct mlx5_ind_table_obj *ind_tbl,
2381 			  uint16_t *queues, const uint32_t queues_n,
2382 			  bool standalone)
2383 {
2384 	struct mlx5_priv *priv = dev->data->dev_private;
2385 	unsigned int i;
2386 	int ret = 0, err;
2387 	const unsigned int n = rte_is_power_of_2(queues_n) ?
2388 			       log2above(queues_n) :
2389 			       log2above(priv->config.ind_table_max_size);
2390 
2391 	MLX5_ASSERT(standalone);
2392 	RTE_SET_USED(standalone);
2393 	if (mlx5_ind_table_obj_check_standalone(dev, ind_tbl) < 0)
2394 		return -rte_errno;
2395 	for (i = 0; i != queues_n; ++i) {
2396 		if (!mlx5_rxq_get(dev, queues[i])) {
2397 			ret = -rte_errno;
2398 			goto error;
2399 		}
2400 	}
2401 	MLX5_ASSERT(priv->obj_ops.ind_table_modify);
2402 	ret = priv->obj_ops.ind_table_modify(dev, n, queues, queues_n, ind_tbl);
2403 	if (ret)
2404 		goto error;
2405 	ind_tbl->queues_n = queues_n;
2406 	ind_tbl->queues = queues;
2407 	return 0;
2408 error:
2409 	err = rte_errno;
2410 	rte_errno = err;
2411 	DRV_LOG(DEBUG, "Port %u cannot setup indirection table.",
2412 		dev->data->port_id);
2413 	return ret;
2414 }
2415 
2416 /**
2417  * Attach an indirection table to its queues.
2418  *
2419  * @param dev
2420  *   Pointer to Ethernet device.
2421  * @param ind_table
2422  *   Indirection table to attach.
2423  *
2424  * @return
2425  *   0 on success, a negative errno value otherwise and rte_errno is set.
2426  */
2427 int
2428 mlx5_ind_table_obj_attach(struct rte_eth_dev *dev,
2429 			  struct mlx5_ind_table_obj *ind_tbl)
2430 {
2431 	unsigned int i;
2432 	int ret;
2433 
2434 	ret = mlx5_ind_table_obj_modify(dev, ind_tbl, ind_tbl->queues,
2435 					ind_tbl->queues_n, true);
2436 	if (ret != 0) {
2437 		DRV_LOG(ERR, "Port %u could not modify indirect table obj %p",
2438 			dev->data->port_id, (void *)ind_tbl);
2439 		return ret;
2440 	}
2441 	for (i = 0; i < ind_tbl->queues_n; i++)
2442 		mlx5_rxq_ref(dev, ind_tbl->queues[i]);
2443 	return 0;
2444 }
2445 
2446 /**
2447  * Detach an indirection table from its queues.
2448  *
2449  * @param dev
2450  *   Pointer to Ethernet device.
2451  * @param ind_table
2452  *   Indirection table to detach.
2453  *
2454  * @return
2455  *   0 on success, a negative errno value otherwise and rte_errno is set.
2456  */
2457 int
2458 mlx5_ind_table_obj_detach(struct rte_eth_dev *dev,
2459 			  struct mlx5_ind_table_obj *ind_tbl)
2460 {
2461 	struct mlx5_priv *priv = dev->data->dev_private;
2462 	const unsigned int n = rte_is_power_of_2(ind_tbl->queues_n) ?
2463 			       log2above(ind_tbl->queues_n) :
2464 			       log2above(priv->config.ind_table_max_size);
2465 	unsigned int i;
2466 	int ret;
2467 
2468 	ret = mlx5_ind_table_obj_check_standalone(dev, ind_tbl);
2469 	if (ret != 0)
2470 		return ret;
2471 	MLX5_ASSERT(priv->obj_ops.ind_table_modify);
2472 	ret = priv->obj_ops.ind_table_modify(dev, n, NULL, 0, ind_tbl);
2473 	if (ret != 0) {
2474 		DRV_LOG(ERR, "Port %u could not modify indirect table obj %p",
2475 			dev->data->port_id, (void *)ind_tbl);
2476 		return ret;
2477 	}
2478 	for (i = 0; i < ind_tbl->queues_n; i++)
2479 		mlx5_rxq_release(dev, ind_tbl->queues[i]);
2480 	return ret;
2481 }
2482 
2483 int
2484 mlx5_hrxq_match_cb(void *tool_ctx __rte_unused, struct mlx5_list_entry *entry,
2485 		   void *cb_ctx)
2486 {
2487 	struct mlx5_flow_cb_ctx *ctx = cb_ctx;
2488 	struct mlx5_flow_rss_desc *rss_desc = ctx->data;
2489 	struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
2490 
2491 	return (hrxq->rss_key_len != rss_desc->key_len ||
2492 	    memcmp(hrxq->rss_key, rss_desc->key, rss_desc->key_len) ||
2493 	    hrxq->hash_fields != rss_desc->hash_fields ||
2494 	    hrxq->ind_table->queues_n != rss_desc->queue_num ||
2495 	    memcmp(hrxq->ind_table->queues, rss_desc->queue,
2496 	    rss_desc->queue_num * sizeof(rss_desc->queue[0])));
2497 }
2498 
2499 /**
2500  * Modify an Rx Hash queue configuration.
2501  *
2502  * @param dev
2503  *   Pointer to Ethernet device.
2504  * @param hrxq
2505  *   Index to Hash Rx queue to modify.
2506  * @param rss_key
2507  *   RSS key for the Rx hash queue.
2508  * @param rss_key_len
2509  *   RSS key length.
2510  * @param hash_fields
2511  *   Verbs protocol hash field to make the RSS on.
2512  * @param queues
2513  *   Queues entering in hash queue. In case of empty hash_fields only the
2514  *   first queue index will be taken for the indirection table.
2515  * @param queues_n
2516  *   Number of queues.
2517  *
2518  * @return
2519  *   0 on success, a negative errno value otherwise and rte_errno is set.
2520  */
2521 int
2522 mlx5_hrxq_modify(struct rte_eth_dev *dev, uint32_t hrxq_idx,
2523 		 const uint8_t *rss_key, uint32_t rss_key_len,
2524 		 uint64_t hash_fields,
2525 		 const uint16_t *queues, uint32_t queues_n)
2526 {
2527 	int err;
2528 	struct mlx5_ind_table_obj *ind_tbl = NULL;
2529 	struct mlx5_priv *priv = dev->data->dev_private;
2530 	struct mlx5_hrxq *hrxq =
2531 		mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
2532 	int ret;
2533 
2534 	if (!hrxq) {
2535 		rte_errno = EINVAL;
2536 		return -rte_errno;
2537 	}
2538 	/* validations */
2539 	if (hrxq->rss_key_len != rss_key_len) {
2540 		/* rss_key_len is fixed size 40 byte & not supposed to change */
2541 		rte_errno = EINVAL;
2542 		return -rte_errno;
2543 	}
2544 	queues_n = hash_fields ? queues_n : 1;
2545 	if (mlx5_ind_table_obj_match_queues(hrxq->ind_table,
2546 					    queues, queues_n)) {
2547 		ind_tbl = hrxq->ind_table;
2548 	} else {
2549 		if (hrxq->standalone) {
2550 			/*
2551 			 * Replacement of indirection table unsupported for
2552 			 * stanalone hrxq objects (used by shared RSS).
2553 			 */
2554 			rte_errno = ENOTSUP;
2555 			return -rte_errno;
2556 		}
2557 		ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
2558 		if (!ind_tbl)
2559 			ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n,
2560 							 hrxq->standalone);
2561 	}
2562 	if (!ind_tbl) {
2563 		rte_errno = ENOMEM;
2564 		return -rte_errno;
2565 	}
2566 	MLX5_ASSERT(priv->obj_ops.hrxq_modify);
2567 	ret = priv->obj_ops.hrxq_modify(dev, hrxq, rss_key,
2568 					hash_fields, ind_tbl);
2569 	if (ret) {
2570 		rte_errno = errno;
2571 		goto error;
2572 	}
2573 	if (ind_tbl != hrxq->ind_table) {
2574 		MLX5_ASSERT(!hrxq->standalone);
2575 		mlx5_ind_table_obj_release(dev, hrxq->ind_table,
2576 					   hrxq->standalone);
2577 		hrxq->ind_table = ind_tbl;
2578 	}
2579 	hrxq->hash_fields = hash_fields;
2580 	memcpy(hrxq->rss_key, rss_key, rss_key_len);
2581 	return 0;
2582 error:
2583 	err = rte_errno;
2584 	if (ind_tbl != hrxq->ind_table) {
2585 		MLX5_ASSERT(!hrxq->standalone);
2586 		mlx5_ind_table_obj_release(dev, ind_tbl, hrxq->standalone);
2587 	}
2588 	rte_errno = err;
2589 	return -rte_errno;
2590 }
2591 
2592 static void
2593 __mlx5_hrxq_remove(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
2594 {
2595 	struct mlx5_priv *priv = dev->data->dev_private;
2596 
2597 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2598 	mlx5_glue->destroy_flow_action(hrxq->action);
2599 #endif
2600 	priv->obj_ops.hrxq_destroy(hrxq);
2601 	if (!hrxq->standalone) {
2602 		mlx5_ind_table_obj_release(dev, hrxq->ind_table,
2603 					   hrxq->standalone);
2604 	}
2605 	mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq->idx);
2606 }
2607 
2608 /**
2609  * Release the hash Rx queue.
2610  *
2611  * @param dev
2612  *   Pointer to Ethernet device.
2613  * @param hrxq
2614  *   Index to Hash Rx queue to release.
2615  *
2616  * @param list
2617  *   mlx5 list pointer.
2618  * @param entry
2619  *   Hash queue entry pointer.
2620  */
2621 void
2622 mlx5_hrxq_remove_cb(void *tool_ctx, struct mlx5_list_entry *entry)
2623 {
2624 	struct rte_eth_dev *dev = tool_ctx;
2625 	struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
2626 
2627 	__mlx5_hrxq_remove(dev, hrxq);
2628 }
2629 
2630 static struct mlx5_hrxq *
2631 __mlx5_hrxq_create(struct rte_eth_dev *dev,
2632 		   struct mlx5_flow_rss_desc *rss_desc)
2633 {
2634 	struct mlx5_priv *priv = dev->data->dev_private;
2635 	const uint8_t *rss_key = rss_desc->key;
2636 	uint32_t rss_key_len =  rss_desc->key_len;
2637 	bool standalone = !!rss_desc->shared_rss;
2638 	const uint16_t *queues =
2639 		standalone ? rss_desc->const_q : rss_desc->queue;
2640 	uint32_t queues_n = rss_desc->queue_num;
2641 	struct mlx5_hrxq *hrxq = NULL;
2642 	uint32_t hrxq_idx = 0;
2643 	struct mlx5_ind_table_obj *ind_tbl = rss_desc->ind_tbl;
2644 	int ret;
2645 
2646 	queues_n = rss_desc->hash_fields ? queues_n : 1;
2647 	if (!ind_tbl)
2648 		ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
2649 	if (!ind_tbl)
2650 		ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n,
2651 						 standalone);
2652 	if (!ind_tbl)
2653 		return NULL;
2654 	hrxq = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_HRXQ], &hrxq_idx);
2655 	if (!hrxq)
2656 		goto error;
2657 	hrxq->standalone = standalone;
2658 	hrxq->idx = hrxq_idx;
2659 	hrxq->ind_table = ind_tbl;
2660 	hrxq->rss_key_len = rss_key_len;
2661 	hrxq->hash_fields = rss_desc->hash_fields;
2662 	memcpy(hrxq->rss_key, rss_key, rss_key_len);
2663 	ret = priv->obj_ops.hrxq_new(dev, hrxq, rss_desc->tunnel);
2664 	if (ret < 0)
2665 		goto error;
2666 	return hrxq;
2667 error:
2668 	if (!rss_desc->ind_tbl)
2669 		mlx5_ind_table_obj_release(dev, ind_tbl, standalone);
2670 	if (hrxq)
2671 		mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
2672 	return NULL;
2673 }
2674 
2675 struct mlx5_list_entry *
2676 mlx5_hrxq_create_cb(void *tool_ctx, void *cb_ctx)
2677 {
2678 	struct rte_eth_dev *dev = tool_ctx;
2679 	struct mlx5_flow_cb_ctx *ctx = cb_ctx;
2680 	struct mlx5_flow_rss_desc *rss_desc = ctx->data;
2681 	struct mlx5_hrxq *hrxq;
2682 
2683 	hrxq = __mlx5_hrxq_create(dev, rss_desc);
2684 	return hrxq ? &hrxq->entry : NULL;
2685 }
2686 
2687 struct mlx5_list_entry *
2688 mlx5_hrxq_clone_cb(void *tool_ctx, struct mlx5_list_entry *entry,
2689 		    void *cb_ctx __rte_unused)
2690 {
2691 	struct rte_eth_dev *dev = tool_ctx;
2692 	struct mlx5_priv *priv = dev->data->dev_private;
2693 	struct mlx5_hrxq *hrxq;
2694 	uint32_t hrxq_idx = 0;
2695 
2696 	hrxq = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_HRXQ], &hrxq_idx);
2697 	if (!hrxq)
2698 		return NULL;
2699 	memcpy(hrxq, entry, sizeof(*hrxq) + MLX5_RSS_HASH_KEY_LEN);
2700 	hrxq->idx = hrxq_idx;
2701 	return &hrxq->entry;
2702 }
2703 
2704 void
2705 mlx5_hrxq_clone_free_cb(void *tool_ctx, struct mlx5_list_entry *entry)
2706 {
2707 	struct rte_eth_dev *dev = tool_ctx;
2708 	struct mlx5_priv *priv = dev->data->dev_private;
2709 	struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
2710 
2711 	mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq->idx);
2712 }
2713 
2714 /**
2715  * Get an Rx Hash queue.
2716  *
2717  * @param dev
2718  *   Pointer to Ethernet device.
2719  * @param rss_desc
2720  *   RSS configuration for the Rx hash queue.
2721  *
2722  * @return
2723  *   An hash Rx queue index on success.
2724  */
2725 uint32_t mlx5_hrxq_get(struct rte_eth_dev *dev,
2726 		       struct mlx5_flow_rss_desc *rss_desc)
2727 {
2728 	struct mlx5_priv *priv = dev->data->dev_private;
2729 	struct mlx5_hrxq *hrxq;
2730 	struct mlx5_list_entry *entry;
2731 	struct mlx5_flow_cb_ctx ctx = {
2732 		.data = rss_desc,
2733 	};
2734 
2735 	if (rss_desc->shared_rss) {
2736 		hrxq = __mlx5_hrxq_create(dev, rss_desc);
2737 	} else {
2738 		entry = mlx5_list_register(priv->hrxqs, &ctx);
2739 		if (!entry)
2740 			return 0;
2741 		hrxq = container_of(entry, typeof(*hrxq), entry);
2742 	}
2743 	if (hrxq)
2744 		return hrxq->idx;
2745 	return 0;
2746 }
2747 
2748 /**
2749  * Release the hash Rx queue.
2750  *
2751  * @param dev
2752  *   Pointer to Ethernet device.
2753  * @param hrxq_idx
2754  *   Index to Hash Rx queue to release.
2755  *
2756  * @return
2757  *   1 while a reference on it exists, 0 when freed.
2758  */
2759 int mlx5_hrxq_release(struct rte_eth_dev *dev, uint32_t hrxq_idx)
2760 {
2761 	struct mlx5_priv *priv = dev->data->dev_private;
2762 	struct mlx5_hrxq *hrxq;
2763 
2764 	hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
2765 	if (!hrxq)
2766 		return 0;
2767 	if (!hrxq->standalone)
2768 		return mlx5_list_unregister(priv->hrxqs, &hrxq->entry);
2769 	__mlx5_hrxq_remove(dev, hrxq);
2770 	return 0;
2771 }
2772 
2773 /**
2774  * Create a drop Rx Hash queue.
2775  *
2776  * @param dev
2777  *   Pointer to Ethernet device.
2778  *
2779  * @return
2780  *   The Verbs/DevX object initialized, NULL otherwise and rte_errno is set.
2781  */
2782 struct mlx5_hrxq *
2783 mlx5_drop_action_create(struct rte_eth_dev *dev)
2784 {
2785 	struct mlx5_priv *priv = dev->data->dev_private;
2786 	struct mlx5_hrxq *hrxq = NULL;
2787 	int ret;
2788 
2789 	if (priv->drop_queue.hrxq)
2790 		return priv->drop_queue.hrxq;
2791 	hrxq = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*hrxq), 0, SOCKET_ID_ANY);
2792 	if (!hrxq) {
2793 		DRV_LOG(WARNING,
2794 			"Port %u cannot allocate memory for drop queue.",
2795 			dev->data->port_id);
2796 		rte_errno = ENOMEM;
2797 		goto error;
2798 	}
2799 	priv->drop_queue.hrxq = hrxq;
2800 	hrxq->ind_table = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*hrxq->ind_table),
2801 				      0, SOCKET_ID_ANY);
2802 	if (!hrxq->ind_table) {
2803 		rte_errno = ENOMEM;
2804 		goto error;
2805 	}
2806 	ret = priv->obj_ops.drop_action_create(dev);
2807 	if (ret < 0)
2808 		goto error;
2809 	return hrxq;
2810 error:
2811 	if (hrxq) {
2812 		if (hrxq->ind_table)
2813 			mlx5_free(hrxq->ind_table);
2814 		priv->drop_queue.hrxq = NULL;
2815 		mlx5_free(hrxq);
2816 	}
2817 	return NULL;
2818 }
2819 
2820 /**
2821  * Release a drop hash Rx queue.
2822  *
2823  * @param dev
2824  *   Pointer to Ethernet device.
2825  */
2826 void
2827 mlx5_drop_action_destroy(struct rte_eth_dev *dev)
2828 {
2829 	struct mlx5_priv *priv = dev->data->dev_private;
2830 	struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
2831 
2832 	if (!priv->drop_queue.hrxq)
2833 		return;
2834 	priv->obj_ops.drop_action_destroy(dev);
2835 	mlx5_free(priv->drop_queue.rxq);
2836 	mlx5_free(hrxq->ind_table);
2837 	mlx5_free(hrxq);
2838 	priv->drop_queue.rxq = NULL;
2839 	priv->drop_queue.hrxq = NULL;
2840 }
2841 
2842 /**
2843  * Verify the Rx Queue list is empty
2844  *
2845  * @param dev
2846  *   Pointer to Ethernet device.
2847  *
2848  * @return
2849  *   The number of object not released.
2850  */
2851 uint32_t
2852 mlx5_hrxq_verify(struct rte_eth_dev *dev)
2853 {
2854 	struct mlx5_priv *priv = dev->data->dev_private;
2855 
2856 	return mlx5_list_get_entry_num(priv->hrxqs);
2857 }
2858 
2859 /**
2860  * Set the Rx queue timestamp conversion parameters
2861  *
2862  * @param[in] dev
2863  *   Pointer to the Ethernet device structure.
2864  */
2865 void
2866 mlx5_rxq_timestamp_set(struct rte_eth_dev *dev)
2867 {
2868 	struct mlx5_priv *priv = dev->data->dev_private;
2869 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2870 	unsigned int i;
2871 
2872 	for (i = 0; i != priv->rxqs_n; ++i) {
2873 		struct mlx5_rxq_data *data = mlx5_rxq_data_get(dev, i);
2874 
2875 		if (data == NULL)
2876 			continue;
2877 		data->sh = sh;
2878 		data->rt_timestamp = priv->config.rt_timestamp;
2879 	}
2880 }
2881