1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
12 #include <sys/queue.h>
15 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
17 #pragma GCC diagnostic ignored "-Wpedantic"
19 #include <infiniband/verbs.h>
20 #include <infiniband/mlx5dv.h>
22 #pragma GCC diagnostic error "-Wpedantic"
26 #include <rte_malloc.h>
27 #include <rte_ethdev_driver.h>
28 #include <rte_common.h>
29 #include <rte_interrupts.h>
30 #include <rte_debug.h>
34 #include "mlx5_rxtx.h"
35 #include "mlx5_utils.h"
36 #include "mlx5_autoconf.h"
37 #include "mlx5_defs.h"
38 #include "mlx5_glue.h"
40 /* Default RSS hash key also used for ConnectX-3. */
41 uint8_t rss_hash_default_key[] = {
42 0x2c, 0xc6, 0x81, 0xd1,
43 0x5b, 0xdb, 0xf4, 0xf7,
44 0xfc, 0xa2, 0x83, 0x19,
45 0xdb, 0x1a, 0x3e, 0x94,
46 0x6b, 0x9e, 0x38, 0xd9,
47 0x2c, 0x9c, 0x03, 0xd1,
48 0xad, 0x99, 0x44, 0xa7,
49 0xd9, 0x56, 0x3d, 0x59,
50 0x06, 0x3c, 0x25, 0xf3,
51 0xfc, 0x1f, 0xdc, 0x2a,
54 /* Length of the default RSS hash key. */
55 const size_t rss_hash_default_key_len = sizeof(rss_hash_default_key);
58 * Allocate RX queue elements.
61 * Pointer to RX queue structure.
64 * 0 on success, a negative errno value otherwise and rte_errno is set.
67 rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
69 const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
70 unsigned int elts_n = 1 << rxq_ctrl->rxq.elts_n;
74 /* Iterate on segments. */
75 for (i = 0; (i != elts_n); ++i) {
78 buf = rte_pktmbuf_alloc(rxq_ctrl->rxq.mp);
80 DRV_LOG(ERR, "port %u empty mbuf pool",
81 PORT_ID(rxq_ctrl->priv));
85 /* Headroom is reserved by rte_pktmbuf_alloc(). */
86 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
87 /* Buffer is supposed to be empty. */
88 assert(rte_pktmbuf_data_len(buf) == 0);
89 assert(rte_pktmbuf_pkt_len(buf) == 0);
91 /* Only the first segment keeps headroom. */
94 PORT(buf) = rxq_ctrl->rxq.port_id;
95 DATA_LEN(buf) = rte_pktmbuf_tailroom(buf);
96 PKT_LEN(buf) = DATA_LEN(buf);
98 (*rxq_ctrl->rxq.elts)[i] = buf;
100 /* If Rx vector is activated. */
101 if (mlx5_rxq_check_vec_support(&rxq_ctrl->rxq) > 0) {
102 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
103 struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
106 /* Initialize default rearm_data for vPMD. */
107 mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
108 rte_mbuf_refcnt_set(mbuf_init, 1);
109 mbuf_init->nb_segs = 1;
110 mbuf_init->port = rxq->port_id;
112 * prevent compiler reordering:
113 * rearm_data covers previous fields.
115 rte_compiler_barrier();
116 rxq->mbuf_initializer =
117 *(uint64_t *)&mbuf_init->rearm_data;
118 /* Padding with a fake mbuf for vectorized Rx. */
119 for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
120 (*rxq->elts)[elts_n + j] = &rxq->fake_mbuf;
123 "port %u Rx queue %u allocated and configured %u segments"
125 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx, elts_n,
126 elts_n / (1 << rxq_ctrl->rxq.sges_n));
129 err = rte_errno; /* Save rte_errno before cleanup. */
131 for (i = 0; (i != elts_n); ++i) {
132 if ((*rxq_ctrl->rxq.elts)[i] != NULL)
133 rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
134 (*rxq_ctrl->rxq.elts)[i] = NULL;
136 DRV_LOG(DEBUG, "port %u Rx queue %u failed, freed everything",
137 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx);
138 rte_errno = err; /* Restore rte_errno. */
143 * Free RX queue elements.
146 * Pointer to RX queue structure.
149 rxq_free_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
151 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
152 const uint16_t q_n = (1 << rxq->elts_n);
153 const uint16_t q_mask = q_n - 1;
154 uint16_t used = q_n - (rxq->rq_ci - rxq->rq_pi);
157 DRV_LOG(DEBUG, "port %u Rx queue %u freeing WRs",
158 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx);
159 if (rxq->elts == NULL)
162 * Some mbuf in the Ring belongs to the application. They cannot be
165 if (mlx5_rxq_check_vec_support(rxq) > 0) {
166 for (i = 0; i < used; ++i)
167 (*rxq->elts)[(rxq->rq_ci + i) & q_mask] = NULL;
168 rxq->rq_pi = rxq->rq_ci;
170 for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
171 if ((*rxq->elts)[i] != NULL)
172 rte_pktmbuf_free_seg((*rxq->elts)[i]);
173 (*rxq->elts)[i] = NULL;
178 * Clean up a RX queue.
180 * Destroy objects, free allocated memory and reset the structure for reuse.
183 * Pointer to RX queue structure.
186 mlx5_rxq_cleanup(struct mlx5_rxq_ctrl *rxq_ctrl)
188 DRV_LOG(DEBUG, "port %u cleaning up Rx queue %u",
189 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx);
191 mlx5_rxq_ibv_release(rxq_ctrl->ibv);
192 memset(rxq_ctrl, 0, sizeof(*rxq_ctrl));
196 * Returns the per-queue supported offloads.
199 * Pointer to Ethernet device.
202 * Supported Rx offloads.
205 mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
207 struct priv *priv = dev->data->dev_private;
208 struct mlx5_dev_config *config = &priv->config;
209 uint64_t offloads = (DEV_RX_OFFLOAD_SCATTER |
210 DEV_RX_OFFLOAD_TIMESTAMP |
211 DEV_RX_OFFLOAD_JUMBO_FRAME);
213 if (config->hw_fcs_strip)
214 offloads |= DEV_RX_OFFLOAD_CRC_STRIP;
216 offloads |= (DEV_RX_OFFLOAD_IPV4_CKSUM |
217 DEV_RX_OFFLOAD_UDP_CKSUM |
218 DEV_RX_OFFLOAD_TCP_CKSUM);
219 if (config->hw_vlan_strip)
220 offloads |= DEV_RX_OFFLOAD_VLAN_STRIP;
226 * Returns the per-port supported offloads.
229 * Supported Rx offloads.
232 mlx5_get_rx_port_offloads(void)
234 uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
242 * Pointer to Ethernet device structure.
246 * Number of descriptors to configure in queue.
248 * NUMA socket on which memory must be allocated.
250 * Thresholds parameters.
252 * Memory pool for buffer allocations.
255 * 0 on success, a negative errno value otherwise and rte_errno is set.
258 mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
259 unsigned int socket, const struct rte_eth_rxconf *conf,
260 struct rte_mempool *mp)
262 struct priv *priv = dev->data->dev_private;
263 struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
264 struct mlx5_rxq_ctrl *rxq_ctrl =
265 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
267 if (!rte_is_power_of_2(desc)) {
268 desc = 1 << log2above(desc);
270 "port %u increased number of descriptors in Rx queue %u"
271 " to the next power of two (%d)",
272 dev->data->port_id, idx, desc);
274 DRV_LOG(DEBUG, "port %u configuring Rx queue %u for %u descriptors",
275 dev->data->port_id, idx, desc);
276 if (idx >= priv->rxqs_n) {
277 DRV_LOG(ERR, "port %u Rx queue index out of range (%u >= %u)",
278 dev->data->port_id, idx, priv->rxqs_n);
279 rte_errno = EOVERFLOW;
282 if (!mlx5_rxq_releasable(dev, idx)) {
283 DRV_LOG(ERR, "port %u unable to release queue index %u",
284 dev->data->port_id, idx);
288 mlx5_rxq_release(dev, idx);
289 rxq_ctrl = mlx5_rxq_new(dev, idx, desc, socket, conf, mp);
291 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
292 dev->data->port_id, idx);
296 DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
297 dev->data->port_id, idx);
298 (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
303 * DPDK callback to release a RX queue.
306 * Generic RX queue pointer.
309 mlx5_rx_queue_release(void *dpdk_rxq)
311 struct mlx5_rxq_data *rxq = (struct mlx5_rxq_data *)dpdk_rxq;
312 struct mlx5_rxq_ctrl *rxq_ctrl;
317 rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
318 priv = rxq_ctrl->priv;
319 if (!mlx5_rxq_releasable(ETH_DEV(priv), rxq_ctrl->rxq.stats.idx))
320 rte_panic("port %u Rx queue %u is still used by a flow and"
321 " cannot be removed\n",
322 PORT_ID(priv), rxq_ctrl->idx);
323 mlx5_rxq_release(ETH_DEV(priv), rxq_ctrl->rxq.stats.idx);
327 * Allocate queue vector and fill epoll fd list for Rx interrupts.
330 * Pointer to Ethernet device.
333 * 0 on success, a negative errno value otherwise and rte_errno is set.
336 mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
338 struct priv *priv = dev->data->dev_private;
340 unsigned int rxqs_n = priv->rxqs_n;
341 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
342 unsigned int count = 0;
343 struct rte_intr_handle *intr_handle = dev->intr_handle;
345 if (!dev->data->dev_conf.intr_conf.rxq)
347 mlx5_rx_intr_vec_disable(dev);
348 intr_handle->intr_vec = malloc(n * sizeof(intr_handle->intr_vec[0]));
349 if (intr_handle->intr_vec == NULL) {
351 "port %u failed to allocate memory for interrupt"
352 " vector, Rx interrupts will not be supported",
357 intr_handle->type = RTE_INTR_HANDLE_EXT;
358 for (i = 0; i != n; ++i) {
359 /* This rxq ibv must not be released in this function. */
360 struct mlx5_rxq_ibv *rxq_ibv = mlx5_rxq_ibv_get(dev, i);
365 /* Skip queues that cannot request interrupts. */
366 if (!rxq_ibv || !rxq_ibv->channel) {
367 /* Use invalid intr_vec[] index to disable entry. */
368 intr_handle->intr_vec[i] =
369 RTE_INTR_VEC_RXTX_OFFSET +
370 RTE_MAX_RXTX_INTR_VEC_ID;
373 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
375 "port %u too many Rx queues for interrupt"
376 " vector size (%d), Rx interrupts cannot be"
378 dev->data->port_id, RTE_MAX_RXTX_INTR_VEC_ID);
379 mlx5_rx_intr_vec_disable(dev);
383 fd = rxq_ibv->channel->fd;
384 flags = fcntl(fd, F_GETFL);
385 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
389 "port %u failed to make Rx interrupt file"
390 " descriptor %d non-blocking for queue index"
392 dev->data->port_id, fd, i);
393 mlx5_rx_intr_vec_disable(dev);
396 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
397 intr_handle->efds[count] = fd;
401 mlx5_rx_intr_vec_disable(dev);
403 intr_handle->nb_efd = count;
408 * Clean up Rx interrupts handler.
411 * Pointer to Ethernet device.
414 mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
416 struct priv *priv = dev->data->dev_private;
417 struct rte_intr_handle *intr_handle = dev->intr_handle;
419 unsigned int rxqs_n = priv->rxqs_n;
420 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
422 if (!dev->data->dev_conf.intr_conf.rxq)
424 if (!intr_handle->intr_vec)
426 for (i = 0; i != n; ++i) {
427 struct mlx5_rxq_ctrl *rxq_ctrl;
428 struct mlx5_rxq_data *rxq_data;
430 if (intr_handle->intr_vec[i] == RTE_INTR_VEC_RXTX_OFFSET +
431 RTE_MAX_RXTX_INTR_VEC_ID)
434 * Need to access directly the queue to release the reference
435 * kept in priv_rx_intr_vec_enable().
437 rxq_data = (*priv->rxqs)[i];
438 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
439 mlx5_rxq_ibv_release(rxq_ctrl->ibv);
442 rte_intr_free_epoll_fd(intr_handle);
443 if (intr_handle->intr_vec)
444 free(intr_handle->intr_vec);
445 intr_handle->nb_efd = 0;
446 intr_handle->intr_vec = NULL;
450 * MLX5 CQ notification .
453 * Pointer to receive queue structure.
455 * Sequence number per receive queue .
458 mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
461 uint32_t doorbell_hi;
463 void *cq_db_reg = (char *)rxq->cq_uar + MLX5_CQ_DOORBELL;
465 sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
466 doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
467 doorbell = (uint64_t)doorbell_hi << 32;
468 doorbell |= rxq->cqn;
469 rxq->cq_db[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
470 rte_write64(rte_cpu_to_be_64(doorbell), cq_db_reg);
474 * DPDK callback for Rx queue interrupt enable.
477 * Pointer to Ethernet device structure.
482 * 0 on success, a negative errno value otherwise and rte_errno is set.
485 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
487 struct priv *priv = dev->data->dev_private;
488 struct mlx5_rxq_data *rxq_data;
489 struct mlx5_rxq_ctrl *rxq_ctrl;
491 rxq_data = (*priv->rxqs)[rx_queue_id];
496 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
498 struct mlx5_rxq_ibv *rxq_ibv;
500 rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
505 mlx5_arm_cq(rxq_data, rxq_data->cq_arm_sn);
506 mlx5_rxq_ibv_release(rxq_ibv);
512 * DPDK callback for Rx queue interrupt disable.
515 * Pointer to Ethernet device structure.
520 * 0 on success, a negative errno value otherwise and rte_errno is set.
523 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
525 struct priv *priv = dev->data->dev_private;
526 struct mlx5_rxq_data *rxq_data;
527 struct mlx5_rxq_ctrl *rxq_ctrl;
528 struct mlx5_rxq_ibv *rxq_ibv = NULL;
529 struct ibv_cq *ev_cq;
533 rxq_data = (*priv->rxqs)[rx_queue_id];
538 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
541 rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
546 ret = mlx5_glue->get_cq_event(rxq_ibv->channel, &ev_cq, &ev_ctx);
547 if (ret || ev_cq != rxq_ibv->cq) {
551 rxq_data->cq_arm_sn++;
552 mlx5_glue->ack_cq_events(rxq_ibv->cq, 1);
555 ret = rte_errno; /* Save rte_errno before cleanup. */
557 mlx5_rxq_ibv_release(rxq_ibv);
558 DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
559 dev->data->port_id, rx_queue_id);
560 rte_errno = ret; /* Restore rte_errno. */
565 * Create the Rx queue Verbs object.
568 * Pointer to Ethernet device.
570 * Queue index in DPDK Rx queue array
573 * The Verbs object initialised, NULL otherwise and rte_errno is set.
575 struct mlx5_rxq_ibv *
576 mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx)
578 struct priv *priv = dev->data->dev_private;
579 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
580 struct mlx5_rxq_ctrl *rxq_ctrl =
581 container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
582 struct ibv_wq_attr mod;
585 struct ibv_cq_init_attr_ex ibv;
586 struct mlx5dv_cq_init_attr mlx5;
588 struct ibv_wq_init_attr wq;
589 struct ibv_cq_ex cq_attr;
591 unsigned int cqe_n = (1 << rxq_data->elts_n) - 1;
592 struct mlx5_rxq_ibv *tmpl;
593 struct mlx5dv_cq cq_info;
594 struct mlx5dv_rwq rwq;
597 struct mlx5dv_obj obj;
598 struct mlx5_dev_config *config = &priv->config;
601 assert(!rxq_ctrl->ibv);
602 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_RX_QUEUE;
603 priv->verbs_alloc_ctx.obj = rxq_ctrl;
604 tmpl = rte_calloc_socket(__func__, 1, sizeof(*tmpl), 0,
608 "port %u Rx queue %u cannot allocate verbs resources",
609 dev->data->port_id, rxq_ctrl->idx);
613 tmpl->rxq_ctrl = rxq_ctrl;
614 /* Use the entire RX mempool as the memory region. */
615 tmpl->mr = mlx5_mr_get(dev, rxq_data->mp);
617 tmpl->mr = mlx5_mr_new(dev, rxq_data->mp);
619 DRV_LOG(ERR, "port %u: memeroy region creation failure",
625 tmpl->channel = mlx5_glue->create_comp_channel(priv->ctx);
626 if (!tmpl->channel) {
627 DRV_LOG(ERR, "port %u: comp channel creation failure",
633 attr.cq.ibv = (struct ibv_cq_init_attr_ex){
635 .channel = tmpl->channel,
638 attr.cq.mlx5 = (struct mlx5dv_cq_init_attr){
641 if (config->cqe_comp && !rxq_data->hw_timestamp) {
642 attr.cq.mlx5.comp_mask |=
643 MLX5DV_CQ_INIT_ATTR_MASK_COMPRESSED_CQE;
644 attr.cq.mlx5.cqe_comp_res_format = MLX5DV_CQE_RES_FORMAT_HASH;
646 * For vectorized Rx, it must not be doubled in order to
647 * make cq_ci and rq_ci aligned.
649 if (mlx5_rxq_check_vec_support(rxq_data) < 0)
650 attr.cq.ibv.cqe *= 2;
651 } else if (config->cqe_comp && rxq_data->hw_timestamp) {
653 "port %u Rx CQE compression is disabled for HW"
657 tmpl->cq = mlx5_glue->cq_ex_to_cq
658 (mlx5_glue->dv_create_cq(priv->ctx, &attr.cq.ibv,
660 if (tmpl->cq == NULL) {
661 DRV_LOG(ERR, "port %u Rx queue %u CQ creation failure",
662 dev->data->port_id, idx);
666 DRV_LOG(DEBUG, "port %u priv->device_attr.max_qp_wr is %d",
667 dev->data->port_id, priv->device_attr.orig_attr.max_qp_wr);
668 DRV_LOG(DEBUG, "port %u priv->device_attr.max_sge is %d",
669 dev->data->port_id, priv->device_attr.orig_attr.max_sge);
670 attr.wq = (struct ibv_wq_init_attr){
671 .wq_context = NULL, /* Could be useful in the future. */
672 .wq_type = IBV_WQT_RQ,
673 /* Max number of outstanding WRs. */
674 .max_wr = (1 << rxq_data->elts_n) >> rxq_data->sges_n,
675 /* Max number of scatter/gather elements in a WR. */
676 .max_sge = 1 << rxq_data->sges_n,
680 IBV_WQ_FLAGS_CVLAN_STRIPPING |
682 .create_flags = (rxq_data->vlan_strip ?
683 IBV_WQ_FLAGS_CVLAN_STRIPPING :
686 /* By default, FCS (CRC) is stripped by hardware. */
687 if (rxq_data->crc_present) {
688 attr.wq.create_flags |= IBV_WQ_FLAGS_SCATTER_FCS;
689 attr.wq.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
691 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
692 if (config->hw_padding) {
693 attr.wq.create_flags |= IBV_WQ_FLAG_RX_END_PADDING;
694 attr.wq.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
697 tmpl->wq = mlx5_glue->create_wq(priv->ctx, &attr.wq);
698 if (tmpl->wq == NULL) {
699 DRV_LOG(ERR, "port %u Rx queue %u WQ creation failure",
700 dev->data->port_id, idx);
705 * Make sure number of WRs*SGEs match expectations since a queue
706 * cannot allocate more than "desc" buffers.
708 if (((int)attr.wq.max_wr !=
709 ((1 << rxq_data->elts_n) >> rxq_data->sges_n)) ||
710 ((int)attr.wq.max_sge != (1 << rxq_data->sges_n))) {
712 "port %u Rx queue %u requested %u*%u but got %u*%u"
714 dev->data->port_id, idx,
715 ((1 << rxq_data->elts_n) >> rxq_data->sges_n),
716 (1 << rxq_data->sges_n),
717 attr.wq.max_wr, attr.wq.max_sge);
721 /* Change queue state to ready. */
722 mod = (struct ibv_wq_attr){
723 .attr_mask = IBV_WQ_ATTR_STATE,
724 .wq_state = IBV_WQS_RDY,
726 ret = mlx5_glue->modify_wq(tmpl->wq, &mod);
729 "port %u Rx queue %u WQ state to IBV_WQS_RDY failed",
730 dev->data->port_id, idx);
734 obj.cq.in = tmpl->cq;
735 obj.cq.out = &cq_info;
736 obj.rwq.in = tmpl->wq;
738 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_RWQ);
743 if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
745 "port %u wrong MLX5_CQE_SIZE environment variable"
746 " value: it should be set to %u",
747 dev->data->port_id, RTE_CACHE_LINE_SIZE);
751 /* Fill the rings. */
752 rxq_data->wqes = (volatile struct mlx5_wqe_data_seg (*)[])
754 for (i = 0; (i != (unsigned int)(1 << rxq_data->elts_n)); ++i) {
755 struct rte_mbuf *buf = (*rxq_data->elts)[i];
756 volatile struct mlx5_wqe_data_seg *scat = &(*rxq_data->wqes)[i];
758 /* scat->addr must be able to store a pointer. */
759 assert(sizeof(scat->addr) >= sizeof(uintptr_t));
760 *scat = (struct mlx5_wqe_data_seg){
761 .addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf,
763 .byte_count = rte_cpu_to_be_32(DATA_LEN(buf)),
764 .lkey = tmpl->mr->lkey,
767 rxq_data->rq_db = rwq.dbrec;
768 rxq_data->cqe_n = log2above(cq_info.cqe_cnt);
772 rxq_data->zip = (struct rxq_zip){
775 rxq_data->cq_db = cq_info.dbrec;
776 rxq_data->cqes = (volatile struct mlx5_cqe (*)[])(uintptr_t)cq_info.buf;
777 rxq_data->cq_uar = cq_info.cq_uar;
778 rxq_data->cqn = cq_info.cqn;
779 rxq_data->cq_arm_sn = 0;
780 /* Update doorbell counter. */
781 rxq_data->rq_ci = (1 << rxq_data->elts_n) >> rxq_data->sges_n;
783 *rxq_data->rq_db = rte_cpu_to_be_32(rxq_data->rq_ci);
784 DRV_LOG(DEBUG, "port %u rxq %u updated with %p", dev->data->port_id,
786 rte_atomic32_inc(&tmpl->refcnt);
787 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u: refcnt %d",
788 dev->data->port_id, idx, rte_atomic32_read(&tmpl->refcnt));
789 LIST_INSERT_HEAD(&priv->rxqsibv, tmpl, next);
790 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
793 ret = rte_errno; /* Save rte_errno before cleanup. */
795 claim_zero(mlx5_glue->destroy_wq(tmpl->wq));
797 claim_zero(mlx5_glue->destroy_cq(tmpl->cq));
799 claim_zero(mlx5_glue->destroy_comp_channel(tmpl->channel));
801 mlx5_mr_release(tmpl->mr);
802 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
803 rte_errno = ret; /* Restore rte_errno. */
808 * Get an Rx queue Verbs object.
811 * Pointer to Ethernet device.
813 * Queue index in DPDK Rx queue array
816 * The Verbs object if it exists.
818 struct mlx5_rxq_ibv *
819 mlx5_rxq_ibv_get(struct rte_eth_dev *dev, uint16_t idx)
821 struct priv *priv = dev->data->dev_private;
822 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
823 struct mlx5_rxq_ctrl *rxq_ctrl;
825 if (idx >= priv->rxqs_n)
829 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
831 mlx5_mr_get(dev, rxq_data->mp);
832 rte_atomic32_inc(&rxq_ctrl->ibv->refcnt);
833 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u: refcnt %d",
834 dev->data->port_id, rxq_ctrl->idx,
835 rte_atomic32_read(&rxq_ctrl->ibv->refcnt));
837 return rxq_ctrl->ibv;
841 * Release an Rx verbs queue object.
844 * Verbs Rx queue object.
847 * 1 while a reference on it exists, 0 when freed.
850 mlx5_rxq_ibv_release(struct mlx5_rxq_ibv *rxq_ibv)
858 ret = mlx5_mr_release(rxq_ibv->mr);
861 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u: refcnt %d",
862 PORT_ID(rxq_ibv->rxq_ctrl->priv),
863 rxq_ibv->rxq_ctrl->idx, rte_atomic32_read(&rxq_ibv->refcnt));
864 if (rte_atomic32_dec_and_test(&rxq_ibv->refcnt)) {
865 rxq_free_elts(rxq_ibv->rxq_ctrl);
866 claim_zero(mlx5_glue->destroy_wq(rxq_ibv->wq));
867 claim_zero(mlx5_glue->destroy_cq(rxq_ibv->cq));
868 if (rxq_ibv->channel)
869 claim_zero(mlx5_glue->destroy_comp_channel
871 LIST_REMOVE(rxq_ibv, next);
879 * Verify the Verbs Rx queue list is empty
882 * Pointer to Ethernet device.
885 * The number of object not released.
888 mlx5_rxq_ibv_verify(struct rte_eth_dev *dev)
890 struct priv *priv = dev->data->dev_private;
892 struct mlx5_rxq_ibv *rxq_ibv;
894 LIST_FOREACH(rxq_ibv, &priv->rxqsibv, next) {
895 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u still referenced",
896 dev->data->port_id, rxq_ibv->rxq_ctrl->idx);
903 * Return true if a single reference exists on the object.
906 * Verbs Rx queue object.
909 mlx5_rxq_ibv_releasable(struct mlx5_rxq_ibv *rxq_ibv)
912 return (rte_atomic32_read(&rxq_ibv->refcnt) == 1);
916 * Create a DPDK Rx queue.
919 * Pointer to Ethernet device.
923 * Number of descriptors to configure in queue.
925 * NUMA socket on which memory must be allocated.
928 * A DPDK queue object on success, NULL otherwise and rte_errno is set.
930 struct mlx5_rxq_ctrl *
931 mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
932 unsigned int socket, const struct rte_eth_rxconf *conf,
933 struct rte_mempool *mp)
935 struct priv *priv = dev->data->dev_private;
936 struct mlx5_rxq_ctrl *tmpl;
937 unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
938 struct mlx5_dev_config *config = &priv->config;
940 * Always allocate extra slots, even if eventually
941 * the vector Rx will not be used.
943 const uint16_t desc_n =
944 desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
945 uint64_t offloads = conf->offloads |
946 dev->data->dev_conf.rxmode.offloads;
948 tmpl = rte_calloc_socket("RXQ", 1,
950 desc_n * sizeof(struct rte_mbuf *),
956 tmpl->socket = socket;
957 if (dev->data->dev_conf.intr_conf.rxq)
959 /* Enable scattered packets support for this queue if necessary. */
960 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
961 if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
962 (mb_len - RTE_PKTMBUF_HEADROOM)) {
963 tmpl->rxq.sges_n = 0;
964 } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
966 RTE_PKTMBUF_HEADROOM +
967 dev->data->dev_conf.rxmode.max_rx_pkt_len;
971 * Determine the number of SGEs needed for a full packet
972 * and round it to the next power of two.
974 sges_n = log2above((size / mb_len) + !!(size % mb_len));
975 tmpl->rxq.sges_n = sges_n;
976 /* Make sure rxq.sges_n did not overflow. */
977 size = mb_len * (1 << tmpl->rxq.sges_n);
978 size -= RTE_PKTMBUF_HEADROOM;
979 if (size < dev->data->dev_conf.rxmode.max_rx_pkt_len) {
981 "port %u too many SGEs (%u) needed to handle"
982 " requested maximum packet size %u",
985 dev->data->dev_conf.rxmode.max_rx_pkt_len);
986 rte_errno = EOVERFLOW;
991 "port %u the requested maximum Rx packet size (%u) is"
992 " larger than a single mbuf (%u) and scattered mode has"
993 " not been requested",
995 dev->data->dev_conf.rxmode.max_rx_pkt_len,
996 mb_len - RTE_PKTMBUF_HEADROOM);
998 DRV_LOG(DEBUG, "port %u maximum number of segments per packet: %u",
999 dev->data->port_id, 1 << tmpl->rxq.sges_n);
1000 if (desc % (1 << tmpl->rxq.sges_n)) {
1002 "port %u number of Rx queue descriptors (%u) is not a"
1003 " multiple of SGEs per packet (%u)",
1006 1 << tmpl->rxq.sges_n);
1010 /* Toggle RX checksum offload if hardware supports it. */
1011 tmpl->rxq.csum = !!(offloads & DEV_RX_OFFLOAD_CHECKSUM);
1012 tmpl->rxq.hw_timestamp = !!(offloads & DEV_RX_OFFLOAD_TIMESTAMP);
1013 /* Configure VLAN stripping. */
1014 tmpl->rxq.vlan_strip = !!(offloads & DEV_RX_OFFLOAD_VLAN_STRIP);
1015 /* By default, FCS (CRC) is stripped by hardware. */
1016 if (offloads & DEV_RX_OFFLOAD_CRC_STRIP) {
1017 tmpl->rxq.crc_present = 0;
1018 } else if (config->hw_fcs_strip) {
1019 tmpl->rxq.crc_present = 1;
1022 "port %u CRC stripping has been disabled but will"
1023 " still be performed by hardware, make sure MLNX_OFED"
1024 " and firmware are up to date",
1025 dev->data->port_id);
1026 tmpl->rxq.crc_present = 0;
1029 "port %u CRC stripping is %s, %u bytes will be subtracted from"
1030 " incoming frames to hide it",
1032 tmpl->rxq.crc_present ? "disabled" : "enabled",
1033 tmpl->rxq.crc_present << 2);
1035 tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
1036 (!!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS));
1037 tmpl->rxq.port_id = dev->data->port_id;
1040 tmpl->rxq.stats.idx = idx;
1041 tmpl->rxq.elts_n = log2above(desc);
1043 (struct rte_mbuf *(*)[1 << tmpl->rxq.elts_n])(tmpl + 1);
1045 rte_atomic32_inc(&tmpl->refcnt);
1046 DRV_LOG(DEBUG, "port %u Rx queue %u: refcnt %d", dev->data->port_id,
1047 idx, rte_atomic32_read(&tmpl->refcnt));
1048 LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1059 * Pointer to Ethernet device.
1064 * A pointer to the queue if it exists, NULL otherwise.
1066 struct mlx5_rxq_ctrl *
1067 mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
1069 struct priv *priv = dev->data->dev_private;
1070 struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1072 if ((*priv->rxqs)[idx]) {
1073 rxq_ctrl = container_of((*priv->rxqs)[idx],
1074 struct mlx5_rxq_ctrl,
1076 mlx5_rxq_ibv_get(dev, idx);
1077 rte_atomic32_inc(&rxq_ctrl->refcnt);
1078 DRV_LOG(DEBUG, "port %u Rx queue %u: refcnt %d",
1079 dev->data->port_id, rxq_ctrl->idx,
1080 rte_atomic32_read(&rxq_ctrl->refcnt));
1086 * Release a Rx queue.
1089 * Pointer to Ethernet device.
1094 * 1 while a reference on it exists, 0 when freed.
1097 mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
1099 struct priv *priv = dev->data->dev_private;
1100 struct mlx5_rxq_ctrl *rxq_ctrl;
1102 if (!(*priv->rxqs)[idx])
1104 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1105 assert(rxq_ctrl->priv);
1106 if (rxq_ctrl->ibv && !mlx5_rxq_ibv_release(rxq_ctrl->ibv))
1107 rxq_ctrl->ibv = NULL;
1108 DRV_LOG(DEBUG, "port %u Rx queue %u: refcnt %d", dev->data->port_id,
1109 rxq_ctrl->idx, rte_atomic32_read(&rxq_ctrl->refcnt));
1110 if (rte_atomic32_dec_and_test(&rxq_ctrl->refcnt)) {
1111 LIST_REMOVE(rxq_ctrl, next);
1113 (*priv->rxqs)[idx] = NULL;
1120 * Verify if the queue can be released.
1123 * Pointer to Ethernet device.
1128 * 1 if the queue can be released, negative errno otherwise and rte_errno is
1132 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
1134 struct priv *priv = dev->data->dev_private;
1135 struct mlx5_rxq_ctrl *rxq_ctrl;
1137 if (!(*priv->rxqs)[idx]) {
1141 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1142 return (rte_atomic32_read(&rxq_ctrl->refcnt) == 1);
1146 * Verify the Rx Queue list is empty
1149 * Pointer to Ethernet device.
1152 * The number of object not released.
1155 mlx5_rxq_verify(struct rte_eth_dev *dev)
1157 struct priv *priv = dev->data->dev_private;
1158 struct mlx5_rxq_ctrl *rxq_ctrl;
1161 LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
1162 DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
1163 dev->data->port_id, rxq_ctrl->idx);
1170 * Create an indirection table.
1173 * Pointer to Ethernet device.
1175 * Queues entering in the indirection table.
1177 * Number of queues in the array.
1180 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1182 struct mlx5_ind_table_ibv *
1183 mlx5_ind_table_ibv_new(struct rte_eth_dev *dev, const uint16_t *queues,
1186 struct priv *priv = dev->data->dev_private;
1187 struct mlx5_ind_table_ibv *ind_tbl;
1188 const unsigned int wq_n = rte_is_power_of_2(queues_n) ?
1189 log2above(queues_n) :
1190 log2above(priv->config.ind_table_max_size);
1191 struct ibv_wq *wq[1 << wq_n];
1195 ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl) +
1196 queues_n * sizeof(uint16_t), 0);
1201 for (i = 0; i != queues_n; ++i) {
1202 struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev, queues[i]);
1206 wq[i] = rxq->ibv->wq;
1207 ind_tbl->queues[i] = queues[i];
1209 ind_tbl->queues_n = queues_n;
1210 /* Finalise indirection table. */
1211 for (j = 0; i != (unsigned int)(1 << wq_n); ++i, ++j)
1213 ind_tbl->ind_table = mlx5_glue->create_rwq_ind_table
1215 &(struct ibv_rwq_ind_table_init_attr){
1216 .log_ind_tbl_size = wq_n,
1220 if (!ind_tbl->ind_table) {
1224 rte_atomic32_inc(&ind_tbl->refcnt);
1225 LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
1226 DEBUG("port %u new indirection table %p: queues:%u refcnt:%d",
1227 dev->data->port_id, (void *)ind_tbl, 1 << wq_n,
1228 rte_atomic32_read(&ind_tbl->refcnt));
1232 DRV_LOG(DEBUG, "port %u cannot create indirection table",
1233 dev->data->port_id);
1238 * Get an indirection table.
1241 * Pointer to Ethernet device.
1243 * Queues entering in the indirection table.
1245 * Number of queues in the array.
1248 * An indirection table if found.
1250 struct mlx5_ind_table_ibv *
1251 mlx5_ind_table_ibv_get(struct rte_eth_dev *dev, const uint16_t *queues,
1254 struct priv *priv = dev->data->dev_private;
1255 struct mlx5_ind_table_ibv *ind_tbl;
1257 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1258 if ((ind_tbl->queues_n == queues_n) &&
1259 (memcmp(ind_tbl->queues, queues,
1260 ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
1267 rte_atomic32_inc(&ind_tbl->refcnt);
1268 DRV_LOG(DEBUG, "port %u indirection table %p: refcnt %d",
1269 dev->data->port_id, (void *)ind_tbl,
1270 rte_atomic32_read(&ind_tbl->refcnt));
1271 for (i = 0; i != ind_tbl->queues_n; ++i)
1272 mlx5_rxq_get(dev, ind_tbl->queues[i]);
1278 * Release an indirection table.
1281 * Pointer to Ethernet device.
1283 * Indirection table to release.
1286 * 1 while a reference on it exists, 0 when freed.
1289 mlx5_ind_table_ibv_release(struct rte_eth_dev *dev,
1290 struct mlx5_ind_table_ibv *ind_tbl)
1294 DRV_LOG(DEBUG, "port %u indirection table %p: refcnt %d",
1295 dev->data->port_id, (void *)ind_tbl,
1296 rte_atomic32_read(&ind_tbl->refcnt));
1297 if (rte_atomic32_dec_and_test(&ind_tbl->refcnt)) {
1298 claim_zero(mlx5_glue->destroy_rwq_ind_table
1299 (ind_tbl->ind_table));
1300 DEBUG("port %u delete indirection table %p: queues: %u",
1301 dev->data->port_id, (void *)ind_tbl, ind_tbl->queues_n);
1303 for (i = 0; i != ind_tbl->queues_n; ++i)
1304 claim_nonzero(mlx5_rxq_release(dev, ind_tbl->queues[i]));
1305 if (!rte_atomic32_read(&ind_tbl->refcnt)) {
1306 LIST_REMOVE(ind_tbl, next);
1314 * Verify the Rx Queue list is empty
1317 * Pointer to Ethernet device.
1320 * The number of object not released.
1323 mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev)
1325 struct priv *priv = dev->data->dev_private;
1326 struct mlx5_ind_table_ibv *ind_tbl;
1329 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1331 "port %u Verbs indirection table %p still referenced",
1332 dev->data->port_id, (void *)ind_tbl);
1339 * Create an Rx Hash queue.
1342 * Pointer to Ethernet device.
1344 * RSS key for the Rx hash queue.
1345 * @param rss_key_len
1347 * @param hash_fields
1348 * Verbs protocol hash field to make the RSS on.
1350 * Queues entering in hash queue. In case of empty hash_fields only the
1351 * first queue index will be taken for the indirection table.
1355 * Tunnel type, implies tunnel offloading like inner checksum if available.
1357 * RSS hash on tunnel level.
1360 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1363 mlx5_hrxq_new(struct rte_eth_dev *dev,
1364 const uint8_t *rss_key, uint32_t rss_key_len,
1365 uint64_t hash_fields,
1366 const uint16_t *queues, uint32_t queues_n,
1367 uint32_t tunnel, uint32_t rss_level)
1369 struct priv *priv = dev->data->dev_private;
1370 struct mlx5_hrxq *hrxq;
1371 struct mlx5_ind_table_ibv *ind_tbl;
1374 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1375 struct mlx5dv_qp_init_attr qp_init_attr = {0};
1378 queues_n = hash_fields ? queues_n : 1;
1379 ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1381 ind_tbl = mlx5_ind_table_ibv_new(dev, queues, queues_n);
1387 rss_key_len = rss_hash_default_key_len;
1388 rss_key = rss_hash_default_key;
1390 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1392 qp_init_attr.comp_mask =
1393 MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
1394 qp_init_attr.create_flags = MLX5DV_QP_CREATE_TUNNEL_OFFLOADS;
1396 qp = mlx5_glue->dv_create_qp
1398 &(struct ibv_qp_init_attr_ex){
1399 .qp_type = IBV_QPT_RAW_PACKET,
1401 IBV_QP_INIT_ATTR_PD |
1402 IBV_QP_INIT_ATTR_IND_TABLE |
1403 IBV_QP_INIT_ATTR_RX_HASH,
1404 .rx_hash_conf = (struct ibv_rx_hash_conf){
1405 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1406 .rx_hash_key_len = rss_key_len ? rss_key_len :
1407 rss_hash_default_key_len,
1408 .rx_hash_key = rss_key ?
1409 (void *)(uintptr_t)rss_key :
1410 rss_hash_default_key,
1411 .rx_hash_fields_mask = hash_fields |
1412 (tunnel && rss_level > 1 ?
1413 (uint32_t)IBV_RX_HASH_INNER : 0),
1415 .rwq_ind_tbl = ind_tbl->ind_table,
1419 DEBUG("port %u new QP:%p ind_tbl:%p hash_fields:0x%" PRIx64
1420 " tunnel:0x%x level:%u dv_attr:comp_mask:0x%" PRIx64
1421 " create_flags:0x%x",
1422 dev->data->port_id, (void *)qp, (void *)ind_tbl,
1423 (tunnel && rss_level == 2 ? (uint32_t)IBV_RX_HASH_INNER : 0) |
1424 hash_fields, tunnel, rss_level,
1425 qp_init_attr.comp_mask, qp_init_attr.create_flags);
1427 qp = mlx5_glue->create_qp_ex
1429 &(struct ibv_qp_init_attr_ex){
1430 .qp_type = IBV_QPT_RAW_PACKET,
1432 IBV_QP_INIT_ATTR_PD |
1433 IBV_QP_INIT_ATTR_IND_TABLE |
1434 IBV_QP_INIT_ATTR_RX_HASH,
1435 .rx_hash_conf = (struct ibv_rx_hash_conf){
1436 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1437 .rx_hash_key_len = rss_key_len ? rss_key_len :
1438 rss_hash_default_key_len,
1439 .rx_hash_key = rss_key ?
1440 (void *)(uintptr_t)rss_key :
1441 rss_hash_default_key,
1442 .rx_hash_fields_mask = hash_fields,
1444 .rwq_ind_tbl = ind_tbl->ind_table,
1447 DEBUG("port %u new QP:%p ind_tbl:%p hash_fields:0x%" PRIx64
1448 " tunnel:0x%x level:%hhu",
1449 dev->data->port_id, (void *)qp, (void *)ind_tbl,
1450 hash_fields, tunnel, rss_level);
1456 hrxq = rte_calloc(__func__, 1, sizeof(*hrxq) + rss_key_len, 0);
1459 hrxq->ind_table = ind_tbl;
1461 hrxq->rss_key_len = rss_key_len;
1462 hrxq->hash_fields = hash_fields;
1463 hrxq->tunnel = tunnel;
1464 hrxq->rss_level = rss_level;
1465 memcpy(hrxq->rss_key, rss_key, rss_key_len);
1466 rte_atomic32_inc(&hrxq->refcnt);
1467 LIST_INSERT_HEAD(&priv->hrxqs, hrxq, next);
1468 DRV_LOG(DEBUG, "port %u hash Rx queue %p: refcnt %d",
1469 dev->data->port_id, (void *)hrxq,
1470 rte_atomic32_read(&hrxq->refcnt));
1473 err = rte_errno; /* Save rte_errno before cleanup. */
1474 mlx5_ind_table_ibv_release(dev, ind_tbl);
1476 claim_zero(mlx5_glue->destroy_qp(qp));
1477 rte_errno = err; /* Restore rte_errno. */
1482 * Get an Rx Hash queue.
1485 * Pointer to Ethernet device.
1487 * RSS configuration for the Rx hash queue.
1489 * Queues entering in hash queue. In case of empty hash_fields only the
1490 * first queue index will be taken for the indirection table.
1494 * Tunnel type, implies tunnel offloading like inner checksum if available.
1496 * RSS hash on tunnel level
1499 * An hash Rx queue on success.
1502 mlx5_hrxq_get(struct rte_eth_dev *dev,
1503 const uint8_t *rss_key, uint32_t rss_key_len,
1504 uint64_t hash_fields,
1505 const uint16_t *queues, uint32_t queues_n,
1506 uint32_t tunnel, uint32_t rss_level)
1508 struct priv *priv = dev->data->dev_private;
1509 struct mlx5_hrxq *hrxq;
1511 queues_n = hash_fields ? queues_n : 1;
1512 LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1513 struct mlx5_ind_table_ibv *ind_tbl;
1515 if (hrxq->rss_key_len != rss_key_len)
1517 if (memcmp(hrxq->rss_key, rss_key, rss_key_len))
1519 if (hrxq->hash_fields != hash_fields)
1521 if (hrxq->tunnel != tunnel)
1523 if (hrxq->rss_level != rss_level)
1525 ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1528 if (ind_tbl != hrxq->ind_table) {
1529 mlx5_ind_table_ibv_release(dev, ind_tbl);
1532 rte_atomic32_inc(&hrxq->refcnt);
1533 DRV_LOG(DEBUG, "port %u hash Rx queue %p: refcnt %d",
1534 dev->data->port_id, (void *)hrxq,
1535 rte_atomic32_read(&hrxq->refcnt));
1542 * Release the hash Rx queue.
1545 * Pointer to Ethernet device.
1547 * Pointer to Hash Rx queue to release.
1550 * 1 while a reference on it exists, 0 when freed.
1553 mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
1555 DRV_LOG(DEBUG, "port %u hash Rx queue %p: refcnt %d",
1556 dev->data->port_id, (void *)hrxq,
1557 rte_atomic32_read(&hrxq->refcnt));
1558 if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
1559 claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
1560 DEBUG("port %u delete QP %p: hash: 0x%" PRIx64 ", tunnel:"
1562 dev->data->port_id, (void *)hrxq, hrxq->hash_fields,
1563 hrxq->tunnel, hrxq->rss_level);
1564 mlx5_ind_table_ibv_release(dev, hrxq->ind_table);
1565 LIST_REMOVE(hrxq, next);
1569 claim_nonzero(mlx5_ind_table_ibv_release(dev, hrxq->ind_table));
1574 * Verify the Rx Queue list is empty
1577 * Pointer to Ethernet device.
1580 * The number of object not released.
1583 mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev)
1585 struct priv *priv = dev->data->dev_private;
1586 struct mlx5_hrxq *hrxq;
1589 LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1591 "port %u Verbs hash Rx queue %p still referenced",
1592 dev->data->port_id, (void *)hrxq);