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 static_assert(MLX5_RSS_HASH_KEY_LEN ==
56 (unsigned int)sizeof(rss_hash_default_key),
57 "wrong RSS default key size.");
60 * Check whether Multi-Packet RQ can be enabled for the device.
63 * Pointer to Ethernet device.
66 * 1 if supported, negative errno value if not.
69 mlx5_check_mprq_support(struct rte_eth_dev *dev)
71 struct mlx5_priv *priv = dev->data->dev_private;
73 if (priv->config.mprq.enabled &&
74 priv->rxqs_n >= priv->config.mprq.min_rxqs_num)
80 * Check whether Multi-Packet RQ is enabled for the Rx queue.
83 * Pointer to receive queue structure.
86 * 0 if disabled, otherwise enabled.
89 mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq)
91 return rxq->strd_num_n > 0;
95 * Check whether Multi-Packet RQ is enabled for the device.
98 * Pointer to Ethernet device.
101 * 0 if disabled, otherwise enabled.
104 mlx5_mprq_enabled(struct rte_eth_dev *dev)
106 struct mlx5_priv *priv = dev->data->dev_private;
110 if (mlx5_check_mprq_support(dev) < 0)
112 /* All the configured queues should be enabled. */
113 for (i = 0; i < priv->rxqs_n; ++i) {
114 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
118 if (mlx5_rxq_mprq_enabled(rxq))
121 /* Multi-Packet RQ can't be partially configured. */
122 assert(n == 0 || n == priv->rxqs_n);
123 return n == priv->rxqs_n;
127 * Allocate RX queue elements for Multi-Packet RQ.
130 * Pointer to RX queue structure.
133 * 0 on success, a negative errno value otherwise and rte_errno is set.
136 rxq_alloc_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
138 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
139 unsigned int wqe_n = 1 << rxq->elts_n;
143 /* Iterate on segments. */
144 for (i = 0; i <= wqe_n; ++i) {
145 struct mlx5_mprq_buf *buf;
147 if (rte_mempool_get(rxq->mprq_mp, (void **)&buf) < 0) {
148 DRV_LOG(ERR, "port %u empty mbuf pool", rxq->port_id);
153 (*rxq->mprq_bufs)[i] = buf;
155 rxq->mprq_repl = buf;
158 "port %u Rx queue %u allocated and configured %u segments",
159 rxq->port_id, rxq->idx, wqe_n);
162 err = rte_errno; /* Save rte_errno before cleanup. */
164 for (i = 0; (i != wqe_n); ++i) {
165 if ((*rxq->mprq_bufs)[i] != NULL)
166 rte_mempool_put(rxq->mprq_mp,
167 (*rxq->mprq_bufs)[i]);
168 (*rxq->mprq_bufs)[i] = NULL;
170 DRV_LOG(DEBUG, "port %u Rx queue %u failed, freed everything",
171 rxq->port_id, rxq->idx);
172 rte_errno = err; /* Restore rte_errno. */
177 * Allocate RX queue elements for Single-Packet RQ.
180 * Pointer to RX queue structure.
183 * 0 on success, errno value on failure.
186 rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
188 const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
189 unsigned int elts_n = 1 << rxq_ctrl->rxq.elts_n;
193 /* Iterate on segments. */
194 for (i = 0; (i != elts_n); ++i) {
195 struct rte_mbuf *buf;
197 buf = rte_pktmbuf_alloc(rxq_ctrl->rxq.mp);
199 DRV_LOG(ERR, "port %u empty mbuf pool",
200 PORT_ID(rxq_ctrl->priv));
204 /* Headroom is reserved by rte_pktmbuf_alloc(). */
205 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
206 /* Buffer is supposed to be empty. */
207 assert(rte_pktmbuf_data_len(buf) == 0);
208 assert(rte_pktmbuf_pkt_len(buf) == 0);
210 /* Only the first segment keeps headroom. */
212 SET_DATA_OFF(buf, 0);
213 PORT(buf) = rxq_ctrl->rxq.port_id;
214 DATA_LEN(buf) = rte_pktmbuf_tailroom(buf);
215 PKT_LEN(buf) = DATA_LEN(buf);
217 (*rxq_ctrl->rxq.elts)[i] = buf;
219 /* If Rx vector is activated. */
220 if (mlx5_rxq_check_vec_support(&rxq_ctrl->rxq) > 0) {
221 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
222 struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
225 /* Initialize default rearm_data for vPMD. */
226 mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
227 rte_mbuf_refcnt_set(mbuf_init, 1);
228 mbuf_init->nb_segs = 1;
229 mbuf_init->port = rxq->port_id;
231 * prevent compiler reordering:
232 * rearm_data covers previous fields.
234 rte_compiler_barrier();
235 rxq->mbuf_initializer =
236 *(uint64_t *)&mbuf_init->rearm_data;
237 /* Padding with a fake mbuf for vectorized Rx. */
238 for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
239 (*rxq->elts)[elts_n + j] = &rxq->fake_mbuf;
242 "port %u Rx queue %u allocated and configured %u segments"
244 PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx, elts_n,
245 elts_n / (1 << rxq_ctrl->rxq.sges_n));
248 err = rte_errno; /* Save rte_errno before cleanup. */
250 for (i = 0; (i != elts_n); ++i) {
251 if ((*rxq_ctrl->rxq.elts)[i] != NULL)
252 rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
253 (*rxq_ctrl->rxq.elts)[i] = NULL;
255 DRV_LOG(DEBUG, "port %u Rx queue %u failed, freed everything",
256 PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx);
257 rte_errno = err; /* Restore rte_errno. */
262 * Allocate RX queue elements.
265 * Pointer to RX queue structure.
268 * 0 on success, errno value on failure.
271 rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
273 return mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
274 rxq_alloc_elts_mprq(rxq_ctrl) : rxq_alloc_elts_sprq(rxq_ctrl);
278 * Free RX queue elements for Multi-Packet RQ.
281 * Pointer to RX queue structure.
284 rxq_free_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
286 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
289 DRV_LOG(DEBUG, "port %u Multi-Packet Rx queue %u freeing WRs",
290 rxq->port_id, rxq->idx);
291 if (rxq->mprq_bufs == NULL)
293 assert(mlx5_rxq_check_vec_support(rxq) < 0);
294 for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
295 if ((*rxq->mprq_bufs)[i] != NULL)
296 mlx5_mprq_buf_free((*rxq->mprq_bufs)[i]);
297 (*rxq->mprq_bufs)[i] = NULL;
299 if (rxq->mprq_repl != NULL) {
300 mlx5_mprq_buf_free(rxq->mprq_repl);
301 rxq->mprq_repl = NULL;
306 * Free RX queue elements for Single-Packet RQ.
309 * Pointer to RX queue structure.
312 rxq_free_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
314 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
315 const uint16_t q_n = (1 << rxq->elts_n);
316 const uint16_t q_mask = q_n - 1;
317 uint16_t used = q_n - (rxq->rq_ci - rxq->rq_pi);
320 DRV_LOG(DEBUG, "port %u Rx queue %u freeing WRs",
321 PORT_ID(rxq_ctrl->priv), rxq->idx);
322 if (rxq->elts == NULL)
325 * Some mbuf in the Ring belongs to the application. They cannot be
328 if (mlx5_rxq_check_vec_support(rxq) > 0) {
329 for (i = 0; i < used; ++i)
330 (*rxq->elts)[(rxq->rq_ci + i) & q_mask] = NULL;
331 rxq->rq_pi = rxq->rq_ci;
333 for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
334 if ((*rxq->elts)[i] != NULL)
335 rte_pktmbuf_free_seg((*rxq->elts)[i]);
336 (*rxq->elts)[i] = NULL;
341 * Free RX queue elements.
344 * Pointer to RX queue structure.
347 rxq_free_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
349 if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
350 rxq_free_elts_mprq(rxq_ctrl);
352 rxq_free_elts_sprq(rxq_ctrl);
356 * Returns the per-queue supported offloads.
359 * Pointer to Ethernet device.
362 * Supported Rx offloads.
365 mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
367 struct mlx5_priv *priv = dev->data->dev_private;
368 struct mlx5_dev_config *config = &priv->config;
369 uint64_t offloads = (DEV_RX_OFFLOAD_SCATTER |
370 DEV_RX_OFFLOAD_TIMESTAMP |
371 DEV_RX_OFFLOAD_JUMBO_FRAME);
373 if (config->hw_fcs_strip)
374 offloads |= DEV_RX_OFFLOAD_KEEP_CRC;
377 offloads |= (DEV_RX_OFFLOAD_IPV4_CKSUM |
378 DEV_RX_OFFLOAD_UDP_CKSUM |
379 DEV_RX_OFFLOAD_TCP_CKSUM);
380 if (config->hw_vlan_strip)
381 offloads |= DEV_RX_OFFLOAD_VLAN_STRIP;
387 * Returns the per-port supported offloads.
390 * Supported Rx offloads.
393 mlx5_get_rx_port_offloads(void)
395 uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
401 * Verify if the queue can be released.
404 * Pointer to Ethernet device.
409 * 1 if the queue can be released
410 * 0 if the queue can not be released, there are references to it.
411 * Negative errno and rte_errno is set if queue doesn't exist.
414 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
416 struct mlx5_priv *priv = dev->data->dev_private;
417 struct mlx5_rxq_ctrl *rxq_ctrl;
419 if (!(*priv->rxqs)[idx]) {
423 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
424 return (rte_atomic32_read(&rxq_ctrl->refcnt) == 1);
430 * Pointer to Ethernet device structure.
434 * Number of descriptors to configure in queue.
436 * NUMA socket on which memory must be allocated.
438 * Thresholds parameters.
440 * Memory pool for buffer allocations.
443 * 0 on success, a negative errno value otherwise and rte_errno is set.
446 mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
447 unsigned int socket, const struct rte_eth_rxconf *conf,
448 struct rte_mempool *mp)
450 struct mlx5_priv *priv = dev->data->dev_private;
451 struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
452 struct mlx5_rxq_ctrl *rxq_ctrl =
453 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
455 if (!rte_is_power_of_2(desc)) {
456 desc = 1 << log2above(desc);
458 "port %u increased number of descriptors in Rx queue %u"
459 " to the next power of two (%d)",
460 dev->data->port_id, idx, desc);
462 DRV_LOG(DEBUG, "port %u configuring Rx queue %u for %u descriptors",
463 dev->data->port_id, idx, desc);
464 if (idx >= priv->rxqs_n) {
465 DRV_LOG(ERR, "port %u Rx queue index out of range (%u >= %u)",
466 dev->data->port_id, idx, priv->rxqs_n);
467 rte_errno = EOVERFLOW;
470 if (!mlx5_rxq_releasable(dev, idx)) {
471 DRV_LOG(ERR, "port %u unable to release queue index %u",
472 dev->data->port_id, idx);
476 mlx5_rxq_release(dev, idx);
477 rxq_ctrl = mlx5_rxq_new(dev, idx, desc, socket, conf, mp);
479 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
480 dev->data->port_id, idx);
484 DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
485 dev->data->port_id, idx);
486 (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
491 * DPDK callback to release a RX queue.
494 * Generic RX queue pointer.
497 mlx5_rx_queue_release(void *dpdk_rxq)
499 struct mlx5_rxq_data *rxq = (struct mlx5_rxq_data *)dpdk_rxq;
500 struct mlx5_rxq_ctrl *rxq_ctrl;
501 struct mlx5_priv *priv;
505 rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
506 priv = rxq_ctrl->priv;
507 if (!mlx5_rxq_releasable(ETH_DEV(priv), rxq_ctrl->rxq.idx))
508 rte_panic("port %u Rx queue %u is still used by a flow and"
509 " cannot be removed\n",
510 PORT_ID(priv), rxq->idx);
511 mlx5_rxq_release(ETH_DEV(priv), rxq_ctrl->rxq.idx);
515 * Get an Rx queue Verbs object.
518 * Pointer to Ethernet device.
520 * Queue index in DPDK Rx queue array
523 * The Verbs object if it exists.
525 static struct mlx5_rxq_ibv *
526 mlx5_rxq_ibv_get(struct rte_eth_dev *dev, uint16_t idx)
528 struct mlx5_priv *priv = dev->data->dev_private;
529 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
530 struct mlx5_rxq_ctrl *rxq_ctrl;
532 if (idx >= priv->rxqs_n)
536 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
538 rte_atomic32_inc(&rxq_ctrl->ibv->refcnt);
539 return rxq_ctrl->ibv;
543 * Release an Rx verbs queue object.
546 * Verbs Rx queue object.
549 * 1 while a reference on it exists, 0 when freed.
552 mlx5_rxq_ibv_release(struct mlx5_rxq_ibv *rxq_ibv)
557 if (rte_atomic32_dec_and_test(&rxq_ibv->refcnt)) {
558 rxq_free_elts(rxq_ibv->rxq_ctrl);
559 claim_zero(mlx5_glue->destroy_wq(rxq_ibv->wq));
560 claim_zero(mlx5_glue->destroy_cq(rxq_ibv->cq));
561 if (rxq_ibv->channel)
562 claim_zero(mlx5_glue->destroy_comp_channel
564 LIST_REMOVE(rxq_ibv, next);
572 * Allocate queue vector and fill epoll fd list for Rx interrupts.
575 * Pointer to Ethernet device.
578 * 0 on success, a negative errno value otherwise and rte_errno is set.
581 mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
583 struct mlx5_priv *priv = dev->data->dev_private;
585 unsigned int rxqs_n = priv->rxqs_n;
586 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
587 unsigned int count = 0;
588 struct rte_intr_handle *intr_handle = dev->intr_handle;
590 if (!dev->data->dev_conf.intr_conf.rxq)
592 mlx5_rx_intr_vec_disable(dev);
593 intr_handle->intr_vec = malloc(n * sizeof(intr_handle->intr_vec[0]));
594 if (intr_handle->intr_vec == NULL) {
596 "port %u failed to allocate memory for interrupt"
597 " vector, Rx interrupts will not be supported",
602 intr_handle->type = RTE_INTR_HANDLE_EXT;
603 for (i = 0; i != n; ++i) {
604 /* This rxq ibv must not be released in this function. */
605 struct mlx5_rxq_ibv *rxq_ibv = mlx5_rxq_ibv_get(dev, i);
610 /* Skip queues that cannot request interrupts. */
611 if (!rxq_ibv || !rxq_ibv->channel) {
612 /* Use invalid intr_vec[] index to disable entry. */
613 intr_handle->intr_vec[i] =
614 RTE_INTR_VEC_RXTX_OFFSET +
615 RTE_MAX_RXTX_INTR_VEC_ID;
618 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
620 "port %u too many Rx queues for interrupt"
621 " vector size (%d), Rx interrupts cannot be"
623 dev->data->port_id, RTE_MAX_RXTX_INTR_VEC_ID);
624 mlx5_rx_intr_vec_disable(dev);
628 fd = rxq_ibv->channel->fd;
629 flags = fcntl(fd, F_GETFL);
630 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
634 "port %u failed to make Rx interrupt file"
635 " descriptor %d non-blocking for queue index"
637 dev->data->port_id, fd, i);
638 mlx5_rx_intr_vec_disable(dev);
641 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
642 intr_handle->efds[count] = fd;
646 mlx5_rx_intr_vec_disable(dev);
648 intr_handle->nb_efd = count;
653 * Clean up Rx interrupts handler.
656 * Pointer to Ethernet device.
659 mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
661 struct mlx5_priv *priv = dev->data->dev_private;
662 struct rte_intr_handle *intr_handle = dev->intr_handle;
664 unsigned int rxqs_n = priv->rxqs_n;
665 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
667 if (!dev->data->dev_conf.intr_conf.rxq)
669 if (!intr_handle->intr_vec)
671 for (i = 0; i != n; ++i) {
672 struct mlx5_rxq_ctrl *rxq_ctrl;
673 struct mlx5_rxq_data *rxq_data;
675 if (intr_handle->intr_vec[i] == RTE_INTR_VEC_RXTX_OFFSET +
676 RTE_MAX_RXTX_INTR_VEC_ID)
679 * Need to access directly the queue to release the reference
680 * kept in mlx5_rx_intr_vec_enable().
682 rxq_data = (*priv->rxqs)[i];
683 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
685 mlx5_rxq_ibv_release(rxq_ctrl->ibv);
688 rte_intr_free_epoll_fd(intr_handle);
689 if (intr_handle->intr_vec)
690 free(intr_handle->intr_vec);
691 intr_handle->nb_efd = 0;
692 intr_handle->intr_vec = NULL;
696 * MLX5 CQ notification .
699 * Pointer to receive queue structure.
701 * Sequence number per receive queue .
704 mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
707 uint32_t doorbell_hi;
709 void *cq_db_reg = (char *)rxq->cq_uar + MLX5_CQ_DOORBELL;
711 sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
712 doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
713 doorbell = (uint64_t)doorbell_hi << 32;
714 doorbell |= rxq->cqn;
715 rxq->cq_db[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
716 mlx5_uar_write64(rte_cpu_to_be_64(doorbell),
717 cq_db_reg, rxq->uar_lock_cq);
721 * DPDK callback for Rx queue interrupt enable.
724 * Pointer to Ethernet device structure.
729 * 0 on success, a negative errno value otherwise and rte_errno is set.
732 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
734 struct mlx5_priv *priv = dev->data->dev_private;
735 struct mlx5_rxq_data *rxq_data;
736 struct mlx5_rxq_ctrl *rxq_ctrl;
738 rxq_data = (*priv->rxqs)[rx_queue_id];
743 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
745 struct mlx5_rxq_ibv *rxq_ibv;
747 rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
752 mlx5_arm_cq(rxq_data, rxq_data->cq_arm_sn);
753 mlx5_rxq_ibv_release(rxq_ibv);
759 * DPDK callback for Rx queue interrupt disable.
762 * Pointer to Ethernet device structure.
767 * 0 on success, a negative errno value otherwise and rte_errno is set.
770 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
772 struct mlx5_priv *priv = dev->data->dev_private;
773 struct mlx5_rxq_data *rxq_data;
774 struct mlx5_rxq_ctrl *rxq_ctrl;
775 struct mlx5_rxq_ibv *rxq_ibv = NULL;
776 struct ibv_cq *ev_cq;
780 rxq_data = (*priv->rxqs)[rx_queue_id];
785 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
788 rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
793 ret = mlx5_glue->get_cq_event(rxq_ibv->channel, &ev_cq, &ev_ctx);
794 if (ret || ev_cq != rxq_ibv->cq) {
798 rxq_data->cq_arm_sn++;
799 mlx5_glue->ack_cq_events(rxq_ibv->cq, 1);
800 mlx5_rxq_ibv_release(rxq_ibv);
803 ret = rte_errno; /* Save rte_errno before cleanup. */
805 mlx5_rxq_ibv_release(rxq_ibv);
806 DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
807 dev->data->port_id, rx_queue_id);
808 rte_errno = ret; /* Restore rte_errno. */
813 * Create the Rx queue Verbs object.
816 * Pointer to Ethernet device.
818 * Queue index in DPDK Rx queue array
821 * The Verbs object initialised, NULL otherwise and rte_errno is set.
823 struct mlx5_rxq_ibv *
824 mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx)
826 struct mlx5_priv *priv = dev->data->dev_private;
827 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
828 struct mlx5_rxq_ctrl *rxq_ctrl =
829 container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
830 struct ibv_wq_attr mod;
833 struct ibv_cq_init_attr_ex ibv;
834 struct mlx5dv_cq_init_attr mlx5;
837 struct ibv_wq_init_attr ibv;
838 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
839 struct mlx5dv_wq_init_attr mlx5;
844 unsigned int wqe_n = 1 << rxq_data->elts_n;
845 struct mlx5_rxq_ibv *tmpl = NULL;
846 struct mlx5dv_cq cq_info;
847 struct mlx5dv_rwq rwq;
849 struct mlx5dv_obj obj;
850 struct mlx5_dev_config *config = &priv->config;
851 const int mprq_en = mlx5_rxq_mprq_enabled(rxq_data);
854 assert(!rxq_ctrl->ibv);
855 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_RX_QUEUE;
856 priv->verbs_alloc_ctx.obj = rxq_ctrl;
857 tmpl = rte_calloc_socket(__func__, 1, sizeof(*tmpl), 0,
861 "port %u Rx queue %u cannot allocate verbs resources",
862 dev->data->port_id, rxq_data->idx);
866 tmpl->rxq_ctrl = rxq_ctrl;
868 tmpl->channel = mlx5_glue->create_comp_channel(priv->sh->ctx);
869 if (!tmpl->channel) {
870 DRV_LOG(ERR, "port %u: comp channel creation failure",
877 cqe_n = wqe_n * (1 << rxq_data->strd_num_n) - 1;
880 attr.cq.ibv = (struct ibv_cq_init_attr_ex){
882 .channel = tmpl->channel,
885 attr.cq.mlx5 = (struct mlx5dv_cq_init_attr){
888 if (config->cqe_comp && !rxq_data->hw_timestamp) {
889 attr.cq.mlx5.comp_mask |=
890 MLX5DV_CQ_INIT_ATTR_MASK_COMPRESSED_CQE;
891 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
892 attr.cq.mlx5.cqe_comp_res_format =
893 mprq_en ? MLX5DV_CQE_RES_FORMAT_CSUM_STRIDX :
894 MLX5DV_CQE_RES_FORMAT_HASH;
896 attr.cq.mlx5.cqe_comp_res_format = MLX5DV_CQE_RES_FORMAT_HASH;
899 * For vectorized Rx, it must not be doubled in order to
900 * make cq_ci and rq_ci aligned.
902 if (mlx5_rxq_check_vec_support(rxq_data) < 0)
903 attr.cq.ibv.cqe *= 2;
904 } else if (config->cqe_comp && rxq_data->hw_timestamp) {
906 "port %u Rx CQE compression is disabled for HW"
910 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
911 if (config->cqe_pad) {
912 attr.cq.mlx5.comp_mask |= MLX5DV_CQ_INIT_ATTR_MASK_FLAGS;
913 attr.cq.mlx5.flags |= MLX5DV_CQ_INIT_ATTR_FLAGS_CQE_PAD;
916 tmpl->cq = mlx5_glue->cq_ex_to_cq
917 (mlx5_glue->dv_create_cq(priv->sh->ctx, &attr.cq.ibv,
919 if (tmpl->cq == NULL) {
920 DRV_LOG(ERR, "port %u Rx queue %u CQ creation failure",
921 dev->data->port_id, idx);
925 DRV_LOG(DEBUG, "port %u device_attr.max_qp_wr is %d",
926 dev->data->port_id, priv->sh->device_attr.orig_attr.max_qp_wr);
927 DRV_LOG(DEBUG, "port %u device_attr.max_sge is %d",
928 dev->data->port_id, priv->sh->device_attr.orig_attr.max_sge);
929 attr.wq.ibv = (struct ibv_wq_init_attr){
930 .wq_context = NULL, /* Could be useful in the future. */
931 .wq_type = IBV_WQT_RQ,
932 /* Max number of outstanding WRs. */
933 .max_wr = wqe_n >> rxq_data->sges_n,
934 /* Max number of scatter/gather elements in a WR. */
935 .max_sge = 1 << rxq_data->sges_n,
939 IBV_WQ_FLAGS_CVLAN_STRIPPING |
941 .create_flags = (rxq_data->vlan_strip ?
942 IBV_WQ_FLAGS_CVLAN_STRIPPING :
945 /* By default, FCS (CRC) is stripped by hardware. */
946 if (rxq_data->crc_present) {
947 attr.wq.ibv.create_flags |= IBV_WQ_FLAGS_SCATTER_FCS;
948 attr.wq.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
950 if (config->hw_padding) {
951 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
952 attr.wq.ibv.create_flags |= IBV_WQ_FLAG_RX_END_PADDING;
953 attr.wq.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
954 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
955 attr.wq.ibv.create_flags |= IBV_WQ_FLAGS_PCI_WRITE_END_PADDING;
956 attr.wq.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
959 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
960 attr.wq.mlx5 = (struct mlx5dv_wq_init_attr){
964 struct mlx5dv_striding_rq_init_attr *mprq_attr =
965 &attr.wq.mlx5.striding_rq_attrs;
967 attr.wq.mlx5.comp_mask |= MLX5DV_WQ_INIT_ATTR_MASK_STRIDING_RQ;
968 *mprq_attr = (struct mlx5dv_striding_rq_init_attr){
969 .single_stride_log_num_of_bytes = rxq_data->strd_sz_n,
970 .single_wqe_log_num_of_strides = rxq_data->strd_num_n,
971 .two_byte_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT,
974 tmpl->wq = mlx5_glue->dv_create_wq(priv->sh->ctx, &attr.wq.ibv,
977 tmpl->wq = mlx5_glue->create_wq(priv->sh->ctx, &attr.wq.ibv);
979 if (tmpl->wq == NULL) {
980 DRV_LOG(ERR, "port %u Rx queue %u WQ creation failure",
981 dev->data->port_id, idx);
986 * Make sure number of WRs*SGEs match expectations since a queue
987 * cannot allocate more than "desc" buffers.
989 if (attr.wq.ibv.max_wr != (wqe_n >> rxq_data->sges_n) ||
990 attr.wq.ibv.max_sge != (1u << rxq_data->sges_n)) {
992 "port %u Rx queue %u requested %u*%u but got %u*%u"
994 dev->data->port_id, idx,
995 wqe_n >> rxq_data->sges_n, (1 << rxq_data->sges_n),
996 attr.wq.ibv.max_wr, attr.wq.ibv.max_sge);
1000 /* Change queue state to ready. */
1001 mod = (struct ibv_wq_attr){
1002 .attr_mask = IBV_WQ_ATTR_STATE,
1003 .wq_state = IBV_WQS_RDY,
1005 ret = mlx5_glue->modify_wq(tmpl->wq, &mod);
1008 "port %u Rx queue %u WQ state to IBV_WQS_RDY failed",
1009 dev->data->port_id, idx);
1013 obj.cq.in = tmpl->cq;
1014 obj.cq.out = &cq_info;
1015 obj.rwq.in = tmpl->wq;
1017 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_RWQ);
1022 if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
1024 "port %u wrong MLX5_CQE_SIZE environment variable"
1025 " value: it should be set to %u",
1026 dev->data->port_id, RTE_CACHE_LINE_SIZE);
1030 /* Fill the rings. */
1031 rxq_data->wqes = rwq.buf;
1032 rxq_data->rq_db = rwq.dbrec;
1033 rxq_data->cqe_n = log2above(cq_info.cqe_cnt);
1034 rxq_data->cq_db = cq_info.dbrec;
1035 rxq_data->cqes = (volatile struct mlx5_cqe (*)[])(uintptr_t)cq_info.buf;
1036 rxq_data->cq_uar = cq_info.cq_uar;
1037 rxq_data->cqn = cq_info.cqn;
1038 rxq_data->cq_arm_sn = 0;
1039 mlx5_rxq_initialize(rxq_data);
1040 rxq_data->cq_ci = 0;
1041 DRV_LOG(DEBUG, "port %u rxq %u updated with %p", dev->data->port_id,
1042 idx, (void *)&tmpl);
1043 rte_atomic32_inc(&tmpl->refcnt);
1044 LIST_INSERT_HEAD(&priv->rxqsibv, tmpl, next);
1045 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1049 ret = rte_errno; /* Save rte_errno before cleanup. */
1051 claim_zero(mlx5_glue->destroy_wq(tmpl->wq));
1053 claim_zero(mlx5_glue->destroy_cq(tmpl->cq));
1055 claim_zero(mlx5_glue->destroy_comp_channel
1058 rte_errno = ret; /* Restore rte_errno. */
1060 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1065 * Verify the Verbs Rx queue list is empty
1068 * Pointer to Ethernet device.
1071 * The number of object not released.
1074 mlx5_rxq_ibv_verify(struct rte_eth_dev *dev)
1076 struct mlx5_priv *priv = dev->data->dev_private;
1078 struct mlx5_rxq_ibv *rxq_ibv;
1080 LIST_FOREACH(rxq_ibv, &priv->rxqsibv, next) {
1081 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u still referenced",
1082 dev->data->port_id, rxq_ibv->rxq_ctrl->rxq.idx);
1089 * Callback function to initialize mbufs for Multi-Packet RQ.
1092 mlx5_mprq_buf_init(struct rte_mempool *mp, void *opaque_arg __rte_unused,
1093 void *_m, unsigned int i __rte_unused)
1095 struct mlx5_mprq_buf *buf = _m;
1097 memset(_m, 0, sizeof(*buf));
1099 rte_atomic16_set(&buf->refcnt, 1);
1103 * Free mempool of Multi-Packet RQ.
1106 * Pointer to Ethernet device.
1109 * 0 on success, negative errno value on failure.
1112 mlx5_mprq_free_mp(struct rte_eth_dev *dev)
1114 struct mlx5_priv *priv = dev->data->dev_private;
1115 struct rte_mempool *mp = priv->mprq_mp;
1120 DRV_LOG(DEBUG, "port %u freeing mempool (%s) for Multi-Packet RQ",
1121 dev->data->port_id, mp->name);
1123 * If a buffer in the pool has been externally attached to a mbuf and it
1124 * is still in use by application, destroying the Rx queue can spoil
1125 * the packet. It is unlikely to happen but if application dynamically
1126 * creates and destroys with holding Rx packets, this can happen.
1128 * TODO: It is unavoidable for now because the mempool for Multi-Packet
1129 * RQ isn't provided by application but managed by PMD.
1131 if (!rte_mempool_full(mp)) {
1133 "port %u mempool for Multi-Packet RQ is still in use",
1134 dev->data->port_id);
1138 rte_mempool_free(mp);
1139 /* Unset mempool for each Rx queue. */
1140 for (i = 0; i != priv->rxqs_n; ++i) {
1141 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1145 rxq->mprq_mp = NULL;
1147 priv->mprq_mp = NULL;
1152 * Allocate a mempool for Multi-Packet RQ. All configured Rx queues share the
1153 * mempool. If already allocated, reuse it if there're enough elements.
1154 * Otherwise, resize it.
1157 * Pointer to Ethernet device.
1160 * 0 on success, negative errno value on failure.
1163 mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
1165 struct mlx5_priv *priv = dev->data->dev_private;
1166 struct rte_mempool *mp = priv->mprq_mp;
1167 char name[RTE_MEMPOOL_NAMESIZE];
1168 unsigned int desc = 0;
1169 unsigned int buf_len;
1170 unsigned int obj_num;
1171 unsigned int obj_size;
1172 unsigned int strd_num_n = 0;
1173 unsigned int strd_sz_n = 0;
1176 if (!mlx5_mprq_enabled(dev))
1178 /* Count the total number of descriptors configured. */
1179 for (i = 0; i != priv->rxqs_n; ++i) {
1180 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1184 desc += 1 << rxq->elts_n;
1185 /* Get the max number of strides. */
1186 if (strd_num_n < rxq->strd_num_n)
1187 strd_num_n = rxq->strd_num_n;
1188 /* Get the max size of a stride. */
1189 if (strd_sz_n < rxq->strd_sz_n)
1190 strd_sz_n = rxq->strd_sz_n;
1192 assert(strd_num_n && strd_sz_n);
1193 buf_len = (1 << strd_num_n) * (1 << strd_sz_n);
1194 obj_size = buf_len + sizeof(struct mlx5_mprq_buf);
1196 * Received packets can be either memcpy'd or externally referenced. In
1197 * case that the packet is attached to an mbuf as an external buffer, as
1198 * it isn't possible to predict how the buffers will be queued by
1199 * application, there's no option to exactly pre-allocate needed buffers
1200 * in advance but to speculatively prepares enough buffers.
1202 * In the data path, if this Mempool is depleted, PMD will try to memcpy
1203 * received packets to buffers provided by application (rxq->mp) until
1204 * this Mempool gets available again.
1207 obj_num = desc + MLX5_MPRQ_MP_CACHE_SZ * priv->rxqs_n;
1209 * rte_mempool_create_empty() has sanity check to refuse large cache
1210 * size compared to the number of elements.
1211 * CACHE_FLUSHTHRESH_MULTIPLIER is defined in a C file, so using a
1212 * constant number 2 instead.
1214 obj_num = RTE_MAX(obj_num, MLX5_MPRQ_MP_CACHE_SZ * 2);
1215 /* Check a mempool is already allocated and if it can be resued. */
1216 if (mp != NULL && mp->elt_size >= obj_size && mp->size >= obj_num) {
1217 DRV_LOG(DEBUG, "port %u mempool %s is being reused",
1218 dev->data->port_id, mp->name);
1221 } else if (mp != NULL) {
1222 DRV_LOG(DEBUG, "port %u mempool %s should be resized, freeing it",
1223 dev->data->port_id, mp->name);
1225 * If failed to free, which means it may be still in use, no way
1226 * but to keep using the existing one. On buffer underrun,
1227 * packets will be memcpy'd instead of external buffer
1230 if (mlx5_mprq_free_mp(dev)) {
1231 if (mp->elt_size >= obj_size)
1237 snprintf(name, sizeof(name), "port-%u-mprq", dev->data->port_id);
1238 mp = rte_mempool_create(name, obj_num, obj_size, MLX5_MPRQ_MP_CACHE_SZ,
1239 0, NULL, NULL, mlx5_mprq_buf_init, NULL,
1240 dev->device->numa_node, 0);
1243 "port %u failed to allocate a mempool for"
1244 " Multi-Packet RQ, count=%u, size=%u",
1245 dev->data->port_id, obj_num, obj_size);
1251 /* Set mempool for each Rx queue. */
1252 for (i = 0; i != priv->rxqs_n; ++i) {
1253 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1259 DRV_LOG(INFO, "port %u Multi-Packet RQ is configured",
1260 dev->data->port_id);
1265 * Create a DPDK Rx queue.
1268 * Pointer to Ethernet device.
1272 * Number of descriptors to configure in queue.
1274 * NUMA socket on which memory must be allocated.
1277 * A DPDK queue object on success, NULL otherwise and rte_errno is set.
1279 struct mlx5_rxq_ctrl *
1280 mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1281 unsigned int socket, const struct rte_eth_rxconf *conf,
1282 struct rte_mempool *mp)
1284 struct mlx5_priv *priv = dev->data->dev_private;
1285 struct mlx5_rxq_ctrl *tmpl;
1286 unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
1287 unsigned int mprq_stride_size;
1288 struct mlx5_dev_config *config = &priv->config;
1290 * Always allocate extra slots, even if eventually
1291 * the vector Rx will not be used.
1294 desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
1295 uint64_t offloads = conf->offloads |
1296 dev->data->dev_conf.rxmode.offloads;
1297 const int mprq_en = mlx5_check_mprq_support(dev) > 0;
1299 tmpl = rte_calloc_socket("RXQ", 1,
1301 desc_n * sizeof(struct rte_mbuf *),
1307 if (mlx5_mr_btree_init(&tmpl->rxq.mr_ctrl.cache_bh,
1308 MLX5_MR_BTREE_CACHE_N, socket)) {
1309 /* rte_errno is already set. */
1312 tmpl->socket = socket;
1313 if (dev->data->dev_conf.intr_conf.rxq)
1316 * This Rx queue can be configured as a Multi-Packet RQ if all of the
1317 * following conditions are met:
1318 * - MPRQ is enabled.
1319 * - The number of descs is more than the number of strides.
1320 * - max_rx_pkt_len plus overhead is less than the max size of a
1322 * Otherwise, enable Rx scatter if necessary.
1324 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
1326 dev->data->dev_conf.rxmode.max_rx_pkt_len +
1327 sizeof(struct rte_mbuf_ext_shared_info) +
1328 RTE_PKTMBUF_HEADROOM;
1330 desc > (1U << config->mprq.stride_num_n) &&
1331 mprq_stride_size <= (1U << config->mprq.max_stride_size_n)) {
1332 /* TODO: Rx scatter isn't supported yet. */
1333 tmpl->rxq.sges_n = 0;
1334 /* Trim the number of descs needed. */
1335 desc >>= config->mprq.stride_num_n;
1336 tmpl->rxq.strd_num_n = config->mprq.stride_num_n;
1337 tmpl->rxq.strd_sz_n = RTE_MAX(log2above(mprq_stride_size),
1338 config->mprq.min_stride_size_n);
1339 tmpl->rxq.strd_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT;
1340 tmpl->rxq.mprq_max_memcpy_len =
1341 RTE_MIN(mb_len - RTE_PKTMBUF_HEADROOM,
1342 config->mprq.max_memcpy_len);
1344 "port %u Rx queue %u: Multi-Packet RQ is enabled"
1345 " strd_num_n = %u, strd_sz_n = %u",
1346 dev->data->port_id, idx,
1347 tmpl->rxq.strd_num_n, tmpl->rxq.strd_sz_n);
1348 } else if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
1349 (mb_len - RTE_PKTMBUF_HEADROOM)) {
1350 tmpl->rxq.sges_n = 0;
1351 } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
1353 RTE_PKTMBUF_HEADROOM +
1354 dev->data->dev_conf.rxmode.max_rx_pkt_len;
1355 unsigned int sges_n;
1358 * Determine the number of SGEs needed for a full packet
1359 * and round it to the next power of two.
1361 sges_n = log2above((size / mb_len) + !!(size % mb_len));
1362 tmpl->rxq.sges_n = sges_n;
1363 /* Make sure rxq.sges_n did not overflow. */
1364 size = mb_len * (1 << tmpl->rxq.sges_n);
1365 size -= RTE_PKTMBUF_HEADROOM;
1366 if (size < dev->data->dev_conf.rxmode.max_rx_pkt_len) {
1368 "port %u too many SGEs (%u) needed to handle"
1369 " requested maximum packet size %u",
1372 dev->data->dev_conf.rxmode.max_rx_pkt_len);
1373 rte_errno = EOVERFLOW;
1378 "port %u the requested maximum Rx packet size (%u) is"
1379 " larger than a single mbuf (%u) and scattered mode has"
1380 " not been requested",
1382 dev->data->dev_conf.rxmode.max_rx_pkt_len,
1383 mb_len - RTE_PKTMBUF_HEADROOM);
1385 if (mprq_en && !mlx5_rxq_mprq_enabled(&tmpl->rxq))
1387 "port %u MPRQ is requested but cannot be enabled"
1388 " (requested: desc = %u, stride_sz = %u,"
1389 " supported: min_stride_num = %u, max_stride_sz = %u).",
1390 dev->data->port_id, desc, mprq_stride_size,
1391 (1 << config->mprq.stride_num_n),
1392 (1 << config->mprq.max_stride_size_n));
1393 DRV_LOG(DEBUG, "port %u maximum number of segments per packet: %u",
1394 dev->data->port_id, 1 << tmpl->rxq.sges_n);
1395 if (desc % (1 << tmpl->rxq.sges_n)) {
1397 "port %u number of Rx queue descriptors (%u) is not a"
1398 " multiple of SGEs per packet (%u)",
1401 1 << tmpl->rxq.sges_n);
1405 /* Toggle RX checksum offload if hardware supports it. */
1406 tmpl->rxq.csum = !!(offloads & DEV_RX_OFFLOAD_CHECKSUM);
1407 tmpl->rxq.hw_timestamp = !!(offloads & DEV_RX_OFFLOAD_TIMESTAMP);
1408 /* Configure VLAN stripping. */
1409 tmpl->rxq.vlan_strip = !!(offloads & DEV_RX_OFFLOAD_VLAN_STRIP);
1410 /* By default, FCS (CRC) is stripped by hardware. */
1411 tmpl->rxq.crc_present = 0;
1412 if (offloads & DEV_RX_OFFLOAD_KEEP_CRC) {
1413 if (config->hw_fcs_strip) {
1414 tmpl->rxq.crc_present = 1;
1417 "port %u CRC stripping has been disabled but will"
1418 " still be performed by hardware, make sure MLNX_OFED"
1419 " and firmware are up to date",
1420 dev->data->port_id);
1424 "port %u CRC stripping is %s, %u bytes will be subtracted from"
1425 " incoming frames to hide it",
1427 tmpl->rxq.crc_present ? "disabled" : "enabled",
1428 tmpl->rxq.crc_present << 2);
1430 tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
1431 (!!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS));
1432 tmpl->rxq.port_id = dev->data->port_id;
1435 tmpl->rxq.elts_n = log2above(desc);
1436 tmpl->rxq.rq_repl_thresh =
1437 MLX5_VPMD_RXQ_RPLNSH_THRESH(1 << tmpl->rxq.elts_n);
1439 (struct rte_mbuf *(*)[1 << tmpl->rxq.elts_n])(tmpl + 1);
1441 tmpl->rxq.uar_lock_cq = &priv->uar_lock_cq;
1443 tmpl->rxq.idx = idx;
1444 rte_atomic32_inc(&tmpl->refcnt);
1445 LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1456 * Pointer to Ethernet device.
1461 * A pointer to the queue if it exists, NULL otherwise.
1463 struct mlx5_rxq_ctrl *
1464 mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
1466 struct mlx5_priv *priv = dev->data->dev_private;
1467 struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1469 if ((*priv->rxqs)[idx]) {
1470 rxq_ctrl = container_of((*priv->rxqs)[idx],
1471 struct mlx5_rxq_ctrl,
1473 mlx5_rxq_ibv_get(dev, idx);
1474 rte_atomic32_inc(&rxq_ctrl->refcnt);
1480 * Release a Rx queue.
1483 * Pointer to Ethernet device.
1488 * 1 while a reference on it exists, 0 when freed.
1491 mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
1493 struct mlx5_priv *priv = dev->data->dev_private;
1494 struct mlx5_rxq_ctrl *rxq_ctrl;
1496 if (!(*priv->rxqs)[idx])
1498 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1499 assert(rxq_ctrl->priv);
1500 if (rxq_ctrl->ibv && !mlx5_rxq_ibv_release(rxq_ctrl->ibv))
1501 rxq_ctrl->ibv = NULL;
1502 if (rte_atomic32_dec_and_test(&rxq_ctrl->refcnt)) {
1503 mlx5_mr_btree_free(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
1504 LIST_REMOVE(rxq_ctrl, next);
1506 (*priv->rxqs)[idx] = NULL;
1513 * Verify the Rx Queue list is empty
1516 * Pointer to Ethernet device.
1519 * The number of object not released.
1522 mlx5_rxq_verify(struct rte_eth_dev *dev)
1524 struct mlx5_priv *priv = dev->data->dev_private;
1525 struct mlx5_rxq_ctrl *rxq_ctrl;
1528 LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
1529 DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
1530 dev->data->port_id, rxq_ctrl->rxq.idx);
1537 * Create an indirection table.
1540 * Pointer to Ethernet device.
1542 * Queues entering in the indirection table.
1544 * Number of queues in the array.
1547 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1549 static struct mlx5_ind_table_ibv *
1550 mlx5_ind_table_ibv_new(struct rte_eth_dev *dev, const uint16_t *queues,
1553 struct mlx5_priv *priv = dev->data->dev_private;
1554 struct mlx5_ind_table_ibv *ind_tbl;
1555 const unsigned int wq_n = rte_is_power_of_2(queues_n) ?
1556 log2above(queues_n) :
1557 log2above(priv->config.ind_table_max_size);
1558 struct ibv_wq *wq[1 << wq_n];
1562 ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl) +
1563 queues_n * sizeof(uint16_t), 0);
1568 for (i = 0; i != queues_n; ++i) {
1569 struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev, queues[i]);
1573 wq[i] = rxq->ibv->wq;
1574 ind_tbl->queues[i] = queues[i];
1576 ind_tbl->queues_n = queues_n;
1577 /* Finalise indirection table. */
1578 for (j = 0; i != (unsigned int)(1 << wq_n); ++i, ++j)
1580 ind_tbl->ind_table = mlx5_glue->create_rwq_ind_table
1582 &(struct ibv_rwq_ind_table_init_attr){
1583 .log_ind_tbl_size = wq_n,
1587 if (!ind_tbl->ind_table) {
1591 rte_atomic32_inc(&ind_tbl->refcnt);
1592 LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
1596 DEBUG("port %u cannot create indirection table", dev->data->port_id);
1601 * Get an indirection table.
1604 * Pointer to Ethernet device.
1606 * Queues entering in the indirection table.
1608 * Number of queues in the array.
1611 * An indirection table if found.
1613 static struct mlx5_ind_table_ibv *
1614 mlx5_ind_table_ibv_get(struct rte_eth_dev *dev, const uint16_t *queues,
1617 struct mlx5_priv *priv = dev->data->dev_private;
1618 struct mlx5_ind_table_ibv *ind_tbl;
1620 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1621 if ((ind_tbl->queues_n == queues_n) &&
1622 (memcmp(ind_tbl->queues, queues,
1623 ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
1630 rte_atomic32_inc(&ind_tbl->refcnt);
1631 for (i = 0; i != ind_tbl->queues_n; ++i)
1632 mlx5_rxq_get(dev, ind_tbl->queues[i]);
1638 * Release an indirection table.
1641 * Pointer to Ethernet device.
1643 * Indirection table to release.
1646 * 1 while a reference on it exists, 0 when freed.
1649 mlx5_ind_table_ibv_release(struct rte_eth_dev *dev,
1650 struct mlx5_ind_table_ibv *ind_tbl)
1654 if (rte_atomic32_dec_and_test(&ind_tbl->refcnt))
1655 claim_zero(mlx5_glue->destroy_rwq_ind_table
1656 (ind_tbl->ind_table));
1657 for (i = 0; i != ind_tbl->queues_n; ++i)
1658 claim_nonzero(mlx5_rxq_release(dev, ind_tbl->queues[i]));
1659 if (!rte_atomic32_read(&ind_tbl->refcnt)) {
1660 LIST_REMOVE(ind_tbl, next);
1668 * Verify the Rx Queue list is empty
1671 * Pointer to Ethernet device.
1674 * The number of object not released.
1677 mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev)
1679 struct mlx5_priv *priv = dev->data->dev_private;
1680 struct mlx5_ind_table_ibv *ind_tbl;
1683 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1685 "port %u Verbs indirection table %p still referenced",
1686 dev->data->port_id, (void *)ind_tbl);
1693 * Create an Rx Hash queue.
1696 * Pointer to Ethernet device.
1698 * RSS key for the Rx hash queue.
1699 * @param rss_key_len
1701 * @param hash_fields
1702 * Verbs protocol hash field to make the RSS on.
1704 * Queues entering in hash queue. In case of empty hash_fields only the
1705 * first queue index will be taken for the indirection table.
1712 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1715 mlx5_hrxq_new(struct rte_eth_dev *dev,
1716 const uint8_t *rss_key, uint32_t rss_key_len,
1717 uint64_t hash_fields,
1718 const uint16_t *queues, uint32_t queues_n,
1719 int tunnel __rte_unused)
1721 struct mlx5_priv *priv = dev->data->dev_private;
1722 struct mlx5_hrxq *hrxq;
1723 struct mlx5_ind_table_ibv *ind_tbl;
1725 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1726 struct mlx5dv_qp_init_attr qp_init_attr;
1730 queues_n = hash_fields ? queues_n : 1;
1731 ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1733 ind_tbl = mlx5_ind_table_ibv_new(dev, queues, queues_n);
1738 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1739 memset(&qp_init_attr, 0, sizeof(qp_init_attr));
1741 qp_init_attr.comp_mask =
1742 MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
1743 qp_init_attr.create_flags = MLX5DV_QP_CREATE_TUNNEL_OFFLOADS;
1745 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1746 if (dev->data->dev_conf.lpbk_mode) {
1747 /* Allow packet sent from NIC loop back w/o source MAC check. */
1748 qp_init_attr.comp_mask |=
1749 MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
1750 qp_init_attr.create_flags |=
1751 MLX5DV_QP_CREATE_TIR_ALLOW_SELF_LOOPBACK_UC;
1754 qp = mlx5_glue->dv_create_qp
1756 &(struct ibv_qp_init_attr_ex){
1757 .qp_type = IBV_QPT_RAW_PACKET,
1759 IBV_QP_INIT_ATTR_PD |
1760 IBV_QP_INIT_ATTR_IND_TABLE |
1761 IBV_QP_INIT_ATTR_RX_HASH,
1762 .rx_hash_conf = (struct ibv_rx_hash_conf){
1763 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1764 .rx_hash_key_len = rss_key_len,
1765 .rx_hash_key = (void *)(uintptr_t)rss_key,
1766 .rx_hash_fields_mask = hash_fields,
1768 .rwq_ind_tbl = ind_tbl->ind_table,
1773 qp = mlx5_glue->create_qp_ex
1775 &(struct ibv_qp_init_attr_ex){
1776 .qp_type = IBV_QPT_RAW_PACKET,
1778 IBV_QP_INIT_ATTR_PD |
1779 IBV_QP_INIT_ATTR_IND_TABLE |
1780 IBV_QP_INIT_ATTR_RX_HASH,
1781 .rx_hash_conf = (struct ibv_rx_hash_conf){
1782 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1783 .rx_hash_key_len = rss_key_len,
1784 .rx_hash_key = (void *)(uintptr_t)rss_key,
1785 .rx_hash_fields_mask = hash_fields,
1787 .rwq_ind_tbl = ind_tbl->ind_table,
1795 hrxq = rte_calloc(__func__, 1, sizeof(*hrxq) + rss_key_len, 0);
1798 hrxq->ind_table = ind_tbl;
1800 hrxq->rss_key_len = rss_key_len;
1801 hrxq->hash_fields = hash_fields;
1802 memcpy(hrxq->rss_key, rss_key, rss_key_len);
1803 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1804 hrxq->action = mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
1805 if (!hrxq->action) {
1810 rte_atomic32_inc(&hrxq->refcnt);
1811 LIST_INSERT_HEAD(&priv->hrxqs, hrxq, next);
1814 err = rte_errno; /* Save rte_errno before cleanup. */
1815 mlx5_ind_table_ibv_release(dev, ind_tbl);
1817 claim_zero(mlx5_glue->destroy_qp(qp));
1818 rte_errno = err; /* Restore rte_errno. */
1823 * Get an Rx Hash queue.
1826 * Pointer to Ethernet device.
1828 * RSS configuration for the Rx hash queue.
1830 * Queues entering in hash queue. In case of empty hash_fields only the
1831 * first queue index will be taken for the indirection table.
1836 * An hash Rx queue on success.
1839 mlx5_hrxq_get(struct rte_eth_dev *dev,
1840 const uint8_t *rss_key, uint32_t rss_key_len,
1841 uint64_t hash_fields,
1842 const uint16_t *queues, uint32_t queues_n)
1844 struct mlx5_priv *priv = dev->data->dev_private;
1845 struct mlx5_hrxq *hrxq;
1847 queues_n = hash_fields ? queues_n : 1;
1848 LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1849 struct mlx5_ind_table_ibv *ind_tbl;
1851 if (hrxq->rss_key_len != rss_key_len)
1853 if (memcmp(hrxq->rss_key, rss_key, rss_key_len))
1855 if (hrxq->hash_fields != hash_fields)
1857 ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1860 if (ind_tbl != hrxq->ind_table) {
1861 mlx5_ind_table_ibv_release(dev, ind_tbl);
1864 rte_atomic32_inc(&hrxq->refcnt);
1871 * Release the hash Rx queue.
1874 * Pointer to Ethernet device.
1876 * Pointer to Hash Rx queue to release.
1879 * 1 while a reference on it exists, 0 when freed.
1882 mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
1884 if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
1885 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
1886 mlx5_glue->destroy_flow_action(hrxq->action);
1888 claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
1889 mlx5_ind_table_ibv_release(dev, hrxq->ind_table);
1890 LIST_REMOVE(hrxq, next);
1894 claim_nonzero(mlx5_ind_table_ibv_release(dev, hrxq->ind_table));
1899 * Verify the Rx Queue list is empty
1902 * Pointer to Ethernet device.
1905 * The number of object not released.
1908 mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev)
1910 struct mlx5_priv *priv = dev->data->dev_private;
1911 struct mlx5_hrxq *hrxq;
1914 LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1916 "port %u Verbs hash Rx queue %p still referenced",
1917 dev->data->port_id, (void *)hrxq);
1924 * Create a drop Rx queue Verbs object.
1927 * Pointer to Ethernet device.
1930 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1932 static struct mlx5_rxq_ibv *
1933 mlx5_rxq_ibv_drop_new(struct rte_eth_dev *dev)
1935 struct mlx5_priv *priv = dev->data->dev_private;
1936 struct ibv_context *ctx = priv->sh->ctx;
1938 struct ibv_wq *wq = NULL;
1939 struct mlx5_rxq_ibv *rxq;
1941 if (priv->drop_queue.rxq)
1942 return priv->drop_queue.rxq;
1943 cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
1945 DEBUG("port %u cannot allocate CQ for drop queue",
1946 dev->data->port_id);
1950 wq = mlx5_glue->create_wq(ctx,
1951 &(struct ibv_wq_init_attr){
1952 .wq_type = IBV_WQT_RQ,
1959 DEBUG("port %u cannot allocate WQ for drop queue",
1960 dev->data->port_id);
1964 rxq = rte_calloc(__func__, 1, sizeof(*rxq), 0);
1966 DEBUG("port %u cannot allocate drop Rx queue memory",
1967 dev->data->port_id);
1973 priv->drop_queue.rxq = rxq;
1977 claim_zero(mlx5_glue->destroy_wq(wq));
1979 claim_zero(mlx5_glue->destroy_cq(cq));
1984 * Release a drop Rx queue Verbs object.
1987 * Pointer to Ethernet device.
1990 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1993 mlx5_rxq_ibv_drop_release(struct rte_eth_dev *dev)
1995 struct mlx5_priv *priv = dev->data->dev_private;
1996 struct mlx5_rxq_ibv *rxq = priv->drop_queue.rxq;
1999 claim_zero(mlx5_glue->destroy_wq(rxq->wq));
2001 claim_zero(mlx5_glue->destroy_cq(rxq->cq));
2003 priv->drop_queue.rxq = NULL;
2007 * Create a drop indirection table.
2010 * Pointer to Ethernet device.
2013 * The Verbs object initialised, NULL otherwise and rte_errno is set.
2015 static struct mlx5_ind_table_ibv *
2016 mlx5_ind_table_ibv_drop_new(struct rte_eth_dev *dev)
2018 struct mlx5_priv *priv = dev->data->dev_private;
2019 struct mlx5_ind_table_ibv *ind_tbl;
2020 struct mlx5_rxq_ibv *rxq;
2021 struct mlx5_ind_table_ibv tmpl;
2023 rxq = mlx5_rxq_ibv_drop_new(dev);
2026 tmpl.ind_table = mlx5_glue->create_rwq_ind_table
2028 &(struct ibv_rwq_ind_table_init_attr){
2029 .log_ind_tbl_size = 0,
2030 .ind_tbl = &rxq->wq,
2033 if (!tmpl.ind_table) {
2034 DEBUG("port %u cannot allocate indirection table for drop"
2036 dev->data->port_id);
2040 ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl), 0);
2045 ind_tbl->ind_table = tmpl.ind_table;
2048 mlx5_rxq_ibv_drop_release(dev);
2053 * Release a drop indirection table.
2056 * Pointer to Ethernet device.
2059 mlx5_ind_table_ibv_drop_release(struct rte_eth_dev *dev)
2061 struct mlx5_priv *priv = dev->data->dev_private;
2062 struct mlx5_ind_table_ibv *ind_tbl = priv->drop_queue.hrxq->ind_table;
2064 claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl->ind_table));
2065 mlx5_rxq_ibv_drop_release(dev);
2067 priv->drop_queue.hrxq->ind_table = NULL;
2071 * Create a drop Rx Hash queue.
2074 * Pointer to Ethernet device.
2077 * The Verbs object initialised, NULL otherwise and rte_errno is set.
2080 mlx5_hrxq_drop_new(struct rte_eth_dev *dev)
2082 struct mlx5_priv *priv = dev->data->dev_private;
2083 struct mlx5_ind_table_ibv *ind_tbl;
2085 struct mlx5_hrxq *hrxq;
2087 if (priv->drop_queue.hrxq) {
2088 rte_atomic32_inc(&priv->drop_queue.hrxq->refcnt);
2089 return priv->drop_queue.hrxq;
2091 ind_tbl = mlx5_ind_table_ibv_drop_new(dev);
2094 qp = mlx5_glue->create_qp_ex(priv->sh->ctx,
2095 &(struct ibv_qp_init_attr_ex){
2096 .qp_type = IBV_QPT_RAW_PACKET,
2098 IBV_QP_INIT_ATTR_PD |
2099 IBV_QP_INIT_ATTR_IND_TABLE |
2100 IBV_QP_INIT_ATTR_RX_HASH,
2101 .rx_hash_conf = (struct ibv_rx_hash_conf){
2103 IBV_RX_HASH_FUNC_TOEPLITZ,
2104 .rx_hash_key_len = MLX5_RSS_HASH_KEY_LEN,
2105 .rx_hash_key = rss_hash_default_key,
2106 .rx_hash_fields_mask = 0,
2108 .rwq_ind_tbl = ind_tbl->ind_table,
2112 DEBUG("port %u cannot allocate QP for drop queue",
2113 dev->data->port_id);
2117 hrxq = rte_calloc(__func__, 1, sizeof(*hrxq), 0);
2120 "port %u cannot allocate memory for drop queue",
2121 dev->data->port_id);
2125 hrxq->ind_table = ind_tbl;
2127 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2128 hrxq->action = mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
2129 if (!hrxq->action) {
2134 priv->drop_queue.hrxq = hrxq;
2135 rte_atomic32_set(&hrxq->refcnt, 1);
2139 mlx5_ind_table_ibv_drop_release(dev);
2144 * Release a drop hash Rx queue.
2147 * Pointer to Ethernet device.
2150 mlx5_hrxq_drop_release(struct rte_eth_dev *dev)
2152 struct mlx5_priv *priv = dev->data->dev_private;
2153 struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
2155 if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
2156 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2157 mlx5_glue->destroy_flow_action(hrxq->action);
2159 claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
2160 mlx5_ind_table_ibv_drop_release(dev);
2162 priv->drop_queue.hrxq = NULL;