1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2015 6WIND S.A.
3 * Copyright 2015 Mellanox Technologies, Ltd
13 #include <ethdev_driver.h>
14 #include <rte_bus_pci.h>
16 #include <rte_common.h>
17 #include <rte_interrupts.h>
18 #include <rte_malloc.h>
19 #include <rte_string_fns.h>
20 #include <rte_rwlock.h>
21 #include <rte_cycles.h>
23 #include <mlx5_malloc.h>
25 #include "mlx5_rxtx.h"
26 #include "mlx5_autoconf.h"
29 * Get the interface index from device name.
32 * Pointer to Ethernet device.
35 * Nonzero interface index on success, zero otherwise and rte_errno is set.
38 mlx5_ifindex(const struct rte_eth_dev *dev)
40 struct mlx5_priv *priv = dev->data->dev_private;
44 MLX5_ASSERT(priv->if_index);
45 if (priv->master && priv->sh->bond.ifindex > 0)
46 ifindex = priv->sh->bond.ifindex;
48 ifindex = priv->if_index;
55 * DPDK callback for Ethernet device configuration.
58 * Pointer to Ethernet device structure.
61 * 0 on success, a negative errno value otherwise and rte_errno is set.
64 mlx5_dev_configure(struct rte_eth_dev *dev)
66 struct mlx5_priv *priv = dev->data->dev_private;
67 unsigned int rxqs_n = dev->data->nb_rx_queues;
68 unsigned int txqs_n = dev->data->nb_tx_queues;
69 const uint8_t use_app_rss_key =
70 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
73 if (use_app_rss_key &&
74 (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
75 MLX5_RSS_HASH_KEY_LEN)) {
76 DRV_LOG(ERR, "port %u RSS key len must be %s Bytes long",
77 dev->data->port_id, RTE_STR(MLX5_RSS_HASH_KEY_LEN));
81 priv->rss_conf.rss_key =
82 mlx5_realloc(priv->rss_conf.rss_key, MLX5_MEM_RTE,
83 MLX5_RSS_HASH_KEY_LEN, 0, SOCKET_ID_ANY);
84 if (!priv->rss_conf.rss_key) {
85 DRV_LOG(ERR, "port %u cannot allocate RSS hash key memory (%u)",
86 dev->data->port_id, rxqs_n);
91 if ((dev->data->dev_conf.txmode.offloads &
92 DEV_TX_OFFLOAD_SEND_ON_TIMESTAMP) &&
93 rte_mbuf_dyn_tx_timestamp_register(NULL, NULL) != 0) {
94 DRV_LOG(ERR, "port %u cannot register Tx timestamp field/flag",
98 memcpy(priv->rss_conf.rss_key,
100 dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
101 rss_hash_default_key,
102 MLX5_RSS_HASH_KEY_LEN);
103 priv->rss_conf.rss_key_len = MLX5_RSS_HASH_KEY_LEN;
104 priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
105 priv->rxqs = (void *)dev->data->rx_queues;
106 priv->txqs = (void *)dev->data->tx_queues;
107 if (txqs_n != priv->txqs_n) {
108 DRV_LOG(INFO, "port %u Tx queues number update: %u -> %u",
109 dev->data->port_id, priv->txqs_n, txqs_n);
110 priv->txqs_n = txqs_n;
112 if (rxqs_n > priv->config.ind_table_max_size) {
113 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
114 dev->data->port_id, rxqs_n);
118 if (rxqs_n != priv->rxqs_n) {
119 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
120 dev->data->port_id, priv->rxqs_n, rxqs_n);
121 priv->rxqs_n = rxqs_n;
123 priv->skip_default_rss_reta = 0;
124 ret = mlx5_proc_priv_init(dev);
131 * Configure default RSS reta.
134 * Pointer to Ethernet device structure.
137 * 0 on success, a negative errno value otherwise and rte_errno is set.
140 mlx5_dev_configure_rss_reta(struct rte_eth_dev *dev)
142 struct mlx5_priv *priv = dev->data->dev_private;
143 unsigned int rxqs_n = dev->data->nb_rx_queues;
146 unsigned int reta_idx_n;
148 unsigned int *rss_queue_arr = NULL;
149 unsigned int rss_queue_n = 0;
151 if (priv->skip_default_rss_reta)
153 rss_queue_arr = mlx5_malloc(0, rxqs_n * sizeof(unsigned int), 0,
155 if (!rss_queue_arr) {
156 DRV_LOG(ERR, "port %u cannot allocate RSS queue list (%u)",
157 dev->data->port_id, rxqs_n);
161 for (i = 0, j = 0; i < rxqs_n; i++) {
162 struct mlx5_rxq_data *rxq_data;
163 struct mlx5_rxq_ctrl *rxq_ctrl;
165 rxq_data = (*priv->rxqs)[i];
166 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
167 if (rxq_ctrl && rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
168 rss_queue_arr[j++] = i;
171 if (rss_queue_n > priv->config.ind_table_max_size) {
172 DRV_LOG(ERR, "port %u cannot handle this many Rx queues (%u)",
173 dev->data->port_id, rss_queue_n);
175 mlx5_free(rss_queue_arr);
178 DRV_LOG(INFO, "port %u Rx queues number update: %u -> %u",
179 dev->data->port_id, priv->rxqs_n, rxqs_n);
180 priv->rxqs_n = rxqs_n;
182 * If the requested number of RX queues is not a power of two,
183 * use the maximum indirection table size for better balancing.
184 * The result is always rounded to the next power of two.
186 reta_idx_n = (1 << log2above((rss_queue_n & (rss_queue_n - 1)) ?
187 priv->config.ind_table_max_size :
189 ret = mlx5_rss_reta_index_resize(dev, reta_idx_n);
191 mlx5_free(rss_queue_arr);
195 * When the number of RX queues is not a power of two,
196 * the remaining table entries are padded with reused WQs
197 * and hashes are not spread uniformly.
199 for (i = 0, j = 0; (i != reta_idx_n); ++i) {
200 (*priv->reta_idx)[i] = rss_queue_arr[j];
201 if (++j == rss_queue_n)
204 mlx5_free(rss_queue_arr);
209 * Sets default tuning parameters.
212 * Pointer to Ethernet device.
214 * Info structure output buffer.
217 mlx5_set_default_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
219 struct mlx5_priv *priv = dev->data->dev_private;
221 /* Minimum CPU utilization. */
222 info->default_rxportconf.ring_size = 256;
223 info->default_txportconf.ring_size = 256;
224 info->default_rxportconf.burst_size = MLX5_RX_DEFAULT_BURST;
225 info->default_txportconf.burst_size = MLX5_TX_DEFAULT_BURST;
226 if ((priv->link_speed_capa & ETH_LINK_SPEED_200G) |
227 (priv->link_speed_capa & ETH_LINK_SPEED_100G)) {
228 info->default_rxportconf.nb_queues = 16;
229 info->default_txportconf.nb_queues = 16;
230 if (dev->data->nb_rx_queues > 2 ||
231 dev->data->nb_tx_queues > 2) {
232 /* Max Throughput. */
233 info->default_rxportconf.ring_size = 2048;
234 info->default_txportconf.ring_size = 2048;
237 info->default_rxportconf.nb_queues = 8;
238 info->default_txportconf.nb_queues = 8;
239 if (dev->data->nb_rx_queues > 2 ||
240 dev->data->nb_tx_queues > 2) {
241 /* Max Throughput. */
242 info->default_rxportconf.ring_size = 4096;
243 info->default_txportconf.ring_size = 4096;
249 * Sets tx mbuf limiting parameters.
252 * Pointer to Ethernet device.
254 * Info structure output buffer.
257 mlx5_set_txlimit_params(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
259 struct mlx5_priv *priv = dev->data->dev_private;
260 struct mlx5_dev_config *config = &priv->config;
264 inlen = (config->txq_inline_max == MLX5_ARG_UNSET) ?
265 MLX5_SEND_DEF_INLINE_LEN :
266 (unsigned int)config->txq_inline_max;
267 MLX5_ASSERT(config->txq_inline_min >= 0);
268 inlen = RTE_MAX(inlen, (unsigned int)config->txq_inline_min);
269 inlen = RTE_MIN(inlen, MLX5_WQE_SIZE_MAX +
270 MLX5_ESEG_MIN_INLINE_SIZE -
273 MLX5_WQE_DSEG_SIZE * 2);
274 nb_max = (MLX5_WQE_SIZE_MAX +
275 MLX5_ESEG_MIN_INLINE_SIZE -
279 inlen) / MLX5_WSEG_SIZE;
280 info->tx_desc_lim.nb_seg_max = nb_max;
281 info->tx_desc_lim.nb_mtu_seg_max = nb_max;
285 * DPDK callback to get information about the device.
288 * Pointer to Ethernet device structure.
290 * Info structure output buffer.
293 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
295 struct mlx5_priv *priv = dev->data->dev_private;
296 struct mlx5_dev_config *config = &priv->config;
299 /* FIXME: we should ask the device for these values. */
300 info->min_rx_bufsize = 32;
301 info->max_rx_pktlen = 65536;
302 info->max_lro_pkt_size = MLX5_MAX_LRO_SIZE;
304 * Since we need one CQ per QP, the limit is the minimum number
305 * between the two values.
307 max = RTE_MIN(priv->sh->device_attr.max_cq,
308 priv->sh->device_attr.max_qp);
309 /* max_rx_queues is uint16_t. */
310 max = RTE_MIN(max, (unsigned int)UINT16_MAX);
311 info->max_rx_queues = max;
312 info->max_tx_queues = max;
313 info->max_mac_addrs = MLX5_MAX_UC_MAC_ADDRESSES;
314 info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
315 info->rx_seg_capa.max_nseg = MLX5_MAX_RXQ_NSEG;
316 info->rx_seg_capa.multi_pools = !config->mprq.enabled;
317 info->rx_seg_capa.offset_allowed = !config->mprq.enabled;
318 info->rx_seg_capa.offset_align_log2 = 0;
319 info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
320 info->rx_queue_offload_capa);
321 info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
322 info->if_index = mlx5_ifindex(dev);
323 info->reta_size = priv->reta_idx_n ?
324 priv->reta_idx_n : config->ind_table_max_size;
325 info->hash_key_size = MLX5_RSS_HASH_KEY_LEN;
326 info->speed_capa = priv->link_speed_capa;
327 info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
328 mlx5_set_default_params(dev, info);
329 mlx5_set_txlimit_params(dev, info);
330 info->switch_info.name = dev->data->name;
331 info->switch_info.domain_id = priv->domain_id;
332 info->switch_info.port_id = priv->representor_id;
333 if (priv->representor) {
336 MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
337 struct mlx5_priv *opriv =
338 rte_eth_devices[port_id].data->dev_private;
341 opriv->representor ||
342 opriv->sh != priv->sh ||
343 opriv->domain_id != priv->domain_id)
346 * Override switch name with that of the master
349 info->switch_info.name = opriv->dev_data->name;
357 * Calculate representor ID from port switch info.
359 * Uint16 representor ID bits definition:
367 * Use this type if port is HPF.
370 * Encoded representor ID.
373 mlx5_representor_id_encode(const struct mlx5_switch_info *info,
374 enum rte_eth_representor_type hpf_type)
376 enum rte_eth_representor_type type = RTE_ETH_REPRESENTOR_VF;
377 uint16_t repr = info->port_name;
379 if (info->representor == 0)
381 if (info->name_type == MLX5_PHYS_PORT_NAME_TYPE_PFSF)
382 type = RTE_ETH_REPRESENTOR_SF;
383 if (info->name_type == MLX5_PHYS_PORT_NAME_TYPE_PFHPF) {
387 return MLX5_REPRESENTOR_ID(info->pf_num, type, repr);
391 * DPDK callback to get information about representor.
393 * Representor ID bits definition:
399 * Pointer to Ethernet device structure.
401 * Nullable info structure output buffer.
404 * negative on error, or the number of representor ranges.
407 mlx5_representor_info_get(struct rte_eth_dev *dev,
408 struct rte_eth_representor_info *info)
410 struct mlx5_priv *priv = dev->data->dev_private;
411 int n_type = 4; /* Representor types, VF, HPF@VF, SF and HPF@SF. */
412 int n_pf = 2; /* Number of PFs. */
417 info->controller = 0;
418 info->pf = priv->pf_bond >= 0 ? priv->pf_bond : 0;
419 for (pf = 0; pf < n_pf; ++pf) {
421 info->ranges[i].type = RTE_ETH_REPRESENTOR_VF;
422 info->ranges[i].controller = 0;
423 info->ranges[i].pf = pf;
424 info->ranges[i].vf = 0;
425 info->ranges[i].id_base =
426 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, 0);
427 info->ranges[i].id_end =
428 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, -1);
429 snprintf(info->ranges[i].name,
430 sizeof(info->ranges[i].name), "pf%dvf", pf);
432 /* HPF range of VF type. */
433 info->ranges[i].type = RTE_ETH_REPRESENTOR_VF;
434 info->ranges[i].controller = 0;
435 info->ranges[i].pf = pf;
436 info->ranges[i].vf = UINT16_MAX;
437 info->ranges[i].id_base =
438 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, -1);
439 info->ranges[i].id_end =
440 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, -1);
441 snprintf(info->ranges[i].name,
442 sizeof(info->ranges[i].name), "pf%dvf", pf);
445 info->ranges[i].type = RTE_ETH_REPRESENTOR_SF;
446 info->ranges[i].controller = 0;
447 info->ranges[i].pf = pf;
448 info->ranges[i].vf = 0;
449 info->ranges[i].id_base =
450 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, 0);
451 info->ranges[i].id_end =
452 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, -1);
453 snprintf(info->ranges[i].name,
454 sizeof(info->ranges[i].name), "pf%dsf", pf);
456 /* HPF range of SF type. */
457 info->ranges[i].type = RTE_ETH_REPRESENTOR_SF;
458 info->ranges[i].controller = 0;
459 info->ranges[i].pf = pf;
460 info->ranges[i].vf = UINT16_MAX;
461 info->ranges[i].id_base =
462 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, -1);
463 info->ranges[i].id_end =
464 MLX5_REPRESENTOR_ID(pf, info->ranges[i].type, -1);
465 snprintf(info->ranges[i].name,
466 sizeof(info->ranges[i].name), "pf%dsf", pf);
470 return n_type * n_pf;
474 * Get firmware version of a device.
477 * Ethernet device port.
479 * String output allocated by caller.
481 * Size of the output string, including terminating null byte.
484 * 0 on success, or the size of the non truncated string if too big.
487 mlx5_fw_version_get(struct rte_eth_dev *dev, char *fw_ver, size_t fw_size)
489 struct mlx5_priv *priv = dev->data->dev_private;
490 struct mlx5_dev_attr *attr = &priv->sh->device_attr;
491 size_t size = strnlen(attr->fw_ver, sizeof(attr->fw_ver)) + 1;
496 strlcpy(fw_ver, attr->fw_ver, fw_size);
501 * Get supported packet types.
504 * Pointer to Ethernet device structure.
507 * A pointer to the supported Packet types array.
510 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
512 static const uint32_t ptypes[] = {
513 /* refers to rxq_cq_to_pkt_type() */
515 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
516 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
517 RTE_PTYPE_L4_NONFRAG,
521 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
522 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
523 RTE_PTYPE_INNER_L4_NONFRAG,
524 RTE_PTYPE_INNER_L4_FRAG,
525 RTE_PTYPE_INNER_L4_TCP,
526 RTE_PTYPE_INNER_L4_UDP,
530 if (dev->rx_pkt_burst == mlx5_rx_burst ||
531 dev->rx_pkt_burst == mlx5_rx_burst_mprq ||
532 dev->rx_pkt_burst == mlx5_rx_burst_vec ||
533 dev->rx_pkt_burst == mlx5_rx_burst_mprq_vec)
539 * DPDK callback to change the MTU.
542 * Pointer to Ethernet device structure.
547 * 0 on success, a negative errno value otherwise and rte_errno is set.
550 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
552 struct mlx5_priv *priv = dev->data->dev_private;
553 uint16_t kern_mtu = 0;
556 ret = mlx5_get_mtu(dev, &kern_mtu);
559 /* Set kernel interface MTU first. */
560 ret = mlx5_set_mtu(dev, mtu);
563 ret = mlx5_get_mtu(dev, &kern_mtu);
566 if (kern_mtu == mtu) {
568 DRV_LOG(DEBUG, "port %u adapter MTU set to %u",
569 dev->data->port_id, mtu);
577 * Configure the RX function to use.
580 * Pointer to private data structure.
583 * Pointer to selected Rx burst function.
586 mlx5_select_rx_function(struct rte_eth_dev *dev)
588 eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
590 MLX5_ASSERT(dev != NULL);
591 if (mlx5_check_vec_rx_support(dev) > 0) {
592 if (mlx5_mprq_enabled(dev)) {
593 rx_pkt_burst = mlx5_rx_burst_mprq_vec;
594 DRV_LOG(DEBUG, "port %u selected vectorized"
595 " MPRQ Rx function", dev->data->port_id);
597 rx_pkt_burst = mlx5_rx_burst_vec;
598 DRV_LOG(DEBUG, "port %u selected vectorized"
599 " SPRQ Rx function", dev->data->port_id);
601 } else if (mlx5_mprq_enabled(dev)) {
602 rx_pkt_burst = mlx5_rx_burst_mprq;
603 DRV_LOG(DEBUG, "port %u selected MPRQ Rx function",
606 DRV_LOG(DEBUG, "port %u selected SPRQ Rx function",
613 * Get the E-Switch parameters by port id.
618 * Device port id is valid, skip check. This flag is useful
619 * when trials are performed from probing and device is not
620 * flagged as valid yet (in attaching process).
621 * @param[out] es_domain_id
622 * E-Switch domain id.
623 * @param[out] es_port_id
624 * The port id of the port in the E-Switch.
627 * pointer to device private data structure containing data needed
628 * on success, NULL otherwise and rte_errno is set.
631 mlx5_port_to_eswitch_info(uint16_t port, bool valid)
633 struct rte_eth_dev *dev;
634 struct mlx5_priv *priv;
636 if (port >= RTE_MAX_ETHPORTS) {
640 if (!valid && !rte_eth_dev_is_valid_port(port)) {
644 dev = &rte_eth_devices[port];
645 priv = dev->data->dev_private;
646 if (!(priv->representor || priv->master)) {
654 * Get the E-Switch parameters by device instance.
658 * @param[out] es_domain_id
659 * E-Switch domain id.
660 * @param[out] es_port_id
661 * The port id of the port in the E-Switch.
664 * pointer to device private data structure containing data needed
665 * on success, NULL otherwise and rte_errno is set.
668 mlx5_dev_to_eswitch_info(struct rte_eth_dev *dev)
670 struct mlx5_priv *priv;
672 priv = dev->data->dev_private;
673 if (!(priv->representor || priv->master)) {
681 * DPDK callback to retrieve hairpin capabilities.
684 * Pointer to Ethernet device structure.
686 * Storage for hairpin capability data.
689 * 0 on success, a negative errno value otherwise and rte_errno is set.
692 mlx5_hairpin_cap_get(struct rte_eth_dev *dev, struct rte_eth_hairpin_cap *cap)
694 struct mlx5_priv *priv = dev->data->dev_private;
695 struct mlx5_dev_config *config = &priv->config;
697 if (!priv->sh->devx || !config->dest_tir || !config->dv_flow_en) {
701 cap->max_nb_queues = UINT16_MAX;
702 cap->max_rx_2_tx = 1;
703 cap->max_tx_2_rx = 1;
704 cap->max_nb_desc = 8192;