1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
11 #include <sys/queue.h>
14 #include <rte_malloc.h>
15 #include <rte_ethdev_driver.h>
16 #include <rte_common.h>
17 #include <rte_interrupts.h>
18 #include <rte_debug.h>
20 #include <rte_eal_paging.h>
22 #include <mlx5_glue.h>
23 #include <mlx5_malloc.h>
25 #include "mlx5_defs.h"
27 #include "mlx5_rxtx.h"
28 #include "mlx5_utils.h"
29 #include "mlx5_autoconf.h"
32 /* Default RSS hash key also used for ConnectX-3. */
33 uint8_t rss_hash_default_key[] = {
34 0x2c, 0xc6, 0x81, 0xd1,
35 0x5b, 0xdb, 0xf4, 0xf7,
36 0xfc, 0xa2, 0x83, 0x19,
37 0xdb, 0x1a, 0x3e, 0x94,
38 0x6b, 0x9e, 0x38, 0xd9,
39 0x2c, 0x9c, 0x03, 0xd1,
40 0xad, 0x99, 0x44, 0xa7,
41 0xd9, 0x56, 0x3d, 0x59,
42 0x06, 0x3c, 0x25, 0xf3,
43 0xfc, 0x1f, 0xdc, 0x2a,
46 /* Length of the default RSS hash key. */
47 static_assert(MLX5_RSS_HASH_KEY_LEN ==
48 (unsigned int)sizeof(rss_hash_default_key),
49 "wrong RSS default key size.");
52 * Check whether Multi-Packet RQ can be enabled for the device.
55 * Pointer to Ethernet device.
58 * 1 if supported, negative errno value if not.
61 mlx5_check_mprq_support(struct rte_eth_dev *dev)
63 struct mlx5_priv *priv = dev->data->dev_private;
65 if (priv->config.mprq.enabled &&
66 priv->rxqs_n >= priv->config.mprq.min_rxqs_num)
72 * Check whether Multi-Packet RQ is enabled for the Rx queue.
75 * Pointer to receive queue structure.
78 * 0 if disabled, otherwise enabled.
81 mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq)
83 return rxq->strd_num_n > 0;
87 * Check whether Multi-Packet RQ is enabled for the device.
90 * Pointer to Ethernet device.
93 * 0 if disabled, otherwise enabled.
96 mlx5_mprq_enabled(struct rte_eth_dev *dev)
98 struct mlx5_priv *priv = dev->data->dev_private;
103 if (mlx5_check_mprq_support(dev) < 0)
105 /* All the configured queues should be enabled. */
106 for (i = 0; i < priv->rxqs_n; ++i) {
107 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
108 struct mlx5_rxq_ctrl *rxq_ctrl = container_of
109 (rxq, struct mlx5_rxq_ctrl, rxq);
111 if (rxq == NULL || rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
114 if (mlx5_rxq_mprq_enabled(rxq))
117 /* Multi-Packet RQ can't be partially configured. */
118 MLX5_ASSERT(n == 0 || n == n_ibv);
123 * Calculate the number of CQEs in CQ for the Rx queue.
126 * Pointer to receive queue structure.
129 * Number of CQEs in CQ.
132 mlx5_rxq_cqe_num(struct mlx5_rxq_data *rxq_data)
135 unsigned int wqe_n = 1 << rxq_data->elts_n;
137 if (mlx5_rxq_mprq_enabled(rxq_data))
138 cqe_n = wqe_n * (1 << rxq_data->strd_num_n) - 1;
145 * Allocate RX queue elements for Multi-Packet RQ.
148 * Pointer to RX queue structure.
151 * 0 on success, a negative errno value otherwise and rte_errno is set.
154 rxq_alloc_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
156 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
157 unsigned int wqe_n = 1 << rxq->elts_n;
161 /* Iterate on segments. */
162 for (i = 0; i <= wqe_n; ++i) {
163 struct mlx5_mprq_buf *buf;
165 if (rte_mempool_get(rxq->mprq_mp, (void **)&buf) < 0) {
166 DRV_LOG(ERR, "port %u empty mbuf pool", rxq->port_id);
171 (*rxq->mprq_bufs)[i] = buf;
173 rxq->mprq_repl = buf;
176 "port %u MPRQ queue %u allocated and configured %u segments",
177 rxq->port_id, rxq->idx, wqe_n);
180 err = rte_errno; /* Save rte_errno before cleanup. */
182 for (i = 0; (i != wqe_n); ++i) {
183 if ((*rxq->mprq_bufs)[i] != NULL)
184 rte_mempool_put(rxq->mprq_mp,
185 (*rxq->mprq_bufs)[i]);
186 (*rxq->mprq_bufs)[i] = NULL;
188 DRV_LOG(DEBUG, "port %u MPRQ queue %u failed, freed everything",
189 rxq->port_id, rxq->idx);
190 rte_errno = err; /* Restore rte_errno. */
195 * Allocate RX queue elements for Single-Packet RQ.
198 * Pointer to RX queue structure.
201 * 0 on success, errno value on failure.
204 rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
206 const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
207 unsigned int elts_n = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
208 (1 << rxq_ctrl->rxq.elts_n) * (1 << rxq_ctrl->rxq.strd_num_n) :
209 (1 << rxq_ctrl->rxq.elts_n);
213 /* Iterate on segments. */
214 for (i = 0; (i != elts_n); ++i) {
215 struct mlx5_eth_rxseg *seg = &rxq_ctrl->rxq.rxseg[i % sges_n];
216 struct rte_mbuf *buf;
218 buf = rte_pktmbuf_alloc(seg->mp);
220 DRV_LOG(ERR, "port %u empty mbuf pool",
221 PORT_ID(rxq_ctrl->priv));
225 /* Headroom is reserved by rte_pktmbuf_alloc(). */
226 MLX5_ASSERT(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
227 /* Buffer is supposed to be empty. */
228 MLX5_ASSERT(rte_pktmbuf_data_len(buf) == 0);
229 MLX5_ASSERT(rte_pktmbuf_pkt_len(buf) == 0);
230 MLX5_ASSERT(!buf->next);
231 SET_DATA_OFF(buf, seg->offset);
232 PORT(buf) = rxq_ctrl->rxq.port_id;
233 DATA_LEN(buf) = seg->length;
234 PKT_LEN(buf) = seg->length;
236 (*rxq_ctrl->rxq.elts)[i] = buf;
238 /* If Rx vector is activated. */
239 if (mlx5_rxq_check_vec_support(&rxq_ctrl->rxq) > 0) {
240 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
241 struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
242 struct rte_pktmbuf_pool_private *priv =
243 (struct rte_pktmbuf_pool_private *)
244 rte_mempool_get_priv(rxq_ctrl->rxq.mp);
247 /* Initialize default rearm_data for vPMD. */
248 mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
249 rte_mbuf_refcnt_set(mbuf_init, 1);
250 mbuf_init->nb_segs = 1;
251 mbuf_init->port = rxq->port_id;
252 if (priv->flags & RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF)
253 mbuf_init->ol_flags = EXT_ATTACHED_MBUF;
255 * prevent compiler reordering:
256 * rearm_data covers previous fields.
258 rte_compiler_barrier();
259 rxq->mbuf_initializer =
260 *(rte_xmm_t *)&mbuf_init->rearm_data;
261 /* Padding with a fake mbuf for vectorized Rx. */
262 for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
263 (*rxq->elts)[elts_n + j] = &rxq->fake_mbuf;
266 "port %u SPRQ queue %u allocated and configured %u segments"
268 PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx, elts_n,
269 elts_n / (1 << rxq_ctrl->rxq.sges_n));
272 err = rte_errno; /* Save rte_errno before cleanup. */
274 for (i = 0; (i != elts_n); ++i) {
275 if ((*rxq_ctrl->rxq.elts)[i] != NULL)
276 rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
277 (*rxq_ctrl->rxq.elts)[i] = NULL;
279 DRV_LOG(DEBUG, "port %u SPRQ queue %u failed, freed everything",
280 PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx);
281 rte_errno = err; /* Restore rte_errno. */
286 * Allocate RX queue elements.
289 * Pointer to RX queue structure.
292 * 0 on success, errno value on failure.
295 rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
300 * For MPRQ we need to allocate both MPRQ buffers
301 * for WQEs and simple mbufs for vector processing.
303 if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
304 ret = rxq_alloc_elts_mprq(rxq_ctrl);
305 return (ret || rxq_alloc_elts_sprq(rxq_ctrl));
309 * Free RX queue elements for Multi-Packet RQ.
312 * Pointer to RX queue structure.
315 rxq_free_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
317 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
320 DRV_LOG(DEBUG, "port %u Multi-Packet Rx queue %u freeing %d WRs",
321 rxq->port_id, rxq->idx, (1u << rxq->elts_n));
322 if (rxq->mprq_bufs == NULL)
324 for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
325 if ((*rxq->mprq_bufs)[i] != NULL)
326 mlx5_mprq_buf_free((*rxq->mprq_bufs)[i]);
327 (*rxq->mprq_bufs)[i] = NULL;
329 if (rxq->mprq_repl != NULL) {
330 mlx5_mprq_buf_free(rxq->mprq_repl);
331 rxq->mprq_repl = NULL;
336 * Free RX queue elements for Single-Packet RQ.
339 * Pointer to RX queue structure.
342 rxq_free_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
344 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
345 const uint16_t q_n = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
346 (1 << rxq->elts_n) * (1 << rxq->strd_num_n) :
348 const uint16_t q_mask = q_n - 1;
349 uint16_t used = q_n - (rxq->rq_ci - rxq->rq_pi);
352 DRV_LOG(DEBUG, "port %u Rx queue %u freeing %d WRs",
353 PORT_ID(rxq_ctrl->priv), rxq->idx, q_n);
354 if (rxq->elts == NULL)
357 * Some mbuf in the Ring belongs to the application.
358 * They cannot be freed.
360 if (mlx5_rxq_check_vec_support(rxq) > 0) {
361 for (i = 0; i < used; ++i)
362 (*rxq->elts)[(rxq->rq_ci + i) & q_mask] = NULL;
363 rxq->rq_pi = rxq->rq_ci;
365 for (i = 0; i != q_n; ++i) {
366 if ((*rxq->elts)[i] != NULL)
367 rte_pktmbuf_free_seg((*rxq->elts)[i]);
368 (*rxq->elts)[i] = NULL;
373 * Free RX queue elements.
376 * Pointer to RX queue structure.
379 rxq_free_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
382 * For MPRQ we need to allocate both MPRQ buffers
383 * for WQEs and simple mbufs for vector processing.
385 if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
386 rxq_free_elts_mprq(rxq_ctrl);
387 rxq_free_elts_sprq(rxq_ctrl);
391 * Returns the per-queue supported offloads.
394 * Pointer to Ethernet device.
397 * Supported Rx offloads.
400 mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
402 struct mlx5_priv *priv = dev->data->dev_private;
403 struct mlx5_dev_config *config = &priv->config;
404 uint64_t offloads = (DEV_RX_OFFLOAD_SCATTER |
405 RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT |
406 DEV_RX_OFFLOAD_TIMESTAMP |
407 DEV_RX_OFFLOAD_JUMBO_FRAME |
408 DEV_RX_OFFLOAD_RSS_HASH);
410 if (config->hw_fcs_strip)
411 offloads |= DEV_RX_OFFLOAD_KEEP_CRC;
414 offloads |= (DEV_RX_OFFLOAD_IPV4_CKSUM |
415 DEV_RX_OFFLOAD_UDP_CKSUM |
416 DEV_RX_OFFLOAD_TCP_CKSUM);
417 if (config->hw_vlan_strip)
418 offloads |= DEV_RX_OFFLOAD_VLAN_STRIP;
419 if (MLX5_LRO_SUPPORTED(dev))
420 offloads |= DEV_RX_OFFLOAD_TCP_LRO;
426 * Returns the per-port supported offloads.
429 * Supported Rx offloads.
432 mlx5_get_rx_port_offloads(void)
434 uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
440 * Verify if the queue can be released.
443 * Pointer to Ethernet device.
448 * 1 if the queue can be released
449 * 0 if the queue can not be released, there are references to it.
450 * Negative errno and rte_errno is set if queue doesn't exist.
453 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
455 struct mlx5_priv *priv = dev->data->dev_private;
456 struct mlx5_rxq_ctrl *rxq_ctrl;
458 if (!(*priv->rxqs)[idx]) {
462 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
463 return (__atomic_load_n(&rxq_ctrl->refcnt, __ATOMIC_RELAXED) == 1);
466 /* Fetches and drops all SW-owned and error CQEs to synchronize CQ. */
468 rxq_sync_cq(struct mlx5_rxq_data *rxq)
470 const uint16_t cqe_n = 1 << rxq->cqe_n;
471 const uint16_t cqe_mask = cqe_n - 1;
472 volatile struct mlx5_cqe *cqe;
477 cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask];
478 ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
479 if (ret == MLX5_CQE_STATUS_HW_OWN)
481 if (ret == MLX5_CQE_STATUS_ERR) {
485 MLX5_ASSERT(ret == MLX5_CQE_STATUS_SW_OWN);
486 if (MLX5_CQE_FORMAT(cqe->op_own) != MLX5_COMPRESSED) {
490 /* Compute the next non compressed CQE. */
491 rxq->cq_ci += rte_be_to_cpu_32(cqe->byte_cnt);
494 /* Move all CQEs to HW ownership, including possible MiniCQEs. */
495 for (i = 0; i < cqe_n; i++) {
496 cqe = &(*rxq->cqes)[i];
497 cqe->op_own = MLX5_CQE_INVALIDATE;
499 /* Resync CQE and WQE (WQ in RESET state). */
501 *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
503 *rxq->rq_db = rte_cpu_to_be_32(0);
508 * Rx queue stop. Device queue goes to the RESET state,
509 * all involved mbufs are freed from WQ.
512 * Pointer to Ethernet device structure.
517 * 0 on success, a negative errno value otherwise and rte_errno is set.
520 mlx5_rx_queue_stop_primary(struct rte_eth_dev *dev, uint16_t idx)
522 struct mlx5_priv *priv = dev->data->dev_private;
523 struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
524 struct mlx5_rxq_ctrl *rxq_ctrl =
525 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
528 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
529 ret = priv->obj_ops.rxq_obj_modify(rxq_ctrl->obj, MLX5_RXQ_MOD_RDY2RST);
531 DRV_LOG(ERR, "Cannot change Rx WQ state to RESET: %s",
536 /* Remove all processes CQEs. */
538 /* Free all involved mbufs. */
539 rxq_free_elts(rxq_ctrl);
540 /* Set the actual queue state. */
541 dev->data->rx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STOPPED;
546 * Rx queue stop. Device queue goes to the RESET state,
547 * all involved mbufs are freed from WQ.
550 * Pointer to Ethernet device structure.
555 * 0 on success, a negative errno value otherwise and rte_errno is set.
558 mlx5_rx_queue_stop(struct rte_eth_dev *dev, uint16_t idx)
560 eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
563 if (rte_eth_dev_is_rx_hairpin_queue(dev, idx)) {
564 DRV_LOG(ERR, "Hairpin queue can't be stopped");
568 if (dev->data->rx_queue_state[idx] == RTE_ETH_QUEUE_STATE_STOPPED)
571 * Vectorized Rx burst requires the CQ and RQ indices
572 * synchronized, that might be broken on RQ restart
573 * and cause Rx malfunction, so queue stopping is
574 * not supported if vectorized Rx burst is engaged.
575 * The routine pointer depends on the process
576 * type, should perform check there.
578 if (pkt_burst == mlx5_rx_burst_vec) {
579 DRV_LOG(ERR, "Rx queue stop is not supported "
580 "for vectorized Rx");
584 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
585 ret = mlx5_mp_os_req_queue_control(dev, idx,
586 MLX5_MP_REQ_QUEUE_RX_STOP);
588 ret = mlx5_rx_queue_stop_primary(dev, idx);
594 * Rx queue start. Device queue goes to the ready state,
595 * all required mbufs are allocated and WQ is replenished.
598 * Pointer to Ethernet device structure.
603 * 0 on success, a negative errno value otherwise and rte_errno is set.
606 mlx5_rx_queue_start_primary(struct rte_eth_dev *dev, uint16_t idx)
608 struct mlx5_priv *priv = dev->data->dev_private;
609 struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
610 struct mlx5_rxq_ctrl *rxq_ctrl =
611 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
614 MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
615 /* Allocate needed buffers. */
616 ret = rxq_alloc_elts(rxq_ctrl);
618 DRV_LOG(ERR, "Cannot reallocate buffers for Rx WQ");
623 *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
625 /* Reset RQ consumer before moving queue ro READY state. */
626 *rxq->rq_db = rte_cpu_to_be_32(0);
628 ret = priv->obj_ops.rxq_obj_modify(rxq_ctrl->obj, MLX5_RXQ_MOD_RST2RDY);
630 DRV_LOG(ERR, "Cannot change Rx WQ state to READY: %s",
635 /* Reinitialize RQ - set WQEs. */
636 mlx5_rxq_initialize(rxq);
637 rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
638 /* Set actual queue state. */
639 dev->data->rx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STARTED;
644 * Rx queue start. Device queue goes to the ready state,
645 * all required mbufs are allocated and WQ is replenished.
648 * Pointer to Ethernet device structure.
653 * 0 on success, a negative errno value otherwise and rte_errno is set.
656 mlx5_rx_queue_start(struct rte_eth_dev *dev, uint16_t idx)
660 if (rte_eth_dev_is_rx_hairpin_queue(dev, idx)) {
661 DRV_LOG(ERR, "Hairpin queue can't be started");
665 if (dev->data->rx_queue_state[idx] == RTE_ETH_QUEUE_STATE_STARTED)
667 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
668 ret = mlx5_mp_os_req_queue_control(dev, idx,
669 MLX5_MP_REQ_QUEUE_RX_START);
671 ret = mlx5_rx_queue_start_primary(dev, idx);
677 * Rx queue presetup checks.
680 * Pointer to Ethernet device structure.
684 * Number of descriptors to configure in queue.
687 * 0 on success, a negative errno value otherwise and rte_errno is set.
690 mlx5_rx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t *desc)
692 struct mlx5_priv *priv = dev->data->dev_private;
694 if (!rte_is_power_of_2(*desc)) {
695 *desc = 1 << log2above(*desc);
697 "port %u increased number of descriptors in Rx queue %u"
698 " to the next power of two (%d)",
699 dev->data->port_id, idx, *desc);
701 DRV_LOG(DEBUG, "port %u configuring Rx queue %u for %u descriptors",
702 dev->data->port_id, idx, *desc);
703 if (idx >= priv->rxqs_n) {
704 DRV_LOG(ERR, "port %u Rx queue index out of range (%u >= %u)",
705 dev->data->port_id, idx, priv->rxqs_n);
706 rte_errno = EOVERFLOW;
709 if (!mlx5_rxq_releasable(dev, idx)) {
710 DRV_LOG(ERR, "port %u unable to release queue index %u",
711 dev->data->port_id, idx);
715 mlx5_rxq_release(dev, idx);
722 * Pointer to Ethernet device structure.
726 * Number of descriptors to configure in queue.
728 * NUMA socket on which memory must be allocated.
730 * Thresholds parameters.
732 * Memory pool for buffer allocations.
735 * 0 on success, a negative errno value otherwise and rte_errno is set.
738 mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
739 unsigned int socket, const struct rte_eth_rxconf *conf,
740 struct rte_mempool *mp)
742 struct mlx5_priv *priv = dev->data->dev_private;
743 struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
744 struct mlx5_rxq_ctrl *rxq_ctrl =
745 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
746 struct rte_eth_rxseg_split *rx_seg =
747 (struct rte_eth_rxseg_split *)conf->rx_seg;
748 struct rte_eth_rxseg_split rx_single = {.mp = mp};
749 uint16_t n_seg = conf->rx_nseg;
754 * The parameters should be checked on rte_eth_dev layer.
755 * If mp is specified it means the compatible configuration
756 * without buffer split feature tuning.
762 uint64_t offloads = conf->offloads |
763 dev->data->dev_conf.rxmode.offloads;
765 /* The offloads should be checked on rte_eth_dev layer. */
766 MLX5_ASSERT(offloads & DEV_RX_OFFLOAD_SCATTER);
767 if (!(offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT)) {
768 DRV_LOG(ERR, "port %u queue index %u split "
769 "offload not configured",
770 dev->data->port_id, idx);
774 MLX5_ASSERT(n_seg < MLX5_MAX_RXQ_NSEG);
776 res = mlx5_rx_queue_pre_setup(dev, idx, &desc);
779 rxq_ctrl = mlx5_rxq_new(dev, idx, desc, socket, conf, rx_seg, n_seg);
781 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
782 dev->data->port_id, idx);
786 DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
787 dev->data->port_id, idx);
788 (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
795 * Pointer to Ethernet device structure.
799 * Number of descriptors to configure in queue.
800 * @param hairpin_conf
801 * Hairpin configuration parameters.
804 * 0 on success, a negative errno value otherwise and rte_errno is set.
807 mlx5_rx_hairpin_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
809 const struct rte_eth_hairpin_conf *hairpin_conf)
811 struct mlx5_priv *priv = dev->data->dev_private;
812 struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
813 struct mlx5_rxq_ctrl *rxq_ctrl =
814 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
817 res = mlx5_rx_queue_pre_setup(dev, idx, &desc);
820 if (hairpin_conf->peer_count != 1) {
822 DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue index %u"
823 " peer count is %u", dev->data->port_id,
824 idx, hairpin_conf->peer_count);
827 if (hairpin_conf->peers[0].port == dev->data->port_id) {
828 if (hairpin_conf->peers[0].queue >= priv->txqs_n) {
830 DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue"
831 " index %u, Tx %u is larger than %u",
832 dev->data->port_id, idx,
833 hairpin_conf->peers[0].queue, priv->txqs_n);
837 if (hairpin_conf->manual_bind == 0 ||
838 hairpin_conf->tx_explicit == 0) {
840 DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue"
841 " index %u peer port %u with attributes %u %u",
842 dev->data->port_id, idx,
843 hairpin_conf->peers[0].port,
844 hairpin_conf->manual_bind,
845 hairpin_conf->tx_explicit);
849 rxq_ctrl = mlx5_rxq_hairpin_new(dev, idx, desc, hairpin_conf);
851 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
852 dev->data->port_id, idx);
856 DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
857 dev->data->port_id, idx);
858 (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
863 * DPDK callback to release a RX queue.
866 * Generic RX queue pointer.
869 mlx5_rx_queue_release(void *dpdk_rxq)
871 struct mlx5_rxq_data *rxq = (struct mlx5_rxq_data *)dpdk_rxq;
872 struct mlx5_rxq_ctrl *rxq_ctrl;
873 struct mlx5_priv *priv;
877 rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
878 priv = rxq_ctrl->priv;
879 if (!mlx5_rxq_releasable(ETH_DEV(priv), rxq_ctrl->rxq.idx))
880 rte_panic("port %u Rx queue %u is still used by a flow and"
881 " cannot be removed\n",
882 PORT_ID(priv), rxq->idx);
883 mlx5_rxq_release(ETH_DEV(priv), rxq_ctrl->rxq.idx);
887 * Allocate queue vector and fill epoll fd list for Rx interrupts.
890 * Pointer to Ethernet device.
893 * 0 on success, a negative errno value otherwise and rte_errno is set.
896 mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
898 struct mlx5_priv *priv = dev->data->dev_private;
900 unsigned int rxqs_n = priv->rxqs_n;
901 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
902 unsigned int count = 0;
903 struct rte_intr_handle *intr_handle = dev->intr_handle;
905 if (!dev->data->dev_conf.intr_conf.rxq)
907 mlx5_rx_intr_vec_disable(dev);
908 intr_handle->intr_vec = mlx5_malloc(0,
909 n * sizeof(intr_handle->intr_vec[0]),
911 if (intr_handle->intr_vec == NULL) {
913 "port %u failed to allocate memory for interrupt"
914 " vector, Rx interrupts will not be supported",
919 intr_handle->type = RTE_INTR_HANDLE_EXT;
920 for (i = 0; i != n; ++i) {
921 /* This rxq obj must not be released in this function. */
922 struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_get(dev, i);
923 struct mlx5_rxq_obj *rxq_obj = rxq_ctrl ? rxq_ctrl->obj : NULL;
926 /* Skip queues that cannot request interrupts. */
927 if (!rxq_obj || (!rxq_obj->ibv_channel &&
928 !rxq_obj->devx_channel)) {
929 /* Use invalid intr_vec[] index to disable entry. */
930 intr_handle->intr_vec[i] =
931 RTE_INTR_VEC_RXTX_OFFSET +
932 RTE_MAX_RXTX_INTR_VEC_ID;
933 /* Decrease the rxq_ctrl's refcnt */
935 mlx5_rxq_release(dev, i);
938 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
940 "port %u too many Rx queues for interrupt"
941 " vector size (%d), Rx interrupts cannot be"
943 dev->data->port_id, RTE_MAX_RXTX_INTR_VEC_ID);
944 mlx5_rx_intr_vec_disable(dev);
948 rc = mlx5_os_set_nonblock_channel_fd(rxq_obj->fd);
952 "port %u failed to make Rx interrupt file"
953 " descriptor %d non-blocking for queue index"
955 dev->data->port_id, rxq_obj->fd, i);
956 mlx5_rx_intr_vec_disable(dev);
959 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
960 intr_handle->efds[count] = rxq_obj->fd;
964 mlx5_rx_intr_vec_disable(dev);
966 intr_handle->nb_efd = count;
971 * Clean up Rx interrupts handler.
974 * Pointer to Ethernet device.
977 mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
979 struct mlx5_priv *priv = dev->data->dev_private;
980 struct rte_intr_handle *intr_handle = dev->intr_handle;
982 unsigned int rxqs_n = priv->rxqs_n;
983 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
985 if (!dev->data->dev_conf.intr_conf.rxq)
987 if (!intr_handle->intr_vec)
989 for (i = 0; i != n; ++i) {
990 if (intr_handle->intr_vec[i] == RTE_INTR_VEC_RXTX_OFFSET +
991 RTE_MAX_RXTX_INTR_VEC_ID)
994 * Need to access directly the queue to release the reference
995 * kept in mlx5_rx_intr_vec_enable().
997 mlx5_rxq_release(dev, i);
1000 rte_intr_free_epoll_fd(intr_handle);
1001 if (intr_handle->intr_vec)
1002 mlx5_free(intr_handle->intr_vec);
1003 intr_handle->nb_efd = 0;
1004 intr_handle->intr_vec = NULL;
1008 * MLX5 CQ notification .
1011 * Pointer to receive queue structure.
1013 * Sequence number per receive queue .
1016 mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
1019 uint32_t doorbell_hi;
1021 void *cq_db_reg = (char *)rxq->cq_uar + MLX5_CQ_DOORBELL;
1023 sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
1024 doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
1025 doorbell = (uint64_t)doorbell_hi << 32;
1026 doorbell |= rxq->cqn;
1027 rxq->cq_db[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
1028 mlx5_uar_write64(rte_cpu_to_be_64(doorbell),
1029 cq_db_reg, rxq->uar_lock_cq);
1033 * DPDK callback for Rx queue interrupt enable.
1036 * Pointer to Ethernet device structure.
1037 * @param rx_queue_id
1041 * 0 on success, a negative errno value otherwise and rte_errno is set.
1044 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1046 struct mlx5_rxq_ctrl *rxq_ctrl;
1048 rxq_ctrl = mlx5_rxq_get(dev, rx_queue_id);
1051 if (rxq_ctrl->irq) {
1052 if (!rxq_ctrl->obj) {
1053 mlx5_rxq_release(dev, rx_queue_id);
1056 mlx5_arm_cq(&rxq_ctrl->rxq, rxq_ctrl->rxq.cq_arm_sn);
1058 mlx5_rxq_release(dev, rx_queue_id);
1066 * DPDK callback for Rx queue interrupt disable.
1069 * Pointer to Ethernet device structure.
1070 * @param rx_queue_id
1074 * 0 on success, a negative errno value otherwise and rte_errno is set.
1077 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1079 struct mlx5_priv *priv = dev->data->dev_private;
1080 struct mlx5_rxq_ctrl *rxq_ctrl;
1083 rxq_ctrl = mlx5_rxq_get(dev, rx_queue_id);
1090 if (rxq_ctrl->irq) {
1091 ret = priv->obj_ops.rxq_event_get(rxq_ctrl->obj);
1094 rxq_ctrl->rxq.cq_arm_sn++;
1096 mlx5_rxq_release(dev, rx_queue_id);
1100 * The ret variable may be EAGAIN which means the get_event function was
1101 * called before receiving one.
1107 ret = rte_errno; /* Save rte_errno before cleanup. */
1108 mlx5_rxq_release(dev, rx_queue_id);
1110 DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
1111 dev->data->port_id, rx_queue_id);
1112 rte_errno = ret; /* Restore rte_errno. */
1117 * Verify the Rx queue objects list is empty
1120 * Pointer to Ethernet device.
1123 * The number of objects not released.
1126 mlx5_rxq_obj_verify(struct rte_eth_dev *dev)
1128 struct mlx5_priv *priv = dev->data->dev_private;
1130 struct mlx5_rxq_obj *rxq_obj;
1132 LIST_FOREACH(rxq_obj, &priv->rxqsobj, next) {
1133 DRV_LOG(DEBUG, "port %u Rx queue %u still referenced",
1134 dev->data->port_id, rxq_obj->rxq_ctrl->rxq.idx);
1141 * Callback function to initialize mbufs for Multi-Packet RQ.
1144 mlx5_mprq_buf_init(struct rte_mempool *mp, void *opaque_arg,
1145 void *_m, unsigned int i __rte_unused)
1147 struct mlx5_mprq_buf *buf = _m;
1148 struct rte_mbuf_ext_shared_info *shinfo;
1149 unsigned int strd_n = (unsigned int)(uintptr_t)opaque_arg;
1152 memset(_m, 0, sizeof(*buf));
1154 __atomic_store_n(&buf->refcnt, 1, __ATOMIC_RELAXED);
1155 for (j = 0; j != strd_n; ++j) {
1156 shinfo = &buf->shinfos[j];
1157 shinfo->free_cb = mlx5_mprq_buf_free_cb;
1158 shinfo->fcb_opaque = buf;
1163 * Free mempool of Multi-Packet RQ.
1166 * Pointer to Ethernet device.
1169 * 0 on success, negative errno value on failure.
1172 mlx5_mprq_free_mp(struct rte_eth_dev *dev)
1174 struct mlx5_priv *priv = dev->data->dev_private;
1175 struct rte_mempool *mp = priv->mprq_mp;
1180 DRV_LOG(DEBUG, "port %u freeing mempool (%s) for Multi-Packet RQ",
1181 dev->data->port_id, mp->name);
1183 * If a buffer in the pool has been externally attached to a mbuf and it
1184 * is still in use by application, destroying the Rx queue can spoil
1185 * the packet. It is unlikely to happen but if application dynamically
1186 * creates and destroys with holding Rx packets, this can happen.
1188 * TODO: It is unavoidable for now because the mempool for Multi-Packet
1189 * RQ isn't provided by application but managed by PMD.
1191 if (!rte_mempool_full(mp)) {
1193 "port %u mempool for Multi-Packet RQ is still in use",
1194 dev->data->port_id);
1198 rte_mempool_free(mp);
1199 /* Unset mempool for each Rx queue. */
1200 for (i = 0; i != priv->rxqs_n; ++i) {
1201 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1205 rxq->mprq_mp = NULL;
1207 priv->mprq_mp = NULL;
1212 * Allocate a mempool for Multi-Packet RQ. All configured Rx queues share the
1213 * mempool. If already allocated, reuse it if there're enough elements.
1214 * Otherwise, resize it.
1217 * Pointer to Ethernet device.
1220 * 0 on success, negative errno value on failure.
1223 mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
1225 struct mlx5_priv *priv = dev->data->dev_private;
1226 struct rte_mempool *mp = priv->mprq_mp;
1227 char name[RTE_MEMPOOL_NAMESIZE];
1228 unsigned int desc = 0;
1229 unsigned int buf_len;
1230 unsigned int obj_num;
1231 unsigned int obj_size;
1232 unsigned int strd_num_n = 0;
1233 unsigned int strd_sz_n = 0;
1235 unsigned int n_ibv = 0;
1237 if (!mlx5_mprq_enabled(dev))
1239 /* Count the total number of descriptors configured. */
1240 for (i = 0; i != priv->rxqs_n; ++i) {
1241 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1242 struct mlx5_rxq_ctrl *rxq_ctrl = container_of
1243 (rxq, struct mlx5_rxq_ctrl, rxq);
1245 if (rxq == NULL || rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
1248 desc += 1 << rxq->elts_n;
1249 /* Get the max number of strides. */
1250 if (strd_num_n < rxq->strd_num_n)
1251 strd_num_n = rxq->strd_num_n;
1252 /* Get the max size of a stride. */
1253 if (strd_sz_n < rxq->strd_sz_n)
1254 strd_sz_n = rxq->strd_sz_n;
1256 MLX5_ASSERT(strd_num_n && strd_sz_n);
1257 buf_len = (1 << strd_num_n) * (1 << strd_sz_n);
1258 obj_size = sizeof(struct mlx5_mprq_buf) + buf_len + (1 << strd_num_n) *
1259 sizeof(struct rte_mbuf_ext_shared_info) + RTE_PKTMBUF_HEADROOM;
1261 * Received packets can be either memcpy'd or externally referenced. In
1262 * case that the packet is attached to an mbuf as an external buffer, as
1263 * it isn't possible to predict how the buffers will be queued by
1264 * application, there's no option to exactly pre-allocate needed buffers
1265 * in advance but to speculatively prepares enough buffers.
1267 * In the data path, if this Mempool is depleted, PMD will try to memcpy
1268 * received packets to buffers provided by application (rxq->mp) until
1269 * this Mempool gets available again.
1272 obj_num = desc + MLX5_MPRQ_MP_CACHE_SZ * n_ibv;
1274 * rte_mempool_create_empty() has sanity check to refuse large cache
1275 * size compared to the number of elements.
1276 * CACHE_FLUSHTHRESH_MULTIPLIER is defined in a C file, so using a
1277 * constant number 2 instead.
1279 obj_num = RTE_MAX(obj_num, MLX5_MPRQ_MP_CACHE_SZ * 2);
1280 /* Check a mempool is already allocated and if it can be resued. */
1281 if (mp != NULL && mp->elt_size >= obj_size && mp->size >= obj_num) {
1282 DRV_LOG(DEBUG, "port %u mempool %s is being reused",
1283 dev->data->port_id, mp->name);
1286 } else if (mp != NULL) {
1287 DRV_LOG(DEBUG, "port %u mempool %s should be resized, freeing it",
1288 dev->data->port_id, mp->name);
1290 * If failed to free, which means it may be still in use, no way
1291 * but to keep using the existing one. On buffer underrun,
1292 * packets will be memcpy'd instead of external buffer
1295 if (mlx5_mprq_free_mp(dev)) {
1296 if (mp->elt_size >= obj_size)
1302 snprintf(name, sizeof(name), "port-%u-mprq", dev->data->port_id);
1303 mp = rte_mempool_create(name, obj_num, obj_size, MLX5_MPRQ_MP_CACHE_SZ,
1304 0, NULL, NULL, mlx5_mprq_buf_init,
1305 (void *)(uintptr_t)(1 << strd_num_n),
1306 dev->device->numa_node, 0);
1309 "port %u failed to allocate a mempool for"
1310 " Multi-Packet RQ, count=%u, size=%u",
1311 dev->data->port_id, obj_num, obj_size);
1317 /* Set mempool for each Rx queue. */
1318 for (i = 0; i != priv->rxqs_n; ++i) {
1319 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1320 struct mlx5_rxq_ctrl *rxq_ctrl = container_of
1321 (rxq, struct mlx5_rxq_ctrl, rxq);
1323 if (rxq == NULL || rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
1327 DRV_LOG(INFO, "port %u Multi-Packet RQ is configured",
1328 dev->data->port_id);
1332 #define MLX5_MAX_TCP_HDR_OFFSET ((unsigned int)(sizeof(struct rte_ether_hdr) + \
1333 sizeof(struct rte_vlan_hdr) * 2 + \
1334 sizeof(struct rte_ipv6_hdr)))
1335 #define MAX_TCP_OPTION_SIZE 40u
1336 #define MLX5_MAX_LRO_HEADER_FIX ((unsigned int)(MLX5_MAX_TCP_HDR_OFFSET + \
1337 sizeof(struct rte_tcp_hdr) + \
1338 MAX_TCP_OPTION_SIZE))
1341 * Adjust the maximum LRO massage size.
1344 * Pointer to Ethernet device.
1347 * @param max_lro_size
1348 * The maximum size for LRO packet.
1351 mlx5_max_lro_msg_size_adjust(struct rte_eth_dev *dev, uint16_t idx,
1352 uint32_t max_lro_size)
1354 struct mlx5_priv *priv = dev->data->dev_private;
1356 if (priv->config.hca_attr.lro_max_msg_sz_mode ==
1357 MLX5_LRO_MAX_MSG_SIZE_START_FROM_L4 && max_lro_size >
1358 MLX5_MAX_TCP_HDR_OFFSET)
1359 max_lro_size -= MLX5_MAX_TCP_HDR_OFFSET;
1360 max_lro_size = RTE_MIN(max_lro_size, MLX5_MAX_LRO_SIZE);
1361 MLX5_ASSERT(max_lro_size >= MLX5_LRO_SEG_CHUNK_SIZE);
1362 max_lro_size /= MLX5_LRO_SEG_CHUNK_SIZE;
1363 if (priv->max_lro_msg_size)
1364 priv->max_lro_msg_size =
1365 RTE_MIN((uint32_t)priv->max_lro_msg_size, max_lro_size);
1367 priv->max_lro_msg_size = max_lro_size;
1369 "port %u Rx Queue %u max LRO message size adjusted to %u bytes",
1370 dev->data->port_id, idx,
1371 priv->max_lro_msg_size * MLX5_LRO_SEG_CHUNK_SIZE);
1375 * Create a DPDK Rx queue.
1378 * Pointer to Ethernet device.
1382 * Number of descriptors to configure in queue.
1384 * NUMA socket on which memory must be allocated.
1387 * A DPDK queue object on success, NULL otherwise and rte_errno is set.
1389 struct mlx5_rxq_ctrl *
1390 mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1391 unsigned int socket, const struct rte_eth_rxconf *conf,
1392 const struct rte_eth_rxseg_split *rx_seg, uint16_t n_seg)
1394 struct mlx5_priv *priv = dev->data->dev_private;
1395 struct mlx5_rxq_ctrl *tmpl;
1396 unsigned int mb_len = rte_pktmbuf_data_room_size(rx_seg[0].mp);
1397 struct mlx5_dev_config *config = &priv->config;
1398 uint64_t offloads = conf->offloads |
1399 dev->data->dev_conf.rxmode.offloads;
1400 unsigned int lro_on_queue = !!(offloads & DEV_RX_OFFLOAD_TCP_LRO);
1401 unsigned int max_rx_pkt_len = lro_on_queue ?
1402 dev->data->dev_conf.rxmode.max_lro_pkt_size :
1403 dev->data->dev_conf.rxmode.max_rx_pkt_len;
1404 unsigned int non_scatter_min_mbuf_size = max_rx_pkt_len +
1405 RTE_PKTMBUF_HEADROOM;
1406 unsigned int max_lro_size = 0;
1407 unsigned int first_mb_free_size = mb_len - RTE_PKTMBUF_HEADROOM;
1408 const int mprq_en = mlx5_check_mprq_support(dev) > 0 && n_seg == 1 &&
1409 !rx_seg[0].offset && !rx_seg[0].length;
1410 unsigned int mprq_stride_nums = config->mprq.stride_num_n ?
1411 config->mprq.stride_num_n : MLX5_MPRQ_STRIDE_NUM_N;
1412 unsigned int mprq_stride_size = non_scatter_min_mbuf_size <=
1413 (1U << config->mprq.max_stride_size_n) ?
1414 log2above(non_scatter_min_mbuf_size) : MLX5_MPRQ_STRIDE_SIZE_N;
1415 unsigned int mprq_stride_cap = (config->mprq.stride_num_n ?
1416 (1U << config->mprq.stride_num_n) : (1U << mprq_stride_nums)) *
1417 (config->mprq.stride_size_n ?
1418 (1U << config->mprq.stride_size_n) : (1U << mprq_stride_size));
1420 * Always allocate extra slots, even if eventually
1421 * the vector Rx will not be used.
1423 uint16_t desc_n = desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
1424 const struct rte_eth_rxseg_split *qs_seg = rx_seg;
1425 unsigned int tail_len;
1427 tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*tmpl) +
1428 desc_n * sizeof(struct rte_mbuf *), 0, socket);
1433 MLX5_ASSERT(n_seg && n_seg <= MLX5_MAX_RXQ_NSEG);
1435 * Build the array of actual buffer offsets and lengths.
1436 * Pad with the buffers from the last memory pool if
1437 * needed to handle max size packets, replace zero length
1438 * with the buffer length from the pool.
1440 tail_len = max_rx_pkt_len;
1442 struct mlx5_eth_rxseg *hw_seg =
1443 &tmpl->rxq.rxseg[tmpl->rxq.rxseg_n];
1444 uint32_t buf_len, offset, seg_len;
1447 * For the buffers beyond descriptions offset is zero,
1448 * the first buffer contains head room.
1450 buf_len = rte_pktmbuf_data_room_size(qs_seg->mp);
1451 offset = (tmpl->rxq.rxseg_n >= n_seg ? 0 : qs_seg->offset) +
1452 (tmpl->rxq.rxseg_n ? 0 : RTE_PKTMBUF_HEADROOM);
1454 * For the buffers beyond descriptions the length is
1455 * pool buffer length, zero lengths are replaced with
1456 * pool buffer length either.
1458 seg_len = tmpl->rxq.rxseg_n >= n_seg ? buf_len :
1462 /* Check is done in long int, now overflows. */
1463 if (buf_len < seg_len + offset) {
1464 DRV_LOG(ERR, "port %u Rx queue %u: Split offset/length "
1465 "%u/%u can't be satisfied",
1466 dev->data->port_id, idx,
1467 qs_seg->length, qs_seg->offset);
1471 if (seg_len > tail_len)
1472 seg_len = buf_len - offset;
1473 if (++tmpl->rxq.rxseg_n > MLX5_MAX_RXQ_NSEG) {
1475 "port %u too many SGEs (%u) needed to handle"
1476 " requested maximum packet size %u, the maximum"
1477 " supported are %u", dev->data->port_id,
1478 tmpl->rxq.rxseg_n, max_rx_pkt_len,
1480 rte_errno = ENOTSUP;
1483 /* Build the actual scattering element in the queue object. */
1484 hw_seg->mp = qs_seg->mp;
1485 MLX5_ASSERT(offset <= UINT16_MAX);
1486 MLX5_ASSERT(seg_len <= UINT16_MAX);
1487 hw_seg->offset = (uint16_t)offset;
1488 hw_seg->length = (uint16_t)seg_len;
1490 * Advance the segment descriptor, the padding is the based
1491 * on the attributes of the last descriptor.
1493 if (tmpl->rxq.rxseg_n < n_seg)
1495 tail_len -= RTE_MIN(tail_len, seg_len);
1496 } while (tail_len || !rte_is_power_of_2(tmpl->rxq.rxseg_n));
1497 MLX5_ASSERT(tmpl->rxq.rxseg_n &&
1498 tmpl->rxq.rxseg_n <= MLX5_MAX_RXQ_NSEG);
1499 if (tmpl->rxq.rxseg_n > 1 && !(offloads & DEV_RX_OFFLOAD_SCATTER)) {
1500 DRV_LOG(ERR, "port %u Rx queue %u: Scatter offload is not"
1501 " configured and no enough mbuf space(%u) to contain "
1502 "the maximum RX packet length(%u) with head-room(%u)",
1503 dev->data->port_id, idx, mb_len, max_rx_pkt_len,
1504 RTE_PKTMBUF_HEADROOM);
1508 tmpl->type = MLX5_RXQ_TYPE_STANDARD;
1509 if (mlx5_mr_btree_init(&tmpl->rxq.mr_ctrl.cache_bh,
1510 MLX5_MR_BTREE_CACHE_N, socket)) {
1511 /* rte_errno is already set. */
1514 tmpl->socket = socket;
1515 if (dev->data->dev_conf.intr_conf.rxq)
1518 * This Rx queue can be configured as a Multi-Packet RQ if all of the
1519 * following conditions are met:
1520 * - MPRQ is enabled.
1521 * - The number of descs is more than the number of strides.
1522 * - max_rx_pkt_len plus overhead is less than the max size
1523 * of a stride or mprq_stride_size is specified by a user.
1524 * Need to make sure that there are enough strides to encap
1525 * the maximum packet size in case mprq_stride_size is set.
1526 * Otherwise, enable Rx scatter if necessary.
1528 if (mprq_en && desc > (1U << mprq_stride_nums) &&
1529 (non_scatter_min_mbuf_size <=
1530 (1U << config->mprq.max_stride_size_n) ||
1531 (config->mprq.stride_size_n &&
1532 non_scatter_min_mbuf_size <= mprq_stride_cap))) {
1533 /* TODO: Rx scatter isn't supported yet. */
1534 tmpl->rxq.sges_n = 0;
1535 /* Trim the number of descs needed. */
1536 desc >>= mprq_stride_nums;
1537 tmpl->rxq.strd_num_n = config->mprq.stride_num_n ?
1538 config->mprq.stride_num_n : mprq_stride_nums;
1539 tmpl->rxq.strd_sz_n = config->mprq.stride_size_n ?
1540 config->mprq.stride_size_n : mprq_stride_size;
1541 tmpl->rxq.strd_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT;
1542 tmpl->rxq.strd_scatter_en =
1543 !!(offloads & DEV_RX_OFFLOAD_SCATTER);
1544 tmpl->rxq.mprq_max_memcpy_len = RTE_MIN(first_mb_free_size,
1545 config->mprq.max_memcpy_len);
1546 max_lro_size = RTE_MIN(max_rx_pkt_len,
1547 (1u << tmpl->rxq.strd_num_n) *
1548 (1u << tmpl->rxq.strd_sz_n));
1550 "port %u Rx queue %u: Multi-Packet RQ is enabled"
1551 " strd_num_n = %u, strd_sz_n = %u",
1552 dev->data->port_id, idx,
1553 tmpl->rxq.strd_num_n, tmpl->rxq.strd_sz_n);
1554 } else if (tmpl->rxq.rxseg_n == 1) {
1555 MLX5_ASSERT(max_rx_pkt_len <= first_mb_free_size);
1556 tmpl->rxq.sges_n = 0;
1557 max_lro_size = max_rx_pkt_len;
1558 } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
1559 unsigned int sges_n;
1561 if (lro_on_queue && first_mb_free_size <
1562 MLX5_MAX_LRO_HEADER_FIX) {
1563 DRV_LOG(ERR, "Not enough space in the first segment(%u)"
1564 " to include the max header size(%u) for LRO",
1565 first_mb_free_size, MLX5_MAX_LRO_HEADER_FIX);
1566 rte_errno = ENOTSUP;
1570 * Determine the number of SGEs needed for a full packet
1571 * and round it to the next power of two.
1573 sges_n = log2above(tmpl->rxq.rxseg_n);
1574 if (sges_n > MLX5_MAX_LOG_RQ_SEGS) {
1576 "port %u too many SGEs (%u) needed to handle"
1577 " requested maximum packet size %u, the maximum"
1578 " supported are %u", dev->data->port_id,
1579 1 << sges_n, max_rx_pkt_len,
1580 1u << MLX5_MAX_LOG_RQ_SEGS);
1581 rte_errno = ENOTSUP;
1584 tmpl->rxq.sges_n = sges_n;
1585 max_lro_size = max_rx_pkt_len;
1587 if (config->mprq.enabled && !mlx5_rxq_mprq_enabled(&tmpl->rxq))
1589 "port %u MPRQ is requested but cannot be enabled\n"
1590 " (requested: pkt_sz = %u, desc_num = %u,"
1591 " rxq_num = %u, stride_sz = %u, stride_num = %u\n"
1592 " supported: min_rxqs_num = %u,"
1593 " min_stride_sz = %u, max_stride_sz = %u).",
1594 dev->data->port_id, non_scatter_min_mbuf_size,
1596 config->mprq.stride_size_n ?
1597 (1U << config->mprq.stride_size_n) :
1598 (1U << mprq_stride_size),
1599 config->mprq.stride_num_n ?
1600 (1U << config->mprq.stride_num_n) :
1601 (1U << mprq_stride_nums),
1602 config->mprq.min_rxqs_num,
1603 (1U << config->mprq.min_stride_size_n),
1604 (1U << config->mprq.max_stride_size_n));
1605 DRV_LOG(DEBUG, "port %u maximum number of segments per packet: %u",
1606 dev->data->port_id, 1 << tmpl->rxq.sges_n);
1607 if (desc % (1 << tmpl->rxq.sges_n)) {
1609 "port %u number of Rx queue descriptors (%u) is not a"
1610 " multiple of SGEs per packet (%u)",
1613 1 << tmpl->rxq.sges_n);
1617 mlx5_max_lro_msg_size_adjust(dev, idx, max_lro_size);
1618 /* Toggle RX checksum offload if hardware supports it. */
1619 tmpl->rxq.csum = !!(offloads & DEV_RX_OFFLOAD_CHECKSUM);
1620 /* Configure Rx timestamp. */
1621 tmpl->rxq.hw_timestamp = !!(offloads & DEV_RX_OFFLOAD_TIMESTAMP);
1622 tmpl->rxq.timestamp_rx_flag = 0;
1623 if (tmpl->rxq.hw_timestamp && rte_mbuf_dyn_rx_timestamp_register(
1624 &tmpl->rxq.timestamp_offset,
1625 &tmpl->rxq.timestamp_rx_flag) != 0) {
1626 DRV_LOG(ERR, "Cannot register Rx timestamp field/flag");
1629 /* Configure VLAN stripping. */
1630 tmpl->rxq.vlan_strip = !!(offloads & DEV_RX_OFFLOAD_VLAN_STRIP);
1631 /* By default, FCS (CRC) is stripped by hardware. */
1632 tmpl->rxq.crc_present = 0;
1633 tmpl->rxq.lro = lro_on_queue;
1634 if (offloads & DEV_RX_OFFLOAD_KEEP_CRC) {
1635 if (config->hw_fcs_strip) {
1637 * RQs used for LRO-enabled TIRs should not be
1638 * configured to scatter the FCS.
1642 "port %u CRC stripping has been "
1643 "disabled but will still be performed "
1644 "by hardware, because LRO is enabled",
1645 dev->data->port_id);
1647 tmpl->rxq.crc_present = 1;
1650 "port %u CRC stripping has been disabled but will"
1651 " still be performed by hardware, make sure MLNX_OFED"
1652 " and firmware are up to date",
1653 dev->data->port_id);
1657 "port %u CRC stripping is %s, %u bytes will be subtracted from"
1658 " incoming frames to hide it",
1660 tmpl->rxq.crc_present ? "disabled" : "enabled",
1661 tmpl->rxq.crc_present << 2);
1663 tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
1664 (!!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS));
1665 tmpl->rxq.port_id = dev->data->port_id;
1667 tmpl->rxq.mp = rx_seg[0].mp;
1668 tmpl->rxq.elts_n = log2above(desc);
1669 tmpl->rxq.rq_repl_thresh =
1670 MLX5_VPMD_RXQ_RPLNSH_THRESH(desc_n);
1672 (struct rte_mbuf *(*)[desc_n])(tmpl + 1);
1673 tmpl->rxq.mprq_bufs =
1674 (struct mlx5_mprq_buf *(*)[desc])(*tmpl->rxq.elts + desc_n);
1676 tmpl->rxq.uar_lock_cq = &priv->sh->uar_lock_cq;
1678 tmpl->rxq.idx = idx;
1679 __atomic_fetch_add(&tmpl->refcnt, 1, __ATOMIC_RELAXED);
1680 LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1688 * Create a DPDK Rx hairpin queue.
1691 * Pointer to Ethernet device.
1695 * Number of descriptors to configure in queue.
1696 * @param hairpin_conf
1697 * The hairpin binding configuration.
1700 * A DPDK queue object on success, NULL otherwise and rte_errno is set.
1702 struct mlx5_rxq_ctrl *
1703 mlx5_rxq_hairpin_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1704 const struct rte_eth_hairpin_conf *hairpin_conf)
1706 struct mlx5_priv *priv = dev->data->dev_private;
1707 struct mlx5_rxq_ctrl *tmpl;
1709 tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*tmpl), 0,
1715 tmpl->type = MLX5_RXQ_TYPE_HAIRPIN;
1716 tmpl->socket = SOCKET_ID_ANY;
1717 tmpl->rxq.rss_hash = 0;
1718 tmpl->rxq.port_id = dev->data->port_id;
1720 tmpl->rxq.mp = NULL;
1721 tmpl->rxq.elts_n = log2above(desc);
1722 tmpl->rxq.elts = NULL;
1723 tmpl->rxq.mr_ctrl.cache_bh = (struct mlx5_mr_btree) { 0 };
1724 tmpl->hairpin_conf = *hairpin_conf;
1725 tmpl->rxq.idx = idx;
1726 __atomic_fetch_add(&tmpl->refcnt, 1, __ATOMIC_RELAXED);
1727 LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1735 * Pointer to Ethernet device.
1740 * A pointer to the queue if it exists, NULL otherwise.
1742 struct mlx5_rxq_ctrl *
1743 mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
1745 struct mlx5_priv *priv = dev->data->dev_private;
1746 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
1747 struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1750 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
1751 __atomic_fetch_add(&rxq_ctrl->refcnt, 1, __ATOMIC_RELAXED);
1757 * Release a Rx queue.
1760 * Pointer to Ethernet device.
1765 * 1 while a reference on it exists, 0 when freed.
1768 mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
1770 struct mlx5_priv *priv = dev->data->dev_private;
1771 struct mlx5_rxq_ctrl *rxq_ctrl;
1773 if (!(*priv->rxqs)[idx])
1775 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1776 if (__atomic_sub_fetch(&rxq_ctrl->refcnt, 1, __ATOMIC_RELAXED) > 1)
1778 if (rxq_ctrl->obj) {
1779 priv->obj_ops.rxq_obj_release(rxq_ctrl->obj);
1780 LIST_REMOVE(rxq_ctrl->obj, next);
1781 mlx5_free(rxq_ctrl->obj);
1782 rxq_ctrl->obj = NULL;
1784 if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
1785 rxq_free_elts(rxq_ctrl);
1786 if (!__atomic_load_n(&rxq_ctrl->refcnt, __ATOMIC_RELAXED)) {
1787 if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
1788 mlx5_mr_btree_free(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
1789 LIST_REMOVE(rxq_ctrl, next);
1790 mlx5_free(rxq_ctrl);
1791 (*priv->rxqs)[idx] = NULL;
1797 * Verify the Rx Queue list is empty
1800 * Pointer to Ethernet device.
1803 * The number of object not released.
1806 mlx5_rxq_verify(struct rte_eth_dev *dev)
1808 struct mlx5_priv *priv = dev->data->dev_private;
1809 struct mlx5_rxq_ctrl *rxq_ctrl;
1812 LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
1813 DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
1814 dev->data->port_id, rxq_ctrl->rxq.idx);
1821 * Get a Rx queue type.
1824 * Pointer to Ethernet device.
1829 * The Rx queue type.
1832 mlx5_rxq_get_type(struct rte_eth_dev *dev, uint16_t idx)
1834 struct mlx5_priv *priv = dev->data->dev_private;
1835 struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1837 if (idx < priv->rxqs_n && (*priv->rxqs)[idx]) {
1838 rxq_ctrl = container_of((*priv->rxqs)[idx],
1839 struct mlx5_rxq_ctrl,
1841 return rxq_ctrl->type;
1843 return MLX5_RXQ_TYPE_UNDEFINED;
1847 * Get a Rx hairpin queue configuration.
1850 * Pointer to Ethernet device.
1855 * Pointer to the configuration if a hairpin RX queue, otherwise NULL.
1857 const struct rte_eth_hairpin_conf *
1858 mlx5_rxq_get_hairpin_conf(struct rte_eth_dev *dev, uint16_t idx)
1860 struct mlx5_priv *priv = dev->data->dev_private;
1861 struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1863 if (idx < priv->rxqs_n && (*priv->rxqs)[idx]) {
1864 rxq_ctrl = container_of((*priv->rxqs)[idx],
1865 struct mlx5_rxq_ctrl,
1867 if (rxq_ctrl->type == MLX5_RXQ_TYPE_HAIRPIN)
1868 return &rxq_ctrl->hairpin_conf;
1874 * Match queues listed in arguments to queues contained in indirection table
1878 * Pointer to indirection table to match.
1880 * Queues to match to ques in indirection table.
1882 * Number of queues in the array.
1885 * 1 if all queues in indirection table match 0 othrwise.
1888 mlx5_ind_table_obj_match_queues(const struct mlx5_ind_table_obj *ind_tbl,
1889 const uint16_t *queues, uint32_t queues_n)
1891 return (ind_tbl->queues_n == queues_n) &&
1892 (!memcmp(ind_tbl->queues, queues,
1893 ind_tbl->queues_n * sizeof(ind_tbl->queues[0])));
1897 * Get an indirection table.
1900 * Pointer to Ethernet device.
1902 * Queues entering in the indirection table.
1904 * Number of queues in the array.
1907 * An indirection table if found.
1909 struct mlx5_ind_table_obj *
1910 mlx5_ind_table_obj_get(struct rte_eth_dev *dev, const uint16_t *queues,
1913 struct mlx5_priv *priv = dev->data->dev_private;
1914 struct mlx5_ind_table_obj *ind_tbl;
1916 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1917 if ((ind_tbl->queues_n == queues_n) &&
1918 (memcmp(ind_tbl->queues, queues,
1919 ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
1926 __atomic_fetch_add(&ind_tbl->refcnt, 1, __ATOMIC_RELAXED);
1927 for (i = 0; i != ind_tbl->queues_n; ++i)
1928 mlx5_rxq_get(dev, ind_tbl->queues[i]);
1934 * Release an indirection table.
1937 * Pointer to Ethernet device.
1939 * Indirection table to release.
1941 * Indirection table for Standalone queue.
1944 * 1 while a reference on it exists, 0 when freed.
1947 mlx5_ind_table_obj_release(struct rte_eth_dev *dev,
1948 struct mlx5_ind_table_obj *ind_tbl,
1951 struct mlx5_priv *priv = dev->data->dev_private;
1954 if (__atomic_sub_fetch(&ind_tbl->refcnt, 1, __ATOMIC_RELAXED) == 0)
1955 priv->obj_ops.ind_table_destroy(ind_tbl);
1956 for (i = 0; i != ind_tbl->queues_n; ++i)
1957 claim_nonzero(mlx5_rxq_release(dev, ind_tbl->queues[i]));
1958 if (__atomic_load_n(&ind_tbl->refcnt, __ATOMIC_RELAXED) == 0) {
1960 LIST_REMOVE(ind_tbl, next);
1968 * Verify the Rx Queue list is empty
1971 * Pointer to Ethernet device.
1974 * The number of object not released.
1977 mlx5_ind_table_obj_verify(struct rte_eth_dev *dev)
1979 struct mlx5_priv *priv = dev->data->dev_private;
1980 struct mlx5_ind_table_obj *ind_tbl;
1983 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1985 "port %u indirection table obj %p still referenced",
1986 dev->data->port_id, (void *)ind_tbl);
1993 * Create an indirection table.
1996 * Pointer to Ethernet device.
1998 * Queues entering in the indirection table.
2000 * Number of queues in the array.
2002 * Indirection table for Standalone queue.
2005 * The Verbs/DevX object initialized, NULL otherwise and rte_errno is set.
2007 static struct mlx5_ind_table_obj *
2008 mlx5_ind_table_obj_new(struct rte_eth_dev *dev, const uint16_t *queues,
2009 uint32_t queues_n, bool standalone)
2011 struct mlx5_priv *priv = dev->data->dev_private;
2012 struct mlx5_ind_table_obj *ind_tbl;
2013 const unsigned int n = rte_is_power_of_2(queues_n) ?
2014 log2above(queues_n) :
2015 log2above(priv->config.ind_table_max_size);
2019 ind_tbl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*ind_tbl) +
2020 queues_n * sizeof(uint16_t), 0, SOCKET_ID_ANY);
2025 ind_tbl->queues_n = queues_n;
2026 for (i = 0; i != queues_n; ++i) {
2027 struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev, queues[i]);
2030 ind_tbl->queues[i] = queues[i];
2032 ret = priv->obj_ops.ind_table_new(dev, n, ind_tbl);
2035 __atomic_fetch_add(&ind_tbl->refcnt, 1, __ATOMIC_RELAXED);
2037 LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
2041 for (j = 0; j < i; j++)
2042 mlx5_rxq_release(dev, ind_tbl->queues[j]);
2045 DEBUG("Port %u cannot create indirection table.", dev->data->port_id);
2050 * Match an Rx Hash queue.
2053 * Cache list pointer.
2055 * Hash queue entry pointer.
2057 * Context of the callback function.
2060 * 0 if match, none zero if not match.
2063 mlx5_hrxq_match_cb(struct mlx5_cache_list *list,
2064 struct mlx5_cache_entry *entry,
2067 struct rte_eth_dev *dev = list->ctx;
2068 struct mlx5_flow_cb_ctx *ctx = cb_ctx;
2069 struct mlx5_flow_rss_desc *rss_desc = ctx->data;
2070 struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
2071 struct mlx5_ind_table_obj *ind_tbl;
2073 if (hrxq->rss_key_len != rss_desc->key_len ||
2074 memcmp(hrxq->rss_key, rss_desc->key, rss_desc->key_len) ||
2075 hrxq->hash_fields != rss_desc->hash_fields)
2077 ind_tbl = mlx5_ind_table_obj_get(dev, rss_desc->queue,
2078 rss_desc->queue_num);
2080 mlx5_ind_table_obj_release(dev, ind_tbl, hrxq->standalone);
2081 return ind_tbl != hrxq->ind_table;
2085 * Modify an Rx Hash queue configuration.
2088 * Pointer to Ethernet device.
2090 * Index to Hash Rx queue to modify.
2092 * RSS key for the Rx hash queue.
2093 * @param rss_key_len
2095 * @param hash_fields
2096 * Verbs protocol hash field to make the RSS on.
2098 * Queues entering in hash queue. In case of empty hash_fields only the
2099 * first queue index will be taken for the indirection table.
2104 * 0 on success, a negative errno value otherwise and rte_errno is set.
2107 mlx5_hrxq_modify(struct rte_eth_dev *dev, uint32_t hrxq_idx,
2108 const uint8_t *rss_key, uint32_t rss_key_len,
2109 uint64_t hash_fields,
2110 const uint16_t *queues, uint32_t queues_n)
2113 struct mlx5_ind_table_obj *ind_tbl = NULL;
2114 struct mlx5_priv *priv = dev->data->dev_private;
2115 struct mlx5_hrxq *hrxq =
2116 mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
2124 if (hrxq->rss_key_len != rss_key_len) {
2125 /* rss_key_len is fixed size 40 byte & not supposed to change */
2129 queues_n = hash_fields ? queues_n : 1;
2130 if (mlx5_ind_table_obj_match_queues(hrxq->ind_table,
2131 queues, queues_n)) {
2132 ind_tbl = hrxq->ind_table;
2134 ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
2136 ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n,
2143 MLX5_ASSERT(priv->obj_ops.hrxq_modify);
2144 ret = priv->obj_ops.hrxq_modify(dev, hrxq, rss_key,
2145 hash_fields, ind_tbl);
2150 if (ind_tbl != hrxq->ind_table) {
2151 mlx5_ind_table_obj_release(dev, hrxq->ind_table,
2153 hrxq->ind_table = ind_tbl;
2155 hrxq->hash_fields = hash_fields;
2156 memcpy(hrxq->rss_key, rss_key, rss_key_len);
2160 if (ind_tbl != hrxq->ind_table)
2161 mlx5_ind_table_obj_release(dev, ind_tbl, hrxq->standalone);
2167 __mlx5_hrxq_remove(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
2169 struct mlx5_priv *priv = dev->data->dev_private;
2171 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2172 mlx5_glue->destroy_flow_action(hrxq->action);
2174 priv->obj_ops.hrxq_destroy(hrxq);
2175 mlx5_ind_table_obj_release(dev, hrxq->ind_table, hrxq->standalone);
2176 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq->idx);
2180 * Release the hash Rx queue.
2183 * Pointer to Ethernet device.
2185 * Index to Hash Rx queue to release.
2188 * Cache list pointer.
2190 * Hash queue entry pointer.
2193 mlx5_hrxq_remove_cb(struct mlx5_cache_list *list,
2194 struct mlx5_cache_entry *entry)
2196 struct rte_eth_dev *dev = list->ctx;
2197 struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
2199 __mlx5_hrxq_remove(dev, hrxq);
2202 static struct mlx5_hrxq *
2203 __mlx5_hrxq_create(struct rte_eth_dev *dev,
2204 struct mlx5_flow_rss_desc *rss_desc)
2206 struct mlx5_priv *priv = dev->data->dev_private;
2207 const uint8_t *rss_key = rss_desc->key;
2208 uint32_t rss_key_len = rss_desc->key_len;
2209 const uint16_t *queues =
2210 rss_desc->standalone ? rss_desc->const_q : rss_desc->queue;
2211 uint32_t queues_n = rss_desc->queue_num;
2212 struct mlx5_hrxq *hrxq = NULL;
2213 uint32_t hrxq_idx = 0;
2214 struct mlx5_ind_table_obj *ind_tbl;
2217 queues_n = rss_desc->hash_fields ? queues_n : 1;
2218 ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
2220 ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n,
2221 rss_desc->standalone);
2224 hrxq = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_HRXQ], &hrxq_idx);
2227 hrxq->standalone = rss_desc->standalone;
2228 hrxq->idx = hrxq_idx;
2229 hrxq->ind_table = ind_tbl;
2230 hrxq->rss_key_len = rss_key_len;
2231 hrxq->hash_fields = rss_desc->hash_fields;
2232 memcpy(hrxq->rss_key, rss_key, rss_key_len);
2233 ret = priv->obj_ops.hrxq_new(dev, hrxq, rss_desc->tunnel);
2238 mlx5_ind_table_obj_release(dev, ind_tbl, rss_desc->standalone);
2240 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
2245 * Create an Rx Hash queue.
2248 * Cache list pointer.
2250 * Hash queue entry pointer.
2252 * Context of the callback function.
2255 * queue entry on success, NULL otherwise.
2257 struct mlx5_cache_entry *
2258 mlx5_hrxq_create_cb(struct mlx5_cache_list *list,
2259 struct mlx5_cache_entry *entry __rte_unused,
2262 struct rte_eth_dev *dev = list->ctx;
2263 struct mlx5_flow_cb_ctx *ctx = cb_ctx;
2264 struct mlx5_flow_rss_desc *rss_desc = ctx->data;
2265 struct mlx5_hrxq *hrxq;
2267 hrxq = __mlx5_hrxq_create(dev, rss_desc);
2268 return hrxq ? &hrxq->entry : NULL;
2272 * Get an Rx Hash queue.
2275 * Pointer to Ethernet device.
2277 * RSS configuration for the Rx hash queue.
2280 * An hash Rx queue index on success.
2282 uint32_t mlx5_hrxq_get(struct rte_eth_dev *dev,
2283 struct mlx5_flow_rss_desc *rss_desc)
2285 struct mlx5_priv *priv = dev->data->dev_private;
2286 struct mlx5_hrxq *hrxq;
2287 struct mlx5_cache_entry *entry;
2288 struct mlx5_flow_cb_ctx ctx = {
2292 if (rss_desc->standalone) {
2293 hrxq = __mlx5_hrxq_create(dev, rss_desc);
2295 entry = mlx5_cache_register(&priv->hrxqs, &ctx);
2298 hrxq = container_of(entry, typeof(*hrxq), entry);
2304 * Release the hash Rx queue.
2307 * Pointer to Ethernet device.
2309 * Index to Hash Rx queue to release.
2312 * 1 while a reference on it exists, 0 when freed.
2314 int mlx5_hrxq_release(struct rte_eth_dev *dev, uint32_t hrxq_idx)
2316 struct mlx5_priv *priv = dev->data->dev_private;
2317 struct mlx5_hrxq *hrxq;
2319 hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
2320 if (!hrxq->standalone)
2321 return mlx5_cache_unregister(&priv->hrxqs, &hrxq->entry);
2322 __mlx5_hrxq_remove(dev, hrxq);
2327 * Create a drop Rx Hash queue.
2330 * Pointer to Ethernet device.
2333 * The Verbs/DevX object initialized, NULL otherwise and rte_errno is set.
2336 mlx5_drop_action_create(struct rte_eth_dev *dev)
2338 struct mlx5_priv *priv = dev->data->dev_private;
2339 struct mlx5_hrxq *hrxq = NULL;
2342 if (priv->drop_queue.hrxq) {
2343 __atomic_fetch_add(&priv->drop_queue.hrxq->refcnt, 1,
2345 return priv->drop_queue.hrxq;
2347 hrxq = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*hrxq), 0, SOCKET_ID_ANY);
2350 "Port %u cannot allocate memory for drop queue.",
2351 dev->data->port_id);
2355 priv->drop_queue.hrxq = hrxq;
2356 hrxq->ind_table = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*hrxq->ind_table),
2358 if (!hrxq->ind_table) {
2362 ret = priv->obj_ops.drop_action_create(dev);
2365 __atomic_store_n(&hrxq->refcnt, 1, __ATOMIC_RELAXED);
2369 if (hrxq->ind_table)
2370 mlx5_free(hrxq->ind_table);
2371 priv->drop_queue.hrxq = NULL;
2378 * Release a drop hash Rx queue.
2381 * Pointer to Ethernet device.
2384 mlx5_drop_action_destroy(struct rte_eth_dev *dev)
2386 struct mlx5_priv *priv = dev->data->dev_private;
2387 struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
2389 if (__atomic_sub_fetch(&hrxq->refcnt, 1, __ATOMIC_RELAXED) == 0) {
2390 priv->obj_ops.drop_action_destroy(dev);
2391 mlx5_free(priv->drop_queue.rxq);
2392 mlx5_free(hrxq->ind_table);
2394 priv->drop_queue.rxq = NULL;
2395 priv->drop_queue.hrxq = NULL;
2400 * Verify the Rx Queue list is empty
2403 * Pointer to Ethernet device.
2406 * The number of object not released.
2409 mlx5_hrxq_verify(struct rte_eth_dev *dev)
2411 struct mlx5_priv *priv = dev->data->dev_private;
2413 return mlx5_cache_list_get_entry_num(&priv->hrxqs);
2417 * Set the Rx queue timestamp conversion parameters
2420 * Pointer to the Ethernet device structure.
2423 mlx5_rxq_timestamp_set(struct rte_eth_dev *dev)
2425 struct mlx5_priv *priv = dev->data->dev_private;
2426 struct mlx5_dev_ctx_shared *sh = priv->sh;
2427 struct mlx5_rxq_data *data;
2430 for (i = 0; i != priv->rxqs_n; ++i) {
2431 if (!(*priv->rxqs)[i])
2433 data = (*priv->rxqs)[i];
2435 data->rt_timestamp = priv->config.rt_timestamp;