xref: /dpdk/drivers/net/mlx5/mlx5_devx.c (revision 1944fbc3c2c8dd0cbaf04ad04cab39c0bb0cd7fc)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2020 Mellanox Technologies, Ltd
3  */
4 
5 #include <stddef.h>
6 #include <errno.h>
7 #include <stdbool.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <sys/queue.h>
11 
12 #include <rte_malloc.h>
13 #include <rte_common.h>
14 #include <rte_eal_paging.h>
15 
16 #include <mlx5_glue.h>
17 #include <mlx5_devx_cmds.h>
18 #include <mlx5_common_devx.h>
19 #include <mlx5_malloc.h>
20 
21 #include "mlx5.h"
22 #include "mlx5_common_os.h"
23 #include "mlx5_tx.h"
24 #include "mlx5_rx.h"
25 #include "mlx5_utils.h"
26 #include "mlx5_devx.h"
27 #include "mlx5_flow.h"
28 #include "mlx5_flow_os.h"
29 
30 /**
31  * Validate given external queue's port is valid or not.
32  *
33  * @param[in] port_id
34  *   The port identifier of the Ethernet device.
35  *
36  * @return
37  *   0 on success, non-0 otherwise
38  */
39 int
40 mlx5_devx_extq_port_validate(uint16_t port_id)
41 {
42 	struct rte_eth_dev *dev;
43 	struct mlx5_priv *priv;
44 
45 	if (rte_eth_dev_is_valid_port(port_id) < 0) {
46 		DRV_LOG(ERR, "There is no Ethernet device for port %u.",
47 			port_id);
48 		rte_errno = ENODEV;
49 		return -rte_errno;
50 	}
51 	dev = &rte_eth_devices[port_id];
52 	priv = dev->data->dev_private;
53 	if (!mlx5_imported_pd_and_ctx(priv->sh->cdev)) {
54 		DRV_LOG(ERR, "Port %u "
55 			"external queue isn't supported on local PD and CTX.",
56 			port_id);
57 		rte_errno = ENOTSUP;
58 		return -rte_errno;
59 	}
60 	if (!mlx5_devx_obj_ops_en(priv->sh)) {
61 		DRV_LOG(ERR,
62 			"Port %u external queue isn't supported by Verbs API.",
63 			port_id);
64 		rte_errno = ENOTSUP;
65 		return -rte_errno;
66 	}
67 	return 0;
68 }
69 
70 /**
71  * Modify RQ vlan stripping offload
72  *
73  * @param rxq
74  *   Rx queue.
75  * @param on
76  *   Enable/disable VLAN stripping.
77  *
78  * @return
79  *   0 on success, non-0 otherwise
80  */
81 static int
82 mlx5_rxq_obj_modify_rq_vlan_strip(struct mlx5_rxq_priv *rxq, int on)
83 {
84 	struct mlx5_devx_modify_rq_attr rq_attr;
85 
86 	memset(&rq_attr, 0, sizeof(rq_attr));
87 	rq_attr.rq_state = MLX5_RQC_STATE_RDY;
88 	rq_attr.state = MLX5_RQC_STATE_RDY;
89 	rq_attr.vsd = (on ? 0 : 1);
90 	rq_attr.modify_bitmask = MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_VSD;
91 	return mlx5_devx_cmd_modify_rq(rxq->devx_rq.rq, &rq_attr);
92 }
93 
94 /**
95  * Modify RQ using DevX API.
96  *
97  * @param rxq
98  *   DevX rx queue.
99  * @param type
100  *   Type of change queue state.
101  *
102  * @return
103  *   0 on success, a negative errno value otherwise and rte_errno is set.
104  */
105 int
106 mlx5_devx_modify_rq(struct mlx5_rxq_priv *rxq, uint8_t type)
107 {
108 	struct mlx5_devx_modify_rq_attr rq_attr;
109 
110 	memset(&rq_attr, 0, sizeof(rq_attr));
111 	switch (type) {
112 	case MLX5_RXQ_MOD_ERR2RST:
113 		rq_attr.rq_state = MLX5_RQC_STATE_ERR;
114 		rq_attr.state = MLX5_RQC_STATE_RST;
115 		break;
116 	case MLX5_RXQ_MOD_RST2RDY:
117 		rq_attr.rq_state = MLX5_RQC_STATE_RST;
118 		rq_attr.state = MLX5_RQC_STATE_RDY;
119 		if (rxq->lwm) {
120 			rq_attr.modify_bitmask |=
121 				MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_WQ_LWM;
122 			rq_attr.lwm = rxq->lwm;
123 		}
124 		break;
125 	case MLX5_RXQ_MOD_RDY2ERR:
126 		rq_attr.rq_state = MLX5_RQC_STATE_RDY;
127 		rq_attr.state = MLX5_RQC_STATE_ERR;
128 		break;
129 	case MLX5_RXQ_MOD_RDY2RST:
130 		rq_attr.rq_state = MLX5_RQC_STATE_RDY;
131 		rq_attr.state = MLX5_RQC_STATE_RST;
132 		break;
133 	case MLX5_RXQ_MOD_RDY2RDY:
134 		rq_attr.rq_state = MLX5_RQC_STATE_RDY;
135 		rq_attr.state = MLX5_RQC_STATE_RDY;
136 		rq_attr.modify_bitmask |= MLX5_MODIFY_RQ_IN_MODIFY_BITMASK_WQ_LWM;
137 		rq_attr.lwm = rxq->lwm;
138 		break;
139 	default:
140 		break;
141 	}
142 	if (rxq->ctrl->is_hairpin)
143 		return mlx5_devx_cmd_modify_rq(rxq->ctrl->obj->rq, &rq_attr);
144 	return mlx5_devx_cmd_modify_rq(rxq->devx_rq.rq, &rq_attr);
145 }
146 
147 /**
148  * Modify SQ using DevX API.
149  *
150  * @param txq_obj
151  *   DevX Tx queue object.
152  * @param type
153  *   Type of change queue state.
154  * @param dev_port
155  *   Unnecessary.
156  *
157  * @return
158  *   0 on success, a negative errno value otherwise and rte_errno is set.
159  */
160 int
161 mlx5_txq_devx_modify(struct mlx5_txq_obj *obj, enum mlx5_txq_modify_type type,
162 		     uint8_t dev_port)
163 {
164 	struct mlx5_devx_modify_sq_attr msq_attr = { 0 };
165 	int ret;
166 
167 	if (type != MLX5_TXQ_MOD_RST2RDY) {
168 		/* Change queue state to reset. */
169 		if (type == MLX5_TXQ_MOD_ERR2RDY)
170 			msq_attr.sq_state = MLX5_SQC_STATE_ERR;
171 		else
172 			msq_attr.sq_state = MLX5_SQC_STATE_RDY;
173 		msq_attr.state = MLX5_SQC_STATE_RST;
174 		ret = mlx5_devx_cmd_modify_sq(obj->sq_obj.sq, &msq_attr);
175 		if (ret) {
176 			DRV_LOG(ERR, "Cannot change the Tx SQ state to RESET"
177 				" %s", strerror(errno));
178 			rte_errno = errno;
179 			return ret;
180 		}
181 	}
182 	if (type != MLX5_TXQ_MOD_RDY2RST) {
183 		/* Change queue state to ready. */
184 		msq_attr.sq_state = MLX5_SQC_STATE_RST;
185 		msq_attr.state = MLX5_SQC_STATE_RDY;
186 		ret = mlx5_devx_cmd_modify_sq(obj->sq_obj.sq, &msq_attr);
187 		if (ret) {
188 			DRV_LOG(ERR, "Cannot change the Tx SQ state to READY"
189 				" %s", strerror(errno));
190 			rte_errno = errno;
191 			return ret;
192 		}
193 	}
194 	/*
195 	 * The dev_port variable is relevant only in Verbs API, and there is a
196 	 * pointer that points to this function and a parallel function in verbs
197 	 * intermittently, so they should have the same parameters.
198 	 */
199 	(void)dev_port;
200 	return 0;
201 }
202 
203 /**
204  * Release an Rx DevX queue object.
205  *
206  * @param rxq
207  *   DevX Rx queue.
208  */
209 static void
210 mlx5_rxq_devx_obj_release(struct mlx5_rxq_priv *rxq)
211 {
212 	struct mlx5_rxq_obj *rxq_obj = rxq->ctrl->obj;
213 
214 	if (rxq_obj == NULL)
215 		return;
216 	if (rxq_obj->rxq_ctrl->is_hairpin) {
217 		if (rxq_obj->rq == NULL)
218 			return;
219 		mlx5_devx_modify_rq(rxq, MLX5_RXQ_MOD_RDY2RST);
220 		claim_zero(mlx5_devx_cmd_destroy(rxq_obj->rq));
221 	} else {
222 		if (rxq->devx_rq.rq == NULL)
223 			return;
224 		mlx5_devx_rq_destroy(&rxq->devx_rq);
225 		if (rxq->devx_rq.rmp != NULL && rxq->devx_rq.rmp->ref_cnt > 0)
226 			return;
227 		mlx5_devx_cq_destroy(&rxq_obj->cq_obj);
228 		memset(&rxq_obj->cq_obj, 0, sizeof(rxq_obj->cq_obj));
229 		if (rxq_obj->devx_channel) {
230 			mlx5_os_devx_destroy_event_channel
231 							(rxq_obj->devx_channel);
232 			rxq_obj->devx_channel = NULL;
233 		}
234 	}
235 	rxq->ctrl->started = false;
236 }
237 
238 /**
239  * Get event for an Rx DevX queue object.
240  *
241  * @param rxq_obj
242  *   DevX Rx queue object.
243  *
244  * @return
245  *   0 on success, a negative errno value otherwise and rte_errno is set.
246  */
247 static int
248 mlx5_rx_devx_get_event(struct mlx5_rxq_obj *rxq_obj)
249 {
250 #ifdef HAVE_IBV_DEVX_EVENT
251 	union {
252 		struct mlx5dv_devx_async_event_hdr event_resp;
253 		uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
254 	} out;
255 	int ret = mlx5_glue->devx_get_event(rxq_obj->devx_channel,
256 					    &out.event_resp,
257 					    sizeof(out.buf));
258 
259 	if (ret < 0) {
260 		rte_errno = errno;
261 		return -rte_errno;
262 	}
263 	if (out.event_resp.cookie != (uint64_t)(uintptr_t)rxq_obj->cq_obj.cq) {
264 		rte_errno = EINVAL;
265 		return -rte_errno;
266 	}
267 	return 0;
268 #else
269 	(void)rxq_obj;
270 	rte_errno = ENOTSUP;
271 	return -rte_errno;
272 #endif /* HAVE_IBV_DEVX_EVENT */
273 }
274 
275 /**
276  * Get LWM event for shared context, return the correct port/rxq for this event.
277  *
278  * @param priv
279  *   Mlx5_priv object.
280  * @param rxq_idx [out]
281  *   Which rxq gets this event.
282  * @param port_id [out]
283  *   Which port gets this event.
284  *
285  * @return
286  *   0 on success, a negative errno value otherwise and rte_errno is set.
287  */
288 static int
289 mlx5_rx_devx_get_event_lwm(struct mlx5_priv *priv, int *rxq_idx, int *port_id)
290 {
291 #ifdef HAVE_IBV_DEVX_EVENT
292 	union {
293 		struct mlx5dv_devx_async_event_hdr event_resp;
294 		uint8_t buf[sizeof(struct mlx5dv_devx_async_event_hdr) + 128];
295 	} out;
296 	int ret;
297 
298 	memset(&out, 0, sizeof(out));
299 	ret = mlx5_glue->devx_get_event(priv->sh->devx_channel_lwm,
300 					&out.event_resp,
301 					sizeof(out.buf));
302 	if (ret < 0) {
303 		rte_errno = errno;
304 		DRV_LOG(WARNING, "%s err\n", __func__);
305 		return -rte_errno;
306 	}
307 	*port_id = (((uint32_t)out.event_resp.cookie) >>
308 		    LWM_COOKIE_PORTID_OFFSET) & LWM_COOKIE_PORTID_MASK;
309 	*rxq_idx = (((uint32_t)out.event_resp.cookie) >>
310 		    LWM_COOKIE_RXQID_OFFSET) & LWM_COOKIE_RXQID_MASK;
311 	return 0;
312 #else
313 	(void)priv;
314 	(void)rxq_idx;
315 	(void)port_id;
316 	rte_errno = ENOTSUP;
317 	return -rte_errno;
318 #endif /* HAVE_IBV_DEVX_EVENT */
319 }
320 
321 /**
322  * Create a RQ object using DevX.
323  *
324  * @param rxq
325  *   Pointer to Rx queue.
326  *
327  * @return
328  *   0 on success, a negative errno value otherwise and rte_errno is set.
329  */
330 static int
331 mlx5_rxq_create_devx_rq_resources(struct mlx5_rxq_priv *rxq)
332 {
333 	struct mlx5_priv *priv = rxq->priv;
334 	struct mlx5_common_device *cdev = priv->sh->cdev;
335 	struct mlx5_rxq_ctrl *rxq_ctrl = rxq->ctrl;
336 	struct mlx5_rxq_data *rxq_data = &rxq->ctrl->rxq;
337 	struct mlx5_devx_create_rq_attr rq_attr = { 0 };
338 	uint16_t log_desc_n = rxq_data->elts_n - rxq_data->sges_n;
339 	uint32_t wqe_size, log_wqe_size;
340 
341 	/* Fill RQ attributes. */
342 	rq_attr.mem_rq_type = MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE;
343 	rq_attr.flush_in_error_en = 1;
344 	rq_attr.vsd = (rxq_data->vlan_strip) ? 0 : 1;
345 	rq_attr.cqn = rxq_ctrl->obj->cq_obj.cq->id;
346 	rq_attr.scatter_fcs = (rxq_data->crc_present) ? 1 : 0;
347 	rq_attr.ts_format =
348 			mlx5_ts_format_conv(cdev->config.hca_attr.rq_ts_format);
349 	/* Fill WQ attributes for this RQ. */
350 	if (mlx5_rxq_mprq_enabled(rxq_data)) {
351 		rq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ;
352 		/*
353 		 * Number of strides in each WQE:
354 		 * 512*2^single_wqe_log_num_of_strides.
355 		 */
356 		rq_attr.wq_attr.single_wqe_log_num_of_strides =
357 				rxq_data->log_strd_num -
358 				MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
359 		/* Stride size = (2^single_stride_log_num_of_bytes)*64B. */
360 		rq_attr.wq_attr.single_stride_log_num_of_bytes =
361 				rxq_data->log_strd_sz -
362 				MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
363 		wqe_size = sizeof(struct mlx5_wqe_mprq);
364 	} else {
365 		rq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC;
366 		wqe_size = sizeof(struct mlx5_wqe_data_seg);
367 	}
368 	log_wqe_size = log2above(wqe_size) + rxq_data->sges_n;
369 	wqe_size = 1 << log_wqe_size; /* round up power of two.*/
370 	rq_attr.wq_attr.log_wq_stride = log_wqe_size;
371 	rq_attr.wq_attr.log_wq_sz = log_desc_n;
372 	rq_attr.wq_attr.end_padding_mode = priv->config.hw_padding ?
373 						MLX5_WQ_END_PAD_MODE_ALIGN :
374 						MLX5_WQ_END_PAD_MODE_NONE;
375 	rq_attr.wq_attr.pd = cdev->pdn;
376 	rq_attr.counter_set_id = priv->counter_set_id;
377 	rq_attr.delay_drop_en = rxq_data->delay_drop;
378 	rq_attr.user_index = rte_cpu_to_be_16(priv->dev_data->port_id);
379 	if (rxq_data->shared) /* Create RMP based RQ. */
380 		rxq->devx_rq.rmp = &rxq_ctrl->obj->devx_rmp;
381 	/* Create RQ using DevX API. */
382 	return mlx5_devx_rq_create(cdev->ctx, &rxq->devx_rq, wqe_size,
383 				   log_desc_n, &rq_attr, rxq_ctrl->socket);
384 }
385 
386 /**
387  * Create a DevX CQ object for an Rx queue.
388  *
389  * @param rxq
390  *   Pointer to Rx queue.
391  *
392  * @return
393  *   0 on success, a negative errno value otherwise and rte_errno is set.
394  */
395 static int
396 mlx5_rxq_create_devx_cq_resources(struct mlx5_rxq_priv *rxq)
397 {
398 	struct mlx5_devx_cq *cq_obj = 0;
399 	struct mlx5_devx_cq_attr cq_attr = { 0 };
400 	struct mlx5_priv *priv = rxq->priv;
401 	struct mlx5_dev_ctx_shared *sh = priv->sh;
402 	uint16_t port_id = priv->dev_data->port_id;
403 	struct mlx5_rxq_ctrl *rxq_ctrl = rxq->ctrl;
404 	struct mlx5_rxq_data *rxq_data = &rxq_ctrl->rxq;
405 	unsigned int cqe_n = mlx5_rxq_cqe_num(rxq_data);
406 	uint32_t log_cqe_n;
407 	uint16_t event_nums[1] = { 0 };
408 	int ret = 0;
409 
410 	if (rxq_ctrl->started)
411 		return 0;
412 	if (priv->config.cqe_comp && !rxq_data->hw_timestamp &&
413 	    !rxq_data->lro) {
414 		cq_attr.cqe_comp_en = 1u;
415 		cq_attr.cqe_comp_layout = priv->config.enh_cqe_comp;
416 		rxq_data->cqe_comp_layout = cq_attr.cqe_comp_layout;
417 		rxq_data->mcqe_format = priv->config.cqe_comp_fmt;
418 		rxq_data->byte_mask = UINT32_MAX;
419 		switch (priv->config.cqe_comp_fmt) {
420 		case MLX5_CQE_RESP_FORMAT_HASH:
421 			/* fallthrough */
422 		case MLX5_CQE_RESP_FORMAT_CSUM:
423 			/*
424 			 * Select CSUM miniCQE format only for non-vectorized
425 			 * MPRQ Rx burst, use HASH miniCQE format for others.
426 			 */
427 			if (mlx5_rxq_check_vec_support(rxq_data) < 0 &&
428 			    mlx5_rxq_mprq_enabled(rxq_data))
429 				cq_attr.mini_cqe_res_format =
430 					MLX5_CQE_RESP_FORMAT_CSUM_STRIDX;
431 			else
432 				cq_attr.mini_cqe_res_format =
433 					MLX5_CQE_RESP_FORMAT_HASH;
434 			rxq_data->mcqe_format = cq_attr.mini_cqe_res_format;
435 			break;
436 		case MLX5_CQE_RESP_FORMAT_FTAG_STRIDX:
437 			rxq_data->byte_mask = MLX5_LEN_WITH_MARK_MASK;
438 			/* fallthrough */
439 		case MLX5_CQE_RESP_FORMAT_CSUM_STRIDX:
440 			cq_attr.mini_cqe_res_format = priv->config.cqe_comp_fmt;
441 			break;
442 		case MLX5_CQE_RESP_FORMAT_L34H_STRIDX:
443 			cq_attr.mini_cqe_res_format = 0;
444 			cq_attr.mini_cqe_res_format_ext = 1;
445 			break;
446 		}
447 		DRV_LOG(DEBUG,
448 			"Port %u Rx CQE compression is enabled, format %d.",
449 			port_id, priv->config.cqe_comp_fmt);
450 		/*
451 		 * For vectorized Rx, it must not be doubled in order to
452 		 * make cq_ci and rq_ci aligned.
453 		 */
454 		if (mlx5_rxq_check_vec_support(rxq_data) < 0)
455 			cqe_n *= 2;
456 	} else if (priv->config.cqe_comp && rxq_data->hw_timestamp) {
457 		DRV_LOG(DEBUG,
458 			"Port %u Rx CQE compression is disabled for HW timestamp.",
459 			port_id);
460 	} else if (priv->config.cqe_comp && rxq_data->lro) {
461 		DRV_LOG(DEBUG,
462 			"Port %u Rx CQE compression is disabled for LRO.",
463 			port_id);
464 	}
465 	cq_attr.uar_page_id = mlx5_os_get_devx_uar_page_id(sh->rx_uar.obj);
466 	log_cqe_n = log2above(cqe_n);
467 	/* Create CQ using DevX API. */
468 	ret = mlx5_devx_cq_create(sh->cdev->ctx, &rxq_ctrl->obj->cq_obj,
469 				  log_cqe_n, &cq_attr, sh->numa_node);
470 	if (ret)
471 		return ret;
472 	cq_obj = &rxq_ctrl->obj->cq_obj;
473 	rxq_data->cqes = (volatile struct mlx5_cqe (*)[])
474 							(uintptr_t)cq_obj->cqes;
475 	rxq_data->cq_db = cq_obj->db_rec;
476 	rxq_data->uar_data = sh->rx_uar.cq_db;
477 	rxq_data->cqe_n = log_cqe_n;
478 	rxq_data->cqn = cq_obj->cq->id;
479 	rxq_data->cq_ci = 0;
480 	if (rxq_ctrl->obj->devx_channel) {
481 		ret = mlx5_os_devx_subscribe_devx_event
482 					      (rxq_ctrl->obj->devx_channel,
483 					       cq_obj->cq->obj,
484 					       sizeof(event_nums),
485 					       event_nums,
486 					       (uint64_t)(uintptr_t)cq_obj->cq);
487 		if (ret) {
488 			DRV_LOG(ERR, "Fail to subscribe CQ to event channel.");
489 			ret = errno;
490 			mlx5_devx_cq_destroy(cq_obj);
491 			memset(cq_obj, 0, sizeof(*cq_obj));
492 			rte_errno = ret;
493 			return -ret;
494 		}
495 	}
496 	return 0;
497 }
498 
499 /**
500  * Create the Rx hairpin queue object.
501  *
502  * @param rxq
503  *   Pointer to Rx queue.
504  *
505  * @return
506  *   0 on success, a negative errno value otherwise and rte_errno is set.
507  */
508 static int
509 mlx5_rxq_obj_hairpin_new(struct mlx5_rxq_priv *rxq)
510 {
511 	uint16_t idx = rxq->idx;
512 	struct mlx5_priv *priv = rxq->priv;
513 	struct mlx5_hca_attr *hca_attr __rte_unused = &priv->sh->cdev->config.hca_attr;
514 	struct mlx5_rxq_ctrl *rxq_ctrl = rxq->ctrl;
515 	struct mlx5_devx_create_rq_attr unlocked_attr = { 0 };
516 	struct mlx5_devx_create_rq_attr locked_attr = { 0 };
517 	struct mlx5_rxq_obj *tmpl = rxq_ctrl->obj;
518 	uint32_t max_wq_data;
519 
520 	MLX5_ASSERT(rxq != NULL && rxq->ctrl != NULL && tmpl != NULL);
521 	tmpl->rxq_ctrl = rxq_ctrl;
522 	unlocked_attr.hairpin = 1;
523 	max_wq_data =
524 		priv->sh->cdev->config.hca_attr.log_max_hairpin_wq_data_sz;
525 	/* Jumbo frames > 9KB should be supported, and more packets. */
526 	if (priv->config.log_hp_size != (uint32_t)MLX5_ARG_UNSET) {
527 		if (priv->config.log_hp_size > max_wq_data) {
528 			DRV_LOG(ERR, "Total data size %u power of 2 is "
529 				"too large for hairpin.",
530 				priv->config.log_hp_size);
531 			rte_errno = ERANGE;
532 			return -rte_errno;
533 		}
534 		unlocked_attr.wq_attr.log_hairpin_data_sz = priv->config.log_hp_size;
535 	} else {
536 		unlocked_attr.wq_attr.log_hairpin_data_sz =
537 				(max_wq_data < MLX5_HAIRPIN_JUMBO_LOG_SIZE) ?
538 				 max_wq_data : MLX5_HAIRPIN_JUMBO_LOG_SIZE;
539 	}
540 	/* Set the packets number to the maximum value for performance. */
541 	unlocked_attr.wq_attr.log_hairpin_num_packets =
542 			unlocked_attr.wq_attr.log_hairpin_data_sz -
543 			MLX5_HAIRPIN_QUEUE_STRIDE;
544 	unlocked_attr.counter_set_id = priv->counter_set_id;
545 	rxq_ctrl->rxq.delay_drop = priv->config.hp_delay_drop;
546 	unlocked_attr.delay_drop_en = priv->config.hp_delay_drop;
547 	unlocked_attr.hairpin_data_buffer_type =
548 			MLX5_RQC_HAIRPIN_DATA_BUFFER_TYPE_UNLOCKED_INTERNAL_BUFFER;
549 	if (rxq->hairpin_conf.use_locked_device_memory) {
550 		/*
551 		 * It is assumed that configuration is verified against capabilities
552 		 * during queue setup.
553 		 */
554 		MLX5_ASSERT(hca_attr->hairpin_data_buffer_locked);
555 		rte_memcpy(&locked_attr, &unlocked_attr, sizeof(locked_attr));
556 		locked_attr.hairpin_data_buffer_type =
557 				MLX5_RQC_HAIRPIN_DATA_BUFFER_TYPE_LOCKED_INTERNAL_BUFFER;
558 		tmpl->rq = mlx5_devx_cmd_create_rq(priv->sh->cdev->ctx, &locked_attr,
559 						   rxq_ctrl->socket);
560 		if (!tmpl->rq && rxq->hairpin_conf.force_memory) {
561 			DRV_LOG(ERR, "Port %u Rx hairpin queue %u can't create RQ object"
562 				     " with locked memory buffer",
563 				     priv->dev_data->port_id, idx);
564 			return -rte_errno;
565 		} else if (!tmpl->rq && !rxq->hairpin_conf.force_memory) {
566 			DRV_LOG(WARNING, "Port %u Rx hairpin queue %u can't create RQ object"
567 					 " with locked memory buffer. Falling back to unlocked"
568 					 " device memory.",
569 					 priv->dev_data->port_id, idx);
570 			rte_errno = 0;
571 			goto create_rq_unlocked;
572 		}
573 		goto create_rq_set_state;
574 	}
575 
576 create_rq_unlocked:
577 	tmpl->rq = mlx5_devx_cmd_create_rq(priv->sh->cdev->ctx, &unlocked_attr,
578 					   rxq_ctrl->socket);
579 	if (!tmpl->rq) {
580 		DRV_LOG(ERR,
581 			"Port %u Rx hairpin queue %u can't create rq object.",
582 			priv->dev_data->port_id, idx);
583 		rte_errno = errno;
584 		return -rte_errno;
585 	}
586 create_rq_set_state:
587 	priv->dev_data->rx_queue_state[idx] = RTE_ETH_QUEUE_STATE_HAIRPIN;
588 	return 0;
589 }
590 
591 /**
592  * Create the Rx queue DevX object.
593  *
594  * @param rxq
595  *   Pointer to Rx queue.
596  *
597  * @return
598  *   0 on success, a negative errno value otherwise and rte_errno is set.
599  */
600 static int
601 mlx5_rxq_devx_obj_new(struct mlx5_rxq_priv *rxq)
602 {
603 	struct mlx5_priv *priv = rxq->priv;
604 	struct mlx5_rxq_ctrl *rxq_ctrl = rxq->ctrl;
605 	struct mlx5_rxq_data *rxq_data = &rxq_ctrl->rxq;
606 	struct mlx5_rxq_obj *tmpl = rxq_ctrl->obj;
607 	int ret = 0;
608 
609 	MLX5_ASSERT(rxq_data);
610 	MLX5_ASSERT(tmpl);
611 	if (rxq_ctrl->is_hairpin)
612 		return mlx5_rxq_obj_hairpin_new(rxq);
613 	tmpl->rxq_ctrl = rxq_ctrl;
614 	if (rxq_ctrl->irq && !rxq_ctrl->started) {
615 		int devx_ev_flag =
616 			  MLX5DV_DEVX_CREATE_EVENT_CHANNEL_FLAGS_OMIT_EV_DATA;
617 
618 		tmpl->devx_channel = mlx5_os_devx_create_event_channel
619 							(priv->sh->cdev->ctx,
620 							 devx_ev_flag);
621 		if (!tmpl->devx_channel) {
622 			rte_errno = errno;
623 			DRV_LOG(ERR, "Failed to create event channel %d.",
624 				rte_errno);
625 			goto error;
626 		}
627 		tmpl->fd = mlx5_os_get_devx_channel_fd(tmpl->devx_channel);
628 	}
629 	/* Create CQ using DevX API. */
630 	ret = mlx5_rxq_create_devx_cq_resources(rxq);
631 	if (ret) {
632 		DRV_LOG(ERR, "Failed to create CQ.");
633 		goto error;
634 	}
635 	rxq_data->delay_drop = priv->config.std_delay_drop;
636 	/* Create RQ using DevX API. */
637 	ret = mlx5_rxq_create_devx_rq_resources(rxq);
638 	if (ret) {
639 		DRV_LOG(ERR, "Port %u Rx queue %u RQ creation failure.",
640 			priv->dev_data->port_id, rxq->idx);
641 		rte_errno = ENOMEM;
642 		goto error;
643 	}
644 	/* Change queue state to ready. */
645 	ret = mlx5_devx_modify_rq(rxq, MLX5_RXQ_MOD_RST2RDY);
646 	if (ret)
647 		goto error;
648 	if (!rxq_data->shared) {
649 		rxq_data->wqes = (void *)(uintptr_t)rxq->devx_rq.wq.umem_buf;
650 		rxq_data->rq_db = (uint32_t *)(uintptr_t)rxq->devx_rq.wq.db_rec;
651 	} else if (!rxq_ctrl->started) {
652 		rxq_data->wqes = (void *)(uintptr_t)tmpl->devx_rmp.wq.umem_buf;
653 		rxq_data->rq_db =
654 				(uint32_t *)(uintptr_t)tmpl->devx_rmp.wq.db_rec;
655 	}
656 	if (!rxq_ctrl->started) {
657 		mlx5_rxq_initialize(rxq_data);
658 		rxq_ctrl->wqn = rxq->devx_rq.rq->id;
659 	}
660 	priv->dev_data->rx_queue_state[rxq->idx] = RTE_ETH_QUEUE_STATE_STARTED;
661 	return 0;
662 error:
663 	ret = rte_errno; /* Save rte_errno before cleanup. */
664 	mlx5_rxq_devx_obj_release(rxq);
665 	rte_errno = ret; /* Restore rte_errno. */
666 	return -rte_errno;
667 }
668 
669 /**
670  * Prepare RQT attribute structure for DevX RQT API.
671  *
672  * @param dev
673  *   Pointer to Ethernet device.
674  * @param log_n
675  *   Log of number of queues in the array.
676  * @param queues
677  *   List of RX queue indices or NULL, in which case
678  *   the attribute will be filled by drop queue ID.
679  * @param queues_n
680  *   Size of @p queues array or 0 if it is NULL.
681  * @param ind_tbl
682  *   DevX indirection table object.
683  *
684  * @return
685  *   The RQT attr object initialized, NULL otherwise and rte_errno is set.
686  */
687 static struct mlx5_devx_rqt_attr *
688 mlx5_devx_ind_table_create_rqt_attr(struct rte_eth_dev *dev,
689 				     const unsigned int log_n,
690 				     const uint16_t *queues,
691 				     const uint32_t queues_n)
692 {
693 	struct mlx5_priv *priv = dev->data->dev_private;
694 	struct mlx5_devx_rqt_attr *rqt_attr = NULL;
695 	const unsigned int rqt_n = 1 << log_n;
696 	unsigned int i, j;
697 
698 	rqt_attr = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*rqt_attr) +
699 			      rqt_n * sizeof(uint32_t), 0, SOCKET_ID_ANY);
700 	if (!rqt_attr) {
701 		DRV_LOG(ERR, "Port %u cannot allocate RQT resources.",
702 			dev->data->port_id);
703 		rte_errno = ENOMEM;
704 		return NULL;
705 	}
706 	rqt_attr->rqt_max_size = priv->sh->dev_cap.ind_table_max_size;
707 	rqt_attr->rqt_actual_size = rqt_n;
708 	if (queues == NULL) {
709 		for (i = 0; i < rqt_n; i++)
710 			rqt_attr->rq_list[i] =
711 					priv->drop_queue.rxq->devx_rq.rq->id;
712 		return rqt_attr;
713 	}
714 	for (i = 0; i != queues_n; ++i) {
715 		if (mlx5_is_external_rxq(dev, queues[i])) {
716 			struct mlx5_external_q *ext_rxq =
717 					mlx5_ext_rxq_get(dev, queues[i]);
718 
719 			rqt_attr->rq_list[i] = ext_rxq->hw_id;
720 		} else {
721 			struct mlx5_rxq_priv *rxq =
722 					mlx5_rxq_get(dev, queues[i]);
723 
724 			MLX5_ASSERT(rxq != NULL);
725 			if (rxq->ctrl->is_hairpin)
726 				rqt_attr->rq_list[i] = rxq->ctrl->obj->rq->id;
727 			else
728 				rqt_attr->rq_list[i] = rxq->devx_rq.rq->id;
729 		}
730 	}
731 	MLX5_ASSERT(i > 0);
732 	for (j = 0; i != rqt_n; ++j, ++i)
733 		rqt_attr->rq_list[i] = rqt_attr->rq_list[j];
734 	return rqt_attr;
735 }
736 
737 /**
738  * Create RQT using DevX API as a filed of indirection table.
739  *
740  * @param dev
741  *   Pointer to Ethernet device.
742  * @param log_n
743  *   Log of number of queues in the array.
744  * @param ind_tbl
745  *   DevX indirection table object.
746  *
747  * @return
748  *   0 on success, a negative errno value otherwise and rte_errno is set.
749  */
750 static int
751 mlx5_devx_ind_table_new(struct rte_eth_dev *dev, const unsigned int log_n,
752 			struct mlx5_ind_table_obj *ind_tbl)
753 {
754 	struct mlx5_priv *priv = dev->data->dev_private;
755 	struct mlx5_devx_rqt_attr *rqt_attr = NULL;
756 	const uint16_t *queues = dev->data->dev_started ? ind_tbl->queues :
757 							  NULL;
758 
759 	MLX5_ASSERT(ind_tbl);
760 	rqt_attr = mlx5_devx_ind_table_create_rqt_attr(dev, log_n, queues,
761 						       ind_tbl->queues_n);
762 	if (!rqt_attr)
763 		return -rte_errno;
764 	ind_tbl->rqt = mlx5_devx_cmd_create_rqt(priv->sh->cdev->ctx, rqt_attr);
765 	mlx5_free(rqt_attr);
766 	if (!ind_tbl->rqt) {
767 		DRV_LOG(ERR, "Port %u cannot create DevX RQT.",
768 			dev->data->port_id);
769 		rte_errno = errno;
770 		return -rte_errno;
771 	}
772 	return 0;
773 }
774 
775 /**
776  * Modify RQT using DevX API as a filed of indirection table.
777  *
778  * @param dev
779  *   Pointer to Ethernet device.
780  * @param log_n
781  *   Log of number of queues in the array.
782  * @param ind_tbl
783  *   DevX indirection table object.
784  *
785  * @return
786  *   0 on success, a negative errno value otherwise and rte_errno is set.
787  */
788 static int
789 mlx5_devx_ind_table_modify(struct rte_eth_dev *dev, const unsigned int log_n,
790 			   const uint16_t *queues, const uint32_t queues_n,
791 			   struct mlx5_ind_table_obj *ind_tbl)
792 {
793 	int ret = 0;
794 	struct mlx5_devx_rqt_attr *rqt_attr = NULL;
795 
796 	MLX5_ASSERT(ind_tbl);
797 	rqt_attr = mlx5_devx_ind_table_create_rqt_attr(dev, log_n,
798 							queues,
799 							queues_n);
800 	if (!rqt_attr)
801 		return -rte_errno;
802 	ret = mlx5_devx_cmd_modify_rqt(ind_tbl->rqt, rqt_attr);
803 	mlx5_free(rqt_attr);
804 	if (ret)
805 		DRV_LOG(ERR, "Port %u cannot modify DevX RQT.",
806 			dev->data->port_id);
807 	return ret;
808 }
809 
810 /**
811  * Destroy the DevX RQT object.
812  *
813  * @param ind_table
814  *   Indirection table to release.
815  */
816 static void
817 mlx5_devx_ind_table_destroy(struct mlx5_ind_table_obj *ind_tbl)
818 {
819 	claim_zero(mlx5_devx_cmd_destroy(ind_tbl->rqt));
820 }
821 
822 /**
823  * Set TIR attribute struct with relevant input values.
824  *
825  * @param[in] dev
826  *   Pointer to Ethernet device.
827  * @param[in] rss_key
828  *   RSS key for the Rx hash queue.
829  * @param[in] hash_fields
830  *   Verbs protocol hash field to make the RSS on.
831  * @param[in] ind_tbl
832  *   Indirection table for TIR. If table queues array is NULL,
833  *   a TIR for drop queue is assumed.
834  * @param[in] tunnel
835  *   Tunnel type.
836  * @param[out] tir_attr
837  *   Parameters structure for TIR creation/modification.
838  *
839  * @return
840  *   The Verbs/DevX object initialised index, 0 otherwise and rte_errno is set.
841  */
842 static void
843 mlx5_devx_tir_attr_set(struct rte_eth_dev *dev, const uint8_t *rss_key,
844 		       uint64_t hash_fields,
845 		       const struct mlx5_ind_table_obj *ind_tbl,
846 		       int tunnel, bool symmetric_hash_function,
847 		       struct mlx5_devx_tir_attr *tir_attr)
848 {
849 	struct mlx5_priv *priv = dev->data->dev_private;
850 	bool is_hairpin;
851 	bool lro = false;
852 	uint32_t i;
853 
854 	/* NULL queues designate drop queue. */
855 	if (ind_tbl->queues == NULL) {
856 		is_hairpin = priv->drop_queue.rxq->ctrl->is_hairpin;
857 	} else if (mlx5_is_external_rxq(dev, ind_tbl->queues[0])) {
858 		/* External RxQ supports neither Hairpin nor LRO. */
859 		is_hairpin = false;
860 	} else {
861 		is_hairpin = mlx5_rxq_is_hairpin(dev, ind_tbl->queues[0]);
862 		lro = true;
863 		/* Enable TIR LRO only if all the queues were configured for. */
864 		for (i = 0; i < ind_tbl->queues_n; ++i) {
865 			struct mlx5_rxq_data *rxq_i =
866 				mlx5_rxq_data_get(dev, ind_tbl->queues[i]);
867 
868 			if (rxq_i != NULL && !rxq_i->lro) {
869 				lro = false;
870 				break;
871 			}
872 		}
873 	}
874 	memset(tir_attr, 0, sizeof(*tir_attr));
875 	tir_attr->disp_type = MLX5_TIRC_DISP_TYPE_INDIRECT;
876 	tir_attr->rx_hash_fn = MLX5_RX_HASH_FN_TOEPLITZ;
877 	tir_attr->tunneled_offload_en = !!tunnel;
878 	tir_attr->rx_hash_symmetric = symmetric_hash_function;
879 	/* If needed, translate hash_fields bitmap to PRM format. */
880 	if (hash_fields) {
881 		struct mlx5_rx_hash_field_select *rx_hash_field_select =
882 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
883 			hash_fields & IBV_RX_HASH_INNER ?
884 				&tir_attr->rx_hash_field_selector_inner :
885 #endif
886 				&tir_attr->rx_hash_field_selector_outer;
887 		/* 1 bit: 0: IPv4, 1: IPv6. */
888 		rx_hash_field_select->l3_prot_type =
889 					!!(hash_fields & MLX5_IPV6_IBV_RX_HASH);
890 		/* 1 bit: 0: TCP, 1: UDP. */
891 		rx_hash_field_select->l4_prot_type =
892 					!!(hash_fields & MLX5_UDP_IBV_RX_HASH);
893 		/* Bitmask which sets which fields to use in RX Hash. */
894 		rx_hash_field_select->selected_fields =
895 			((!!(hash_fields & MLX5_L3_SRC_IBV_RX_HASH)) <<
896 			 MLX5_RX_HASH_FIELD_SELECT_SELECTED_FIELDS_SRC_IP) |
897 			(!!(hash_fields & MLX5_L3_DST_IBV_RX_HASH)) <<
898 			 MLX5_RX_HASH_FIELD_SELECT_SELECTED_FIELDS_DST_IP |
899 			(!!(hash_fields & MLX5_L4_SRC_IBV_RX_HASH)) <<
900 			 MLX5_RX_HASH_FIELD_SELECT_SELECTED_FIELDS_L4_SPORT |
901 			(!!(hash_fields & MLX5_L4_DST_IBV_RX_HASH)) <<
902 			 MLX5_RX_HASH_FIELD_SELECT_SELECTED_FIELDS_L4_DPORT |
903 			(!!(hash_fields & IBV_RX_HASH_IPSEC_SPI)) <<
904 			 MLX5_RX_HASH_FIELD_SELECT_SELECTED_FIELDS_IPSEC_SPI;
905 	}
906 	if (is_hairpin)
907 		tir_attr->transport_domain = priv->sh->td->id;
908 	else
909 		tir_attr->transport_domain = priv->sh->tdn;
910 	memcpy(tir_attr->rx_hash_toeplitz_key, rss_key, MLX5_RSS_HASH_KEY_LEN);
911 	tir_attr->indirect_table = ind_tbl->rqt->id;
912 	if (dev->data->dev_conf.lpbk_mode)
913 		tir_attr->self_lb_block = MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
914 	if (lro) {
915 		MLX5_ASSERT(priv->sh->config.lro_allowed);
916 		tir_attr->lro_timeout_period_usecs = priv->config.lro_timeout;
917 		tir_attr->lro_max_msg_sz =
918 			priv->max_lro_msg_size / MLX5_LRO_SEG_CHUNK_SIZE;
919 		tir_attr->lro_enable_mask =
920 				MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
921 				MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO;
922 	}
923 }
924 
925 /**
926  * Create an Rx Hash queue.
927  *
928  * @param dev
929  *   Pointer to Ethernet device.
930  * @param hrxq
931  *   Pointer to Rx Hash queue.
932  * @param tunnel
933  *   Tunnel type.
934  *
935  * @return
936  *   0 on success, a negative errno value otherwise and rte_errno is set.
937  */
938 static int
939 mlx5_devx_hrxq_new(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq,
940 		   int tunnel __rte_unused)
941 {
942 	struct mlx5_priv *priv = dev->data->dev_private;
943 	struct mlx5_devx_tir_attr tir_attr = {0};
944 	int err;
945 
946 	mlx5_devx_tir_attr_set(dev, hrxq->rss_key, hrxq->hash_fields,
947 			       hrxq->ind_table, tunnel, hrxq->symmetric_hash_function,
948 			       &tir_attr);
949 	hrxq->tir = mlx5_devx_cmd_create_tir(priv->sh->cdev->ctx, &tir_attr);
950 	if (!hrxq->tir) {
951 		DRV_LOG(ERR, "Port %u cannot create DevX TIR.",
952 			dev->data->port_id);
953 		rte_errno = errno;
954 		goto error;
955 	}
956 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
957 #ifdef HAVE_MLX5_HWS_SUPPORT
958 	if (hrxq->hws_flags) {
959 		hrxq->action = mlx5dr_action_create_dest_tir
960 			(priv->dr_ctx,
961 			 (struct mlx5dr_devx_obj *)hrxq->tir, hrxq->hws_flags, true);
962 		if (!hrxq->action)
963 			goto error;
964 		return 0;
965 	}
966 #endif
967 	if (mlx5_flow_os_create_flow_action_dest_devx_tir(hrxq->tir,
968 							  &hrxq->action)) {
969 		rte_errno = errno;
970 		goto error;
971 	}
972 #endif
973 	return 0;
974 error:
975 	err = rte_errno; /* Save rte_errno before cleanup. */
976 	if (hrxq->tir)
977 		claim_zero(mlx5_devx_cmd_destroy(hrxq->tir));
978 	rte_errno = err; /* Restore rte_errno. */
979 	return -rte_errno;
980 }
981 
982 /**
983  * Destroy a DevX TIR object.
984  *
985  * @param hrxq
986  *   Hash Rx queue to release its tir.
987  */
988 static void
989 mlx5_devx_tir_destroy(struct mlx5_hrxq *hrxq)
990 {
991 	claim_zero(mlx5_devx_cmd_destroy(hrxq->tir));
992 }
993 
994 /**
995  * Modify an Rx Hash queue configuration.
996  *
997  * @param dev
998  *   Pointer to Ethernet device.
999  * @param hrxq
1000  *   Hash Rx queue to modify.
1001  * @param rss_key
1002  *   RSS key for the Rx hash queue.
1003  * @param hash_fields
1004  *   Verbs protocol hash field to make the RSS on.
1005  * @param[in] ind_tbl
1006  *   Indirection table for TIR.
1007  *
1008  * @return
1009  *   0 on success, a negative errno value otherwise and rte_errno is set.
1010  */
1011 static int
1012 mlx5_devx_hrxq_modify(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq,
1013 		       const uint8_t *rss_key,
1014 		       uint64_t hash_fields,
1015 		       bool symmetric_hash_function,
1016 		       const struct mlx5_ind_table_obj *ind_tbl)
1017 {
1018 	struct mlx5_devx_modify_tir_attr modify_tir = {0};
1019 
1020 	/*
1021 	 * untested for modification fields:
1022 	 * - rx_hash_fn set hard-coded in hrxq_new(),
1023 	 * - lro_xxx not set after rxq setup
1024 	 */
1025 	if (ind_tbl != hrxq->ind_table)
1026 		modify_tir.modify_bitmask |=
1027 			MLX5_MODIFY_TIR_IN_MODIFY_BITMASK_INDIRECT_TABLE;
1028 	if (hash_fields != hrxq->hash_fields ||
1029 			symmetric_hash_function != hrxq->symmetric_hash_function ||
1030 			memcmp(hrxq->rss_key, rss_key, MLX5_RSS_HASH_KEY_LEN))
1031 		modify_tir.modify_bitmask |=
1032 			MLX5_MODIFY_TIR_IN_MODIFY_BITMASK_HASH;
1033 	mlx5_devx_tir_attr_set(dev, rss_key, hash_fields, ind_tbl,
1034 			       0, /* N/A - tunnel modification unsupported */
1035 			       symmetric_hash_function,
1036 			       &modify_tir.tir);
1037 	modify_tir.tirn = hrxq->tir->id;
1038 	if (mlx5_devx_cmd_modify_tir(hrxq->tir, &modify_tir)) {
1039 		DRV_LOG(ERR, "port %u cannot modify DevX TIR",
1040 			dev->data->port_id);
1041 		rte_errno = errno;
1042 		return -rte_errno;
1043 	}
1044 	return 0;
1045 }
1046 
1047 /**
1048  * Create a DevX drop Rx queue.
1049  *
1050  * @param dev
1051  *   Pointer to Ethernet device.
1052  *
1053  * @return
1054  *   0 on success, a negative errno value otherwise and rte_errno is set.
1055  */
1056 static int
1057 mlx5_rxq_devx_obj_drop_create(struct rte_eth_dev *dev)
1058 {
1059 	struct mlx5_priv *priv = dev->data->dev_private;
1060 	int socket_id = dev->device->numa_node;
1061 	struct mlx5_rxq_priv *rxq;
1062 	struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1063 	struct mlx5_rxq_obj *rxq_obj = NULL;
1064 	int ret;
1065 
1066 	/*
1067 	 * Initialize dummy control structures.
1068 	 * They are required to hold pointers for cleanup
1069 	 * and are only accessible via drop queue DevX objects.
1070 	 */
1071 	rxq = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*rxq), 0, socket_id);
1072 	if (rxq == NULL) {
1073 		DRV_LOG(ERR, "Port %u could not allocate drop queue private",
1074 			dev->data->port_id);
1075 		rte_errno = ENOMEM;
1076 		goto error;
1077 	}
1078 	rxq_ctrl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*rxq_ctrl),
1079 			       0, socket_id);
1080 	if (rxq_ctrl == NULL) {
1081 		DRV_LOG(ERR, "Port %u could not allocate drop queue control",
1082 			dev->data->port_id);
1083 		rte_errno = ENOMEM;
1084 		goto error;
1085 	}
1086 	rxq_obj = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*rxq_obj), 0, socket_id);
1087 	if (rxq_obj == NULL) {
1088 		DRV_LOG(ERR, "Port %u could not allocate drop queue object",
1089 			dev->data->port_id);
1090 		rte_errno = ENOMEM;
1091 		goto error;
1092 	}
1093 	/* set the CPU socket ID where the rxq_ctrl was allocated */
1094 	rxq_ctrl->socket = socket_id;
1095 	rxq_obj->rxq_ctrl = rxq_ctrl;
1096 	rxq_ctrl->is_hairpin = false;
1097 	rxq_ctrl->sh = priv->sh;
1098 	rxq_ctrl->obj = rxq_obj;
1099 	rxq->ctrl = rxq_ctrl;
1100 	rxq->priv = priv;
1101 	LIST_INSERT_HEAD(&rxq_ctrl->owners, rxq, owner_entry);
1102 	/* Create CQ using DevX API. */
1103 	ret = mlx5_rxq_create_devx_cq_resources(rxq);
1104 	if (ret != 0) {
1105 		DRV_LOG(ERR, "Port %u drop queue CQ creation failed.",
1106 			dev->data->port_id);
1107 		goto error;
1108 	}
1109 	rxq_ctrl->rxq.delay_drop = 0;
1110 	/* Create RQ using DevX API. */
1111 	ret = mlx5_rxq_create_devx_rq_resources(rxq);
1112 	if (ret != 0) {
1113 		DRV_LOG(ERR, "Port %u drop queue RQ creation failed.",
1114 			dev->data->port_id);
1115 		rte_errno = ENOMEM;
1116 		goto error;
1117 	}
1118 	/* Change queue state to ready. */
1119 	ret = mlx5_devx_modify_rq(rxq, MLX5_RXQ_MOD_RST2RDY);
1120 	if (ret != 0)
1121 		goto error;
1122 	/* Initialize drop queue. */
1123 	priv->drop_queue.rxq = rxq;
1124 	return 0;
1125 error:
1126 	ret = rte_errno; /* Save rte_errno before cleanup. */
1127 	if (rxq != NULL && rxq->devx_rq.rq != NULL)
1128 		mlx5_devx_rq_destroy(&rxq->devx_rq);
1129 	if (rxq_obj != NULL) {
1130 		if (rxq_obj->cq_obj.cq != NULL)
1131 			mlx5_devx_cq_destroy(&rxq_obj->cq_obj);
1132 		if (rxq_obj->devx_channel)
1133 			mlx5_os_devx_destroy_event_channel
1134 							(rxq_obj->devx_channel);
1135 		mlx5_free(rxq_obj);
1136 	}
1137 	if (rxq_ctrl != NULL)
1138 		mlx5_free(rxq_ctrl);
1139 	if (rxq != NULL)
1140 		mlx5_free(rxq);
1141 	rte_errno = ret; /* Restore rte_errno. */
1142 	return -rte_errno;
1143 }
1144 
1145 /**
1146  * Release drop Rx queue resources.
1147  *
1148  * @param dev
1149  *   Pointer to Ethernet device.
1150  */
1151 static void
1152 mlx5_rxq_devx_obj_drop_release(struct rte_eth_dev *dev)
1153 {
1154 	struct mlx5_priv *priv = dev->data->dev_private;
1155 	struct mlx5_rxq_priv *rxq = priv->drop_queue.rxq;
1156 	struct mlx5_rxq_ctrl *rxq_ctrl = rxq->ctrl;
1157 
1158 	mlx5_rxq_devx_obj_release(rxq);
1159 	mlx5_free(rxq_ctrl->obj);
1160 	mlx5_free(rxq_ctrl);
1161 	mlx5_free(rxq);
1162 	priv->drop_queue.rxq = NULL;
1163 }
1164 
1165 /**
1166  * Release a drop hash Rx queue.
1167  *
1168  * @param dev
1169  *   Pointer to Ethernet device.
1170  */
1171 static void
1172 mlx5_devx_drop_action_destroy(struct rte_eth_dev *dev)
1173 {
1174 	struct mlx5_priv *priv = dev->data->dev_private;
1175 	struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
1176 
1177 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1178 	if (hrxq->action != NULL)
1179 		mlx5_flow_os_destroy_flow_action(hrxq->action);
1180 #endif
1181 	if (hrxq->tir != NULL)
1182 		mlx5_devx_tir_destroy(hrxq);
1183 	if (hrxq->ind_table->ind_table != NULL)
1184 		mlx5_devx_ind_table_destroy(hrxq->ind_table);
1185 	if (priv->drop_queue.rxq->devx_rq.rq != NULL)
1186 		mlx5_rxq_devx_obj_drop_release(dev);
1187 }
1188 
1189 /**
1190  * Create a DevX drop action for Rx Hash queue.
1191  *
1192  * @param dev
1193  *   Pointer to Ethernet device.
1194  *
1195  * @return
1196  *   0 on success, a negative errno value otherwise and rte_errno is set.
1197  */
1198 static int
1199 mlx5_devx_drop_action_create(struct rte_eth_dev *dev)
1200 {
1201 	struct mlx5_priv *priv = dev->data->dev_private;
1202 	struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
1203 	int ret;
1204 
1205 	ret = mlx5_rxq_devx_obj_drop_create(dev);
1206 	if (ret != 0) {
1207 		DRV_LOG(ERR, "Cannot create drop RX queue");
1208 		return ret;
1209 	}
1210 	if (priv->sh->config.dv_flow_en == 2)
1211 		return 0;
1212 	/* hrxq->ind_table queues are NULL, drop RX queue ID will be used */
1213 	ret = mlx5_devx_ind_table_new(dev, 0, hrxq->ind_table);
1214 	if (ret != 0) {
1215 		DRV_LOG(ERR, "Cannot create drop hash RX queue indirection table");
1216 		goto error;
1217 	}
1218 	ret = mlx5_devx_hrxq_new(dev, hrxq, /* tunnel */ false);
1219 	if (ret != 0) {
1220 		DRV_LOG(ERR, "Cannot create drop hash RX queue");
1221 		goto error;
1222 	}
1223 	return 0;
1224 error:
1225 	mlx5_devx_drop_action_destroy(dev);
1226 	return ret;
1227 }
1228 
1229 /**
1230  * Select TXQ TIS number.
1231  *
1232  * @param dev
1233  *   Pointer to Ethernet device.
1234  * @param queue_idx
1235  *   Queue index in DPDK Tx queue array.
1236  *
1237  * @return
1238  *   > 0 on success, a negative errno value otherwise.
1239  */
1240 static uint32_t
1241 mlx5_get_txq_tis_num(struct rte_eth_dev *dev, uint16_t queue_idx)
1242 {
1243 	struct mlx5_priv *priv = dev->data->dev_private;
1244 	struct mlx5_txq_data *txq_data = (*priv->txqs)[queue_idx];
1245 	int tis_idx = 0;
1246 
1247 	if (priv->sh->bond.n_port) {
1248 		if (txq_data->tx_aggr_affinity) {
1249 			tis_idx = txq_data->tx_aggr_affinity;
1250 		} else if (priv->sh->lag.affinity_mode == MLX5_LAG_MODE_TIS) {
1251 			tis_idx = (priv->lag_affinity_idx + queue_idx) %
1252 				priv->sh->bond.n_port + 1;
1253 			DRV_LOG(INFO, "port %d txq %d gets affinity %d and maps to PF %d.",
1254 				dev->data->port_id, queue_idx, tis_idx,
1255 				priv->sh->lag.tx_remap_affinity[tis_idx - 1]);
1256 		}
1257 	}
1258 	MLX5_ASSERT(priv->sh->tis[tis_idx]);
1259 	return priv->sh->tis[tis_idx]->id;
1260 }
1261 
1262 /**
1263  * Create the Tx hairpin queue object.
1264  *
1265  * @param dev
1266  *   Pointer to Ethernet device.
1267  * @param idx
1268  *   Queue index in DPDK Tx queue array.
1269  *
1270  * @return
1271  *   0 on success, a negative errno value otherwise and rte_errno is set.
1272  */
1273 static int
1274 mlx5_txq_obj_hairpin_new(struct rte_eth_dev *dev, uint16_t idx)
1275 {
1276 	struct mlx5_priv *priv = dev->data->dev_private;
1277 	struct mlx5_hca_attr *hca_attr = &priv->sh->cdev->config.hca_attr;
1278 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
1279 	struct mlx5_txq_ctrl *txq_ctrl =
1280 		container_of(txq_data, struct mlx5_txq_ctrl, txq);
1281 	struct mlx5_devx_create_sq_attr dev_mem_attr = { 0 };
1282 	struct mlx5_devx_create_sq_attr host_mem_attr = { 0 };
1283 	struct mlx5_txq_obj *tmpl = txq_ctrl->obj;
1284 	void *umem_buf = NULL;
1285 	void *umem_obj = NULL;
1286 	uint32_t max_wq_data;
1287 
1288 	MLX5_ASSERT(txq_data);
1289 	MLX5_ASSERT(tmpl);
1290 	tmpl->txq_ctrl = txq_ctrl;
1291 	dev_mem_attr.hairpin = 1;
1292 	dev_mem_attr.tis_lst_sz = 1;
1293 	dev_mem_attr.tis_num = mlx5_get_txq_tis_num(dev, idx);
1294 	max_wq_data =
1295 		priv->sh->cdev->config.hca_attr.log_max_hairpin_wq_data_sz;
1296 	/* Jumbo frames > 9KB should be supported, and more packets. */
1297 	if (priv->config.log_hp_size != (uint32_t)MLX5_ARG_UNSET) {
1298 		if (priv->config.log_hp_size > max_wq_data) {
1299 			DRV_LOG(ERR, "Total data size %u power of 2 is "
1300 				"too large for hairpin.",
1301 				priv->config.log_hp_size);
1302 			rte_errno = ERANGE;
1303 			return -rte_errno;
1304 		}
1305 		dev_mem_attr.wq_attr.log_hairpin_data_sz = priv->config.log_hp_size;
1306 	} else {
1307 		dev_mem_attr.wq_attr.log_hairpin_data_sz =
1308 				(max_wq_data < MLX5_HAIRPIN_JUMBO_LOG_SIZE) ?
1309 				 max_wq_data : MLX5_HAIRPIN_JUMBO_LOG_SIZE;
1310 	}
1311 	/* Set the packets number to the maximum value for performance. */
1312 	dev_mem_attr.wq_attr.log_hairpin_num_packets =
1313 			dev_mem_attr.wq_attr.log_hairpin_data_sz -
1314 			MLX5_HAIRPIN_QUEUE_STRIDE;
1315 	dev_mem_attr.hairpin_wq_buffer_type = MLX5_SQC_HAIRPIN_WQ_BUFFER_TYPE_INTERNAL_BUFFER;
1316 	if (txq_ctrl->hairpin_conf.use_rte_memory) {
1317 		uint32_t umem_size;
1318 		uint32_t umem_dbrec;
1319 		size_t alignment = MLX5_WQE_BUF_ALIGNMENT;
1320 
1321 		if (alignment == (size_t)-1) {
1322 			DRV_LOG(ERR, "Failed to get WQE buf alignment.");
1323 			rte_errno = ENOMEM;
1324 			return -rte_errno;
1325 		}
1326 		/*
1327 		 * It is assumed that configuration is verified against capabilities
1328 		 * during queue setup.
1329 		 */
1330 		MLX5_ASSERT(hca_attr->hairpin_sq_wq_in_host_mem);
1331 		MLX5_ASSERT(hca_attr->hairpin_sq_wqe_bb_size > 0);
1332 		rte_memcpy(&host_mem_attr, &dev_mem_attr, sizeof(host_mem_attr));
1333 		umem_size = MLX5_WQE_SIZE *
1334 			RTE_BIT32(host_mem_attr.wq_attr.log_hairpin_num_packets);
1335 		umem_dbrec = RTE_ALIGN(umem_size, MLX5_DBR_SIZE);
1336 		umem_size += MLX5_DBR_SIZE;
1337 		umem_buf = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, umem_size,
1338 				       alignment, priv->sh->numa_node);
1339 		if (umem_buf == NULL && txq_ctrl->hairpin_conf.force_memory) {
1340 			DRV_LOG(ERR, "Failed to allocate memory for hairpin TX queue");
1341 			rte_errno = ENOMEM;
1342 			return -rte_errno;
1343 		} else if (umem_buf == NULL && !txq_ctrl->hairpin_conf.force_memory) {
1344 			DRV_LOG(WARNING, "Failed to allocate memory for hairpin TX queue."
1345 					 " Falling back to TX queue located on the device.");
1346 			goto create_sq_on_device;
1347 		}
1348 		umem_obj = mlx5_os_umem_reg(priv->sh->cdev->ctx,
1349 					    (void *)(uintptr_t)umem_buf,
1350 					    umem_size,
1351 					    IBV_ACCESS_LOCAL_WRITE);
1352 		if (umem_obj == NULL && txq_ctrl->hairpin_conf.force_memory) {
1353 			DRV_LOG(ERR, "Failed to register UMEM for hairpin TX queue");
1354 			mlx5_free(umem_buf);
1355 			return -rte_errno;
1356 		} else if (umem_obj == NULL && !txq_ctrl->hairpin_conf.force_memory) {
1357 			DRV_LOG(WARNING, "Failed to register UMEM for hairpin TX queue."
1358 					 " Falling back to TX queue located on the device.");
1359 			rte_errno = 0;
1360 			mlx5_free(umem_buf);
1361 			goto create_sq_on_device;
1362 		}
1363 		host_mem_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC;
1364 		host_mem_attr.wq_attr.wq_umem_valid = 1;
1365 		host_mem_attr.wq_attr.wq_umem_id = mlx5_os_get_umem_id(umem_obj);
1366 		host_mem_attr.wq_attr.wq_umem_offset = 0;
1367 		host_mem_attr.wq_attr.dbr_umem_valid = 1;
1368 		host_mem_attr.wq_attr.dbr_umem_id = host_mem_attr.wq_attr.wq_umem_id;
1369 		host_mem_attr.wq_attr.dbr_addr = umem_dbrec;
1370 		host_mem_attr.wq_attr.log_wq_stride = rte_log2_u32(MLX5_WQE_SIZE);
1371 		host_mem_attr.wq_attr.log_wq_sz =
1372 				host_mem_attr.wq_attr.log_hairpin_num_packets *
1373 				hca_attr->hairpin_sq_wqe_bb_size;
1374 		host_mem_attr.wq_attr.log_wq_pg_sz = MLX5_LOG_PAGE_SIZE;
1375 		host_mem_attr.hairpin_wq_buffer_type = MLX5_SQC_HAIRPIN_WQ_BUFFER_TYPE_HOST_MEMORY;
1376 		tmpl->sq = mlx5_devx_cmd_create_sq(priv->sh->cdev->ctx, &host_mem_attr);
1377 		if (!tmpl->sq && txq_ctrl->hairpin_conf.force_memory) {
1378 			DRV_LOG(ERR,
1379 				"Port %u tx hairpin queue %u can't create SQ object.",
1380 				dev->data->port_id, idx);
1381 			claim_zero(mlx5_os_umem_dereg(umem_obj));
1382 			mlx5_free(umem_buf);
1383 			return -rte_errno;
1384 		} else if (!tmpl->sq && !txq_ctrl->hairpin_conf.force_memory) {
1385 			DRV_LOG(WARNING,
1386 				"Port %u tx hairpin queue %u failed to allocate SQ object"
1387 				" using host memory. Falling back to TX queue located"
1388 				" on the device",
1389 				dev->data->port_id, idx);
1390 			rte_errno = 0;
1391 			claim_zero(mlx5_os_umem_dereg(umem_obj));
1392 			mlx5_free(umem_buf);
1393 			goto create_sq_on_device;
1394 		}
1395 		tmpl->umem_buf_wq_buffer = umem_buf;
1396 		tmpl->umem_obj_wq_buffer = umem_obj;
1397 		return 0;
1398 	}
1399 
1400 create_sq_on_device:
1401 	tmpl->sq = mlx5_devx_cmd_create_sq(priv->sh->cdev->ctx, &dev_mem_attr);
1402 	if (!tmpl->sq) {
1403 		DRV_LOG(ERR,
1404 			"Port %u tx hairpin queue %u can't create SQ object.",
1405 			dev->data->port_id, idx);
1406 		rte_errno = errno;
1407 		return -rte_errno;
1408 	}
1409 	return 0;
1410 }
1411 
1412 #if defined(HAVE_MLX5DV_DEVX_UAR_OFFSET) || !defined(HAVE_INFINIBAND_VERBS_H)
1413 /**
1414  * Destroy the Tx queue DevX object.
1415  *
1416  * @param txq_obj
1417  *   Txq object to destroy.
1418  */
1419 static void
1420 mlx5_txq_release_devx_resources(struct mlx5_txq_obj *txq_obj)
1421 {
1422 	mlx5_devx_sq_destroy(&txq_obj->sq_obj);
1423 	memset(&txq_obj->sq_obj, 0, sizeof(txq_obj->sq_obj));
1424 	mlx5_devx_cq_destroy(&txq_obj->cq_obj);
1425 	memset(&txq_obj->cq_obj, 0, sizeof(txq_obj->cq_obj));
1426 }
1427 
1428 /**
1429  * Create a SQ object and its resources using DevX.
1430  *
1431  * @param dev
1432  *   Pointer to Ethernet device.
1433  * @param idx
1434  *   Queue index in DPDK Tx queue array.
1435  * @param[in] log_desc_n
1436  *   Log of number of descriptors in queue.
1437  *
1438  * @return
1439  *   0 on success, a negative errno value otherwise and rte_errno is set.
1440  */
1441 static int
1442 mlx5_txq_create_devx_sq_resources(struct rte_eth_dev *dev, uint16_t idx,
1443 				  uint16_t log_desc_n)
1444 {
1445 	struct mlx5_priv *priv = dev->data->dev_private;
1446 	struct mlx5_common_device *cdev = priv->sh->cdev;
1447 	struct mlx5_uar *uar = &priv->sh->tx_uar;
1448 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
1449 	struct mlx5_txq_ctrl *txq_ctrl =
1450 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
1451 	struct mlx5_txq_obj *txq_obj = txq_ctrl->obj;
1452 	struct mlx5_devx_create_sq_attr sq_attr = {
1453 		.flush_in_error_en = 1,
1454 		.allow_multi_pkt_send_wqe = !!priv->config.mps,
1455 		.min_wqe_inline_mode = cdev->config.hca_attr.vport_inline_mode,
1456 		.allow_swp = !!priv->sh->dev_cap.swp,
1457 		.cqn = txq_obj->cq_obj.cq->id,
1458 		.tis_lst_sz = 1,
1459 		.wq_attr = (struct mlx5_devx_wq_attr){
1460 			.pd = cdev->pdn,
1461 			.uar_page = mlx5_os_get_devx_uar_page_id(uar->obj),
1462 		},
1463 		.ts_format =
1464 			mlx5_ts_format_conv(cdev->config.hca_attr.sq_ts_format),
1465 		.tis_num = mlx5_get_txq_tis_num(dev, idx),
1466 	};
1467 
1468 	/* Create Send Queue object with DevX. */
1469 	return mlx5_devx_sq_create(cdev->ctx, &txq_obj->sq_obj,
1470 				   log_desc_n, &sq_attr, priv->sh->numa_node);
1471 }
1472 #endif
1473 
1474 /**
1475  * Create the Tx queue DevX object.
1476  *
1477  * @param dev
1478  *   Pointer to Ethernet device.
1479  * @param idx
1480  *   Queue index in DPDK Tx queue array.
1481  *
1482  * @return
1483  *   0 on success, a negative errno value otherwise and rte_errno is set.
1484  */
1485 int
1486 mlx5_txq_devx_obj_new(struct rte_eth_dev *dev, uint16_t idx)
1487 {
1488 	struct mlx5_priv *priv = dev->data->dev_private;
1489 	struct mlx5_txq_data *txq_data = (*priv->txqs)[idx];
1490 	struct mlx5_txq_ctrl *txq_ctrl =
1491 			container_of(txq_data, struct mlx5_txq_ctrl, txq);
1492 
1493 	if (txq_ctrl->is_hairpin)
1494 		return mlx5_txq_obj_hairpin_new(dev, idx);
1495 #if !defined(HAVE_MLX5DV_DEVX_UAR_OFFSET) && defined(HAVE_INFINIBAND_VERBS_H)
1496 	DRV_LOG(ERR, "Port %u Tx queue %u cannot create with DevX, no UAR.",
1497 		     dev->data->port_id, idx);
1498 	rte_errno = ENOMEM;
1499 	return -rte_errno;
1500 #else
1501 	struct mlx5_proc_priv *ppriv = MLX5_PROC_PRIV(PORT_ID(priv));
1502 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1503 	struct mlx5_txq_obj *txq_obj = txq_ctrl->obj;
1504 	struct mlx5_devx_cq_attr cq_attr = {
1505 		.uar_page_id = mlx5_os_get_devx_uar_page_id(sh->tx_uar.obj),
1506 	};
1507 	uint32_t cqe_n, log_desc_n;
1508 	uint32_t wqe_n, wqe_size;
1509 	int ret = 0;
1510 
1511 	MLX5_ASSERT(txq_data);
1512 	MLX5_ASSERT(txq_obj);
1513 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1514 	MLX5_ASSERT(ppriv);
1515 	txq_obj->txq_ctrl = txq_ctrl;
1516 	txq_obj->dev = dev;
1517 	if (__rte_trace_point_fp_is_enabled() &&
1518 	    txq_data->offloads & RTE_ETH_TX_OFFLOAD_SEND_ON_TIMESTAMP)
1519 		cqe_n = UINT16_MAX / 2 - 1;
1520 	else
1521 		cqe_n = (1UL << txq_data->elts_n) / MLX5_TX_COMP_THRESH +
1522 			1 + MLX5_TX_COMP_THRESH_INLINE_DIV;
1523 	log_desc_n = log2above(cqe_n);
1524 	cqe_n = 1UL << log_desc_n;
1525 	if (cqe_n > UINT16_MAX) {
1526 		DRV_LOG(ERR, "Port %u Tx queue %u requests to many CQEs %u.",
1527 			dev->data->port_id, txq_data->idx, cqe_n);
1528 		rte_errno = EINVAL;
1529 		return 0;
1530 	}
1531 	/* Create completion queue object with DevX. */
1532 	ret = mlx5_devx_cq_create(sh->cdev->ctx, &txq_obj->cq_obj, log_desc_n,
1533 				  &cq_attr, priv->sh->numa_node);
1534 	if (ret) {
1535 		DRV_LOG(ERR, "Port %u Tx queue %u CQ creation failure.",
1536 			dev->data->port_id, idx);
1537 		goto error;
1538 	}
1539 	txq_data->cqe_n = log_desc_n;
1540 	txq_data->cqe_s = cqe_n;
1541 	txq_data->cqe_m = txq_data->cqe_s - 1;
1542 	txq_data->cqes = txq_obj->cq_obj.cqes;
1543 	txq_data->cq_ci = 0;
1544 	txq_data->cq_pi = 0;
1545 	txq_data->cq_db = txq_obj->cq_obj.db_rec;
1546 	*txq_data->cq_db = 0;
1547 	/*
1548 	 * Adjust the amount of WQEs depending on inline settings.
1549 	 * The number of descriptors should be enough to handle
1550 	 * the specified number of packets. If queue is being created
1551 	 * with Verbs the rdma-core does queue size adjustment
1552 	 * internally in the mlx5_calc_sq_size(), we do the same
1553 	 * for the queue being created with DevX at this point.
1554 	 */
1555 	wqe_size = txq_data->tso_en ?
1556 		   RTE_ALIGN(txq_ctrl->max_tso_header, MLX5_WSEG_SIZE) : 0;
1557 	wqe_size += sizeof(struct mlx5_wqe_cseg) +
1558 		    sizeof(struct mlx5_wqe_eseg) +
1559 		    sizeof(struct mlx5_wqe_dseg);
1560 	if (txq_data->inlen_send)
1561 		wqe_size = RTE_MAX(wqe_size, sizeof(struct mlx5_wqe_cseg) +
1562 					     sizeof(struct mlx5_wqe_eseg) +
1563 					     RTE_ALIGN(txq_data->inlen_send +
1564 						       sizeof(uint32_t),
1565 						       MLX5_WSEG_SIZE));
1566 	wqe_size = RTE_ALIGN(wqe_size, MLX5_WQE_SIZE) / MLX5_WQE_SIZE;
1567 	/* Create Send Queue object with DevX. */
1568 	wqe_n = RTE_MIN((1UL << txq_data->elts_n) * wqe_size,
1569 			(uint32_t)priv->sh->dev_cap.max_qp_wr);
1570 	log_desc_n = log2above(wqe_n);
1571 	ret = mlx5_txq_create_devx_sq_resources(dev, idx, log_desc_n);
1572 	if (ret) {
1573 		DRV_LOG(ERR, "Port %u Tx queue %u SQ creation failure.",
1574 			dev->data->port_id, idx);
1575 		rte_errno = errno;
1576 		goto error;
1577 	}
1578 	/* Create the Work Queue. */
1579 	txq_data->wqe_n = log_desc_n;
1580 	txq_data->wqe_s = 1 << txq_data->wqe_n;
1581 	txq_data->wqe_m = txq_data->wqe_s - 1;
1582 	txq_data->wqes = (struct mlx5_wqe *)(uintptr_t)txq_obj->sq_obj.wqes;
1583 	txq_data->wqes_end = txq_data->wqes + txq_data->wqe_s;
1584 	txq_data->wqe_ci = 0;
1585 	txq_data->wqe_pi = 0;
1586 	txq_data->wqe_comp = 0;
1587 	txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV;
1588 	txq_data->qp_db = &txq_obj->sq_obj.db_rec[MLX5_SND_DBR];
1589 	*txq_data->qp_db = 0;
1590 	txq_data->qp_num_8s = txq_obj->sq_obj.sq->id << 8;
1591 	txq_data->db_heu = sh->cdev->config.dbnc == MLX5_SQ_DB_HEURISTIC;
1592 	txq_data->db_nc = sh->tx_uar.dbnc;
1593 	txq_data->wait_on_time = !!(!sh->config.tx_pp &&
1594 				    sh->cdev->config.hca_attr.wait_on_time);
1595 	/* Change Send Queue state to Ready-to-Send. */
1596 	ret = mlx5_txq_devx_modify(txq_obj, MLX5_TXQ_MOD_RST2RDY, 0);
1597 	if (ret) {
1598 		rte_errno = errno;
1599 		DRV_LOG(ERR,
1600 			"Port %u Tx queue %u SQ state to SQC_STATE_RDY failed.",
1601 			dev->data->port_id, idx);
1602 		goto error;
1603 	}
1604 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1605 	/*
1606 	 * If using DevX need to query and store TIS transport domain value.
1607 	 * This is done once per port.
1608 	 * Will use this value on Rx, when creating matching TIR.
1609 	 */
1610 	if (!priv->sh->tdn)
1611 		priv->sh->tdn = priv->sh->td->id;
1612 #endif
1613 	txq_ctrl->uar_mmap_offset =
1614 			mlx5_os_get_devx_uar_mmap_offset(sh->tx_uar.obj);
1615 	ppriv->uar_table[txq_data->idx] = sh->tx_uar.bf_db;
1616 	dev->data->tx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STARTED;
1617 	return 0;
1618 error:
1619 	ret = rte_errno; /* Save rte_errno before cleanup. */
1620 	mlx5_txq_release_devx_resources(txq_obj);
1621 	rte_errno = ret; /* Restore rte_errno. */
1622 	return -rte_errno;
1623 #endif
1624 }
1625 
1626 /**
1627  * Release an Tx DevX queue object.
1628  *
1629  * @param txq_obj
1630  *   DevX Tx queue object.
1631  */
1632 void
1633 mlx5_txq_devx_obj_release(struct mlx5_txq_obj *txq_obj)
1634 {
1635 	MLX5_ASSERT(txq_obj);
1636 	if (txq_obj->txq_ctrl->is_hairpin) {
1637 		if (txq_obj->sq) {
1638 			claim_zero(mlx5_devx_cmd_destroy(txq_obj->sq));
1639 			txq_obj->sq = NULL;
1640 		}
1641 		if (txq_obj->tis)
1642 			claim_zero(mlx5_devx_cmd_destroy(txq_obj->tis));
1643 		if (txq_obj->umem_obj_wq_buffer) {
1644 			claim_zero(mlx5_os_umem_dereg(txq_obj->umem_obj_wq_buffer));
1645 			txq_obj->umem_obj_wq_buffer = NULL;
1646 		}
1647 		if (txq_obj->umem_buf_wq_buffer) {
1648 			mlx5_free(txq_obj->umem_buf_wq_buffer);
1649 			txq_obj->umem_buf_wq_buffer = NULL;
1650 		}
1651 #if defined(HAVE_MLX5DV_DEVX_UAR_OFFSET) || !defined(HAVE_INFINIBAND_VERBS_H)
1652 	} else {
1653 		mlx5_txq_release_devx_resources(txq_obj);
1654 #endif
1655 	}
1656 }
1657 
1658 struct mlx5_obj_ops devx_obj_ops = {
1659 	.rxq_obj_modify_vlan_strip = mlx5_rxq_obj_modify_rq_vlan_strip,
1660 	.rxq_obj_new = mlx5_rxq_devx_obj_new,
1661 	.rxq_event_get = mlx5_rx_devx_get_event,
1662 	.rxq_obj_modify = mlx5_devx_modify_rq,
1663 	.rxq_obj_release = mlx5_rxq_devx_obj_release,
1664 	.rxq_event_get_lwm = mlx5_rx_devx_get_event_lwm,
1665 	.ind_table_new = mlx5_devx_ind_table_new,
1666 	.ind_table_modify = mlx5_devx_ind_table_modify,
1667 	.ind_table_destroy = mlx5_devx_ind_table_destroy,
1668 	.hrxq_new = mlx5_devx_hrxq_new,
1669 	.hrxq_destroy = mlx5_devx_tir_destroy,
1670 	.hrxq_modify = mlx5_devx_hrxq_modify,
1671 	.drop_action_create = mlx5_devx_drop_action_create,
1672 	.drop_action_destroy = mlx5_devx_drop_action_destroy,
1673 	.txq_obj_new = mlx5_txq_devx_obj_new,
1674 	.txq_obj_modify = mlx5_txq_devx_modify,
1675 	.txq_obj_release = mlx5_txq_devx_obj_release,
1676 	.lb_dummy_queue_create = NULL,
1677 	.lb_dummy_queue_release = NULL,
1678 };
1679