drivers/net/mlx5/mlx5_rxq.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2015 6WIND S.A.
   3  * Copyright 2015 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <stddef.h>
   7 #include <assert.h>
   8 #include <errno.h>
   9 #include <string.h>
  10 #include <stdint.h>
  11 #include <fcntl.h>
  12 #include <sys/queue.h>
  13
  14 /* Verbs header. */
  15 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  16 #ifdef PEDANTIC
  17 #pragma GCC diagnostic ignored "-Wpedantic"
  18 #endif
  19 #include <infiniband/verbs.h>
  20 #include <infiniband/mlx5dv.h>
  21 #ifdef PEDANTIC
  22 #pragma GCC diagnostic error "-Wpedantic"
  23 #endif
  24
  25 #include <rte_mbuf.h>
  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>
  31 #include <rte_io.h>
  32
  33 #include "mlx5.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"
  39
  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,
  52 };
  53
  54 /* Length of the default RSS hash key. */
  55 const size_t rss_hash_default_key_len = sizeof(rss_hash_default_key);
  56
  57 /**
  58  * Allocate RX queue elements.
  59  *
  60  * @param rxq_ctrl
  61  *   Pointer to RX queue structure.
  62  *
  63  * @return
  64  *   0 on success, a negative errno value otherwise and rte_errno is set.
  65  */
  66 int
  67 rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
  68 {
  69         const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
  70         unsigned int elts_n = 1 << rxq_ctrl->rxq.elts_n;
  71         unsigned int i;
  72         int err;
  73
  74         /* Iterate on segments. */
  75         for (i = 0; (i != elts_n); ++i) {
  76                 struct rte_mbuf *buf;
  77
  78                 buf = rte_pktmbuf_alloc(rxq_ctrl->rxq.mp);
  79                 if (buf == NULL) {
  80                         DRV_LOG(ERR, "port %u empty mbuf pool",
  81                                 PORT_ID(rxq_ctrl->priv));
  82                         rte_errno = ENOMEM;
  83                         goto error;
  84                 }
  85                 /* Headroom is reserved by rte_pktmbuf_alloc(). */
  86                 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
  87                 /* Buffer is supposed to be empty. */
  88                 assert(rte_pktmbuf_data_len(buf) == 0);
  89                 assert(rte_pktmbuf_pkt_len(buf) == 0);
  90                 assert(!buf->next);
  91                 /* Only the first segment keeps headroom. */
  92                 if (i % sges_n)
  93                         SET_DATA_OFF(buf, 0);
  94                 PORT(buf) = rxq_ctrl->rxq.port_id;
  95                 DATA_LEN(buf) = rte_pktmbuf_tailroom(buf);
  96                 PKT_LEN(buf) = DATA_LEN(buf);
  97                 NB_SEGS(buf) = 1;
  98                 (*rxq_ctrl->rxq.elts)[i] = buf;
  99         }
 100         /* If Rx vector is activated. */
 101         if (mlx5_rxq_check_vec_support(&rxq_ctrl->rxq) > 0) {
 102                 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
 103                 struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
 104                 int j;
 105
 106                 /* Initialize default rearm_data for vPMD. */
 107                 mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
 108                 rte_mbuf_refcnt_set(mbuf_init, 1);
 109                 mbuf_init->nb_segs = 1;
 110                 mbuf_init->port = rxq->port_id;
 111                 /*
 112                  * prevent compiler reordering:
 113                  * rearm_data covers previous fields.
 114                  */
 115                 rte_compiler_barrier();
 116                 rxq->mbuf_initializer =
 117                         *(uint64_t *)&mbuf_init->rearm_data;
 118                 /* Padding with a fake mbuf for vectorized Rx. */
 119                 for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
 120                         (*rxq->elts)[elts_n + j] = &rxq->fake_mbuf;
 121         }
 122         DRV_LOG(DEBUG,
 123                 "port %u Rx queue %u allocated and configured %u segments"
 124                 " (max %u packets)",
 125                 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx, elts_n,
 126                 elts_n / (1 << rxq_ctrl->rxq.sges_n));
 127         return 0;
 128 error:
 129         err = rte_errno; /* Save rte_errno before cleanup. */
 130         elts_n = i;
 131         for (i = 0; (i != elts_n); ++i) {
 132                 if ((*rxq_ctrl->rxq.elts)[i] != NULL)
 133                         rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
 134                 (*rxq_ctrl->rxq.elts)[i] = NULL;
 135         }
 136         DRV_LOG(DEBUG, "port %u Rx queue %u failed, freed everything",
 137                 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx);
 138         rte_errno = err; /* Restore rte_errno. */
 139         return -rte_errno;
 140 }
 141
 142 /**
 143  * Free RX queue elements.
 144  *
 145  * @param rxq_ctrl
 146  *   Pointer to RX queue structure.
 147  */
 148 static void
 149 rxq_free_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
 150 {
 151         struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
 152         const uint16_t q_n = (1 << rxq->elts_n);
 153         const uint16_t q_mask = q_n - 1;
 154         uint16_t used = q_n - (rxq->rq_ci - rxq->rq_pi);
 155         uint16_t i;
 156
 157         DRV_LOG(DEBUG, "port %u Rx queue %u freeing WRs",
 158                 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx);
 159         if (rxq->elts == NULL)
 160                 return;
 161         /**
 162          * Some mbuf in the Ring belongs to the application.  They cannot be
 163          * freed.
 164          */
 165         if (mlx5_rxq_check_vec_support(rxq) > 0) {
 166                 for (i = 0; i < used; ++i)
 167                         (*rxq->elts)[(rxq->rq_ci + i) & q_mask] = NULL;
 168                 rxq->rq_pi = rxq->rq_ci;
 169         }
 170         for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
 171                 if ((*rxq->elts)[i] != NULL)
 172                         rte_pktmbuf_free_seg((*rxq->elts)[i]);
 173                 (*rxq->elts)[i] = NULL;
 174         }
 175 }
 176
 177 /**
 178  * Clean up a RX queue.
 179  *
 180  * Destroy objects, free allocated memory and reset the structure for reuse.
 181  *
 182  * @param rxq_ctrl
 183  *   Pointer to RX queue structure.
 184  */
 185 void
 186 mlx5_rxq_cleanup(struct mlx5_rxq_ctrl *rxq_ctrl)
 187 {
 188         DRV_LOG(DEBUG, "port %u cleaning up Rx queue %u",
 189                 PORT_ID(rxq_ctrl->priv), rxq_ctrl->idx);
 190         if (rxq_ctrl->ibv)
 191                 mlx5_rxq_ibv_release(rxq_ctrl->ibv);
 192         memset(rxq_ctrl, 0, sizeof(*rxq_ctrl));
 193 }
 194
 195 /**
 196  * Returns the per-queue supported offloads.
 197  *
 198  * @param dev
 199  *   Pointer to Ethernet device.
 200  *
 201  * @return
 202  *   Supported Rx offloads.
 203  */
 204 uint64_t
 205 mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
 206 {
 207         struct priv *priv = dev->data->dev_private;
 208         struct mlx5_dev_config *config = &priv->config;
 209         uint64_t offloads = (DEV_RX_OFFLOAD_SCATTER |
 210                              DEV_RX_OFFLOAD_TIMESTAMP |
 211                              DEV_RX_OFFLOAD_JUMBO_FRAME);
 212
 213         if (config->hw_fcs_strip)
 214                 offloads |= DEV_RX_OFFLOAD_CRC_STRIP;
 215         if (config->hw_csum)
 216                 offloads |= (DEV_RX_OFFLOAD_IPV4_CKSUM |
 217                              DEV_RX_OFFLOAD_UDP_CKSUM |
 218                              DEV_RX_OFFLOAD_TCP_CKSUM);
 219         if (config->hw_vlan_strip)
 220                 offloads |= DEV_RX_OFFLOAD_VLAN_STRIP;
 221         return offloads;
 222 }
 223
 224
 225 /**
 226  * Returns the per-port supported offloads.
 227  *
 228  * @return
 229  *   Supported Rx offloads.
 230  */
 231 uint64_t
 232 mlx5_get_rx_port_offloads(void)
 233 {
 234         uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
 235
 236         return offloads;
 237 }
 238
 239 /**
 240  *
 241  * @param dev
 242  *   Pointer to Ethernet device structure.
 243  * @param idx
 244  *   RX queue index.
 245  * @param desc
 246  *   Number of descriptors to configure in queue.
 247  * @param socket
 248  *   NUMA socket on which memory must be allocated.
 249  * @param[in] conf
 250  *   Thresholds parameters.
 251  * @param mp
 252  *   Memory pool for buffer allocations.
 253  *
 254  * @return
 255  *   0 on success, a negative errno value otherwise and rte_errno is set.
 256  */
 257 int
 258 mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 259                     unsigned int socket, const struct rte_eth_rxconf *conf,
 260                     struct rte_mempool *mp)
 261 {
 262         struct priv *priv = dev->data->dev_private;
 263         struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
 264         struct mlx5_rxq_ctrl *rxq_ctrl =
 265                 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
 266
 267         if (!rte_is_power_of_2(desc)) {
 268                 desc = 1 << log2above(desc);
 269                 DRV_LOG(WARNING,
 270                         "port %u increased number of descriptors in Rx queue %u"
 271                         " to the next power of two (%d)",
 272                         dev->data->port_id, idx, desc);
 273         }
 274         DRV_LOG(DEBUG, "port %u configuring Rx queue %u for %u descriptors",
 275                 dev->data->port_id, idx, desc);
 276         if (idx >= priv->rxqs_n) {
 277                 DRV_LOG(ERR, "port %u Rx queue index out of range (%u >= %u)",
 278                         dev->data->port_id, idx, priv->rxqs_n);
 279                 rte_errno = EOVERFLOW;
 280                 return -rte_errno;
 281         }
 282         if (!mlx5_rxq_releasable(dev, idx)) {
 283                 DRV_LOG(ERR, "port %u unable to release queue index %u",
 284                         dev->data->port_id, idx);
 285                 rte_errno = EBUSY;
 286                 return -rte_errno;
 287         }
 288         mlx5_rxq_release(dev, idx);
 289         rxq_ctrl = mlx5_rxq_new(dev, idx, desc, socket, conf, mp);
 290         if (!rxq_ctrl) {
 291                 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
 292                         dev->data->port_id, idx);
 293                 rte_errno = ENOMEM;
 294                 return -rte_errno;
 295         }
 296         DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
 297                 dev->data->port_id, idx);
 298         (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
 299         return 0;
 300 }
 301
 302 /**
 303  * DPDK callback to release a RX queue.
 304  *
 305  * @param dpdk_rxq
 306  *   Generic RX queue pointer.
 307  */
 308 void
 309 mlx5_rx_queue_release(void *dpdk_rxq)
 310 {
 311         struct mlx5_rxq_data *rxq = (struct mlx5_rxq_data *)dpdk_rxq;
 312         struct mlx5_rxq_ctrl *rxq_ctrl;
 313         struct priv *priv;
 314
 315         if (rxq == NULL)
 316                 return;
 317         rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
 318         priv = rxq_ctrl->priv;
 319         if (!mlx5_rxq_releasable(ETH_DEV(priv), rxq_ctrl->rxq.stats.idx))
 320                 rte_panic("port %u Rx queue %u is still used by a flow and"
 321                           " cannot be removed\n",
 322                           PORT_ID(priv), rxq_ctrl->idx);
 323         mlx5_rxq_release(ETH_DEV(priv), rxq_ctrl->rxq.stats.idx);
 324 }
 325
 326 /**
 327  * Allocate queue vector and fill epoll fd list for Rx interrupts.
 328  *
 329  * @param dev
 330  *   Pointer to Ethernet device.
 331  *
 332  * @return
 333  *   0 on success, a negative errno value otherwise and rte_errno is set.
 334  */
 335 int
 336 mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
 337 {
 338         struct priv *priv = dev->data->dev_private;
 339         unsigned int i;
 340         unsigned int rxqs_n = priv->rxqs_n;
 341         unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
 342         unsigned int count = 0;
 343         struct rte_intr_handle *intr_handle = dev->intr_handle;
 344
 345         if (!dev->data->dev_conf.intr_conf.rxq)
 346                 return 0;
 347         mlx5_rx_intr_vec_disable(dev);
 348         intr_handle->intr_vec = malloc(n * sizeof(intr_handle->intr_vec[0]));
 349         if (intr_handle->intr_vec == NULL) {
 350                 DRV_LOG(ERR,
 351                         "port %u failed to allocate memory for interrupt"
 352                         " vector, Rx interrupts will not be supported",
 353                         dev->data->port_id);
 354                 rte_errno = ENOMEM;
 355                 return -rte_errno;
 356         }
 357         intr_handle->type = RTE_INTR_HANDLE_EXT;
 358         for (i = 0; i != n; ++i) {
 359                 /* This rxq ibv must not be released in this function. */
 360                 struct mlx5_rxq_ibv *rxq_ibv = mlx5_rxq_ibv_get(dev, i);
 361                 int fd;
 362                 int flags;
 363                 int rc;
 364
 365                 /* Skip queues that cannot request interrupts. */
 366                 if (!rxq_ibv || !rxq_ibv->channel) {
 367                         /* Use invalid intr_vec[] index to disable entry. */
 368                         intr_handle->intr_vec[i] =
 369                                 RTE_INTR_VEC_RXTX_OFFSET +
 370                                 RTE_MAX_RXTX_INTR_VEC_ID;
 371                         continue;
 372                 }
 373                 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
 374                         DRV_LOG(ERR,
 375                                 "port %u too many Rx queues for interrupt"
 376                                 " vector size (%d), Rx interrupts cannot be"
 377                                 " enabled",
 378                                 dev->data->port_id, RTE_MAX_RXTX_INTR_VEC_ID);
 379                         mlx5_rx_intr_vec_disable(dev);
 380                         rte_errno = ENOMEM;
 381                         return -rte_errno;
 382                 }
 383                 fd = rxq_ibv->channel->fd;
 384                 flags = fcntl(fd, F_GETFL);
 385                 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
 386                 if (rc < 0) {
 387                         rte_errno = errno;
 388                         DRV_LOG(ERR,
 389                                 "port %u failed to make Rx interrupt file"
 390                                 " descriptor %d non-blocking for queue index"
 391                                 " %d",
 392                                 dev->data->port_id, fd, i);
 393                         mlx5_rx_intr_vec_disable(dev);
 394                         return -rte_errno;
 395                 }
 396                 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
 397                 intr_handle->efds[count] = fd;
 398                 count++;
 399         }
 400         if (!count)
 401                 mlx5_rx_intr_vec_disable(dev);
 402         else
 403                 intr_handle->nb_efd = count;
 404         return 0;
 405 }
 406
 407 /**
 408  * Clean up Rx interrupts handler.
 409  *
 410  * @param dev
 411  *   Pointer to Ethernet device.
 412  */
 413 void
 414 mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
 415 {
 416         struct priv *priv = dev->data->dev_private;
 417         struct rte_intr_handle *intr_handle = dev->intr_handle;
 418         unsigned int i;
 419         unsigned int rxqs_n = priv->rxqs_n;
 420         unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
 421
 422         if (!dev->data->dev_conf.intr_conf.rxq)
 423                 return;
 424         if (!intr_handle->intr_vec)
 425                 goto free;
 426         for (i = 0; i != n; ++i) {
 427                 struct mlx5_rxq_ctrl *rxq_ctrl;
 428                 struct mlx5_rxq_data *rxq_data;
 429
 430                 if (intr_handle->intr_vec[i] == RTE_INTR_VEC_RXTX_OFFSET +
 431                     RTE_MAX_RXTX_INTR_VEC_ID)
 432                         continue;
 433                 /**
 434                  * Need to access directly the queue to release the reference
 435                  * kept in priv_rx_intr_vec_enable().
 436                  */
 437                 rxq_data = (*priv->rxqs)[i];
 438                 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 439                 mlx5_rxq_ibv_release(rxq_ctrl->ibv);
 440         }
 441 free:
 442         rte_intr_free_epoll_fd(intr_handle);
 443         if (intr_handle->intr_vec)
 444                 free(intr_handle->intr_vec);
 445         intr_handle->nb_efd = 0;
 446         intr_handle->intr_vec = NULL;
 447 }
 448
 449 /**
 450  *  MLX5 CQ notification .
 451  *
 452  *  @param rxq
 453  *     Pointer to receive queue structure.
 454  *  @param sq_n_rxq
 455  *     Sequence number per receive queue .
 456  */
 457 static inline void
 458 mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
 459 {
 460         int sq_n = 0;
 461         uint32_t doorbell_hi;
 462         uint64_t doorbell;
 463         void *cq_db_reg = (char *)rxq->cq_uar + MLX5_CQ_DOORBELL;
 464
 465         sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
 466         doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
 467         doorbell = (uint64_t)doorbell_hi << 32;
 468         doorbell |=  rxq->cqn;
 469         rxq->cq_db[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
 470         rte_write64(rte_cpu_to_be_64(doorbell), cq_db_reg);
 471 }
 472
 473 /**
 474  * DPDK callback for Rx queue interrupt enable.
 475  *
 476  * @param dev
 477  *   Pointer to Ethernet device structure.
 478  * @param rx_queue_id
 479  *   Rx queue number.
 480  *
 481  * @return
 482  *   0 on success, a negative errno value otherwise and rte_errno is set.
 483  */
 484 int
 485 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 486 {
 487         struct priv *priv = dev->data->dev_private;
 488         struct mlx5_rxq_data *rxq_data;
 489         struct mlx5_rxq_ctrl *rxq_ctrl;
 490
 491         rxq_data = (*priv->rxqs)[rx_queue_id];
 492         if (!rxq_data) {
 493                 rte_errno = EINVAL;
 494                 return -rte_errno;
 495         }
 496         rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 497         if (rxq_ctrl->irq) {
 498                 struct mlx5_rxq_ibv *rxq_ibv;
 499
 500                 rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
 501                 if (!rxq_ibv) {
 502                         rte_errno = EINVAL;
 503                         return -rte_errno;
 504                 }
 505                 mlx5_arm_cq(rxq_data, rxq_data->cq_arm_sn);
 506                 mlx5_rxq_ibv_release(rxq_ibv);
 507         }
 508         return 0;
 509 }
 510
 511 /**
 512  * DPDK callback for Rx queue interrupt disable.
 513  *
 514  * @param dev
 515  *   Pointer to Ethernet device structure.
 516  * @param rx_queue_id
 517  *   Rx queue number.
 518  *
 519  * @return
 520  *   0 on success, a negative errno value otherwise and rte_errno is set.
 521  */
 522 int
 523 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 524 {
 525         struct priv *priv = dev->data->dev_private;
 526         struct mlx5_rxq_data *rxq_data;
 527         struct mlx5_rxq_ctrl *rxq_ctrl;
 528         struct mlx5_rxq_ibv *rxq_ibv = NULL;
 529         struct ibv_cq *ev_cq;
 530         void *ev_ctx;
 531         int ret;
 532
 533         rxq_data = (*priv->rxqs)[rx_queue_id];
 534         if (!rxq_data) {
 535                 rte_errno = EINVAL;
 536                 return -rte_errno;
 537         }
 538         rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 539         if (!rxq_ctrl->irq)
 540                 return 0;
 541         rxq_ibv = mlx5_rxq_ibv_get(dev, rx_queue_id);
 542         if (!rxq_ibv) {
 543                 rte_errno = EINVAL;
 544                 return -rte_errno;
 545         }
 546         ret = mlx5_glue->get_cq_event(rxq_ibv->channel, &ev_cq, &ev_ctx);
 547         if (ret || ev_cq != rxq_ibv->cq) {
 548                 rte_errno = EINVAL;
 549                 goto exit;
 550         }
 551         rxq_data->cq_arm_sn++;
 552         mlx5_glue->ack_cq_events(rxq_ibv->cq, 1);
 553         return 0;
 554 exit:
 555         ret = rte_errno; /* Save rte_errno before cleanup. */
 556         if (rxq_ibv)
 557                 mlx5_rxq_ibv_release(rxq_ibv);
 558         DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
 559                 dev->data->port_id, rx_queue_id);
 560         rte_errno = ret; /* Restore rte_errno. */
 561         return -rte_errno;
 562 }
 563
 564 /**
 565  * Create the Rx queue Verbs object.
 566  *
 567  * @param dev
 568  *   Pointer to Ethernet device.
 569  * @param idx
 570  *   Queue index in DPDK Rx queue array
 571  *
 572  * @return
 573  *   The Verbs object initialised, NULL otherwise and rte_errno is set.
 574  */
 575 struct mlx5_rxq_ibv *
 576 mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx)
 577 {
 578         struct priv *priv = dev->data->dev_private;
 579         struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
 580         struct mlx5_rxq_ctrl *rxq_ctrl =
 581                 container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 582         struct ibv_wq_attr mod;
 583         union {
 584                 struct {
 585                         struct ibv_cq_init_attr_ex ibv;
 586                         struct mlx5dv_cq_init_attr mlx5;
 587                 } cq;
 588                 struct ibv_wq_init_attr wq;
 589                 struct ibv_cq_ex cq_attr;
 590         } attr;
 591         unsigned int cqe_n = (1 << rxq_data->elts_n) - 1;
 592         struct mlx5_rxq_ibv *tmpl;
 593         struct mlx5dv_cq cq_info;
 594         struct mlx5dv_rwq rwq;
 595         unsigned int i;
 596         int ret = 0;
 597         struct mlx5dv_obj obj;
 598         struct mlx5_dev_config *config = &priv->config;
 599
 600         assert(rxq_data);
 601         assert(!rxq_ctrl->ibv);
 602         priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_RX_QUEUE;
 603         priv->verbs_alloc_ctx.obj = rxq_ctrl;
 604         tmpl = rte_calloc_socket(__func__, 1, sizeof(*tmpl), 0,
 605                                  rxq_ctrl->socket);
 606         if (!tmpl) {
 607                 DRV_LOG(ERR,
 608                         "port %u Rx queue %u cannot allocate verbs resources",
 609                         dev->data->port_id, rxq_ctrl->idx);
 610                 rte_errno = ENOMEM;
 611                 goto error;
 612         }
 613         tmpl->rxq_ctrl = rxq_ctrl;
 614         if (rxq_ctrl->irq) {
 615                 tmpl->channel = mlx5_glue->create_comp_channel(priv->ctx);
 616                 if (!tmpl->channel) {
 617                         DRV_LOG(ERR, "port %u: comp channel creation failure",
 618                                 dev->data->port_id);
 619                         rte_errno = ENOMEM;
 620                         goto error;
 621                 }
 622         }
 623         attr.cq.ibv = (struct ibv_cq_init_attr_ex){
 624                 .cqe = cqe_n,
 625                 .channel = tmpl->channel,
 626                 .comp_mask = 0,
 627         };
 628         attr.cq.mlx5 = (struct mlx5dv_cq_init_attr){
 629                 .comp_mask = 0,
 630         };
 631         if (config->cqe_comp && !rxq_data->hw_timestamp) {
 632                 attr.cq.mlx5.comp_mask |=
 633                         MLX5DV_CQ_INIT_ATTR_MASK_COMPRESSED_CQE;
 634                 attr.cq.mlx5.cqe_comp_res_format = MLX5DV_CQE_RES_FORMAT_HASH;
 635                 /*
 636                  * For vectorized Rx, it must not be doubled in order to
 637                  * make cq_ci and rq_ci aligned.
 638                  */
 639                 if (mlx5_rxq_check_vec_support(rxq_data) < 0)
 640                         attr.cq.ibv.cqe *= 2;
 641         } else if (config->cqe_comp && rxq_data->hw_timestamp) {
 642                 DRV_LOG(DEBUG,
 643                         "port %u Rx CQE compression is disabled for HW"
 644                         " timestamp",
 645                         dev->data->port_id);
 646         }
 647         tmpl->cq = mlx5_glue->cq_ex_to_cq
 648                 (mlx5_glue->dv_create_cq(priv->ctx, &attr.cq.ibv,
 649                                          &attr.cq.mlx5));
 650         if (tmpl->cq == NULL) {
 651                 DRV_LOG(ERR, "port %u Rx queue %u CQ creation failure",
 652                         dev->data->port_id, idx);
 653                 rte_errno = ENOMEM;
 654                 goto error;
 655         }
 656         DRV_LOG(DEBUG, "port %u priv->device_attr.max_qp_wr is %d",
 657                 dev->data->port_id, priv->device_attr.orig_attr.max_qp_wr);
 658         DRV_LOG(DEBUG, "port %u priv->device_attr.max_sge is %d",
 659                 dev->data->port_id, priv->device_attr.orig_attr.max_sge);
 660         attr.wq = (struct ibv_wq_init_attr){
 661                 .wq_context = NULL, /* Could be useful in the future. */
 662                 .wq_type = IBV_WQT_RQ,
 663                 /* Max number of outstanding WRs. */
 664                 .max_wr = (1 << rxq_data->elts_n) >> rxq_data->sges_n,
 665                 /* Max number of scatter/gather elements in a WR. */
 666                 .max_sge = 1 << rxq_data->sges_n,
 667                 .pd = priv->pd,
 668                 .cq = tmpl->cq,
 669                 .comp_mask =
 670                         IBV_WQ_FLAGS_CVLAN_STRIPPING |
 671                         0,
 672                 .create_flags = (rxq_data->vlan_strip ?
 673                                  IBV_WQ_FLAGS_CVLAN_STRIPPING :
 674                                  0),
 675         };
 676         /* By default, FCS (CRC) is stripped by hardware. */
 677         if (rxq_data->crc_present) {
 678                 attr.wq.create_flags |= IBV_WQ_FLAGS_SCATTER_FCS;
 679                 attr.wq.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
 680         }
 681 #ifdef HAVE_IBV_WQ_FLAG_RX_END_PADDING
 682         if (config->hw_padding) {
 683                 attr.wq.create_flags |= IBV_WQ_FLAG_RX_END_PADDING;
 684                 attr.wq.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
 685         }
 686 #endif
 687         tmpl->wq = mlx5_glue->create_wq(priv->ctx, &attr.wq);
 688         if (tmpl->wq == NULL) {
 689                 DRV_LOG(ERR, "port %u Rx queue %u WQ creation failure",
 690                         dev->data->port_id, idx);
 691                 rte_errno = ENOMEM;
 692                 goto error;
 693         }
 694         /*
 695          * Make sure number of WRs*SGEs match expectations since a queue
 696          * cannot allocate more than "desc" buffers.
 697          */
 698         if (((int)attr.wq.max_wr !=
 699              ((1 << rxq_data->elts_n) >> rxq_data->sges_n)) ||
 700             ((int)attr.wq.max_sge != (1 << rxq_data->sges_n))) {
 701                 DRV_LOG(ERR,
 702                         "port %u Rx queue %u requested %u*%u but got %u*%u"
 703                         " WRs*SGEs",
 704                         dev->data->port_id, idx,
 705                         ((1 << rxq_data->elts_n) >> rxq_data->sges_n),
 706                         (1 << rxq_data->sges_n),
 707                         attr.wq.max_wr, attr.wq.max_sge);
 708                 rte_errno = EINVAL;
 709                 goto error;
 710         }
 711         /* Change queue state to ready. */
 712         mod = (struct ibv_wq_attr){
 713                 .attr_mask = IBV_WQ_ATTR_STATE,
 714                 .wq_state = IBV_WQS_RDY,
 715         };
 716         ret = mlx5_glue->modify_wq(tmpl->wq, &mod);
 717         if (ret) {
 718                 DRV_LOG(ERR,
 719                         "port %u Rx queue %u WQ state to IBV_WQS_RDY failed",
 720                         dev->data->port_id, idx);
 721                 rte_errno = ret;
 722                 goto error;
 723         }
 724         obj.cq.in = tmpl->cq;
 725         obj.cq.out = &cq_info;
 726         obj.rwq.in = tmpl->wq;
 727         obj.rwq.out = &rwq;
 728         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_RWQ);
 729         if (ret) {
 730                 rte_errno = ret;
 731                 goto error;
 732         }
 733         if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
 734                 DRV_LOG(ERR,
 735                         "port %u wrong MLX5_CQE_SIZE environment variable"
 736                         " value: it should be set to %u",
 737                         dev->data->port_id, RTE_CACHE_LINE_SIZE);
 738                 rte_errno = EINVAL;
 739                 goto error;
 740         }
 741         /* Fill the rings. */
 742         rxq_data->wqes = (volatile struct mlx5_wqe_data_seg (*)[])
 743                 (uintptr_t)rwq.buf;
 744         for (i = 0; (i != (unsigned int)(1 << rxq_data->elts_n)); ++i) {
 745                 struct rte_mbuf *buf = (*rxq_data->elts)[i];
 746                 volatile struct mlx5_wqe_data_seg *scat = &(*rxq_data->wqes)[i];
 747
 748                 /* scat->addr must be able to store a pointer. */
 749                 assert(sizeof(scat->addr) >= sizeof(uintptr_t));
 750                 *scat = (struct mlx5_wqe_data_seg){
 751                         .addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf,
 752                                                                   uintptr_t)),
 753                         .byte_count = rte_cpu_to_be_32(DATA_LEN(buf)),
 754                         .lkey = UINT32_MAX,
 755                 };
 756         }
 757         rxq_data->rq_db = rwq.dbrec;
 758         rxq_data->cqe_n = log2above(cq_info.cqe_cnt);
 759         rxq_data->cq_ci = 0;
 760         rxq_data->rq_ci = 0;
 761         rxq_data->rq_pi = 0;
 762         rxq_data->zip = (struct rxq_zip){
 763                 .ai = 0,
 764         };
 765         rxq_data->cq_db = cq_info.dbrec;
 766         rxq_data->cqes = (volatile struct mlx5_cqe (*)[])(uintptr_t)cq_info.buf;
 767         rxq_data->cq_uar = cq_info.cq_uar;
 768         rxq_data->cqn = cq_info.cqn;
 769         rxq_data->cq_arm_sn = 0;
 770         /* Update doorbell counter. */
 771         rxq_data->rq_ci = (1 << rxq_data->elts_n) >> rxq_data->sges_n;
 772         rte_wmb();
 773         *rxq_data->rq_db = rte_cpu_to_be_32(rxq_data->rq_ci);
 774         DRV_LOG(DEBUG, "port %u rxq %u updated with %p", dev->data->port_id,
 775                 idx, (void *)&tmpl);
 776         rte_atomic32_inc(&tmpl->refcnt);
 777         DRV_LOG(DEBUG, "port %u Verbs Rx queue %u: refcnt %d",
 778                 dev->data->port_id, idx, rte_atomic32_read(&tmpl->refcnt));
 779         LIST_INSERT_HEAD(&priv->rxqsibv, tmpl, next);
 780         priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
 781         return tmpl;
 782 error:
 783         ret = rte_errno; /* Save rte_errno before cleanup. */
 784         if (tmpl->wq)
 785                 claim_zero(mlx5_glue->destroy_wq(tmpl->wq));
 786         if (tmpl->cq)
 787                 claim_zero(mlx5_glue->destroy_cq(tmpl->cq));
 788         if (tmpl->channel)
 789                 claim_zero(mlx5_glue->destroy_comp_channel(tmpl->channel));
 790         priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
 791         rte_errno = ret; /* Restore rte_errno. */
 792         return NULL;
 793 }
 794
 795 /**
 796  * Get an Rx queue Verbs object.
 797  *
 798  * @param dev
 799  *   Pointer to Ethernet device.
 800  * @param idx
 801  *   Queue index in DPDK Rx queue array
 802  *
 803  * @return
 804  *   The Verbs object if it exists.
 805  */
 806 struct mlx5_rxq_ibv *
 807 mlx5_rxq_ibv_get(struct rte_eth_dev *dev, uint16_t idx)
 808 {
 809         struct priv *priv = dev->data->dev_private;
 810         struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
 811         struct mlx5_rxq_ctrl *rxq_ctrl;
 812
 813         if (idx >= priv->rxqs_n)
 814                 return NULL;
 815         if (!rxq_data)
 816                 return NULL;
 817         rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 818         if (rxq_ctrl->ibv) {
 819                 rte_atomic32_inc(&rxq_ctrl->ibv->refcnt);
 820                 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u: refcnt %d",
 821                         dev->data->port_id, rxq_ctrl->idx,
 822                         rte_atomic32_read(&rxq_ctrl->ibv->refcnt));
 823         }
 824         return rxq_ctrl->ibv;
 825 }
 826
 827 /**
 828  * Release an Rx verbs queue object.
 829  *
 830  * @param rxq_ibv
 831  *   Verbs Rx queue object.
 832  *
 833  * @return
 834  *   1 while a reference on it exists, 0 when freed.
 835  */
 836 int
 837 mlx5_rxq_ibv_release(struct mlx5_rxq_ibv *rxq_ibv)
 838 {
 839         assert(rxq_ibv);
 840         assert(rxq_ibv->wq);
 841         assert(rxq_ibv->cq);
 842         DRV_LOG(DEBUG, "port %u Verbs Rx queue %u: refcnt %d",
 843                 PORT_ID(rxq_ibv->rxq_ctrl->priv),
 844                 rxq_ibv->rxq_ctrl->idx, rte_atomic32_read(&rxq_ibv->refcnt));
 845         if (rte_atomic32_dec_and_test(&rxq_ibv->refcnt)) {
 846                 rxq_free_elts(rxq_ibv->rxq_ctrl);
 847                 claim_zero(mlx5_glue->destroy_wq(rxq_ibv->wq));
 848                 claim_zero(mlx5_glue->destroy_cq(rxq_ibv->cq));
 849                 if (rxq_ibv->channel)
 850                         claim_zero(mlx5_glue->destroy_comp_channel
 851                                    (rxq_ibv->channel));
 852                 LIST_REMOVE(rxq_ibv, next);
 853                 rte_free(rxq_ibv);
 854                 return 0;
 855         }
 856         return 1;
 857 }
 858
 859 /**
 860  * Verify the Verbs Rx queue list is empty
 861  *
 862  * @param dev
 863  *   Pointer to Ethernet device.
 864  *
 865  * @return
 866  *   The number of object not released.
 867  */
 868 int
 869 mlx5_rxq_ibv_verify(struct rte_eth_dev *dev)
 870 {
 871         struct priv *priv = dev->data->dev_private;
 872         int ret = 0;
 873         struct mlx5_rxq_ibv *rxq_ibv;
 874
 875         LIST_FOREACH(rxq_ibv, &priv->rxqsibv, next) {
 876                 DRV_LOG(DEBUG, "port %u Verbs Rx queue %u still referenced",
 877                         dev->data->port_id, rxq_ibv->rxq_ctrl->idx);
 878                 ++ret;
 879         }
 880         return ret;
 881 }
 882
 883 /**
 884  * Return true if a single reference exists on the object.
 885  *
 886  * @param rxq_ibv
 887  *   Verbs Rx queue object.
 888  */
 889 int
 890 mlx5_rxq_ibv_releasable(struct mlx5_rxq_ibv *rxq_ibv)
 891 {
 892         assert(rxq_ibv);
 893         return (rte_atomic32_read(&rxq_ibv->refcnt) == 1);
 894 }
 895
 896 /**
 897  * Create a DPDK Rx queue.
 898  *
 899  * @param dev
 900  *   Pointer to Ethernet device.
 901  * @param idx
 902  *   TX queue index.
 903  * @param desc
 904  *   Number of descriptors to configure in queue.
 905  * @param socket
 906  *   NUMA socket on which memory must be allocated.
 907  *
 908  * @return
 909  *   A DPDK queue object on success, NULL otherwise and rte_errno is set.
 910  */
 911 struct mlx5_rxq_ctrl *
 912 mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 913              unsigned int socket, const struct rte_eth_rxconf *conf,
 914              struct rte_mempool *mp)
 915 {
 916         struct priv *priv = dev->data->dev_private;
 917         struct mlx5_rxq_ctrl *tmpl;
 918         unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
 919         struct mlx5_dev_config *config = &priv->config;
 920         /*
 921          * Always allocate extra slots, even if eventually
 922          * the vector Rx will not be used.
 923          */
 924         const uint16_t desc_n =
 925                 desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
 926         uint64_t offloads = conf->offloads |
 927                            dev->data->dev_conf.rxmode.offloads;
 928
 929         tmpl = rte_calloc_socket("RXQ", 1,
 930                                  sizeof(*tmpl) +
 931                                  desc_n * sizeof(struct rte_mbuf *),
 932                                  0, socket);
 933         if (!tmpl) {
 934                 rte_errno = ENOMEM;
 935                 return NULL;
 936         }
 937         tmpl->socket = socket;
 938         if (dev->data->dev_conf.intr_conf.rxq)
 939                 tmpl->irq = 1;
 940         /* Enable scattered packets support for this queue if necessary. */
 941         assert(mb_len >= RTE_PKTMBUF_HEADROOM);
 942         if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
 943             (mb_len - RTE_PKTMBUF_HEADROOM)) {
 944                 tmpl->rxq.sges_n = 0;
 945         } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
 946                 unsigned int size =
 947                         RTE_PKTMBUF_HEADROOM +
 948                         dev->data->dev_conf.rxmode.max_rx_pkt_len;
 949                 unsigned int sges_n;
 950
 951                 /*
 952                  * Determine the number of SGEs needed for a full packet
 953                  * and round it to the next power of two.
 954                  */
 955                 sges_n = log2above((size / mb_len) + !!(size % mb_len));
 956                 tmpl->rxq.sges_n = sges_n;
 957                 /* Make sure rxq.sges_n did not overflow. */
 958                 size = mb_len * (1 << tmpl->rxq.sges_n);
 959                 size -= RTE_PKTMBUF_HEADROOM;
 960                 if (size < dev->data->dev_conf.rxmode.max_rx_pkt_len) {
 961                         DRV_LOG(ERR,
 962                                 "port %u too many SGEs (%u) needed to handle"
 963                                 " requested maximum packet size %u",
 964                                 dev->data->port_id,
 965                                 1 << sges_n,
 966                                 dev->data->dev_conf.rxmode.max_rx_pkt_len);
 967                         rte_errno = EOVERFLOW;
 968                         goto error;
 969                 }
 970         } else {
 971                 DRV_LOG(WARNING,
 972                         "port %u the requested maximum Rx packet size (%u) is"
 973                         " larger than a single mbuf (%u) and scattered mode has"
 974                         " not been requested",
 975                         dev->data->port_id,
 976                         dev->data->dev_conf.rxmode.max_rx_pkt_len,
 977                         mb_len - RTE_PKTMBUF_HEADROOM);
 978         }
 979         DRV_LOG(DEBUG, "port %u maximum number of segments per packet: %u",
 980                 dev->data->port_id, 1 << tmpl->rxq.sges_n);
 981         if (desc % (1 << tmpl->rxq.sges_n)) {
 982                 DRV_LOG(ERR,
 983                         "port %u number of Rx queue descriptors (%u) is not a"
 984                         " multiple of SGEs per packet (%u)",
 985                         dev->data->port_id,
 986                         desc,
 987                         1 << tmpl->rxq.sges_n);
 988                 rte_errno = EINVAL;
 989                 goto error;
 990         }
 991         /* Toggle RX checksum offload if hardware supports it. */
 992         tmpl->rxq.csum = !!(offloads & DEV_RX_OFFLOAD_CHECKSUM);
 993         tmpl->rxq.hw_timestamp = !!(offloads & DEV_RX_OFFLOAD_TIMESTAMP);
 994         /* Configure VLAN stripping. */
 995         tmpl->rxq.vlan_strip = !!(offloads & DEV_RX_OFFLOAD_VLAN_STRIP);
 996         /* By default, FCS (CRC) is stripped by hardware. */
 997         if (offloads & DEV_RX_OFFLOAD_CRC_STRIP) {
 998                 tmpl->rxq.crc_present = 0;
 999         } else if (config->hw_fcs_strip) {
1000                 tmpl->rxq.crc_present = 1;
1001         } else {
1002                 DRV_LOG(WARNING,
1003                         "port %u CRC stripping has been disabled but will"
1004                         " still be performed by hardware, make sure MLNX_OFED"
1005                         " and firmware are up to date",
1006                         dev->data->port_id);
1007                 tmpl->rxq.crc_present = 0;
1008         }
1009         DRV_LOG(DEBUG,
1010                 "port %u CRC stripping is %s, %u bytes will be subtracted from"
1011                 " incoming frames to hide it",
1012                 dev->data->port_id,
1013                 tmpl->rxq.crc_present ? "disabled" : "enabled",
1014                 tmpl->rxq.crc_present << 2);
1015         /* Save port ID. */
1016         tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
1017                 (!!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS));
1018         tmpl->rxq.port_id = dev->data->port_id;
1019         tmpl->priv = priv;
1020         tmpl->rxq.mp = mp;
1021         tmpl->rxq.stats.idx = idx;
1022         tmpl->rxq.elts_n = log2above(desc);
1023         tmpl->rxq.elts =
1024                 (struct rte_mbuf *(*)[1 << tmpl->rxq.elts_n])(tmpl + 1);
1025         tmpl->idx = idx;
1026         rte_atomic32_inc(&tmpl->refcnt);
1027         DRV_LOG(DEBUG, "port %u Rx queue %u: refcnt %d", dev->data->port_id,
1028                 idx, rte_atomic32_read(&tmpl->refcnt));
1029         LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1030         return tmpl;
1031 error:
1032         rte_free(tmpl);
1033         return NULL;
1034 }
1035
1036 /**
1037  * Get a Rx queue.
1038  *
1039  * @param dev
1040  *   Pointer to Ethernet device.
1041  * @param idx
1042  *   TX queue index.
1043  *
1044  * @return
1045  *   A pointer to the queue if it exists, NULL otherwise.
1046  */
1047 struct mlx5_rxq_ctrl *
1048 mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
1049 {
1050         struct priv *priv = dev->data->dev_private;
1051         struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1052
1053         if ((*priv->rxqs)[idx]) {
1054                 rxq_ctrl = container_of((*priv->rxqs)[idx],
1055                                         struct mlx5_rxq_ctrl,
1056                                         rxq);
1057                 mlx5_rxq_ibv_get(dev, idx);
1058                 rte_atomic32_inc(&rxq_ctrl->refcnt);
1059                 DRV_LOG(DEBUG, "port %u Rx queue %u: refcnt %d",
1060                         dev->data->port_id, rxq_ctrl->idx,
1061                         rte_atomic32_read(&rxq_ctrl->refcnt));
1062         }
1063         return rxq_ctrl;
1064 }
1065
1066 /**
1067  * Release a Rx queue.
1068  *
1069  * @param dev
1070  *   Pointer to Ethernet device.
1071  * @param idx
1072  *   TX queue index.
1073  *
1074  * @return
1075  *   1 while a reference on it exists, 0 when freed.
1076  */
1077 int
1078 mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
1079 {
1080         struct priv *priv = dev->data->dev_private;
1081         struct mlx5_rxq_ctrl *rxq_ctrl;
1082
1083         if (!(*priv->rxqs)[idx])
1084                 return 0;
1085         rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1086         assert(rxq_ctrl->priv);
1087         if (rxq_ctrl->ibv && !mlx5_rxq_ibv_release(rxq_ctrl->ibv))
1088                 rxq_ctrl->ibv = NULL;
1089         DRV_LOG(DEBUG, "port %u Rx queue %u: refcnt %d", dev->data->port_id,
1090                 rxq_ctrl->idx, rte_atomic32_read(&rxq_ctrl->refcnt));
1091         if (rte_atomic32_dec_and_test(&rxq_ctrl->refcnt)) {
1092                 LIST_REMOVE(rxq_ctrl, next);
1093                 rte_free(rxq_ctrl);
1094                 (*priv->rxqs)[idx] = NULL;
1095                 return 0;
1096         }
1097         return 1;
1098 }
1099
1100 /**
1101  * Verify if the queue can be released.
1102  *
1103  * @param dev
1104  *   Pointer to Ethernet device.
1105  * @param idx
1106  *   TX queue index.
1107  *
1108  * @return
1109  *   1 if the queue can be released, negative errno otherwise and rte_errno is
1110  *   set.
1111  */
1112 int
1113 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
1114 {
1115         struct priv *priv = dev->data->dev_private;
1116         struct mlx5_rxq_ctrl *rxq_ctrl;
1117
1118         if (!(*priv->rxqs)[idx]) {
1119                 rte_errno = EINVAL;
1120                 return -rte_errno;
1121         }
1122         rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1123         return (rte_atomic32_read(&rxq_ctrl->refcnt) == 1);
1124 }
1125
1126 /**
1127  * Verify the Rx Queue list is empty
1128  *
1129  * @param dev
1130  *   Pointer to Ethernet device.
1131  *
1132  * @return
1133  *   The number of object not released.
1134  */
1135 int
1136 mlx5_rxq_verify(struct rte_eth_dev *dev)
1137 {
1138         struct priv *priv = dev->data->dev_private;
1139         struct mlx5_rxq_ctrl *rxq_ctrl;
1140         int ret = 0;
1141
1142         LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
1143                 DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
1144                         dev->data->port_id, rxq_ctrl->idx);
1145                 ++ret;
1146         }
1147         return ret;
1148 }
1149
1150 /**
1151  * Create an indirection table.
1152  *
1153  * @param dev
1154  *   Pointer to Ethernet device.
1155  * @param queues
1156  *   Queues entering in the indirection table.
1157  * @param queues_n
1158  *   Number of queues in the array.
1159  *
1160  * @return
1161  *   The Verbs object initialised, NULL otherwise and rte_errno is set.
1162  */
1163 struct mlx5_ind_table_ibv *
1164 mlx5_ind_table_ibv_new(struct rte_eth_dev *dev, const uint16_t *queues,
1165                        uint32_t queues_n)
1166 {
1167         struct priv *priv = dev->data->dev_private;
1168         struct mlx5_ind_table_ibv *ind_tbl;
1169         const unsigned int wq_n = rte_is_power_of_2(queues_n) ?
1170                 log2above(queues_n) :
1171                 log2above(priv->config.ind_table_max_size);
1172         struct ibv_wq *wq[1 << wq_n];
1173         unsigned int i;
1174         unsigned int j;
1175
1176         ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl) +
1177                              queues_n * sizeof(uint16_t), 0);
1178         if (!ind_tbl) {
1179                 rte_errno = ENOMEM;
1180                 return NULL;
1181         }
1182         for (i = 0; i != queues_n; ++i) {
1183                 struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev, queues[i]);
1184
1185                 if (!rxq)
1186                         goto error;
1187                 wq[i] = rxq->ibv->wq;
1188                 ind_tbl->queues[i] = queues[i];
1189         }
1190         ind_tbl->queues_n = queues_n;
1191         /* Finalise indirection table. */
1192         for (j = 0; i != (unsigned int)(1 << wq_n); ++i, ++j)
1193                 wq[i] = wq[j];
1194         ind_tbl->ind_table = mlx5_glue->create_rwq_ind_table
1195                 (priv->ctx,
1196                  &(struct ibv_rwq_ind_table_init_attr){
1197                         .log_ind_tbl_size = wq_n,
1198                         .ind_tbl = wq,
1199                         .comp_mask = 0,
1200                  });
1201         if (!ind_tbl->ind_table) {
1202                 rte_errno = errno;
1203                 goto error;
1204         }
1205         rte_atomic32_inc(&ind_tbl->refcnt);
1206         LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
1207         DEBUG("port %u new indirection table %p: queues:%u refcnt:%d",
1208               dev->data->port_id, (void *)ind_tbl, 1 << wq_n,
1209               rte_atomic32_read(&ind_tbl->refcnt));
1210         return ind_tbl;
1211 error:
1212         rte_free(ind_tbl);
1213         DRV_LOG(DEBUG, "port %u cannot create indirection table",
1214                 dev->data->port_id);
1215         return NULL;
1216 }
1217
1218 /**
1219  * Get an indirection table.
1220  *
1221  * @param dev
1222  *   Pointer to Ethernet device.
1223  * @param queues
1224  *   Queues entering in the indirection table.
1225  * @param queues_n
1226  *   Number of queues in the array.
1227  *
1228  * @return
1229  *   An indirection table if found.
1230  */
1231 struct mlx5_ind_table_ibv *
1232 mlx5_ind_table_ibv_get(struct rte_eth_dev *dev, const uint16_t *queues,
1233                        uint32_t queues_n)
1234 {
1235         struct priv *priv = dev->data->dev_private;
1236         struct mlx5_ind_table_ibv *ind_tbl;
1237
1238         LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1239                 if ((ind_tbl->queues_n == queues_n) &&
1240                     (memcmp(ind_tbl->queues, queues,
1241                             ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
1242                      == 0))
1243                         break;
1244         }
1245         if (ind_tbl) {
1246                 unsigned int i;
1247
1248                 rte_atomic32_inc(&ind_tbl->refcnt);
1249                 DRV_LOG(DEBUG, "port %u indirection table %p: refcnt %d",
1250                         dev->data->port_id, (void *)ind_tbl,
1251                         rte_atomic32_read(&ind_tbl->refcnt));
1252                 for (i = 0; i != ind_tbl->queues_n; ++i)
1253                         mlx5_rxq_get(dev, ind_tbl->queues[i]);
1254         }
1255         return ind_tbl;
1256 }
1257
1258 /**
1259  * Release an indirection table.
1260  *
1261  * @param dev
1262  *   Pointer to Ethernet device.
1263  * @param ind_table
1264  *   Indirection table to release.
1265  *
1266  * @return
1267  *   1 while a reference on it exists, 0 when freed.
1268  */
1269 int
1270 mlx5_ind_table_ibv_release(struct rte_eth_dev *dev,
1271                            struct mlx5_ind_table_ibv *ind_tbl)
1272 {
1273         unsigned int i;
1274
1275         DRV_LOG(DEBUG, "port %u indirection table %p: refcnt %d",
1276                 dev->data->port_id, (void *)ind_tbl,
1277                 rte_atomic32_read(&ind_tbl->refcnt));
1278         if (rte_atomic32_dec_and_test(&ind_tbl->refcnt)) {
1279                 claim_zero(mlx5_glue->destroy_rwq_ind_table
1280                            (ind_tbl->ind_table));
1281                 DEBUG("port %u delete indirection table %p: queues: %u",
1282                       dev->data->port_id, (void *)ind_tbl, ind_tbl->queues_n);
1283         }
1284         for (i = 0; i != ind_tbl->queues_n; ++i)
1285                 claim_nonzero(mlx5_rxq_release(dev, ind_tbl->queues[i]));
1286         if (!rte_atomic32_read(&ind_tbl->refcnt)) {
1287                 LIST_REMOVE(ind_tbl, next);
1288                 rte_free(ind_tbl);
1289                 return 0;
1290         }
1291         return 1;
1292 }
1293
1294 /**
1295  * Verify the Rx Queue list is empty
1296  *
1297  * @param dev
1298  *   Pointer to Ethernet device.
1299  *
1300  * @return
1301  *   The number of object not released.
1302  */
1303 int
1304 mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev)
1305 {
1306         struct priv *priv = dev->data->dev_private;
1307         struct mlx5_ind_table_ibv *ind_tbl;
1308         int ret = 0;
1309
1310         LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
1311                 DRV_LOG(DEBUG,
1312                         "port %u Verbs indirection table %p still referenced",
1313                         dev->data->port_id, (void *)ind_tbl);
1314                 ++ret;
1315         }
1316         return ret;
1317 }
1318
1319 /**
1320  * Create an Rx Hash queue.
1321  *
1322  * @param dev
1323  *   Pointer to Ethernet device.
1324  * @param rss_key
1325  *   RSS key for the Rx hash queue.
1326  * @param rss_key_len
1327  *   RSS key length.
1328  * @param hash_fields
1329  *   Verbs protocol hash field to make the RSS on.
1330  * @param queues
1331  *   Queues entering in hash queue. In case of empty hash_fields only the
1332  *   first queue index will be taken for the indirection table.
1333  * @param queues_n
1334  *   Number of queues.
1335  * @param tunnel
1336  *   Tunnel type, implies tunnel offloading like inner checksum if available.
1337  * @param rss_level
1338  *   RSS hash on tunnel level.
1339  *
1340  * @return
1341  *   The Verbs object initialised, NULL otherwise and rte_errno is set.
1342  */
1343 struct mlx5_hrxq *
1344 mlx5_hrxq_new(struct rte_eth_dev *dev,
1345               const uint8_t *rss_key, uint32_t rss_key_len,
1346               uint64_t hash_fields,
1347               const uint16_t *queues, uint32_t queues_n,
1348               uint32_t tunnel, uint32_t rss_level)
1349 {
1350         struct priv *priv = dev->data->dev_private;
1351         struct mlx5_hrxq *hrxq;
1352         struct mlx5_ind_table_ibv *ind_tbl;
1353         struct ibv_qp *qp;
1354         int err;
1355 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1356         struct mlx5dv_qp_init_attr qp_init_attr = {0};
1357 #endif
1358
1359         queues_n = hash_fields ? queues_n : 1;
1360         ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1361         if (!ind_tbl)
1362                 ind_tbl = mlx5_ind_table_ibv_new(dev, queues, queues_n);
1363         if (!ind_tbl) {
1364                 rte_errno = ENOMEM;
1365                 return NULL;
1366         }
1367         if (!rss_key_len) {
1368                 rss_key_len = rss_hash_default_key_len;
1369                 rss_key = rss_hash_default_key;
1370         }
1371 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
1372         if (tunnel) {
1373                 qp_init_attr.comp_mask =
1374                                 MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
1375                 qp_init_attr.create_flags = MLX5DV_QP_CREATE_TUNNEL_OFFLOADS;
1376         }
1377         qp = mlx5_glue->dv_create_qp
1378                 (priv->ctx,
1379                  &(struct ibv_qp_init_attr_ex){
1380                         .qp_type = IBV_QPT_RAW_PACKET,
1381                         .comp_mask =
1382                                 IBV_QP_INIT_ATTR_PD |
1383                                 IBV_QP_INIT_ATTR_IND_TABLE |
1384                                 IBV_QP_INIT_ATTR_RX_HASH,
1385                         .rx_hash_conf = (struct ibv_rx_hash_conf){
1386                                 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1387                                 .rx_hash_key_len = rss_key_len ? rss_key_len :
1388                                                    rss_hash_default_key_len,
1389                                 .rx_hash_key = rss_key ?
1390                                                (void *)(uintptr_t)rss_key :
1391                                                rss_hash_default_key,
1392                                 .rx_hash_fields_mask = hash_fields |
1393                                         (tunnel && rss_level > 1 ?
1394                                         (uint32_t)IBV_RX_HASH_INNER : 0),
1395                         },
1396                         .rwq_ind_tbl = ind_tbl->ind_table,
1397                         .pd = priv->pd,
1398                  },
1399                  &qp_init_attr);
1400         DEBUG("port %u new QP:%p ind_tbl:%p hash_fields:0x%" PRIx64
1401               " tunnel:0x%x level:%u dv_attr:comp_mask:0x%" PRIx64
1402               " create_flags:0x%x",
1403               dev->data->port_id, (void *)qp, (void *)ind_tbl,
1404               (tunnel && rss_level == 2 ? (uint32_t)IBV_RX_HASH_INNER : 0) |
1405               hash_fields, tunnel, rss_level,
1406               qp_init_attr.comp_mask, qp_init_attr.create_flags);
1407 #else
1408         qp = mlx5_glue->create_qp_ex
1409                 (priv->ctx,
1410                  &(struct ibv_qp_init_attr_ex){
1411                         .qp_type = IBV_QPT_RAW_PACKET,
1412                         .comp_mask =
1413                                 IBV_QP_INIT_ATTR_PD |
1414                                 IBV_QP_INIT_ATTR_IND_TABLE |
1415                                 IBV_QP_INIT_ATTR_RX_HASH,
1416                         .rx_hash_conf = (struct ibv_rx_hash_conf){
1417                                 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
1418                                 .rx_hash_key_len = rss_key_len ? rss_key_len :
1419                                                    rss_hash_default_key_len,
1420                                 .rx_hash_key = rss_key ?
1421                                                (void *)(uintptr_t)rss_key :
1422                                                rss_hash_default_key,
1423                                 .rx_hash_fields_mask = hash_fields,
1424                         },
1425                         .rwq_ind_tbl = ind_tbl->ind_table,
1426                         .pd = priv->pd,
1427                  });
1428         DEBUG("port %u new QP:%p ind_tbl:%p hash_fields:0x%" PRIx64
1429               " tunnel:0x%x level:%hhu",
1430               dev->data->port_id, (void *)qp, (void *)ind_tbl,
1431               hash_fields, tunnel, rss_level);
1432 #endif
1433         if (!qp) {
1434                 rte_errno = errno;
1435                 goto error;
1436         }
1437         hrxq = rte_calloc(__func__, 1, sizeof(*hrxq) + rss_key_len, 0);
1438         if (!hrxq)
1439                 goto error;
1440         hrxq->ind_table = ind_tbl;
1441         hrxq->qp = qp;
1442         hrxq->rss_key_len = rss_key_len;
1443         hrxq->hash_fields = hash_fields;
1444         hrxq->tunnel = tunnel;
1445         hrxq->rss_level = rss_level;
1446         memcpy(hrxq->rss_key, rss_key, rss_key_len);
1447         rte_atomic32_inc(&hrxq->refcnt);
1448         LIST_INSERT_HEAD(&priv->hrxqs, hrxq, next);
1449         DRV_LOG(DEBUG, "port %u hash Rx queue %p: refcnt %d",
1450                 dev->data->port_id, (void *)hrxq,
1451                 rte_atomic32_read(&hrxq->refcnt));
1452         return hrxq;
1453 error:
1454         err = rte_errno; /* Save rte_errno before cleanup. */
1455         mlx5_ind_table_ibv_release(dev, ind_tbl);
1456         if (qp)
1457                 claim_zero(mlx5_glue->destroy_qp(qp));
1458         rte_errno = err; /* Restore rte_errno. */
1459         return NULL;
1460 }
1461
1462 /**
1463  * Get an Rx Hash queue.
1464  *
1465  * @param dev
1466  *   Pointer to Ethernet device.
1467  * @param rss_conf
1468  *   RSS configuration for the Rx hash queue.
1469  * @param queues
1470  *   Queues entering in hash queue. In case of empty hash_fields only the
1471  *   first queue index will be taken for the indirection table.
1472  * @param queues_n
1473  *   Number of queues.
1474  * @param tunnel
1475  *   Tunnel type, implies tunnel offloading like inner checksum if available.
1476  * @param rss_level
1477  *   RSS hash on tunnel level
1478  *
1479  * @return
1480  *   An hash Rx queue on success.
1481  */
1482 struct mlx5_hrxq *
1483 mlx5_hrxq_get(struct rte_eth_dev *dev,
1484               const uint8_t *rss_key, uint32_t rss_key_len,
1485               uint64_t hash_fields,
1486               const uint16_t *queues, uint32_t queues_n,
1487               uint32_t tunnel, uint32_t rss_level)
1488 {
1489         struct priv *priv = dev->data->dev_private;
1490         struct mlx5_hrxq *hrxq;
1491
1492         queues_n = hash_fields ? queues_n : 1;
1493         LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1494                 struct mlx5_ind_table_ibv *ind_tbl;
1495
1496                 if (hrxq->rss_key_len != rss_key_len)
1497                         continue;
1498                 if (memcmp(hrxq->rss_key, rss_key, rss_key_len))
1499                         continue;
1500                 if (hrxq->hash_fields != hash_fields)
1501                         continue;
1502                 if (hrxq->tunnel != tunnel)
1503                         continue;
1504                 if (hrxq->rss_level != rss_level)
1505                         continue;
1506                 ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
1507                 if (!ind_tbl)
1508                         continue;
1509                 if (ind_tbl != hrxq->ind_table) {
1510                         mlx5_ind_table_ibv_release(dev, ind_tbl);
1511                         continue;
1512                 }
1513                 rte_atomic32_inc(&hrxq->refcnt);
1514                 DRV_LOG(DEBUG, "port %u hash Rx queue %p: refcnt %d",
1515                         dev->data->port_id, (void *)hrxq,
1516                         rte_atomic32_read(&hrxq->refcnt));
1517                 return hrxq;
1518         }
1519         return NULL;
1520 }
1521
1522 /**
1523  * Release the hash Rx queue.
1524  *
1525  * @param dev
1526  *   Pointer to Ethernet device.
1527  * @param hrxq
1528  *   Pointer to Hash Rx queue to release.
1529  *
1530  * @return
1531  *   1 while a reference on it exists, 0 when freed.
1532  */
1533 int
1534 mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
1535 {
1536         DRV_LOG(DEBUG, "port %u hash Rx queue %p: refcnt %d",
1537                 dev->data->port_id, (void *)hrxq,
1538                 rte_atomic32_read(&hrxq->refcnt));
1539         if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
1540                 claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
1541                 DEBUG("port %u delete QP %p: hash: 0x%" PRIx64 ", tunnel:"
1542                       " 0x%x, level: %u",
1543                       dev->data->port_id, (void *)hrxq, hrxq->hash_fields,
1544                       hrxq->tunnel, hrxq->rss_level);
1545                 mlx5_ind_table_ibv_release(dev, hrxq->ind_table);
1546                 LIST_REMOVE(hrxq, next);
1547                 rte_free(hrxq);
1548                 return 0;
1549         }
1550         claim_nonzero(mlx5_ind_table_ibv_release(dev, hrxq->ind_table));
1551         return 1;
1552 }
1553
1554 /**
1555  * Verify the Rx Queue list is empty
1556  *
1557  * @param dev
1558  *   Pointer to Ethernet device.
1559  *
1560  * @return
1561  *   The number of object not released.
1562  */
1563 int
1564 mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev)
1565 {
1566         struct priv *priv = dev->data->dev_private;
1567         struct mlx5_hrxq *hrxq;
1568         int ret = 0;
1569
1570         LIST_FOREACH(hrxq, &priv->hrxqs, next) {
1571                 DRV_LOG(DEBUG,
1572                         "port %u Verbs hash Rx queue %p still referenced",
1573                         dev->data->port_id, (void *)hrxq);
1574                 ++ret;
1575         }
1576         return ret;
1577 }