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 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 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_ctrl->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_ctrl->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->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->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_ctrl->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_ctrl->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 * Clean up a RX queue.
358 * Destroy objects, free allocated memory and reset the structure for reuse.
361 * Pointer to RX queue structure.
364 mlx5_rxq_cleanup(struct mlx5_rxq_ctrl *rxq_ctrl)
366 DRV_LOG(DEBUG, "port %u cleaning up Rx queue %u",
367 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx);
369 mlx5_rxq_ibv_release(rxq_ctrl->ibv);
370 memset(rxq_ctrl, 0, sizeof(*rxq_ctrl));
374 * Returns the per-queue supported offloads.
377 * Pointer to Ethernet device.
380 * Supported Rx offloads.
383 mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
385 struct priv *priv = dev->data->dev_private;
386 struct mlx5_dev_config *config = &priv->config;
387 uint64_t offloads = (DEV_RX_OFFLOAD_SCATTER |
388 DEV_RX_OFFLOAD_TIMESTAMP |
389 DEV_RX_OFFLOAD_JUMBO_FRAME);
391 if (config->hw_fcs_strip)
392 offloads |= DEV_RX_OFFLOAD_KEEP_CRC;
395 offloads |= (DEV_RX_OFFLOAD_IPV4_CKSUM |
396 DEV_RX_OFFLOAD_UDP_CKSUM |
397 DEV_RX_OFFLOAD_TCP_CKSUM);
398 if (config->hw_vlan_strip)
399 offloads |= DEV_RX_OFFLOAD_VLAN_STRIP;
405 * Returns the per-port supported offloads.
408 * Supported Rx offloads.
411 mlx5_get_rx_port_offloads(void)
413 uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
421 * Pointer to Ethernet device structure.
425 * Number of descriptors to configure in queue.
427 * NUMA socket on which memory must be allocated.
429 * Thresholds parameters.
431 * Memory pool for buffer allocations.
434 * 0 on success, a negative errno value otherwise and rte_errno is set.
437 mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
438 unsigned int socket, const struct rte_eth_rxconf *conf,
439 struct rte_mempool *mp)
441 struct priv *priv = dev->data->dev_private;
442 struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
443 struct mlx5_rxq_ctrl *rxq_ctrl =
444 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
446 if (!rte_is_power_of_2(desc)) {
447 desc = 1 << log2above(desc);
449 "port %u increased number of descriptors in Rx queue %u"
450 " to the next power of two (%d)",
451 dev->data->port_id, idx, desc);
453 DRV_LOG(DEBUG, "port %u configuring Rx queue %u for %u descriptors",
454 dev->data->port_id, idx, desc);
455 if (idx >= priv->rxqs_n) {
456 DRV_LOG(ERR, "port %u Rx queue index out of range (%u >= %u)",
457 dev->data->port_id, idx, priv->rxqs_n);
458 rte_errno = EOVERFLOW;
461 if (!mlx5_rxq_releasable(dev, idx)) {
462 DRV_LOG(ERR, "port %u unable to release queue index %u",
463 dev->data->port_id, idx);
467 mlx5_rxq_release(dev, idx);
468 rxq_ctrl = mlx5_rxq_new(dev, idx, desc, socket, conf, mp);
470 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
471 dev->data->port_id, idx);
475 DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
476 dev->data->port_id, idx);
477 (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
482 * DPDK callback to release a RX queue.
485 * Generic RX queue pointer.
488 mlx5_rx_queue_release(void *dpdk_rxq)
490 struct mlx5_rxq_data *rxq = (struct mlx5_rxq_data *)dpdk_rxq;
491 struct mlx5_rxq_ctrl *rxq_ctrl;
496 rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
497 priv = rxq_ctrl->priv;
498 if (!mlx5_rxq_releasable(ETH_DEV(priv), rxq_ctrl->rxq.stats.idx))
499 rte_panic("port %u Rx queue %u is still used by a flow and"
500 " cannot be removed\n",
501 PORT_ID(priv), rxq_ctrl->idx);
502 mlx5_rxq_release(ETH_DEV(priv), rxq_ctrl->rxq.stats.idx);
506 * Allocate queue vector and fill epoll fd list for Rx interrupts.
509 * Pointer to Ethernet device.
512 * 0 on success, a negative errno value otherwise and rte_errno is set.
515 mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
517 struct priv *priv = dev->data->dev_private;
519 unsigned int rxqs_n = priv->rxqs_n;
520 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
521 unsigned int count = 0;
522 struct rte_intr_handle *intr_handle = dev->intr_handle;
524 if (!dev->data->dev_conf.intr_conf.rxq)
526 mlx5_rx_intr_vec_disable(dev);
527 intr_handle->intr_vec = malloc(n * sizeof(intr_handle->intr_vec[0]));
528 if (intr_handle->intr_vec == NULL) {
530 "port %u failed to allocate memory for interrupt"
531 " vector, Rx interrupts will not be supported",
536 intr_handle->type = RTE_INTR_HANDLE_EXT;
537 for (i = 0; i != n; ++i) {
538 /* This rxq ibv must not be released in this function. */
539 struct mlx5_rxq_ibv *rxq_ibv = mlx5_rxq_ibv_get(dev, i);
544 /* Skip queues that cannot request interrupts. */
545 if (!rxq_ibv || !rxq_ibv->channel) {
546 /* Use invalid intr_vec[] index to disable entry. */
547 intr_handle->intr_vec[i] =
548 RTE_INTR_VEC_RXTX_OFFSET +
549 RTE_MAX_RXTX_INTR_VEC_ID;
552 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
554 "port %u too many Rx queues for interrupt"
555 " vector size (%d), Rx interrupts cannot be"
557 dev->data->port_id, RTE_MAX_RXTX_INTR_VEC_ID);
558 mlx5_rx_intr_vec_disable(dev);
562 fd = rxq_ibv->channel->fd;
563 flags = fcntl(fd, F_GETFL);
564 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
568 "port %u failed to make Rx interrupt file"
569 " descriptor %d non-blocking for queue index"
571 dev->data->port_id, fd, i);
572 mlx5_rx_intr_vec_disable(dev);
575 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
576 intr_handle->efds[count] = fd;
580 mlx5_rx_intr_vec_disable(dev);
582 intr_handle->nb_efd = count;
587 * Clean up Rx interrupts handler.
590 * Pointer to Ethernet device.
593 mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
595 struct priv *priv = dev->data->dev_private;
596 struct rte_intr_handle *intr_handle = dev->intr_handle;
598 unsigned int rxqs_n = priv->rxqs_n;
599 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
601 if (!dev->data->dev_conf.intr_conf.rxq)
603 if (!intr_handle->intr_vec)
605 for (i = 0; i != n; ++i) {
606 struct mlx5_rxq_ctrl *rxq_ctrl;
607 struct mlx5_rxq_data *rxq_data;
609 if (intr_handle->intr_vec[i] == RTE_INTR_VEC_RXTX_OFFSET +
610 RTE_MAX_RXTX_INTR_VEC_ID)
613 * Need to access directly the queue to release the reference
614 * kept in priv_rx_intr_vec_enable().
616 rxq_data = (*priv->rxqs)[i];
617 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
618 mlx5_rxq_ibv_release(rxq_ctrl->ibv);
621 rte_intr_free_epoll_fd(intr_handle);
622 if (intr_handle->intr_vec)
623 free(intr_handle->intr_vec);
624 intr_handle->nb_efd = 0;
625 intr_handle->intr_vec = NULL;
629 * MLX5 CQ notification .
632 * Pointer to receive queue structure.
634 * Sequence number per receive queue .
637 mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
640 uint32_t doorbell_hi;
642 void *cq_db_reg = (char *)rxq->cq_uar + MLX5_CQ_DOORBELL;
644 sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
645 doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
646 doorbell = (uint64_t)doorbell_hi << 32;
647 doorbell |= rxq->cqn;
648 rxq->cq_db[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
649 mlx5_uar_write64(rte_cpu_to_be_64(doorbell),
650 cq_db_reg, rxq->uar_lock_cq);
654 * DPDK callback for Rx queue interrupt enable.
657 * Pointer to Ethernet device structure.
662 * 0 on success, a negative errno value otherwise and rte_errno is set.
665 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
667 struct priv *priv = dev->data->dev_private;
668 struct mlx5_rxq_data *rxq_data;
669 struct mlx5_rxq_ctrl *rxq_ctrl;
671 rxq_data = (*priv->rxqs)[rx_queue_id];
676 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
678 struct mlx5_rxq_ibv *rxq_ibv;
680 rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
685 mlx5_arm_cq(rxq_data, rxq_data->cq_arm_sn);
686 mlx5_rxq_ibv_release(rxq_ibv);
692 * DPDK callback for Rx queue interrupt disable.
695 * Pointer to Ethernet device structure.
700 * 0 on success, a negative errno value otherwise and rte_errno is set.
703 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
705 struct priv *priv = dev->data->dev_private;
706 struct mlx5_rxq_data *rxq_data;
707 struct mlx5_rxq_ctrl *rxq_ctrl;
708 struct mlx5_rxq_ibv *rxq_ibv = NULL;
709 struct ibv_cq *ev_cq;
713 rxq_data = (*priv->rxqs)[rx_queue_id];
718 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
721 rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
726 ret = mlx5_glue->get_cq_event(rxq_ibv->channel, &ev_cq, &ev_ctx);
727 if (ret || ev_cq != rxq_ibv->cq) {
731 rxq_data->cq_arm_sn++;
732 mlx5_glue->ack_cq_events(rxq_ibv->cq, 1);
735 ret = rte_errno; /* Save rte_errno before cleanup. */
737 mlx5_rxq_ibv_release(rxq_ibv);
738 DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
739 dev->data->port_id, rx_queue_id);
740 rte_errno = ret; /* Restore rte_errno. */
745 * Create the Rx queue Verbs object.
748 * Pointer to Ethernet device.
750 * Queue index in DPDK Rx queue array
753 * The Verbs object initialised, NULL otherwise and rte_errno is set.
755 struct mlx5_rxq_ibv *
756 mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx)
758 struct priv *priv = dev->data->dev_private;
759 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
760 struct mlx5_rxq_ctrl *rxq_ctrl =
761 container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
762 struct ibv_wq_attr mod;
765 struct ibv_cq_init_attr_ex ibv;
766 struct mlx5dv_cq_init_attr mlx5;
769 struct ibv_wq_init_attr ibv;
770 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
771 struct mlx5dv_wq_init_attr mlx5;
774 struct ibv_cq_ex cq_attr;
777 unsigned int wqe_n = 1 << rxq_data->elts_n;
778 struct mlx5_rxq_ibv *tmpl;
779 struct mlx5dv_cq cq_info;
780 struct mlx5dv_rwq rwq;
783 struct mlx5dv_obj obj;
784 struct mlx5_dev_config *config = &priv->config;
785 const int mprq_en = mlx5_rxq_mprq_enabled(rxq_data);
788 assert(!rxq_ctrl->ibv);
789 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_RX_QUEUE;
790 priv->verbs_alloc_ctx.obj = rxq_ctrl;
791 tmpl = rte_calloc_socket(__func__, 1, sizeof(*tmpl), 0,
795 "port %u Rx queue %u cannot allocate verbs resources",
796 dev->data->port_id, rxq_ctrl->idx);
800 tmpl->rxq_ctrl = rxq_ctrl;
802 tmpl->channel = mlx5_glue->create_comp_channel(priv->ctx);
803 if (!tmpl->channel) {
804 DRV_LOG(ERR, "port %u: comp channel creation failure",
811 cqe_n = wqe_n * (1 << rxq_data->strd_num_n) - 1;
814 attr.cq.ibv = (struct ibv_cq_init_attr_ex){
816 .channel = tmpl->channel,
819 attr.cq.mlx5 = (struct mlx5dv_cq_init_attr){
822 if (config->cqe_comp && !rxq_data->hw_timestamp) {
823 attr.cq.mlx5.comp_mask |=
824 MLX5DV_CQ_INIT_ATTR_MASK_COMPRESSED_CQE;
825 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
826 attr.cq.mlx5.cqe_comp_res_format =
827 mprq_en ? MLX5DV_CQE_RES_FORMAT_CSUM_STRIDX :
828 MLX5DV_CQE_RES_FORMAT_HASH;
830 attr.cq.mlx5.cqe_comp_res_format = MLX5DV_CQE_RES_FORMAT_HASH;
833 * For vectorized Rx, it must not be doubled in order to
834 * make cq_ci and rq_ci aligned.
836 if (mlx5_rxq_check_vec_support(rxq_data) < 0)
837 attr.cq.ibv.cqe *= 2;
838 } else if (config->cqe_comp && rxq_data->hw_timestamp) {
840 "port %u Rx CQE compression is disabled for HW"
844 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
845 if (config->cqe_pad) {
846 attr.cq.mlx5.comp_mask |= MLX5DV_CQ_INIT_ATTR_MASK_FLAGS;
847 attr.cq.mlx5.flags |= MLX5DV_CQ_INIT_ATTR_FLAGS_CQE_PAD;
850 tmpl->cq = mlx5_glue->cq_ex_to_cq
851 (mlx5_glue->dv_create_cq(priv->ctx, &attr.cq.ibv,
853 if (tmpl->cq == NULL) {
854 DRV_LOG(ERR, "port %u Rx queue %u CQ creation failure",
855 dev->data->port_id, idx);
859 DRV_LOG(DEBUG, "port %u priv->device_attr.max_qp_wr is %d",
860 dev->data->port_id, priv->device_attr.orig_attr.max_qp_wr);
861 DRV_LOG(DEBUG, "port %u priv->device_attr.max_sge is %d",
862 dev->data->port_id, priv->device_attr.orig_attr.max_sge);
863 attr.wq.ibv = (struct ibv_wq_init_attr){
864 .wq_context = NULL, /* Could be useful in the future. */
865 .wq_type = IBV_WQT_RQ,
866 /* Max number of outstanding WRs. */
867 .max_wr = wqe_n >> rxq_data->sges_n,
868 /* Max number of scatter/gather elements in a WR. */
869 .max_sge = 1 << rxq_data->sges_n,
873 IBV_WQ_FLAGS_CVLAN_STRIPPING |
875 .create_flags = (rxq_data->vlan_strip ?
876 IBV_WQ_FLAGS_CVLAN_STRIPPING :
879 /* By default, FCS (CRC) is stripped by hardware. */
880 if (rxq_data->crc_present) {
881 attr.wq.ibv.create_flags |= IBV_WQ_FLAGS_SCATTER_FCS;
882 attr.wq.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
884 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
885 if (config->hw_padding) {
886 attr.wq.ibv.create_flags |= IBV_WQ_FLAG_RX_END_PADDING;
887 attr.wq.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
890 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
891 attr.wq.mlx5 = (struct mlx5dv_wq_init_attr){
895 struct mlx5dv_striding_rq_init_attr *mprq_attr =
896 &attr.wq.mlx5.striding_rq_attrs;
898 attr.wq.mlx5.comp_mask |= MLX5DV_WQ_INIT_ATTR_MASK_STRIDING_RQ;
899 *mprq_attr = (struct mlx5dv_striding_rq_init_attr){
900 .single_stride_log_num_of_bytes = rxq_data->strd_sz_n,
901 .single_wqe_log_num_of_strides = rxq_data->strd_num_n,
902 .two_byte_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT,
905 tmpl->wq = mlx5_glue->dv_create_wq(priv->ctx, &attr.wq.ibv,
908 tmpl->wq = mlx5_glue->create_wq(priv->ctx, &attr.wq.ibv);
910 if (tmpl->wq == NULL) {
911 DRV_LOG(ERR, "port %u Rx queue %u WQ creation failure",
912 dev->data->port_id, idx);
917 * Make sure number of WRs*SGEs match expectations since a queue
918 * cannot allocate more than "desc" buffers.
920 if (attr.wq.ibv.max_wr != (wqe_n >> rxq_data->sges_n) ||
921 attr.wq.ibv.max_sge != (1u << rxq_data->sges_n)) {
923 "port %u Rx queue %u requested %u*%u but got %u*%u"
925 dev->data->port_id, idx,
926 wqe_n >> rxq_data->sges_n, (1 << rxq_data->sges_n),
927 attr.wq.ibv.max_wr, attr.wq.ibv.max_sge);
931 /* Change queue state to ready. */
932 mod = (struct ibv_wq_attr){
933 .attr_mask = IBV_WQ_ATTR_STATE,
934 .wq_state = IBV_WQS_RDY,
936 ret = mlx5_glue->modify_wq(tmpl->wq, &mod);
939 "port %u Rx queue %u WQ state to IBV_WQS_RDY failed",
940 dev->data->port_id, idx);
944 obj.cq.in = tmpl->cq;
945 obj.cq.out = &cq_info;
946 obj.rwq.in = tmpl->wq;
948 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_RWQ);
953 if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
955 "port %u wrong MLX5_CQE_SIZE environment variable"
956 " value: it should be set to %u",
957 dev->data->port_id, RTE_CACHE_LINE_SIZE);
961 /* Fill the rings. */
962 rxq_data->wqes = rwq.buf;
963 for (i = 0; (i != wqe_n); ++i) {
964 volatile struct mlx5_wqe_data_seg *scat;
969 struct mlx5_mprq_buf *buf = (*rxq_data->mprq_bufs)[i];
971 scat = &((volatile struct mlx5_wqe_mprq *)
972 rxq_data->wqes)[i].dseg;
973 addr = (uintptr_t)mlx5_mprq_buf_addr(buf);
974 byte_count = (1 << rxq_data->strd_sz_n) *
975 (1 << rxq_data->strd_num_n);
977 struct rte_mbuf *buf = (*rxq_data->elts)[i];
979 scat = &((volatile struct mlx5_wqe_data_seg *)
981 addr = rte_pktmbuf_mtod(buf, uintptr_t);
982 byte_count = DATA_LEN(buf);
984 /* scat->addr must be able to store a pointer. */
985 assert(sizeof(scat->addr) >= sizeof(uintptr_t));
986 *scat = (struct mlx5_wqe_data_seg){
987 .addr = rte_cpu_to_be_64(addr),
988 .byte_count = rte_cpu_to_be_32(byte_count),
989 .lkey = mlx5_rx_addr2mr(rxq_data, addr),
992 rxq_data->rq_db = rwq.dbrec;
993 rxq_data->cqe_n = log2above(cq_info.cqe_cnt);
995 rxq_data->consumed_strd = 0;
997 rxq_data->zip = (struct rxq_zip){
1000 rxq_data->cq_db = cq_info.dbrec;
1001 rxq_data->cqes = (volatile struct mlx5_cqe (*)[])(uintptr_t)cq_info.buf;
1002 rxq_data->cq_uar = cq_info.cq_uar;
1003 rxq_data->cqn = cq_info.cqn;
1004 rxq_data->cq_arm_sn = 0;
1005 /* Update doorbell counter. */
1006 rxq_data->rq_ci = wqe_n >> rxq_data->sges_n;
1008 *rxq_data->rq_db = rte_cpu_to_be_32(rxq_data->rq_ci);
1009 DRV_LOG(DEBUG, "port %u rxq %u updated with %p", dev->data->port_id,
1010 idx, (void *)&tmpl);
1011 rte_atomic32_inc(&tmpl->refcnt);
1012 LIST_INSERT_HEAD(&priv->rxqsibv, tmpl, next);
1013 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1016 ret = rte_errno; /* Save rte_errno before cleanup. */
1018 claim_zero(mlx5_glue->destroy_wq(tmpl->wq));
1020 claim_zero(mlx5_glue->destroy_cq(tmpl->cq));
1022 claim_zero(mlx5_glue->destroy_comp_channel(tmpl->channel));
1023 priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1024 rte_errno = ret; /* Restore rte_errno. */
1029 * Get an Rx queue Verbs object.
1032 * Pointer to Ethernet device.
1034 * Queue index in DPDK Rx queue array
1037 * The Verbs object if it exists.
1039 struct mlx5_rxq_ibv *
1040 mlx5_rxq_ibv_get(struct rte_eth_dev *dev, uint16_t idx)
1042 struct priv *priv = dev->data->dev_private;
1043 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
1044 struct mlx5_rxq_ctrl *rxq_ctrl;
1046 if (idx >= priv->rxqs_n)
1050 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
1051 if (rxq_ctrl->ibv) {
1052 rte_atomic32_inc(&rxq_ctrl->ibv->refcnt);
1054 return rxq_ctrl->ibv;
1058 * Release an Rx verbs queue object.
1061 * Verbs Rx queue object.
1064 * 1 while a reference on it exists, 0 when freed.
1067 mlx5_rxq_ibv_release(struct mlx5_rxq_ibv *rxq_ibv)
1070 assert(rxq_ibv->wq);
1071 assert(rxq_ibv->cq);
1072 if (rte_atomic32_dec_and_test(&rxq_ibv->refcnt)) {
1073 rxq_free_elts(rxq_ibv->rxq_ctrl);
1074 claim_zero(mlx5_glue->destroy_wq(rxq_ibv->wq));
1075 claim_zero(mlx5_glue->destroy_cq(rxq_ibv->cq));
1076 if (rxq_ibv->channel)
1077 claim_zero(mlx5_glue->destroy_comp_channel
1078 (rxq_ibv->channel));
1079 LIST_REMOVE(rxq_ibv, next);
1087 * Verify the Verbs Rx queue list is empty
1090 * Pointer to Ethernet device.
1093 * The number of object not released.
1096 mlx5_rxq_ibv_verify(struct rte_eth_dev *dev)
1098 struct priv *priv = dev->data->dev_private;
1100 struct mlx5_rxq_ibv *rxq_ibv;
1102 LIST_FOREACH(rxq_ibv, &priv->rxqsibv, next) {
1103 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u still referenced",
1104 dev->data->port_id, rxq_ibv->rxq_ctrl->idx);
1111 * Return true if a single reference exists on the object.
1114 * Verbs Rx queue object.
1117 mlx5_rxq_ibv_releasable(struct mlx5_rxq_ibv *rxq_ibv)
1120 return (rte_atomic32_read(&rxq_ibv->refcnt) == 1);
1124 * Callback function to initialize mbufs for Multi-Packet RQ.
1127 mlx5_mprq_buf_init(struct rte_mempool *mp, void *opaque_arg __rte_unused,
1128 void *_m, unsigned int i __rte_unused)
1130 struct mlx5_mprq_buf *buf = _m;
1132 memset(_m, 0, sizeof(*buf));
1134 rte_atomic16_set(&buf->refcnt, 1);
1138 * Free mempool of Multi-Packet RQ.
1141 * Pointer to Ethernet device.
1144 * 0 on success, negative errno value on failure.
1147 mlx5_mprq_free_mp(struct rte_eth_dev *dev)
1149 struct priv *priv = dev->data->dev_private;
1150 struct rte_mempool *mp = priv->mprq_mp;
1155 DRV_LOG(DEBUG, "port %u freeing mempool (%s) for Multi-Packet RQ",
1156 dev->data->port_id, mp->name);
1158 * If a buffer in the pool has been externally attached to a mbuf and it
1159 * is still in use by application, destroying the Rx qeueue can spoil
1160 * the packet. It is unlikely to happen but if application dynamically
1161 * creates and destroys with holding Rx packets, this can happen.
1163 * TODO: It is unavoidable for now because the mempool for Multi-Packet
1164 * RQ isn't provided by application but managed by PMD.
1166 if (!rte_mempool_full(mp)) {
1168 "port %u mempool for Multi-Packet RQ is still in use",
1169 dev->data->port_id);
1173 rte_mempool_free(mp);
1174 /* Unset mempool for each Rx queue. */
1175 for (i = 0; i != priv->rxqs_n; ++i) {
1176 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1180 rxq->mprq_mp = NULL;
1186 * Allocate a mempool for Multi-Packet RQ. All configured Rx queues share the
1187 * mempool. If already allocated, reuse it if there're enough elements.
1188 * Otherwise, resize it.
1191 * Pointer to Ethernet device.
1194 * 0 on success, negative errno value on failure.
1197 mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
1199 struct priv *priv = dev->data->dev_private;
1200 struct rte_mempool *mp = priv->mprq_mp;
1201 char name[RTE_MEMPOOL_NAMESIZE];
1202 unsigned int desc = 0;
1203 unsigned int buf_len;
1204 unsigned int obj_num;
1205 unsigned int obj_size;
1206 unsigned int strd_num_n = 0;
1207 unsigned int strd_sz_n = 0;
1210 if (!mlx5_mprq_enabled(dev))
1212 /* Count the total number of descriptors configured. */
1213 for (i = 0; i != priv->rxqs_n; ++i) {
1214 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1218 desc += 1 << rxq->elts_n;
1219 /* Get the max number of strides. */
1220 if (strd_num_n < rxq->strd_num_n)
1221 strd_num_n = rxq->strd_num_n;
1222 /* Get the max size of a stride. */
1223 if (strd_sz_n < rxq->strd_sz_n)
1224 strd_sz_n = rxq->strd_sz_n;
1226 assert(strd_num_n && strd_sz_n);
1227 buf_len = (1 << strd_num_n) * (1 << strd_sz_n);
1228 obj_size = buf_len + sizeof(struct mlx5_mprq_buf);
1230 * Received packets can be either memcpy'd or externally referenced. In
1231 * case that the packet is attached to an mbuf as an external buffer, as
1232 * it isn't possible to predict how the buffers will be queued by
1233 * application, there's no option to exactly pre-allocate needed buffers
1234 * in advance but to speculatively prepares enough buffers.
1236 * In the data path, if this Mempool is depleted, PMD will try to memcpy
1237 * received packets to buffers provided by application (rxq->mp) until
1238 * this Mempool gets available again.
1241 obj_num = desc + MLX5_MPRQ_MP_CACHE_SZ * priv->rxqs_n;
1243 * rte_mempool_create_empty() has sanity check to refuse large cache
1244 * size compared to the number of elements.
1245 * CACHE_FLUSHTHRESH_MULTIPLIER is defined in a C file, so using a
1246 * constant number 2 instead.
1248 obj_num = RTE_MAX(obj_num, MLX5_MPRQ_MP_CACHE_SZ * 2);
1249 /* Check a mempool is already allocated and if it can be resued. */
1250 if (mp != NULL && mp->elt_size >= obj_size && mp->size >= obj_num) {
1251 DRV_LOG(DEBUG, "port %u mempool %s is being reused",
1252 dev->data->port_id, mp->name);
1255 } else if (mp != NULL) {
1256 DRV_LOG(DEBUG, "port %u mempool %s should be resized, freeing it",
1257 dev->data->port_id, mp->name);
1259 * If failed to free, which means it may be still in use, no way
1260 * but to keep using the existing one. On buffer underrun,
1261 * packets will be memcpy'd instead of external buffer
1264 if (mlx5_mprq_free_mp(dev)) {
1265 if (mp->elt_size >= obj_size)
1271 snprintf(name, sizeof(name), "%s-mprq", dev->device->name);
1272 mp = rte_mempool_create(name, obj_num, obj_size, MLX5_MPRQ_MP_CACHE_SZ,
1273 0, NULL, NULL, mlx5_mprq_buf_init, NULL,
1274 dev->device->numa_node, 0);
1277 "port %u failed to allocate a mempool for"
1278 " Multi-Packet RQ, count=%u, size=%u",
1279 dev->data->port_id, obj_num, obj_size);
1285 /* Set mempool for each Rx queue. */
1286 for (i = 0; i != priv->rxqs_n; ++i) {
1287 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1293 DRV_LOG(INFO, "port %u Multi-Packet RQ is configured",
1294 dev->data->port_id);
1299 * Create a DPDK Rx queue.
1302 * Pointer to Ethernet device.
1306 * Number of descriptors to configure in queue.
1308 * NUMA socket on which memory must be allocated.
1311 * A DPDK queue object on success, NULL otherwise and rte_errno is set.
1313 struct mlx5_rxq_ctrl *
1314 mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1315 unsigned int socket, const struct rte_eth_rxconf *conf,
1316 struct rte_mempool *mp)
1318 struct priv *priv = dev->data->dev_private;
1319 struct mlx5_rxq_ctrl *tmpl;
1320 unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
1321 unsigned int mprq_stride_size;
1322 struct mlx5_dev_config *config = &priv->config;
1324 * Always allocate extra slots, even if eventually
1325 * the vector Rx will not be used.
1328 desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
1329 uint64_t offloads = conf->offloads |
1330 dev->data->dev_conf.rxmode.offloads;
1331 const int mprq_en = mlx5_check_mprq_support(dev) > 0;
1333 tmpl = rte_calloc_socket("RXQ", 1,
1335 desc_n * sizeof(struct rte_mbuf *),
1341 if (mlx5_mr_btree_init(&tmpl->rxq.mr_ctrl.cache_bh,
1342 MLX5_MR_BTREE_CACHE_N, socket)) {
1343 /* rte_errno is already set. */
1346 tmpl->socket = socket;
1347 if (dev->data->dev_conf.intr_conf.rxq)
1350 * This Rx queue can be configured as a Multi-Packet RQ if all of the
1351 * following conditions are met:
1352 * - MPRQ is enabled.
1353 * - The number of descs is more than the number of strides.
1354 * - max_rx_pkt_len plus overhead is less than the max size of a
1356 * Otherwise, enable Rx scatter if necessary.
1358 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
1360 dev->data->dev_conf.rxmode.max_rx_pkt_len +
1361 sizeof(struct rte_mbuf_ext_shared_info) +
1362 RTE_PKTMBUF_HEADROOM;
1364 desc > (1U << config->mprq.stride_num_n) &&
1365 mprq_stride_size <= (1U << config->mprq.max_stride_size_n)) {
1366 /* TODO: Rx scatter isn't supported yet. */
1367 tmpl->rxq.sges_n = 0;
1368 /* Trim the number of descs needed. */
1369 desc >>= config->mprq.stride_num_n;
1370 tmpl->rxq.strd_num_n = config->mprq.stride_num_n;
1371 tmpl->rxq.strd_sz_n = RTE_MAX(log2above(mprq_stride_size),
1372 config->mprq.min_stride_size_n);
1373 tmpl->rxq.strd_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT;
1374 tmpl->rxq.mprq_max_memcpy_len =
1375 RTE_MIN(mb_len - RTE_PKTMBUF_HEADROOM,
1376 config->mprq.max_memcpy_len);
1378 "port %u Rx queue %u: Multi-Packet RQ is enabled"
1379 " strd_num_n = %u, strd_sz_n = %u",
1380 dev->data->port_id, idx,
1381 tmpl->rxq.strd_num_n, tmpl->rxq.strd_sz_n);
1382 } else if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
1383 (mb_len - RTE_PKTMBUF_HEADROOM)) {
1384 tmpl->rxq.sges_n = 0;
1385 } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
1387 RTE_PKTMBUF_HEADROOM +
1388 dev->data->dev_conf.rxmode.max_rx_pkt_len;
1389 unsigned int sges_n;
1392 * Determine the number of SGEs needed for a full packet
1393 * and round it to the next power of two.
1395 sges_n = log2above((size / mb_len) + !!(size % mb_len));
1396 tmpl->rxq.sges_n = sges_n;
1397 /* Make sure rxq.sges_n did not overflow. */
1398 size = mb_len * (1 << tmpl->rxq.sges_n);
1399 size -= RTE_PKTMBUF_HEADROOM;
1400 if (size < dev->data->dev_conf.rxmode.max_rx_pkt_len) {
1402 "port %u too many SGEs (%u) needed to handle"
1403 " requested maximum packet size %u",
1406 dev->data->dev_conf.rxmode.max_rx_pkt_len);
1407 rte_errno = EOVERFLOW;
1412 "port %u the requested maximum Rx packet size (%u) is"
1413 " larger than a single mbuf (%u) and scattered mode has"
1414 " not been requested",
1416 dev->data->dev_conf.rxmode.max_rx_pkt_len,
1417 mb_len - RTE_PKTMBUF_HEADROOM);
1419 if (mprq_en && !mlx5_rxq_mprq_enabled(&tmpl->rxq))
1421 "port %u MPRQ is requested but cannot be enabled"
1422 " (requested: desc = %u, stride_sz = %u,"
1423 " supported: min_stride_num = %u, max_stride_sz = %u).",
1424 dev->data->port_id, desc, mprq_stride_size,
1425 (1 << config->mprq.stride_num_n),
1426 (1 << config->mprq.max_stride_size_n));
1427 DRV_LOG(DEBUG, "port %u maximum number of segments per packet: %u",
1428 dev->data->port_id, 1 << tmpl->rxq.sges_n);
1429 if (desc % (1 << tmpl->rxq.sges_n)) {
1431 "port %u number of Rx queue descriptors (%u) is not a"
1432 " multiple of SGEs per packet (%u)",
1435 1 << tmpl->rxq.sges_n);
1439 /* Toggle RX checksum offload if hardware supports it. */
1440 tmpl->rxq.csum = !!(offloads & DEV_RX_OFFLOAD_CHECKSUM);
1441 tmpl->rxq.hw_timestamp = !!(offloads & DEV_RX_OFFLOAD_TIMESTAMP);
1442 /* Configure VLAN stripping. */
1443 tmpl->rxq.vlan_strip = !!(offloads & DEV_RX_OFFLOAD_VLAN_STRIP);
1444 /* By default, FCS (CRC) is stripped by hardware. */
1445 tmpl->rxq.crc_present = 0;
1446 if (offloads & DEV_RX_OFFLOAD_KEEP_CRC) {
1447 if (config->hw_fcs_strip) {
1448 tmpl->rxq.crc_present = 1;
1451 "port %u CRC stripping has been disabled but will"
1452 " still be performed by hardware, make sure MLNX_OFED"
1453 " and firmware are up to date",
1454 dev->data->port_id);
1458 "port %u CRC stripping is %s, %u bytes will be subtracted from"
1459 " incoming frames to hide it",
1461 tmpl->rxq.crc_present ? "disabled" : "enabled",
1462 tmpl->rxq.crc_present << 2);
1464 tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
1465 (!!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS));
1466 tmpl->rxq.port_id = dev->data->port_id;
1469 tmpl->rxq.stats.idx = idx;
1470 tmpl->rxq.elts_n = log2above(desc);
1472 (struct rte_mbuf *(*)[1 << tmpl->rxq.elts_n])(tmpl + 1);
1474 tmpl->rxq.uar_lock_cq = &priv->uar_lock_cq;
1477 rte_atomic32_inc(&tmpl->refcnt);
1478 LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1489 * Pointer to Ethernet device.
1494 * A pointer to the queue if it exists, NULL otherwise.
1496 struct mlx5_rxq_ctrl *
1497 mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
1499 struct priv *priv = dev->data->dev_private;
1500 struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1502 if ((*priv->rxqs)[idx]) {
1503 rxq_ctrl = container_of((*priv->rxqs)[idx],
1504 struct mlx5_rxq_ctrl,
1506 mlx5_rxq_ibv_get(dev, idx);
1507 rte_atomic32_inc(&rxq_ctrl->refcnt);
1513 * Release a Rx queue.
1516 * Pointer to Ethernet device.
1521 * 1 while a reference on it exists, 0 when freed.
1524 mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
1526 struct priv *priv = dev->data->dev_private;
1527 struct mlx5_rxq_ctrl *rxq_ctrl;
1529 if (!(*priv->rxqs)[idx])
1531 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1532 assert(rxq_ctrl->priv);
1533 if (rxq_ctrl->ibv && !mlx5_rxq_ibv_release(rxq_ctrl->ibv))
1534 rxq_ctrl->ibv = NULL;
1535 if (rte_atomic32_dec_and_test(&rxq_ctrl->refcnt)) {
1536 mlx5_mr_btree_free(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
1537 LIST_REMOVE(rxq_ctrl, next);
1539 (*priv->rxqs)[idx] = NULL;
1546 * Verify if the queue can be released.
1549 * Pointer to Ethernet device.
1554 * 1 if the queue can be released, negative errno otherwise and rte_errno is
1558 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
1560 struct priv *priv = dev->data->dev_private;
1561 struct mlx5_rxq_ctrl *rxq_ctrl;
1563 if (!(*priv->rxqs)[idx]) {
1567 rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1568 return (rte_atomic32_read(&rxq_ctrl->refcnt) == 1);
1572 * Verify the Rx Queue list is empty
1575 * Pointer to Ethernet device.
1578 * The number of object not released.
1581 mlx5_rxq_verify(struct rte_eth_dev *dev)
1583 struct priv *priv = dev->data->dev_private;
1584 struct mlx5_rxq_ctrl *rxq_ctrl;
1587 LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
1588 DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
1589 dev->data->port_id, rxq_ctrl->idx);
1596 * Create an indirection table.
1599 * Pointer to Ethernet device.
1601 * Queues entering in the indirection table.
1603 * Number of queues in the array.
1606 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1608 struct mlx5_ind_table_ibv *
1609 mlx5_ind_table_ibv_new(struct rte_eth_dev *dev, const uint16_t *queues,
1612 struct priv *priv = dev->data->dev_private;
1613 struct mlx5_ind_table_ibv *ind_tbl;
1614 const unsigned int wq_n = rte_is_power_of_2(queues_n) ?
1615 log2above(queues_n) :
1616 log2above(priv->config.ind_table_max_size);
1617 struct ibv_wq *wq[1 << wq_n];
1621 ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl) +
1622 queues_n * sizeof(uint16_t), 0);
1627 for (i = 0; i != queues_n; ++i) {
1628 struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev, queues[i]);
1632 wq[i] = rxq->ibv->wq;
1633 ind_tbl->queues[i] = queues[i];
1635 ind_tbl->queues_n = queues_n;
1636 /* Finalise indirection table. */
1637 for (j = 0; i != (unsigned int)(1 << wq_n); ++i, ++j)
1639 ind_tbl->ind_table = mlx5_glue->create_rwq_ind_table
1641 &(struct ibv_rwq_ind_table_init_attr){
1642 .log_ind_tbl_size = wq_n,
1646 if (!ind_tbl->ind_table) {
1650 rte_atomic32_inc(&ind_tbl->refcnt);
1651 LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
1655 DEBUG("port %u cannot create indirection table", dev->data->port_id);
1660 * Get an indirection table.
1663 * Pointer to Ethernet device.
1665 * Queues entering in the indirection table.
1667 * Number of queues in the array.
1670 * An indirection table if found.
1672 struct mlx5_ind_table_ibv *
1673 mlx5_ind_table_ibv_get(struct rte_eth_dev *dev, const uint16_t *queues,
1676 struct priv *priv = dev->data->dev_private;
1677 struct mlx5_ind_table_ibv *ind_tbl;
1679 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1680 if ((ind_tbl->queues_n == queues_n) &&
1681 (memcmp(ind_tbl->queues, queues,
1682 ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
1689 rte_atomic32_inc(&ind_tbl->refcnt);
1690 for (i = 0; i != ind_tbl->queues_n; ++i)
1691 mlx5_rxq_get(dev, ind_tbl->queues[i]);
1697 * Release an indirection table.
1700 * Pointer to Ethernet device.
1702 * Indirection table to release.
1705 * 1 while a reference on it exists, 0 when freed.
1708 mlx5_ind_table_ibv_release(struct rte_eth_dev *dev,
1709 struct mlx5_ind_table_ibv *ind_tbl)
1713 if (rte_atomic32_dec_and_test(&ind_tbl->refcnt))
1714 claim_zero(mlx5_glue->destroy_rwq_ind_table
1715 (ind_tbl->ind_table));
1716 for (i = 0; i != ind_tbl->queues_n; ++i)
1717 claim_nonzero(mlx5_rxq_release(dev, ind_tbl->queues[i]));
1718 if (!rte_atomic32_read(&ind_tbl->refcnt)) {
1719 LIST_REMOVE(ind_tbl, next);
1727 * Verify the Rx Queue list is empty
1730 * Pointer to Ethernet device.
1733 * The number of object not released.
1736 mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev)
1738 struct priv *priv = dev->data->dev_private;
1739 struct mlx5_ind_table_ibv *ind_tbl;
1742 LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1744 "port %u Verbs indirection table %p still referenced",
1745 dev->data->port_id, (void *)ind_tbl);
1752 * Create an Rx Hash queue.
1755 * Pointer to Ethernet device.
1757 * RSS key for the Rx hash queue.
1758 * @param rss_key_len
1760 * @param hash_fields
1761 * Verbs protocol hash field to make the RSS on.
1763 * Queues entering in hash queue. In case of empty hash_fields only the
1764 * first queue index will be taken for the indirection table.
1769 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1772 mlx5_hrxq_new(struct rte_eth_dev *dev,
1773 const uint8_t *rss_key, uint32_t rss_key_len,
1774 uint64_t hash_fields,
1775 const uint16_t *queues, uint32_t queues_n,
1776 int tunnel __rte_unused)
1778 struct priv *priv = dev->data->dev_private;
1779 struct mlx5_hrxq *hrxq;
1780 struct mlx5_ind_table_ibv *ind_tbl;
1784 queues_n = hash_fields ? queues_n : 1;
1785 ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1787 ind_tbl = mlx5_ind_table_ibv_new(dev, queues, queues_n);
1793 rss_key_len = MLX5_RSS_HASH_KEY_LEN;
1794 rss_key = rss_hash_default_key;
1796 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1797 qp = mlx5_glue->dv_create_qp
1799 &(struct ibv_qp_init_attr_ex){
1800 .qp_type = IBV_QPT_RAW_PACKET,
1802 IBV_QP_INIT_ATTR_PD |
1803 IBV_QP_INIT_ATTR_IND_TABLE |
1804 IBV_QP_INIT_ATTR_RX_HASH,
1805 .rx_hash_conf = (struct ibv_rx_hash_conf){
1806 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1807 .rx_hash_key_len = rss_key_len ? rss_key_len :
1808 MLX5_RSS_HASH_KEY_LEN,
1809 .rx_hash_key = rss_key ?
1810 (void *)(uintptr_t)rss_key :
1811 rss_hash_default_key,
1812 .rx_hash_fields_mask = hash_fields,
1814 .rwq_ind_tbl = ind_tbl->ind_table,
1817 &(struct mlx5dv_qp_init_attr){
1818 .comp_mask = tunnel ?
1819 MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS : 0,
1820 .create_flags = MLX5DV_QP_CREATE_TUNNEL_OFFLOADS,
1823 qp = mlx5_glue->create_qp_ex
1825 &(struct ibv_qp_init_attr_ex){
1826 .qp_type = IBV_QPT_RAW_PACKET,
1828 IBV_QP_INIT_ATTR_PD |
1829 IBV_QP_INIT_ATTR_IND_TABLE |
1830 IBV_QP_INIT_ATTR_RX_HASH,
1831 .rx_hash_conf = (struct ibv_rx_hash_conf){
1832 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1833 .rx_hash_key_len = rss_key_len ? rss_key_len :
1834 MLX5_RSS_HASH_KEY_LEN,
1835 .rx_hash_key = rss_key ?
1836 (void *)(uintptr_t)rss_key :
1837 rss_hash_default_key,
1838 .rx_hash_fields_mask = hash_fields,
1840 .rwq_ind_tbl = ind_tbl->ind_table,
1848 hrxq = rte_calloc(__func__, 1, sizeof(*hrxq) + rss_key_len, 0);
1851 hrxq->ind_table = ind_tbl;
1853 hrxq->rss_key_len = rss_key_len;
1854 hrxq->hash_fields = hash_fields;
1855 memcpy(hrxq->rss_key, rss_key, rss_key_len);
1856 rte_atomic32_inc(&hrxq->refcnt);
1857 LIST_INSERT_HEAD(&priv->hrxqs, hrxq, next);
1860 err = rte_errno; /* Save rte_errno before cleanup. */
1861 mlx5_ind_table_ibv_release(dev, ind_tbl);
1863 claim_zero(mlx5_glue->destroy_qp(qp));
1864 rte_errno = err; /* Restore rte_errno. */
1869 * Get an Rx Hash queue.
1872 * Pointer to Ethernet device.
1874 * RSS configuration for the Rx hash queue.
1876 * Queues entering in hash queue. In case of empty hash_fields only the
1877 * first queue index will be taken for the indirection table.
1882 * An hash Rx queue on success.
1885 mlx5_hrxq_get(struct rte_eth_dev *dev,
1886 const uint8_t *rss_key, uint32_t rss_key_len,
1887 uint64_t hash_fields,
1888 const uint16_t *queues, uint32_t queues_n)
1890 struct priv *priv = dev->data->dev_private;
1891 struct mlx5_hrxq *hrxq;
1893 queues_n = hash_fields ? queues_n : 1;
1894 LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1895 struct mlx5_ind_table_ibv *ind_tbl;
1897 if (hrxq->rss_key_len != rss_key_len)
1899 if (memcmp(hrxq->rss_key, rss_key, rss_key_len))
1901 if (hrxq->hash_fields != hash_fields)
1903 ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1906 if (ind_tbl != hrxq->ind_table) {
1907 mlx5_ind_table_ibv_release(dev, ind_tbl);
1910 rte_atomic32_inc(&hrxq->refcnt);
1917 * Release the hash Rx queue.
1920 * Pointer to Ethernet device.
1922 * Pointer to Hash Rx queue to release.
1925 * 1 while a reference on it exists, 0 when freed.
1928 mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
1930 if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
1931 claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
1932 mlx5_ind_table_ibv_release(dev, hrxq->ind_table);
1933 LIST_REMOVE(hrxq, next);
1937 claim_nonzero(mlx5_ind_table_ibv_release(dev, hrxq->ind_table));
1942 * Verify the Rx Queue list is empty
1945 * Pointer to Ethernet device.
1948 * The number of object not released.
1951 mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev)
1953 struct priv *priv = dev->data->dev_private;
1954 struct mlx5_hrxq *hrxq;
1957 LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1959 "port %u Verbs hash Rx queue %p still referenced",
1960 dev->data->port_id, (void *)hrxq);
1967 * Create a drop Rx queue Verbs object.
1970 * Pointer to Ethernet device.
1973 * The Verbs object initialised, NULL otherwise and rte_errno is set.
1975 struct mlx5_rxq_ibv *
1976 mlx5_rxq_ibv_drop_new(struct rte_eth_dev *dev)
1978 struct priv *priv = dev->data->dev_private;
1980 struct ibv_wq *wq = NULL;
1981 struct mlx5_rxq_ibv *rxq;
1983 if (priv->drop_queue.rxq)
1984 return priv->drop_queue.rxq;
1985 cq = mlx5_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
1987 DEBUG("port %u cannot allocate CQ for drop queue",
1988 dev->data->port_id);
1992 wq = mlx5_glue->create_wq(priv->ctx,
1993 &(struct ibv_wq_init_attr){
1994 .wq_type = IBV_WQT_RQ,
2001 DEBUG("port %u cannot allocate WQ for drop queue",
2002 dev->data->port_id);
2006 rxq = rte_calloc(__func__, 1, sizeof(*rxq), 0);
2008 DEBUG("port %u cannot allocate drop Rx queue memory",
2009 dev->data->port_id);
2015 priv->drop_queue.rxq = rxq;
2019 claim_zero(mlx5_glue->destroy_wq(wq));
2021 claim_zero(mlx5_glue->destroy_cq(cq));
2026 * Release a drop Rx queue Verbs object.
2029 * Pointer to Ethernet device.
2032 * The Verbs object initialised, NULL otherwise and rte_errno is set.
2035 mlx5_rxq_ibv_drop_release(struct rte_eth_dev *dev)
2037 struct priv *priv = dev->data->dev_private;
2038 struct mlx5_rxq_ibv *rxq = priv->drop_queue.rxq;
2041 claim_zero(mlx5_glue->destroy_wq(rxq->wq));
2043 claim_zero(mlx5_glue->destroy_cq(rxq->cq));
2045 priv->drop_queue.rxq = NULL;
2049 * Create a drop indirection table.
2052 * Pointer to Ethernet device.
2055 * The Verbs object initialised, NULL otherwise and rte_errno is set.
2057 struct mlx5_ind_table_ibv *
2058 mlx5_ind_table_ibv_drop_new(struct rte_eth_dev *dev)
2060 struct priv *priv = dev->data->dev_private;
2061 struct mlx5_ind_table_ibv *ind_tbl;
2062 struct mlx5_rxq_ibv *rxq;
2063 struct mlx5_ind_table_ibv tmpl;
2065 rxq = mlx5_rxq_ibv_drop_new(dev);
2068 tmpl.ind_table = mlx5_glue->create_rwq_ind_table
2070 &(struct ibv_rwq_ind_table_init_attr){
2071 .log_ind_tbl_size = 0,
2072 .ind_tbl = &rxq->wq,
2075 if (!tmpl.ind_table) {
2076 DEBUG("port %u cannot allocate indirection table for drop"
2078 dev->data->port_id);
2082 ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl), 0);
2087 ind_tbl->ind_table = tmpl.ind_table;
2090 mlx5_rxq_ibv_drop_release(dev);
2095 * Release a drop indirection table.
2098 * Pointer to Ethernet device.
2101 mlx5_ind_table_ibv_drop_release(struct rte_eth_dev *dev)
2103 struct priv *priv = dev->data->dev_private;
2104 struct mlx5_ind_table_ibv *ind_tbl = priv->drop_queue.hrxq->ind_table;
2106 claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl->ind_table));
2107 mlx5_rxq_ibv_drop_release(dev);
2109 priv->drop_queue.hrxq->ind_table = NULL;
2113 * Create a drop Rx Hash queue.
2116 * Pointer to Ethernet device.
2119 * The Verbs object initialised, NULL otherwise and rte_errno is set.
2122 mlx5_hrxq_drop_new(struct rte_eth_dev *dev)
2124 struct priv *priv = dev->data->dev_private;
2125 struct mlx5_ind_table_ibv *ind_tbl;
2127 struct mlx5_hrxq *hrxq;
2129 if (priv->drop_queue.hrxq) {
2130 rte_atomic32_inc(&priv->drop_queue.hrxq->refcnt);
2131 return priv->drop_queue.hrxq;
2133 ind_tbl = mlx5_ind_table_ibv_drop_new(dev);
2136 qp = mlx5_glue->create_qp_ex(priv->ctx,
2137 &(struct ibv_qp_init_attr_ex){
2138 .qp_type = IBV_QPT_RAW_PACKET,
2140 IBV_QP_INIT_ATTR_PD |
2141 IBV_QP_INIT_ATTR_IND_TABLE |
2142 IBV_QP_INIT_ATTR_RX_HASH,
2143 .rx_hash_conf = (struct ibv_rx_hash_conf){
2145 IBV_RX_HASH_FUNC_TOEPLITZ,
2146 .rx_hash_key_len = MLX5_RSS_HASH_KEY_LEN,
2147 .rx_hash_key = rss_hash_default_key,
2148 .rx_hash_fields_mask = 0,
2150 .rwq_ind_tbl = ind_tbl->ind_table,
2154 DEBUG("port %u cannot allocate QP for drop queue",
2155 dev->data->port_id);
2159 hrxq = rte_calloc(__func__, 1, sizeof(*hrxq), 0);
2162 "port %u cannot allocate memory for drop queue",
2163 dev->data->port_id);
2167 hrxq->ind_table = ind_tbl;
2169 priv->drop_queue.hrxq = hrxq;
2170 rte_atomic32_set(&hrxq->refcnt, 1);
2174 mlx5_ind_table_ibv_drop_release(dev);
2179 * Release a drop hash Rx queue.
2182 * Pointer to Ethernet device.
2185 mlx5_hrxq_drop_release(struct rte_eth_dev *dev)
2187 struct priv *priv = dev->data->dev_private;
2188 struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
2190 if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
2191 claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
2192 mlx5_ind_table_ibv_drop_release(dev);
2194 priv->drop_queue.hrxq = NULL;