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 static_assert(MLX5_RSS_HASH_KEY_LEN ==
  56               (unsigned int)sizeof(rss_hash_default_key),
  57               "wrong RSS default key size.");
  58
  59 /**
  60  * Check whether Multi-Packet RQ can be enabled for the device.
  61  *
  62  * @param dev
  63  *   Pointer to Ethernet device.
  64  *
  65  * @return
  66  *   1 if supported, negative errno value if not.
  67  */
  68 inline int
  69 mlx5_check_mprq_support(struct rte_eth_dev *dev)
  70 {
  71         struct mlx5_priv *priv = dev->data->dev_private;
  72
  73         if (priv->config.mprq.enabled &&
  74             priv->rxqs_n >= priv->config.mprq.min_rxqs_num)
  75                 return 1;
  76         return -ENOTSUP;
  77 }
  78
  79 /**
  80  * Check whether Multi-Packet RQ is enabled for the Rx queue.
  81  *
  82  *  @param rxq
  83  *     Pointer to receive queue structure.
  84  *
  85  * @return
  86  *   0 if disabled, otherwise enabled.
  87  */
  88 inline int
  89 mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq)
  90 {
  91         return rxq->strd_num_n > 0;
  92 }
  93
  94 /**
  95  * Check whether Multi-Packet RQ is enabled for the device.
  96  * MPRQ can be enabled explicitly, or implicitly by enabling LRO.
  97  *
  98  * @param dev
  99  *   Pointer to Ethernet device.
 100  *
 101  * @return
 102  *   0 if disabled, otherwise enabled.
 103  */
 104 inline int
 105 mlx5_mprq_enabled(struct rte_eth_dev *dev)
 106 {
 107         struct mlx5_priv *priv = dev->data->dev_private;
 108         uint16_t i;
 109         uint16_t n = 0;
 110
 111         if (mlx5_check_mprq_support(dev) < 0)
 112                 return 0;
 113         /* All the configured queues should be enabled. */
 114         for (i = 0; i < priv->rxqs_n; ++i) {
 115                 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
 116
 117                 if (!rxq)
 118                         continue;
 119                 if (mlx5_rxq_mprq_enabled(rxq))
 120                         ++n;
 121         }
 122         /* Multi-Packet RQ can't be partially configured. */
 123         assert(n == 0 || n == priv->rxqs_n);
 124         return n == priv->rxqs_n;
 125 }
 126
 127 /**
 128  * Check whether LRO is supported and enabled for the device.
 129  *
 130  * @param dev
 131  *   Pointer to Ethernet device.
 132  *
 133  * @return
 134  *   0 if disabled, 1 if enabled.
 135  */
 136 inline int
 137 mlx5_lro_on(struct rte_eth_dev *dev)
 138 {
 139         return (MLX5_LRO_SUPPORTED(dev) && MLX5_LRO_ENABLED(dev));
 140 }
 141
 142 /**
 143  * Allocate RX queue elements for Multi-Packet RQ.
 144  *
 145  * @param rxq_ctrl
 146  *   Pointer to RX queue structure.
 147  *
 148  * @return
 149  *   0 on success, a negative errno value otherwise and rte_errno is set.
 150  */
 151 static int
 152 rxq_alloc_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
 153 {
 154         struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
 155         unsigned int wqe_n = 1 << rxq->elts_n;
 156         unsigned int i;
 157         int err;
 158
 159         /* Iterate on segments. */
 160         for (i = 0; i <= wqe_n; ++i) {
 161                 struct mlx5_mprq_buf *buf;
 162
 163                 if (rte_mempool_get(rxq->mprq_mp, (void **)&buf) < 0) {
 164                         DRV_LOG(ERR, "port %u empty mbuf pool", rxq->port_id);
 165                         rte_errno = ENOMEM;
 166                         goto error;
 167                 }
 168                 if (i < wqe_n)
 169                         (*rxq->mprq_bufs)[i] = buf;
 170                 else
 171                         rxq->mprq_repl = buf;
 172         }
 173         DRV_LOG(DEBUG,
 174                 "port %u Rx queue %u allocated and configured %u segments",
 175                 rxq->port_id, rxq->idx, wqe_n);
 176         return 0;
 177 error:
 178         err = rte_errno; /* Save rte_errno before cleanup. */
 179         wqe_n = i;
 180         for (i = 0; (i != wqe_n); ++i) {
 181                 if ((*rxq->mprq_bufs)[i] != NULL)
 182                         rte_mempool_put(rxq->mprq_mp,
 183                                         (*rxq->mprq_bufs)[i]);
 184                 (*rxq->mprq_bufs)[i] = NULL;
 185         }
 186         DRV_LOG(DEBUG, "port %u Rx queue %u failed, freed everything",
 187                 rxq->port_id, rxq->idx);
 188         rte_errno = err; /* Restore rte_errno. */
 189         return -rte_errno;
 190 }
 191
 192 /**
 193  * Allocate RX queue elements for Single-Packet RQ.
 194  *
 195  * @param rxq_ctrl
 196  *   Pointer to RX queue structure.
 197  *
 198  * @return
 199  *   0 on success, errno value on failure.
 200  */
 201 static int
 202 rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
 203 {
 204         const unsigned int sges_n = 1 << rxq_ctrl->rxq.sges_n;
 205         unsigned int elts_n = 1 << rxq_ctrl->rxq.elts_n;
 206         unsigned int i;
 207         int err;
 208
 209         /* Iterate on segments. */
 210         for (i = 0; (i != elts_n); ++i) {
 211                 struct rte_mbuf *buf;
 212
 213                 buf = rte_pktmbuf_alloc(rxq_ctrl->rxq.mp);
 214                 if (buf == NULL) {
 215                         DRV_LOG(ERR, "port %u empty mbuf pool",
 216                                 PORT_ID(rxq_ctrl->priv));
 217                         rte_errno = ENOMEM;
 218                         goto error;
 219                 }
 220                 /* Headroom is reserved by rte_pktmbuf_alloc(). */
 221                 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
 222                 /* Buffer is supposed to be empty. */
 223                 assert(rte_pktmbuf_data_len(buf) == 0);
 224                 assert(rte_pktmbuf_pkt_len(buf) == 0);
 225                 assert(!buf->next);
 226                 /* Only the first segment keeps headroom. */
 227                 if (i % sges_n)
 228                         SET_DATA_OFF(buf, 0);
 229                 PORT(buf) = rxq_ctrl->rxq.port_id;
 230                 DATA_LEN(buf) = rte_pktmbuf_tailroom(buf);
 231                 PKT_LEN(buf) = DATA_LEN(buf);
 232                 NB_SEGS(buf) = 1;
 233                 (*rxq_ctrl->rxq.elts)[i] = buf;
 234         }
 235         /* If Rx vector is activated. */
 236         if (mlx5_rxq_check_vec_support(&rxq_ctrl->rxq) > 0) {
 237                 struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
 238                 struct rte_mbuf *mbuf_init = &rxq->fake_mbuf;
 239                 int j;
 240
 241                 /* Initialize default rearm_data for vPMD. */
 242                 mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
 243                 rte_mbuf_refcnt_set(mbuf_init, 1);
 244                 mbuf_init->nb_segs = 1;
 245                 mbuf_init->port = rxq->port_id;
 246                 /*
 247                  * prevent compiler reordering:
 248                  * rearm_data covers previous fields.
 249                  */
 250                 rte_compiler_barrier();
 251                 rxq->mbuf_initializer =
 252                         *(uint64_t *)&mbuf_init->rearm_data;
 253                 /* Padding with a fake mbuf for vectorized Rx. */
 254                 for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
 255                         (*rxq->elts)[elts_n + j] = &rxq->fake_mbuf;
 256         }
 257         DRV_LOG(DEBUG,
 258                 "port %u Rx queue %u allocated and configured %u segments"
 259                 " (max %u packets)",
 260                 PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx, elts_n,
 261                 elts_n / (1 << rxq_ctrl->rxq.sges_n));
 262         return 0;
 263 error:
 264         err = rte_errno; /* Save rte_errno before cleanup. */
 265         elts_n = i;
 266         for (i = 0; (i != elts_n); ++i) {
 267                 if ((*rxq_ctrl->rxq.elts)[i] != NULL)
 268                         rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
 269                 (*rxq_ctrl->rxq.elts)[i] = NULL;
 270         }
 271         DRV_LOG(DEBUG, "port %u Rx queue %u failed, freed everything",
 272                 PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx);
 273         rte_errno = err; /* Restore rte_errno. */
 274         return -rte_errno;
 275 }
 276
 277 /**
 278  * Allocate RX queue elements.
 279  *
 280  * @param rxq_ctrl
 281  *   Pointer to RX queue structure.
 282  *
 283  * @return
 284  *   0 on success, errno value on failure.
 285  */
 286 int
 287 rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
 288 {
 289         return mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
 290                rxq_alloc_elts_mprq(rxq_ctrl) : rxq_alloc_elts_sprq(rxq_ctrl);
 291 }
 292
 293 /**
 294  * Free RX queue elements for Multi-Packet RQ.
 295  *
 296  * @param rxq_ctrl
 297  *   Pointer to RX queue structure.
 298  */
 299 static void
 300 rxq_free_elts_mprq(struct mlx5_rxq_ctrl *rxq_ctrl)
 301 {
 302         struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
 303         uint16_t i;
 304
 305         DRV_LOG(DEBUG, "port %u Multi-Packet Rx queue %u freeing WRs",
 306                 rxq->port_id, rxq->idx);
 307         if (rxq->mprq_bufs == NULL)
 308                 return;
 309         assert(mlx5_rxq_check_vec_support(rxq) < 0);
 310         for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
 311                 if ((*rxq->mprq_bufs)[i] != NULL)
 312                         mlx5_mprq_buf_free((*rxq->mprq_bufs)[i]);
 313                 (*rxq->mprq_bufs)[i] = NULL;
 314         }
 315         if (rxq->mprq_repl != NULL) {
 316                 mlx5_mprq_buf_free(rxq->mprq_repl);
 317                 rxq->mprq_repl = NULL;
 318         }
 319 }
 320
 321 /**
 322  * Free RX queue elements for Single-Packet RQ.
 323  *
 324  * @param rxq_ctrl
 325  *   Pointer to RX queue structure.
 326  */
 327 static void
 328 rxq_free_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
 329 {
 330         struct mlx5_rxq_data *rxq = &rxq_ctrl->rxq;
 331         const uint16_t q_n = (1 << rxq->elts_n);
 332         const uint16_t q_mask = q_n - 1;
 333         uint16_t used = q_n - (rxq->rq_ci - rxq->rq_pi);
 334         uint16_t i;
 335
 336         DRV_LOG(DEBUG, "port %u Rx queue %u freeing WRs",
 337                 PORT_ID(rxq_ctrl->priv), rxq->idx);
 338         if (rxq->elts == NULL)
 339                 return;
 340         /**
 341          * Some mbuf in the Ring belongs to the application.  They cannot be
 342          * freed.
 343          */
 344         if (mlx5_rxq_check_vec_support(rxq) > 0) {
 345                 for (i = 0; i < used; ++i)
 346                         (*rxq->elts)[(rxq->rq_ci + i) & q_mask] = NULL;
 347                 rxq->rq_pi = rxq->rq_ci;
 348         }
 349         for (i = 0; (i != (1u << rxq->elts_n)); ++i) {
 350                 if ((*rxq->elts)[i] != NULL)
 351                         rte_pktmbuf_free_seg((*rxq->elts)[i]);
 352                 (*rxq->elts)[i] = NULL;
 353         }
 354 }
 355
 356 /**
 357  * Free RX queue elements.
 358  *
 359  * @param rxq_ctrl
 360  *   Pointer to RX queue structure.
 361  */
 362 static void
 363 rxq_free_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
 364 {
 365         if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
 366                 rxq_free_elts_mprq(rxq_ctrl);
 367         else
 368                 rxq_free_elts_sprq(rxq_ctrl);
 369 }
 370
 371 /**
 372  * Returns the per-queue supported offloads.
 373  *
 374  * @param dev
 375  *   Pointer to Ethernet device.
 376  *
 377  * @return
 378  *   Supported Rx offloads.
 379  */
 380 uint64_t
 381 mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
 382 {
 383         struct mlx5_priv *priv = dev->data->dev_private;
 384         struct mlx5_dev_config *config = &priv->config;
 385         uint64_t offloads = (DEV_RX_OFFLOAD_SCATTER |
 386                              DEV_RX_OFFLOAD_TIMESTAMP |
 387                              DEV_RX_OFFLOAD_JUMBO_FRAME);
 388
 389         if (config->hw_fcs_strip)
 390                 offloads |= DEV_RX_OFFLOAD_KEEP_CRC;
 391
 392         if (config->hw_csum)
 393                 offloads |= (DEV_RX_OFFLOAD_IPV4_CKSUM |
 394                              DEV_RX_OFFLOAD_UDP_CKSUM |
 395                              DEV_RX_OFFLOAD_TCP_CKSUM);
 396         if (config->hw_vlan_strip)
 397                 offloads |= DEV_RX_OFFLOAD_VLAN_STRIP;
 398         return offloads;
 399 }
 400
 401
 402 /**
 403  * Returns the per-port supported offloads.
 404  *
 405  * @param dev
 406  *   Pointer to Ethernet device.
 407  *
 408  * @return
 409  *   Supported Rx offloads.
 410  */
 411 uint64_t
 412 mlx5_get_rx_port_offloads(struct rte_eth_dev *dev)
 413 {
 414         uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
 415
 416         if (MLX5_LRO_SUPPORTED(dev))
 417                 offloads |= DEV_RX_OFFLOAD_TCP_LRO;
 418         return offloads;
 419 }
 420
 421 /**
 422  * Verify if the queue can be released.
 423  *
 424  * @param dev
 425  *   Pointer to Ethernet device.
 426  * @param idx
 427  *   RX queue index.
 428  *
 429  * @return
 430  *   1 if the queue can be released
 431  *   0 if the queue can not be released, there are references to it.
 432  *   Negative errno and rte_errno is set if queue doesn't exist.
 433  */
 434 static int
 435 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
 436 {
 437         struct mlx5_priv *priv = dev->data->dev_private;
 438         struct mlx5_rxq_ctrl *rxq_ctrl;
 439
 440         if (!(*priv->rxqs)[idx]) {
 441                 rte_errno = EINVAL;
 442                 return -rte_errno;
 443         }
 444         rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
 445         return (rte_atomic32_read(&rxq_ctrl->refcnt) == 1);
 446 }
 447
 448 /**
 449  *
 450  * @param dev
 451  *   Pointer to Ethernet device structure.
 452  * @param idx
 453  *   RX queue index.
 454  * @param desc
 455  *   Number of descriptors to configure in queue.
 456  * @param socket
 457  *   NUMA socket on which memory must be allocated.
 458  * @param[in] conf
 459  *   Thresholds parameters.
 460  * @param mp
 461  *   Memory pool for buffer allocations.
 462  *
 463  * @return
 464  *   0 on success, a negative errno value otherwise and rte_errno is set.
 465  */
 466 int
 467 mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 468                     unsigned int socket, const struct rte_eth_rxconf *conf,
 469                     struct rte_mempool *mp)
 470 {
 471         struct mlx5_priv *priv = dev->data->dev_private;
 472         struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
 473         struct mlx5_rxq_ctrl *rxq_ctrl =
 474                 container_of(rxq, struct mlx5_rxq_ctrl, rxq);
 475
 476         if (!rte_is_power_of_2(desc)) {
 477                 desc = 1 << log2above(desc);
 478                 DRV_LOG(WARNING,
 479                         "port %u increased number of descriptors in Rx queue %u"
 480                         " to the next power of two (%d)",
 481                         dev->data->port_id, idx, desc);
 482         }
 483         DRV_LOG(DEBUG, "port %u configuring Rx queue %u for %u descriptors",
 484                 dev->data->port_id, idx, desc);
 485         if (idx >= priv->rxqs_n) {
 486                 DRV_LOG(ERR, "port %u Rx queue index out of range (%u >= %u)",
 487                         dev->data->port_id, idx, priv->rxqs_n);
 488                 rte_errno = EOVERFLOW;
 489                 return -rte_errno;
 490         }
 491         if (!mlx5_rxq_releasable(dev, idx)) {
 492                 DRV_LOG(ERR, "port %u unable to release queue index %u",
 493                         dev->data->port_id, idx);
 494                 rte_errno = EBUSY;
 495                 return -rte_errno;
 496         }
 497         mlx5_rxq_release(dev, idx);
 498         rxq_ctrl = mlx5_rxq_new(dev, idx, desc, socket, conf, mp);
 499         if (!rxq_ctrl) {
 500                 DRV_LOG(ERR, "port %u unable to allocate queue index %u",
 501                         dev->data->port_id, idx);
 502                 rte_errno = ENOMEM;
 503                 return -rte_errno;
 504         }
 505         DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
 506                 dev->data->port_id, idx);
 507         (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
 508         return 0;
 509 }
 510
 511 /**
 512  * DPDK callback to release a RX queue.
 513  *
 514  * @param dpdk_rxq
 515  *   Generic RX queue pointer.
 516  */
 517 void
 518 mlx5_rx_queue_release(void *dpdk_rxq)
 519 {
 520         struct mlx5_rxq_data *rxq = (struct mlx5_rxq_data *)dpdk_rxq;
 521         struct mlx5_rxq_ctrl *rxq_ctrl;
 522         struct mlx5_priv *priv;
 523
 524         if (rxq == NULL)
 525                 return;
 526         rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
 527         priv = rxq_ctrl->priv;
 528         if (!mlx5_rxq_releasable(ETH_DEV(priv), rxq_ctrl->rxq.idx))
 529                 rte_panic("port %u Rx queue %u is still used by a flow and"
 530                           " cannot be removed\n",
 531                           PORT_ID(priv), rxq->idx);
 532         mlx5_rxq_release(ETH_DEV(priv), rxq_ctrl->rxq.idx);
 533 }
 534
 535 /**
 536  * Get an Rx queue Verbs/DevX object.
 537  *
 538  * @param dev
 539  *   Pointer to Ethernet device.
 540  * @param idx
 541  *   Queue index in DPDK Rx queue array
 542  *
 543  * @return
 544  *   The Verbs/DevX object if it exists.
 545  */
 546 static struct mlx5_rxq_obj *
 547 mlx5_rxq_obj_get(struct rte_eth_dev *dev, uint16_t idx)
 548 {
 549         struct mlx5_priv *priv = dev->data->dev_private;
 550         struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
 551         struct mlx5_rxq_ctrl *rxq_ctrl;
 552
 553         if (idx >= priv->rxqs_n)
 554                 return NULL;
 555         if (!rxq_data)
 556                 return NULL;
 557         rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 558         if (rxq_ctrl->obj)
 559                 rte_atomic32_inc(&rxq_ctrl->obj->refcnt);
 560         return rxq_ctrl->obj;
 561 }
 562
 563 /**
 564  * Release the resources allocated for an RQ DevX object.
 565  *
 566  * @param rxq_ctrl
 567  *   DevX Rx queue object.
 568  */
 569 static void
 570 rxq_release_rq_resources(struct mlx5_rxq_ctrl *rxq_ctrl)
 571 {
 572         if (rxq_ctrl->rxq.wqes) {
 573                 rte_free((void *)(uintptr_t)rxq_ctrl->rxq.wqes);
 574                 rxq_ctrl->rxq.wqes = NULL;
 575         }
 576         if (rxq_ctrl->wq_umem) {
 577                 mlx5_glue->devx_umem_dereg(rxq_ctrl->wq_umem);
 578                 rxq_ctrl->wq_umem = NULL;
 579         }
 580 }
 581
 582 /**
 583  * Release an Rx verbs/DevX queue object.
 584  *
 585  * @param rxq_obj
 586  *   Verbs/DevX Rx queue object.
 587  *
 588  * @return
 589  *   1 while a reference on it exists, 0 when freed.
 590  */
 591 static int
 592 mlx5_rxq_obj_release(struct mlx5_rxq_obj *rxq_obj)
 593 {
 594         assert(rxq_obj);
 595         if (rxq_obj->type == MLX5_RXQ_OBJ_TYPE_IBV)
 596                 assert(rxq_obj->wq);
 597         assert(rxq_obj->cq);
 598         if (rte_atomic32_dec_and_test(&rxq_obj->refcnt)) {
 599                 rxq_free_elts(rxq_obj->rxq_ctrl);
 600                 if (rxq_obj->type == MLX5_RXQ_OBJ_TYPE_IBV) {
 601                         claim_zero(mlx5_glue->destroy_wq(rxq_obj->wq));
 602                 } else if (rxq_obj->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ) {
 603                         claim_zero(mlx5_devx_cmd_destroy(rxq_obj->rq));
 604                         rxq_release_rq_resources(rxq_obj->rxq_ctrl);
 605                 }
 606                 claim_zero(mlx5_glue->destroy_cq(rxq_obj->cq));
 607                 if (rxq_obj->channel)
 608                         claim_zero(mlx5_glue->destroy_comp_channel
 609                                    (rxq_obj->channel));
 610                 LIST_REMOVE(rxq_obj, next);
 611                 rte_free(rxq_obj);
 612                 return 0;
 613         }
 614         return 1;
 615 }
 616
 617 /**
 618  * Allocate queue vector and fill epoll fd list for Rx interrupts.
 619  *
 620  * @param dev
 621  *   Pointer to Ethernet device.
 622  *
 623  * @return
 624  *   0 on success, a negative errno value otherwise and rte_errno is set.
 625  */
 626 int
 627 mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
 628 {
 629         struct mlx5_priv *priv = dev->data->dev_private;
 630         unsigned int i;
 631         unsigned int rxqs_n = priv->rxqs_n;
 632         unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
 633         unsigned int count = 0;
 634         struct rte_intr_handle *intr_handle = dev->intr_handle;
 635
 636         if (!dev->data->dev_conf.intr_conf.rxq)
 637                 return 0;
 638         mlx5_rx_intr_vec_disable(dev);
 639         intr_handle->intr_vec = malloc(n * sizeof(intr_handle->intr_vec[0]));
 640         if (intr_handle->intr_vec == NULL) {
 641                 DRV_LOG(ERR,
 642                         "port %u failed to allocate memory for interrupt"
 643                         " vector, Rx interrupts will not be supported",
 644                         dev->data->port_id);
 645                 rte_errno = ENOMEM;
 646                 return -rte_errno;
 647         }
 648         intr_handle->type = RTE_INTR_HANDLE_EXT;
 649         for (i = 0; i != n; ++i) {
 650                 /* This rxq obj must not be released in this function. */
 651                 struct mlx5_rxq_obj *rxq_obj = mlx5_rxq_obj_get(dev, i);
 652                 int fd;
 653                 int flags;
 654                 int rc;
 655
 656                 /* Skip queues that cannot request interrupts. */
 657                 if (!rxq_obj || !rxq_obj->channel) {
 658                         /* Use invalid intr_vec[] index to disable entry. */
 659                         intr_handle->intr_vec[i] =
 660                                 RTE_INTR_VEC_RXTX_OFFSET +
 661                                 RTE_MAX_RXTX_INTR_VEC_ID;
 662                         continue;
 663                 }
 664                 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
 665                         DRV_LOG(ERR,
 666                                 "port %u too many Rx queues for interrupt"
 667                                 " vector size (%d), Rx interrupts cannot be"
 668                                 " enabled",
 669                                 dev->data->port_id, RTE_MAX_RXTX_INTR_VEC_ID);
 670                         mlx5_rx_intr_vec_disable(dev);
 671                         rte_errno = ENOMEM;
 672                         return -rte_errno;
 673                 }
 674                 fd = rxq_obj->channel->fd;
 675                 flags = fcntl(fd, F_GETFL);
 676                 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
 677                 if (rc < 0) {
 678                         rte_errno = errno;
 679                         DRV_LOG(ERR,
 680                                 "port %u failed to make Rx interrupt file"
 681                                 " descriptor %d non-blocking for queue index"
 682                                 " %d",
 683                                 dev->data->port_id, fd, i);
 684                         mlx5_rx_intr_vec_disable(dev);
 685                         return -rte_errno;
 686                 }
 687                 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
 688                 intr_handle->efds[count] = fd;
 689                 count++;
 690         }
 691         if (!count)
 692                 mlx5_rx_intr_vec_disable(dev);
 693         else
 694                 intr_handle->nb_efd = count;
 695         return 0;
 696 }
 697
 698 /**
 699  * Clean up Rx interrupts handler.
 700  *
 701  * @param dev
 702  *   Pointer to Ethernet device.
 703  */
 704 void
 705 mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
 706 {
 707         struct mlx5_priv *priv = dev->data->dev_private;
 708         struct rte_intr_handle *intr_handle = dev->intr_handle;
 709         unsigned int i;
 710         unsigned int rxqs_n = priv->rxqs_n;
 711         unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
 712
 713         if (!dev->data->dev_conf.intr_conf.rxq)
 714                 return;
 715         if (!intr_handle->intr_vec)
 716                 goto free;
 717         for (i = 0; i != n; ++i) {
 718                 struct mlx5_rxq_ctrl *rxq_ctrl;
 719                 struct mlx5_rxq_data *rxq_data;
 720
 721                 if (intr_handle->intr_vec[i] == RTE_INTR_VEC_RXTX_OFFSET +
 722                     RTE_MAX_RXTX_INTR_VEC_ID)
 723                         continue;
 724                 /**
 725                  * Need to access directly the queue to release the reference
 726                  * kept in mlx5_rx_intr_vec_enable().
 727                  */
 728                 rxq_data = (*priv->rxqs)[i];
 729                 rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 730                 if (rxq_ctrl->obj)
 731                         mlx5_rxq_obj_release(rxq_ctrl->obj);
 732         }
 733 free:
 734         rte_intr_free_epoll_fd(intr_handle);
 735         if (intr_handle->intr_vec)
 736                 free(intr_handle->intr_vec);
 737         intr_handle->nb_efd = 0;
 738         intr_handle->intr_vec = NULL;
 739 }
 740
 741 /**
 742  *  MLX5 CQ notification .
 743  *
 744  *  @param rxq
 745  *     Pointer to receive queue structure.
 746  *  @param sq_n_rxq
 747  *     Sequence number per receive queue .
 748  */
 749 static inline void
 750 mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
 751 {
 752         int sq_n = 0;
 753         uint32_t doorbell_hi;
 754         uint64_t doorbell;
 755         void *cq_db_reg = (char *)rxq->cq_uar + MLX5_CQ_DOORBELL;
 756
 757         sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
 758         doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
 759         doorbell = (uint64_t)doorbell_hi << 32;
 760         doorbell |=  rxq->cqn;
 761         rxq->cq_db[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
 762         mlx5_uar_write64(rte_cpu_to_be_64(doorbell),
 763                          cq_db_reg, rxq->uar_lock_cq);
 764 }
 765
 766 /**
 767  * DPDK callback for Rx queue interrupt enable.
 768  *
 769  * @param dev
 770  *   Pointer to Ethernet device structure.
 771  * @param rx_queue_id
 772  *   Rx queue number.
 773  *
 774  * @return
 775  *   0 on success, a negative errno value otherwise and rte_errno is set.
 776  */
 777 int
 778 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 779 {
 780         struct mlx5_priv *priv = dev->data->dev_private;
 781         struct mlx5_rxq_data *rxq_data;
 782         struct mlx5_rxq_ctrl *rxq_ctrl;
 783
 784         rxq_data = (*priv->rxqs)[rx_queue_id];
 785         if (!rxq_data) {
 786                 rte_errno = EINVAL;
 787                 return -rte_errno;
 788         }
 789         rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 790         if (rxq_ctrl->irq) {
 791                 struct mlx5_rxq_obj *rxq_obj;
 792
 793                 rxq_obj = mlx5_rxq_obj_get(dev, rx_queue_id);
 794                 if (!rxq_obj) {
 795                         rte_errno = EINVAL;
 796                         return -rte_errno;
 797                 }
 798                 mlx5_arm_cq(rxq_data, rxq_data->cq_arm_sn);
 799                 mlx5_rxq_obj_release(rxq_obj);
 800         }
 801         return 0;
 802 }
 803
 804 /**
 805  * DPDK callback for Rx queue interrupt disable.
 806  *
 807  * @param dev
 808  *   Pointer to Ethernet device structure.
 809  * @param rx_queue_id
 810  *   Rx queue number.
 811  *
 812  * @return
 813  *   0 on success, a negative errno value otherwise and rte_errno is set.
 814  */
 815 int
 816 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 817 {
 818         struct mlx5_priv *priv = dev->data->dev_private;
 819         struct mlx5_rxq_data *rxq_data;
 820         struct mlx5_rxq_ctrl *rxq_ctrl;
 821         struct mlx5_rxq_obj *rxq_obj = NULL;
 822         struct ibv_cq *ev_cq;
 823         void *ev_ctx;
 824         int ret;
 825
 826         rxq_data = (*priv->rxqs)[rx_queue_id];
 827         if (!rxq_data) {
 828                 rte_errno = EINVAL;
 829                 return -rte_errno;
 830         }
 831         rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
 832         if (!rxq_ctrl->irq)
 833                 return 0;
 834         rxq_obj = mlx5_rxq_obj_get(dev, rx_queue_id);
 835         if (!rxq_obj) {
 836                 rte_errno = EINVAL;
 837                 return -rte_errno;
 838         }
 839         ret = mlx5_glue->get_cq_event(rxq_obj->channel, &ev_cq, &ev_ctx);
 840         if (ret || ev_cq != rxq_obj->cq) {
 841                 rte_errno = EINVAL;
 842                 goto exit;
 843         }
 844         rxq_data->cq_arm_sn++;
 845         mlx5_glue->ack_cq_events(rxq_obj->cq, 1);
 846         mlx5_rxq_obj_release(rxq_obj);
 847         return 0;
 848 exit:
 849         ret = rte_errno; /* Save rte_errno before cleanup. */
 850         if (rxq_obj)
 851                 mlx5_rxq_obj_release(rxq_obj);
 852         DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
 853                 dev->data->port_id, rx_queue_id);
 854         rte_errno = ret; /* Restore rte_errno. */
 855         return -rte_errno;
 856 }
 857
 858 /**
 859  * Create a CQ Verbs object.
 860  *
 861  * @param dev
 862  *   Pointer to Ethernet device.
 863  * @param priv
 864  *   Pointer to device private data.
 865  * @param rxq_data
 866  *   Pointer to Rx queue data.
 867  * @param cqe_n
 868  *   Number of CQEs in CQ.
 869  * @param rxq_obj
 870  *   Pointer to Rx queue object data.
 871  *
 872  * @return
 873  *   The Verbs object initialised, NULL otherwise and rte_errno is set.
 874  */
 875 static struct ibv_cq *
 876 mlx5_ibv_cq_new(struct rte_eth_dev *dev, struct mlx5_priv *priv,
 877                 struct mlx5_rxq_data *rxq_data,
 878                 unsigned int cqe_n, struct mlx5_rxq_obj *rxq_obj)
 879 {
 880         struct {
 881                 struct ibv_cq_init_attr_ex ibv;
 882                 struct mlx5dv_cq_init_attr mlx5;
 883         } cq_attr;
 884
 885         cq_attr.ibv = (struct ibv_cq_init_attr_ex){
 886                 .cqe = cqe_n,
 887                 .channel = rxq_obj->channel,
 888                 .comp_mask = 0,
 889         };
 890         cq_attr.mlx5 = (struct mlx5dv_cq_init_attr){
 891                 .comp_mask = 0,
 892         };
 893         if (priv->config.cqe_comp && !rxq_data->hw_timestamp) {
 894                 cq_attr.mlx5.comp_mask |=
 895                                 MLX5DV_CQ_INIT_ATTR_MASK_COMPRESSED_CQE;
 896 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 897                 cq_attr.mlx5.cqe_comp_res_format =
 898                                 mlx5_rxq_mprq_enabled(rxq_data) ?
 899                                 MLX5DV_CQE_RES_FORMAT_CSUM_STRIDX :
 900                                 MLX5DV_CQE_RES_FORMAT_HASH;
 901 #else
 902                 cq_attr.mlx5.cqe_comp_res_format = MLX5DV_CQE_RES_FORMAT_HASH;
 903 #endif
 904                 /*
 905                  * For vectorized Rx, it must not be doubled in order to
 906                  * make cq_ci and rq_ci aligned.
 907                  */
 908                 if (mlx5_rxq_check_vec_support(rxq_data) < 0)
 909                         cq_attr.ibv.cqe *= 2;
 910         } else if (priv->config.cqe_comp && rxq_data->hw_timestamp) {
 911                 DRV_LOG(DEBUG,
 912                         "port %u Rx CQE compression is disabled for HW"
 913                         " timestamp",
 914                         dev->data->port_id);
 915         }
 916 #ifdef HAVE_IBV_MLX5_MOD_CQE_128B_PAD
 917         if (priv->config.cqe_pad) {
 918                 cq_attr.mlx5.comp_mask |= MLX5DV_CQ_INIT_ATTR_MASK_FLAGS;
 919                 cq_attr.mlx5.flags |= MLX5DV_CQ_INIT_ATTR_FLAGS_CQE_PAD;
 920         }
 921 #endif
 922         return mlx5_glue->cq_ex_to_cq(mlx5_glue->dv_create_cq(priv->sh->ctx,
 923                                                               &cq_attr.ibv,
 924                                                               &cq_attr.mlx5));
 925 }
 926
 927 /**
 928  * Create a WQ Verbs object.
 929  *
 930  * @param dev
 931  *   Pointer to Ethernet device.
 932  * @param priv
 933  *   Pointer to device private data.
 934  * @param rxq_data
 935  *   Pointer to Rx queue data.
 936  * @param idx
 937  *   Queue index in DPDK Rx queue array
 938  * @param wqe_n
 939  *   Number of WQEs in WQ.
 940  * @param rxq_obj
 941  *   Pointer to Rx queue object data.
 942  *
 943  * @return
 944  *   The Verbs object initialised, NULL otherwise and rte_errno is set.
 945  */
 946 static struct ibv_wq *
 947 mlx5_ibv_wq_new(struct rte_eth_dev *dev, struct mlx5_priv *priv,
 948                 struct mlx5_rxq_data *rxq_data, uint16_t idx,
 949                 unsigned int wqe_n, struct mlx5_rxq_obj *rxq_obj)
 950 {
 951         struct {
 952                 struct ibv_wq_init_attr ibv;
 953 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 954                 struct mlx5dv_wq_init_attr mlx5;
 955 #endif
 956         } wq_attr;
 957
 958         wq_attr.ibv = (struct ibv_wq_init_attr){
 959                 .wq_context = NULL, /* Could be useful in the future. */
 960                 .wq_type = IBV_WQT_RQ,
 961                 /* Max number of outstanding WRs. */
 962                 .max_wr = wqe_n >> rxq_data->sges_n,
 963                 /* Max number of scatter/gather elements in a WR. */
 964                 .max_sge = 1 << rxq_data->sges_n,
 965                 .pd = priv->sh->pd,
 966                 .cq = rxq_obj->cq,
 967                 .comp_mask = IBV_WQ_FLAGS_CVLAN_STRIPPING | 0,
 968                 .create_flags = (rxq_data->vlan_strip ?
 969                                  IBV_WQ_FLAGS_CVLAN_STRIPPING : 0),
 970         };
 971         /* By default, FCS (CRC) is stripped by hardware. */
 972         if (rxq_data->crc_present) {
 973                 wq_attr.ibv.create_flags |= IBV_WQ_FLAGS_SCATTER_FCS;
 974                 wq_attr.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
 975         }
 976         if (priv->config.hw_padding) {
 977 #if defined(HAVE_IBV_WQ_FLAG_RX_END_PADDING)
 978                 wq_attr.ibv.create_flags |= IBV_WQ_FLAG_RX_END_PADDING;
 979                 wq_attr.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
 980 #elif defined(HAVE_IBV_WQ_FLAGS_PCI_WRITE_END_PADDING)
 981                 wq_attr.ibv.create_flags |= IBV_WQ_FLAGS_PCI_WRITE_END_PADDING;
 982                 wq_attr.ibv.comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
 983 #endif
 984         }
 985 #ifdef HAVE_IBV_DEVICE_STRIDING_RQ_SUPPORT
 986         wq_attr.mlx5 = (struct mlx5dv_wq_init_attr){
 987                 .comp_mask = 0,
 988         };
 989         if (mlx5_rxq_mprq_enabled(rxq_data)) {
 990                 struct mlx5dv_striding_rq_init_attr *mprq_attr =
 991                                                 &wq_attr.mlx5.striding_rq_attrs;
 992
 993                 wq_attr.mlx5.comp_mask |= MLX5DV_WQ_INIT_ATTR_MASK_STRIDING_RQ;
 994                 *mprq_attr = (struct mlx5dv_striding_rq_init_attr){
 995                         .single_stride_log_num_of_bytes = rxq_data->strd_sz_n,
 996                         .single_wqe_log_num_of_strides = rxq_data->strd_num_n,
 997                         .two_byte_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT,
 998                 };
 999         }
1000         rxq_obj->wq = mlx5_glue->dv_create_wq(priv->sh->ctx, &wq_attr.ibv,
1001                                               &wq_attr.mlx5);
1002 #else
1003         rxq_obj->wq = mlx5_glue->create_wq(priv->sh->ctx, &wq_attr.ibv);
1004 #endif
1005         if (rxq_obj->wq) {
1006                 /*
1007                  * Make sure number of WRs*SGEs match expectations since a queue
1008                  * cannot allocate more than "desc" buffers.
1009                  */
1010                 if (wq_attr.ibv.max_wr != (wqe_n >> rxq_data->sges_n) ||
1011                     wq_attr.ibv.max_sge != (1u << rxq_data->sges_n)) {
1012                         DRV_LOG(ERR,
1013                                 "port %u Rx queue %u requested %u*%u but got"
1014                                 " %u*%u WRs*SGEs",
1015                                 dev->data->port_id, idx,
1016                                 wqe_n >> rxq_data->sges_n,
1017                                 (1 << rxq_data->sges_n),
1018                                 wq_attr.ibv.max_wr, wq_attr.ibv.max_sge);
1019                         claim_zero(mlx5_glue->destroy_wq(rxq_obj->wq));
1020                         rxq_obj->wq = NULL;
1021                         rte_errno = EINVAL;
1022                 }
1023         }
1024         return rxq_obj->wq;
1025 }
1026
1027 /**
1028  * Fill common fields of create RQ attributes structure.
1029  *
1030  * @param rxq_data
1031  *   Pointer to Rx queue data.
1032  * @param cqn
1033  *   CQ number to use with this RQ.
1034  * @param rq_attr
1035  *   RQ attributes structure to fill..
1036  */
1037 static void
1038 mlx5_devx_create_rq_attr_fill(struct mlx5_rxq_data *rxq_data, uint32_t cqn,
1039                               struct mlx5_devx_create_rq_attr *rq_attr)
1040 {
1041         rq_attr->state = MLX5_RQC_STATE_RST;
1042         rq_attr->vsd = (rxq_data->vlan_strip) ? 0 : 1;
1043         rq_attr->cqn = cqn;
1044         rq_attr->scatter_fcs = (rxq_data->crc_present) ? 1 : 0;
1045 }
1046
1047 /**
1048  * Fill common fields of DevX WQ attributes structure.
1049  *
1050  * @param priv
1051  *   Pointer to device private data.
1052  * @param rxq_ctrl
1053  *   Pointer to Rx queue control structure.
1054  * @param wq_attr
1055  *   WQ attributes structure to fill..
1056  */
1057 static void
1058 mlx5_devx_wq_attr_fill(struct mlx5_priv *priv, struct mlx5_rxq_ctrl *rxq_ctrl,
1059                        struct mlx5_devx_wq_attr *wq_attr)
1060 {
1061         wq_attr->end_padding_mode = priv->config.cqe_pad ?
1062                                         MLX5_WQ_END_PAD_MODE_ALIGN :
1063                                         MLX5_WQ_END_PAD_MODE_NONE;
1064         wq_attr->pd = priv->sh->pdn;
1065         wq_attr->dbr_addr = rxq_ctrl->dbr_offset;
1066         wq_attr->dbr_umem_id = rxq_ctrl->dbr_umem_id;
1067         wq_attr->dbr_umem_valid = 1;
1068         wq_attr->wq_umem_id = rxq_ctrl->wq_umem->umem_id;
1069         wq_attr->wq_umem_valid = 1;
1070 }
1071
1072 /**
1073  * Create a RQ object using DevX.
1074  *
1075  * @param dev
1076  *   Pointer to Ethernet device.
1077  * @param idx
1078  *   Queue index in DPDK Rx queue array
1079  * @param cqn
1080  *   CQ number to use with this RQ.
1081  *
1082  * @return
1083  *   The DevX object initialised, NULL otherwise and rte_errno is set.
1084  */
1085 static struct mlx5_devx_obj *
1086 mlx5_devx_rq_new(struct rte_eth_dev *dev, uint16_t idx, uint32_t cqn)
1087 {
1088         struct mlx5_priv *priv = dev->data->dev_private;
1089         struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
1090         struct mlx5_rxq_ctrl *rxq_ctrl =
1091                 container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
1092         struct mlx5_devx_create_rq_attr rq_attr;
1093         uint32_t wqe_n = 1 << rxq_data->elts_n;
1094         uint32_t wq_size = 0;
1095         uint32_t wqe_size = 0;
1096         uint32_t log_wqe_size = 0;
1097         void *buf = NULL;
1098         struct mlx5_devx_obj *rq;
1099
1100         memset(&rq_attr, 0, sizeof(rq_attr));
1101         /* Fill RQ attributes. */
1102         rq_attr.mem_rq_type = MLX5_RQC_MEM_RQ_TYPE_MEMORY_RQ_INLINE;
1103         rq_attr.flush_in_error_en = 1;
1104         mlx5_devx_create_rq_attr_fill(rxq_data, cqn, &rq_attr);
1105         /* Fill WQ attributes for this RQ. */
1106         if (mlx5_rxq_mprq_enabled(rxq_data)) {
1107                 rq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC_STRIDING_RQ;
1108                 /*
1109                  * Number of strides in each WQE:
1110                  * 512*2^single_wqe_log_num_of_strides.
1111                  */
1112                 rq_attr.wq_attr.single_wqe_log_num_of_strides =
1113                                 rxq_data->strd_num_n -
1114                                 MLX5_MIN_SINGLE_WQE_LOG_NUM_STRIDES;
1115                 /* Stride size = (2^single_stride_log_num_of_bytes)*64B. */
1116                 rq_attr.wq_attr.single_stride_log_num_of_bytes =
1117                                 rxq_data->strd_sz_n -
1118                                 MLX5_MIN_SINGLE_STRIDE_LOG_NUM_BYTES;
1119                 wqe_size = sizeof(struct mlx5_wqe_mprq);
1120         } else {
1121                 int max_sge = 0;
1122                 int num_scatter = 0;
1123
1124                 rq_attr.wq_attr.wq_type = MLX5_WQ_TYPE_CYCLIC;
1125                 max_sge = 1 << rxq_data->sges_n;
1126                 num_scatter = RTE_MAX(max_sge, 1);
1127                 wqe_size = sizeof(struct mlx5_wqe_data_seg) * num_scatter;
1128         }
1129         log_wqe_size = log2above(wqe_size);
1130         rq_attr.wq_attr.log_wq_stride = log_wqe_size;
1131         rq_attr.wq_attr.log_wq_sz = rxq_data->elts_n;
1132         /* Calculate and allocate WQ memory space. */
1133         wqe_size = 1 << log_wqe_size; /* round up power of two.*/
1134         wq_size = wqe_n * wqe_size;
1135         buf = rte_calloc_socket(__func__, 1, wq_size, RTE_CACHE_LINE_SIZE,
1136                                 rxq_ctrl->socket);
1137         if (!buf)
1138                 return NULL;
1139         rxq_data->wqes = buf;
1140         rxq_ctrl->wq_umem = mlx5_glue->devx_umem_reg(priv->sh->ctx,
1141                                                      buf, wq_size, 0);
1142         if (!rxq_ctrl->wq_umem) {
1143                 rte_free(buf);
1144                 return NULL;
1145         }
1146         mlx5_devx_wq_attr_fill(priv, rxq_ctrl, &rq_attr.wq_attr);
1147         rq = mlx5_devx_cmd_create_rq(priv->sh->ctx, &rq_attr, rxq_ctrl->socket);
1148         if (!rq)
1149                 rxq_release_rq_resources(rxq_ctrl);
1150         return rq;
1151 }
1152
1153 /**
1154  * Create the Rx queue Verbs/DevX object.
1155  *
1156  * @param dev
1157  *   Pointer to Ethernet device.
1158  * @param idx
1159  *   Queue index in DPDK Rx queue array
1160  * @param type
1161  *   Type of Rx queue object to create.
1162  *
1163  * @return
1164  *   The Verbs/DevX object initialised, NULL otherwise and rte_errno is set.
1165  */
1166 struct mlx5_rxq_obj *
1167 mlx5_rxq_obj_new(struct rte_eth_dev *dev, uint16_t idx,
1168                  enum mlx5_rxq_obj_type type)
1169 {
1170         struct mlx5_priv *priv = dev->data->dev_private;
1171         struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
1172         struct mlx5_rxq_ctrl *rxq_ctrl =
1173                 container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
1174         struct ibv_wq_attr mod;
1175         unsigned int cqe_n;
1176         unsigned int wqe_n = 1 << rxq_data->elts_n;
1177         struct mlx5_rxq_obj *tmpl = NULL;
1178         struct mlx5dv_cq cq_info;
1179         struct mlx5dv_rwq rwq;
1180         int ret = 0;
1181         struct mlx5dv_obj obj;
1182
1183         assert(rxq_data);
1184         assert(!rxq_ctrl->obj);
1185         priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_RX_QUEUE;
1186         priv->verbs_alloc_ctx.obj = rxq_ctrl;
1187         tmpl = rte_calloc_socket(__func__, 1, sizeof(*tmpl), 0,
1188                                  rxq_ctrl->socket);
1189         if (!tmpl) {
1190                 DRV_LOG(ERR,
1191                         "port %u Rx queue %u cannot allocate verbs resources",
1192                         dev->data->port_id, rxq_data->idx);
1193                 rte_errno = ENOMEM;
1194                 goto error;
1195         }
1196         tmpl->type = type;
1197         tmpl->rxq_ctrl = rxq_ctrl;
1198         if (rxq_ctrl->irq) {
1199                 tmpl->channel = mlx5_glue->create_comp_channel(priv->sh->ctx);
1200                 if (!tmpl->channel) {
1201                         DRV_LOG(ERR, "port %u: comp channel creation failure",
1202                                 dev->data->port_id);
1203                         rte_errno = ENOMEM;
1204                         goto error;
1205                 }
1206         }
1207         if (mlx5_rxq_mprq_enabled(rxq_data))
1208                 cqe_n = wqe_n * (1 << rxq_data->strd_num_n) - 1;
1209         else
1210                 cqe_n = wqe_n  - 1;
1211         tmpl->cq = mlx5_ibv_cq_new(dev, priv, rxq_data, cqe_n, tmpl);
1212         if (!tmpl->cq) {
1213                 DRV_LOG(ERR, "port %u Rx queue %u CQ creation failure",
1214                         dev->data->port_id, idx);
1215                 rte_errno = ENOMEM;
1216                 goto error;
1217         }
1218         obj.cq.in = tmpl->cq;
1219         obj.cq.out = &cq_info;
1220         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_CQ);
1221         if (ret) {
1222                 rte_errno = ret;
1223                 goto error;
1224         }
1225         if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
1226                 DRV_LOG(ERR,
1227                         "port %u wrong MLX5_CQE_SIZE environment variable"
1228                         " value: it should be set to %u",
1229                         dev->data->port_id, RTE_CACHE_LINE_SIZE);
1230                 rte_errno = EINVAL;
1231                 goto error;
1232         }
1233         DRV_LOG(DEBUG, "port %u device_attr.max_qp_wr is %d",
1234                 dev->data->port_id, priv->sh->device_attr.orig_attr.max_qp_wr);
1235         DRV_LOG(DEBUG, "port %u device_attr.max_sge is %d",
1236                 dev->data->port_id, priv->sh->device_attr.orig_attr.max_sge);
1237         /* Allocate door-bell for types created with DevX. */
1238         if (tmpl->type != MLX5_RXQ_OBJ_TYPE_IBV) {
1239                 struct mlx5_devx_dbr_page *dbr_page;
1240                 int64_t dbr_offset;
1241
1242                 dbr_offset = mlx5_get_dbr(dev, &dbr_page);
1243                 if (dbr_offset < 0)
1244                         goto error;
1245                 rxq_ctrl->dbr_offset = dbr_offset;
1246                 rxq_ctrl->dbr_umem_id = dbr_page->umem->umem_id;
1247                 rxq_ctrl->dbr_umem_id_valid = 1;
1248                 rxq_data->rq_db = (uint32_t *)((uintptr_t)dbr_page->dbrs +
1249                                                (uintptr_t)rxq_ctrl->dbr_offset);
1250         }
1251         if (tmpl->type == MLX5_RXQ_OBJ_TYPE_IBV) {
1252                 tmpl->wq = mlx5_ibv_wq_new(dev, priv, rxq_data, idx, wqe_n,
1253                                            tmpl);
1254                 if (!tmpl->wq) {
1255                         DRV_LOG(ERR, "port %u Rx queue %u WQ creation failure",
1256                                 dev->data->port_id, idx);
1257                         rte_errno = ENOMEM;
1258                         goto error;
1259                 }
1260                 /* Change queue state to ready. */
1261                 mod = (struct ibv_wq_attr){
1262                         .attr_mask = IBV_WQ_ATTR_STATE,
1263                         .wq_state = IBV_WQS_RDY,
1264                 };
1265                 ret = mlx5_glue->modify_wq(tmpl->wq, &mod);
1266                 if (ret) {
1267                         DRV_LOG(ERR,
1268                                 "port %u Rx queue %u WQ state to IBV_WQS_RDY"
1269                                 " failed", dev->data->port_id, idx);
1270                         rte_errno = ret;
1271                         goto error;
1272                 }
1273                 obj.rwq.in = tmpl->wq;
1274                 obj.rwq.out = &rwq;
1275                 ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_RWQ);
1276                 if (ret) {
1277                         rte_errno = ret;
1278                         goto error;
1279                 }
1280                 rxq_data->wqes = rwq.buf;
1281                 rxq_data->rq_db = rwq.dbrec;
1282         } else if (tmpl->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ) {
1283                 struct mlx5_devx_modify_rq_attr rq_attr;
1284
1285                 memset(&rq_attr, 0, sizeof(rq_attr));
1286                 tmpl->rq = mlx5_devx_rq_new(dev, idx, cq_info.cqn);
1287                 if (!tmpl->rq) {
1288                         DRV_LOG(ERR, "port %u Rx queue %u RQ creation failure",
1289                                 dev->data->port_id, idx);
1290                         rte_errno = ENOMEM;
1291                         goto error;
1292                 }
1293                 /* Change queue state to ready. */
1294                 rq_attr.rq_state = MLX5_RQC_STATE_RST;
1295                 rq_attr.state = MLX5_RQC_STATE_RDY;
1296                 ret = mlx5_devx_cmd_modify_rq(tmpl->rq, &rq_attr);
1297                 if (ret)
1298                         goto error;
1299         }
1300         /* Fill the rings. */
1301         rxq_data->cqe_n = log2above(cq_info.cqe_cnt);
1302         rxq_data->cq_db = cq_info.dbrec;
1303         rxq_data->cqes = (volatile struct mlx5_cqe (*)[])(uintptr_t)cq_info.buf;
1304         rxq_data->cq_uar = cq_info.cq_uar;
1305         rxq_data->cqn = cq_info.cqn;
1306         rxq_data->cq_arm_sn = 0;
1307         mlx5_rxq_initialize(rxq_data);
1308         rxq_data->cq_ci = 0;
1309         DRV_LOG(DEBUG, "port %u rxq %u updated with %p", dev->data->port_id,
1310                 idx, (void *)&tmpl);
1311         rte_atomic32_inc(&tmpl->refcnt);
1312         LIST_INSERT_HEAD(&priv->rxqsobj, tmpl, next);
1313         priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1314         return tmpl;
1315 error:
1316         if (tmpl) {
1317                 ret = rte_errno; /* Save rte_errno before cleanup. */
1318                 if (tmpl->type == MLX5_RXQ_OBJ_TYPE_IBV && tmpl->wq)
1319                         claim_zero(mlx5_glue->destroy_wq(tmpl->wq));
1320                 else if (tmpl->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ && tmpl->rq)
1321                         claim_zero(mlx5_devx_cmd_destroy(tmpl->rq));
1322                 if (tmpl->cq)
1323                         claim_zero(mlx5_glue->destroy_cq(tmpl->cq));
1324                 if (tmpl->channel)
1325                         claim_zero(mlx5_glue->destroy_comp_channel
1326                                                         (tmpl->channel));
1327                 rte_free(tmpl);
1328                 rte_errno = ret; /* Restore rte_errno. */
1329         }
1330         if (type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ)
1331                 rxq_release_rq_resources(rxq_ctrl);
1332         priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
1333         return NULL;
1334 }
1335
1336 /**
1337  * Verify the Rx queue objects list is empty
1338  *
1339  * @param dev
1340  *   Pointer to Ethernet device.
1341  *
1342  * @return
1343  *   The number of objects not released.
1344  */
1345 int
1346 mlx5_rxq_obj_verify(struct rte_eth_dev *dev)
1347 {
1348         struct mlx5_priv *priv = dev->data->dev_private;
1349         int ret = 0;
1350         struct mlx5_rxq_obj *rxq_obj;
1351
1352         LIST_FOREACH(rxq_obj, &priv->rxqsobj, next) {
1353                 DRV_LOG(DEBUG, "port %u Rx queue %u still referenced",
1354                         dev->data->port_id, rxq_obj->rxq_ctrl->rxq.idx);
1355                 ++ret;
1356         }
1357         return ret;
1358 }
1359
1360 /**
1361  * Callback function to initialize mbufs for Multi-Packet RQ.
1362  */
1363 static inline void
1364 mlx5_mprq_buf_init(struct rte_mempool *mp, void *opaque_arg,
1365                     void *_m, unsigned int i __rte_unused)
1366 {
1367         struct mlx5_mprq_buf *buf = _m;
1368         struct rte_mbuf_ext_shared_info *shinfo;
1369         unsigned int strd_n = (unsigned int)(uintptr_t)opaque_arg;
1370         unsigned int j;
1371
1372         memset(_m, 0, sizeof(*buf));
1373         buf->mp = mp;
1374         rte_atomic16_set(&buf->refcnt, 1);
1375         for (j = 0; j != strd_n; ++j) {
1376                 shinfo = &buf->shinfos[j];
1377                 shinfo->free_cb = mlx5_mprq_buf_free_cb;
1378                 shinfo->fcb_opaque = buf;
1379         }
1380 }
1381
1382 /**
1383  * Free mempool of Multi-Packet RQ.
1384  *
1385  * @param dev
1386  *   Pointer to Ethernet device.
1387  *
1388  * @return
1389  *   0 on success, negative errno value on failure.
1390  */
1391 int
1392 mlx5_mprq_free_mp(struct rte_eth_dev *dev)
1393 {
1394         struct mlx5_priv *priv = dev->data->dev_private;
1395         struct rte_mempool *mp = priv->mprq_mp;
1396         unsigned int i;
1397
1398         if (mp == NULL)
1399                 return 0;
1400         DRV_LOG(DEBUG, "port %u freeing mempool (%s) for Multi-Packet RQ",
1401                 dev->data->port_id, mp->name);
1402         /*
1403          * If a buffer in the pool has been externally attached to a mbuf and it
1404          * is still in use by application, destroying the Rx queue can spoil
1405          * the packet. It is unlikely to happen but if application dynamically
1406          * creates and destroys with holding Rx packets, this can happen.
1407          *
1408          * TODO: It is unavoidable for now because the mempool for Multi-Packet
1409          * RQ isn't provided by application but managed by PMD.
1410          */
1411         if (!rte_mempool_full(mp)) {
1412                 DRV_LOG(ERR,
1413                         "port %u mempool for Multi-Packet RQ is still in use",
1414                         dev->data->port_id);
1415                 rte_errno = EBUSY;
1416                 return -rte_errno;
1417         }
1418         rte_mempool_free(mp);
1419         /* Unset mempool for each Rx queue. */
1420         for (i = 0; i != priv->rxqs_n; ++i) {
1421                 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1422
1423                 if (rxq == NULL)
1424                         continue;
1425                 rxq->mprq_mp = NULL;
1426         }
1427         priv->mprq_mp = NULL;
1428         return 0;
1429 }
1430
1431 /**
1432  * Allocate a mempool for Multi-Packet RQ. All configured Rx queues share the
1433  * mempool. If already allocated, reuse it if there're enough elements.
1434  * Otherwise, resize it.
1435  *
1436  * @param dev
1437  *   Pointer to Ethernet device.
1438  *
1439  * @return
1440  *   0 on success, negative errno value on failure.
1441  */
1442 int
1443 mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
1444 {
1445         struct mlx5_priv *priv = dev->data->dev_private;
1446         struct rte_mempool *mp = priv->mprq_mp;
1447         char name[RTE_MEMPOOL_NAMESIZE];
1448         unsigned int desc = 0;
1449         unsigned int buf_len;
1450         unsigned int obj_num;
1451         unsigned int obj_size;
1452         unsigned int strd_num_n = 0;
1453         unsigned int strd_sz_n = 0;
1454         unsigned int i;
1455
1456         if (!mlx5_mprq_enabled(dev))
1457                 return 0;
1458         /* Count the total number of descriptors configured. */
1459         for (i = 0; i != priv->rxqs_n; ++i) {
1460                 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1461
1462                 if (rxq == NULL)
1463                         continue;
1464                 desc += 1 << rxq->elts_n;
1465                 /* Get the max number of strides. */
1466                 if (strd_num_n < rxq->strd_num_n)
1467                         strd_num_n = rxq->strd_num_n;
1468                 /* Get the max size of a stride. */
1469                 if (strd_sz_n < rxq->strd_sz_n)
1470                         strd_sz_n = rxq->strd_sz_n;
1471         }
1472         assert(strd_num_n && strd_sz_n);
1473         buf_len = (1 << strd_num_n) * (1 << strd_sz_n);
1474         obj_size = sizeof(struct mlx5_mprq_buf) + buf_len + (1 << strd_num_n) *
1475                 sizeof(struct rte_mbuf_ext_shared_info) + RTE_PKTMBUF_HEADROOM;
1476         /*
1477          * Received packets can be either memcpy'd or externally referenced. In
1478          * case that the packet is attached to an mbuf as an external buffer, as
1479          * it isn't possible to predict how the buffers will be queued by
1480          * application, there's no option to exactly pre-allocate needed buffers
1481          * in advance but to speculatively prepares enough buffers.
1482          *
1483          * In the data path, if this Mempool is depleted, PMD will try to memcpy
1484          * received packets to buffers provided by application (rxq->mp) until
1485          * this Mempool gets available again.
1486          */
1487         desc *= 4;
1488         obj_num = desc + MLX5_MPRQ_MP_CACHE_SZ * priv->rxqs_n;
1489         /*
1490          * rte_mempool_create_empty() has sanity check to refuse large cache
1491          * size compared to the number of elements.
1492          * CACHE_FLUSHTHRESH_MULTIPLIER is defined in a C file, so using a
1493          * constant number 2 instead.
1494          */
1495         obj_num = RTE_MAX(obj_num, MLX5_MPRQ_MP_CACHE_SZ * 2);
1496         /* Check a mempool is already allocated and if it can be resued. */
1497         if (mp != NULL && mp->elt_size >= obj_size && mp->size >= obj_num) {
1498                 DRV_LOG(DEBUG, "port %u mempool %s is being reused",
1499                         dev->data->port_id, mp->name);
1500                 /* Reuse. */
1501                 goto exit;
1502         } else if (mp != NULL) {
1503                 DRV_LOG(DEBUG, "port %u mempool %s should be resized, freeing it",
1504                         dev->data->port_id, mp->name);
1505                 /*
1506                  * If failed to free, which means it may be still in use, no way
1507                  * but to keep using the existing one. On buffer underrun,
1508                  * packets will be memcpy'd instead of external buffer
1509                  * attachment.
1510                  */
1511                 if (mlx5_mprq_free_mp(dev)) {
1512                         if (mp->elt_size >= obj_size)
1513                                 goto exit;
1514                         else
1515                                 return -rte_errno;
1516                 }
1517         }
1518         snprintf(name, sizeof(name), "port-%u-mprq", dev->data->port_id);
1519         mp = rte_mempool_create(name, obj_num, obj_size, MLX5_MPRQ_MP_CACHE_SZ,
1520                                 0, NULL, NULL, mlx5_mprq_buf_init,
1521                                 (void *)(uintptr_t)(1 << strd_num_n),
1522                                 dev->device->numa_node, 0);
1523         if (mp == NULL) {
1524                 DRV_LOG(ERR,
1525                         "port %u failed to allocate a mempool for"
1526                         " Multi-Packet RQ, count=%u, size=%u",
1527                         dev->data->port_id, obj_num, obj_size);
1528                 rte_errno = ENOMEM;
1529                 return -rte_errno;
1530         }
1531         priv->mprq_mp = mp;
1532 exit:
1533         /* Set mempool for each Rx queue. */
1534         for (i = 0; i != priv->rxqs_n; ++i) {
1535                 struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
1536
1537                 if (rxq == NULL)
1538                         continue;
1539                 rxq->mprq_mp = mp;
1540         }
1541         DRV_LOG(INFO, "port %u Multi-Packet RQ is configured",
1542                 dev->data->port_id);
1543         return 0;
1544 }
1545
1546 #define MLX5_MAX_LRO_SIZE (UINT8_MAX * 256u)
1547 #define MLX5_MAX_TCP_HDR_OFFSET ((unsigned int)(sizeof(struct rte_ether_hdr) + \
1548                                         sizeof(struct rte_vlan_hdr) * 2 + \
1549                                         sizeof(struct rte_ipv6_hdr)))
1550 /**
1551  * Adjust the maximum LRO massage size.
1552  *
1553  * @param dev
1554  *   Pointer to Ethernet device.
1555  * @param max_lro_size
1556  *   The maximum size for LRO packet.
1557  */
1558 static void
1559 mlx5_max_lro_msg_size_adjust(struct rte_eth_dev *dev, uint32_t max_lro_size)
1560 {
1561         struct mlx5_priv *priv = dev->data->dev_private;
1562
1563         if (priv->config.hca_attr.lro_max_msg_sz_mode ==
1564             MLX5_LRO_MAX_MSG_SIZE_START_FROM_L4 && max_lro_size >
1565             MLX5_MAX_TCP_HDR_OFFSET)
1566                 max_lro_size -= MLX5_MAX_TCP_HDR_OFFSET;
1567         max_lro_size = RTE_MIN(max_lro_size, MLX5_MAX_LRO_SIZE);
1568         assert(max_lro_size >= 256u);
1569         max_lro_size /= 256u;
1570         if (priv->max_lro_msg_size)
1571                 priv->max_lro_msg_size =
1572                         RTE_MIN((uint32_t)priv->max_lro_msg_size, max_lro_size);
1573         else
1574                 priv->max_lro_msg_size = max_lro_size;
1575 }
1576
1577 /**
1578  * Create a DPDK Rx queue.
1579  *
1580  * @param dev
1581  *   Pointer to Ethernet device.
1582  * @param idx
1583  *   RX queue index.
1584  * @param desc
1585  *   Number of descriptors to configure in queue.
1586  * @param socket
1587  *   NUMA socket on which memory must be allocated.
1588  *
1589  * @return
1590  *   A DPDK queue object on success, NULL otherwise and rte_errno is set.
1591  */
1592 struct mlx5_rxq_ctrl *
1593 mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1594              unsigned int socket, const struct rte_eth_rxconf *conf,
1595              struct rte_mempool *mp)
1596 {
1597         struct mlx5_priv *priv = dev->data->dev_private;
1598         struct mlx5_rxq_ctrl *tmpl;
1599         unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
1600         unsigned int mprq_stride_size;
1601         struct mlx5_dev_config *config = &priv->config;
1602         unsigned int strd_headroom_en;
1603         /*
1604          * Always allocate extra slots, even if eventually
1605          * the vector Rx will not be used.
1606          */
1607         uint16_t desc_n =
1608                 desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
1609         uint64_t offloads = conf->offloads |
1610                            dev->data->dev_conf.rxmode.offloads;
1611         const int mprq_en = mlx5_check_mprq_support(dev) > 0;
1612         unsigned int max_rx_pkt_len = dev->data->dev_conf.rxmode.max_rx_pkt_len;
1613         unsigned int non_scatter_min_mbuf_size = max_rx_pkt_len +
1614                                                         RTE_PKTMBUF_HEADROOM;
1615
1616         if (non_scatter_min_mbuf_size > mb_len && !(offloads &
1617                                                     DEV_RX_OFFLOAD_SCATTER)) {
1618                 DRV_LOG(ERR, "port %u Rx queue %u: Scatter offload is not"
1619                         " configured and no enough mbuf space(%u) to contain "
1620                         "the maximum RX packet length(%u) with head-room(%u)",
1621                         dev->data->port_id, idx, mb_len, max_rx_pkt_len,
1622                         RTE_PKTMBUF_HEADROOM);
1623                 rte_errno = ENOSPC;
1624                 return NULL;
1625         }
1626         tmpl = rte_calloc_socket("RXQ", 1,
1627                                  sizeof(*tmpl) +
1628                                  desc_n * sizeof(struct rte_mbuf *),
1629                                  0, socket);
1630         if (!tmpl) {
1631                 rte_errno = ENOMEM;
1632                 return NULL;
1633         }
1634         if (mlx5_mr_btree_init(&tmpl->rxq.mr_ctrl.cache_bh,
1635                                MLX5_MR_BTREE_CACHE_N, socket)) {
1636                 /* rte_errno is already set. */
1637                 goto error;
1638         }
1639         tmpl->socket = socket;
1640         if (dev->data->dev_conf.intr_conf.rxq)
1641                 tmpl->irq = 1;
1642         /*
1643          * LRO packet may consume all the stride memory, hence we cannot
1644          * guaranty head-room near the packet memory in the stride.
1645          * In this case scatter is, for sure, enabled and an empty mbuf may be
1646          * added in the start for the head-room.
1647          */
1648         if (mlx5_lro_on(dev) && RTE_PKTMBUF_HEADROOM > 0 &&
1649             non_scatter_min_mbuf_size > mb_len) {
1650                 strd_headroom_en = 0;
1651                 mprq_stride_size = RTE_MIN(max_rx_pkt_len,
1652                                         1u << config->mprq.max_stride_size_n);
1653         } else {
1654                 strd_headroom_en = 1;
1655                 mprq_stride_size = non_scatter_min_mbuf_size;
1656         }
1657         /*
1658          * This Rx queue can be configured as a Multi-Packet RQ if all of the
1659          * following conditions are met:
1660          *  - MPRQ is enabled.
1661          *  - The number of descs is more than the number of strides.
1662          *  - max_rx_pkt_len plus overhead is less than the max size of a
1663          *    stride.
1664          *  Otherwise, enable Rx scatter if necessary.
1665          */
1666         if (mprq_en &&
1667             desc > (1U << config->mprq.stride_num_n) &&
1668             mprq_stride_size <= (1U << config->mprq.max_stride_size_n)) {
1669                 /* TODO: Rx scatter isn't supported yet. */
1670                 tmpl->rxq.sges_n = 0;
1671                 /* Trim the number of descs needed. */
1672                 desc >>= config->mprq.stride_num_n;
1673                 tmpl->rxq.strd_num_n = config->mprq.stride_num_n;
1674                 tmpl->rxq.strd_sz_n = RTE_MAX(log2above(mprq_stride_size),
1675                                               config->mprq.min_stride_size_n);
1676                 tmpl->rxq.strd_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT;
1677                 tmpl->rxq.strd_headroom_en = strd_headroom_en;
1678                 tmpl->rxq.mprq_max_memcpy_len = RTE_MIN(mb_len -
1679                             RTE_PKTMBUF_HEADROOM, config->mprq.max_memcpy_len);
1680                 mlx5_max_lro_msg_size_adjust(dev, RTE_MIN(max_rx_pkt_len,
1681                    (1u << tmpl->rxq.strd_num_n) * (1u << tmpl->rxq.strd_sz_n)));
1682                 DRV_LOG(DEBUG,
1683                         "port %u Rx queue %u: Multi-Packet RQ is enabled"
1684                         " strd_num_n = %u, strd_sz_n = %u",
1685                         dev->data->port_id, idx,
1686                         tmpl->rxq.strd_num_n, tmpl->rxq.strd_sz_n);
1687         } else if (max_rx_pkt_len <= (mb_len - RTE_PKTMBUF_HEADROOM)) {
1688                 tmpl->rxq.sges_n = 0;
1689         } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
1690                 unsigned int size = non_scatter_min_mbuf_size;
1691                 unsigned int sges_n;
1692
1693                 /*
1694                  * Determine the number of SGEs needed for a full packet
1695                  * and round it to the next power of two.
1696                  */
1697                 sges_n = log2above((size / mb_len) + !!(size % mb_len));
1698                 tmpl->rxq.sges_n = sges_n;
1699                 /* Make sure rxq.sges_n did not overflow. */
1700                 size = mb_len * (1 << tmpl->rxq.sges_n);
1701                 size -= RTE_PKTMBUF_HEADROOM;
1702                 if (size < max_rx_pkt_len) {
1703                         DRV_LOG(ERR,
1704                                 "port %u too many SGEs (%u) needed to handle"
1705                                 " requested maximum packet size %u",
1706                                 dev->data->port_id,
1707                                 1 << sges_n,
1708                                 max_rx_pkt_len);
1709                         rte_errno = EOVERFLOW;
1710                         goto error;
1711                 }
1712         }
1713         if (mprq_en && !mlx5_rxq_mprq_enabled(&tmpl->rxq))
1714                 DRV_LOG(WARNING,
1715                         "port %u MPRQ is requested but cannot be enabled"
1716                         " (requested: desc = %u, stride_sz = %u,"
1717                         " supported: min_stride_num = %u, max_stride_sz = %u).",
1718                         dev->data->port_id, desc, mprq_stride_size,
1719                         (1 << config->mprq.stride_num_n),
1720                         (1 << config->mprq.max_stride_size_n));
1721         DRV_LOG(DEBUG, "port %u maximum number of segments per packet: %u",
1722                 dev->data->port_id, 1 << tmpl->rxq.sges_n);
1723         if (desc % (1 << tmpl->rxq.sges_n)) {
1724                 DRV_LOG(ERR,
1725                         "port %u number of Rx queue descriptors (%u) is not a"
1726                         " multiple of SGEs per packet (%u)",
1727                         dev->data->port_id,
1728                         desc,
1729                         1 << tmpl->rxq.sges_n);
1730                 rte_errno = EINVAL;
1731                 goto error;
1732         }
1733         /* Toggle RX checksum offload if hardware supports it. */
1734         tmpl->rxq.csum = !!(offloads & DEV_RX_OFFLOAD_CHECKSUM);
1735         tmpl->rxq.hw_timestamp = !!(offloads & DEV_RX_OFFLOAD_TIMESTAMP);
1736         /* Configure VLAN stripping. */
1737         tmpl->rxq.vlan_strip = !!(offloads & DEV_RX_OFFLOAD_VLAN_STRIP);
1738         /* By default, FCS (CRC) is stripped by hardware. */
1739         tmpl->rxq.crc_present = 0;
1740         if (offloads & DEV_RX_OFFLOAD_KEEP_CRC) {
1741                 if (config->hw_fcs_strip) {
1742                         /*
1743                          * RQs used for LRO-enabled TIRs should not be
1744                          * configured to scatter the FCS.
1745                          */
1746                         if (mlx5_lro_on(dev))
1747                                 DRV_LOG(WARNING,
1748                                         "port %u CRC stripping has been "
1749                                         "disabled but will still be performed "
1750                                         "by hardware, because LRO is enabled",
1751                                         dev->data->port_id);
1752                         else
1753                                 tmpl->rxq.crc_present = 1;
1754                 } else {
1755                         DRV_LOG(WARNING,
1756                                 "port %u CRC stripping has been disabled but will"
1757                                 " still be performed by hardware, make sure MLNX_OFED"
1758                                 " and firmware are up to date",
1759                                 dev->data->port_id);
1760                 }
1761         }
1762         DRV_LOG(DEBUG,
1763                 "port %u CRC stripping is %s, %u bytes will be subtracted from"
1764                 " incoming frames to hide it",
1765                 dev->data->port_id,
1766                 tmpl->rxq.crc_present ? "disabled" : "enabled",
1767                 tmpl->rxq.crc_present << 2);
1768         /* Save port ID. */
1769         tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
1770                 (!!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS));
1771         tmpl->rxq.port_id = dev->data->port_id;
1772         tmpl->priv = priv;
1773         tmpl->rxq.mp = mp;
1774         tmpl->rxq.elts_n = log2above(desc);
1775         tmpl->rxq.rq_repl_thresh =
1776                 MLX5_VPMD_RXQ_RPLNSH_THRESH(1 << tmpl->rxq.elts_n);
1777         tmpl->rxq.elts =
1778                 (struct rte_mbuf *(*)[1 << tmpl->rxq.elts_n])(tmpl + 1);
1779 #ifndef RTE_ARCH_64
1780         tmpl->rxq.uar_lock_cq = &priv->uar_lock_cq;
1781 #endif
1782         tmpl->rxq.idx = idx;
1783         rte_atomic32_inc(&tmpl->refcnt);
1784         LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
1785         return tmpl;
1786 error:
1787         rte_free(tmpl);
1788         return NULL;
1789 }
1790
1791 /**
1792  * Get a Rx queue.
1793  *
1794  * @param dev
1795  *   Pointer to Ethernet device.
1796  * @param idx
1797  *   RX queue index.
1798  *
1799  * @return
1800  *   A pointer to the queue if it exists, NULL otherwise.
1801  */
1802 struct mlx5_rxq_ctrl *
1803 mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
1804 {
1805         struct mlx5_priv *priv = dev->data->dev_private;
1806         struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
1807
1808         if ((*priv->rxqs)[idx]) {
1809                 rxq_ctrl = container_of((*priv->rxqs)[idx],
1810                                         struct mlx5_rxq_ctrl,
1811                                         rxq);
1812                 mlx5_rxq_obj_get(dev, idx);
1813                 rte_atomic32_inc(&rxq_ctrl->refcnt);
1814         }
1815         return rxq_ctrl;
1816 }
1817
1818 /**
1819  * Release a Rx queue.
1820  *
1821  * @param dev
1822  *   Pointer to Ethernet device.
1823  * @param idx
1824  *   RX queue index.
1825  *
1826  * @return
1827  *   1 while a reference on it exists, 0 when freed.
1828  */
1829 int
1830 mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
1831 {
1832         struct mlx5_priv *priv = dev->data->dev_private;
1833         struct mlx5_rxq_ctrl *rxq_ctrl;
1834
1835         if (!(*priv->rxqs)[idx])
1836                 return 0;
1837         rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
1838         assert(rxq_ctrl->priv);
1839         if (rxq_ctrl->obj && !mlx5_rxq_obj_release(rxq_ctrl->obj))
1840                 rxq_ctrl->obj = NULL;
1841         if (rte_atomic32_dec_and_test(&rxq_ctrl->refcnt)) {
1842                 if (rxq_ctrl->dbr_umem_id_valid)
1843                         claim_zero(mlx5_release_dbr(dev, rxq_ctrl->dbr_umem_id,
1844                                                     rxq_ctrl->dbr_offset));
1845                 mlx5_mr_btree_free(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
1846                 LIST_REMOVE(rxq_ctrl, next);
1847                 rte_free(rxq_ctrl);
1848                 (*priv->rxqs)[idx] = NULL;
1849                 return 0;
1850         }
1851         return 1;
1852 }
1853
1854 /**
1855  * Verify the Rx Queue list is empty
1856  *
1857  * @param dev
1858  *   Pointer to Ethernet device.
1859  *
1860  * @return
1861  *   The number of object not released.
1862  */
1863 int
1864 mlx5_rxq_verify(struct rte_eth_dev *dev)
1865 {
1866         struct mlx5_priv *priv = dev->data->dev_private;
1867         struct mlx5_rxq_ctrl *rxq_ctrl;
1868         int ret = 0;
1869
1870         LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
1871                 DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
1872                         dev->data->port_id, rxq_ctrl->rxq.idx);
1873                 ++ret;
1874         }
1875         return ret;
1876 }
1877
1878 /**
1879  * Create an indirection table.
1880  *
1881  * @param dev
1882  *   Pointer to Ethernet device.
1883  * @param queues
1884  *   Queues entering in the indirection table.
1885  * @param queues_n
1886  *   Number of queues in the array.
1887  *
1888  * @return
1889  *   The Verbs/DevX object initialised, NULL otherwise and rte_errno is set.
1890  */
1891 static struct mlx5_ind_table_obj *
1892 mlx5_ind_table_obj_new(struct rte_eth_dev *dev, const uint16_t *queues,
1893                        uint32_t queues_n, enum mlx5_ind_tbl_type type)
1894 {
1895         struct mlx5_priv *priv = dev->data->dev_private;
1896         struct mlx5_ind_table_obj *ind_tbl;
1897         unsigned int i = 0, j = 0, k = 0;
1898
1899         ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl) +
1900                              queues_n * sizeof(uint16_t), 0);
1901         if (!ind_tbl) {
1902                 rte_errno = ENOMEM;
1903                 return NULL;
1904         }
1905         ind_tbl->type = type;
1906         if (ind_tbl->type == MLX5_IND_TBL_TYPE_IBV) {
1907                 const unsigned int wq_n = rte_is_power_of_2(queues_n) ?
1908                         log2above(queues_n) :
1909                         log2above(priv->config.ind_table_max_size);
1910                 struct ibv_wq *wq[1 << wq_n];
1911
1912                 for (i = 0; i != queues_n; ++i) {
1913                         struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev,
1914                                                                  queues[i]);
1915                         if (!rxq)
1916                                 goto error;
1917                         wq[i] = rxq->obj->wq;
1918                         ind_tbl->queues[i] = queues[i];
1919                 }
1920                 ind_tbl->queues_n = queues_n;
1921                 /* Finalise indirection table. */
1922                 k = i; /* Retain value of i for use in error case. */
1923                 for (j = 0; k != (unsigned int)(1 << wq_n); ++k, ++j)
1924                         wq[k] = wq[j];
1925                 ind_tbl->ind_table = mlx5_glue->create_rwq_ind_table
1926                         (priv->sh->ctx,
1927                          &(struct ibv_rwq_ind_table_init_attr){
1928                                 .log_ind_tbl_size = wq_n,
1929                                 .ind_tbl = wq,
1930                                 .comp_mask = 0,
1931                         });
1932                 if (!ind_tbl->ind_table) {
1933                         rte_errno = errno;
1934                         goto error;
1935                 }
1936         } else { /* ind_tbl->type == MLX5_IND_TBL_TYPE_DEVX */
1937                 struct mlx5_devx_rqt_attr *rqt_attr = NULL;
1938
1939                 rqt_attr = rte_calloc(__func__, 1, sizeof(*rqt_attr) +
1940                                       queues_n * sizeof(uint16_t), 0);
1941                 if (!rqt_attr) {
1942                         DRV_LOG(ERR, "port %u cannot allocate RQT resources",
1943                                 dev->data->port_id);
1944                         rte_errno = ENOMEM;
1945                         goto error;
1946                 }
1947                 rqt_attr->rqt_max_size = priv->config.ind_table_max_size;
1948                 rqt_attr->rqt_actual_size = queues_n;
1949                 for (i = 0; i != queues_n; ++i) {
1950                         struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev,
1951                                                                  queues[i]);
1952                         if (!rxq)
1953                                 goto error;
1954                         rqt_attr->rq_list[i] = rxq->obj->rq->id;
1955                         ind_tbl->queues[i] = queues[i];
1956                 }
1957                 ind_tbl->rqt = mlx5_devx_cmd_create_rqt(priv->sh->ctx,
1958                                                         rqt_attr);
1959                 rte_free(rqt_attr);
1960                 if (!ind_tbl->rqt) {
1961                         DRV_LOG(ERR, "port %u cannot create DevX RQT",
1962                                 dev->data->port_id);
1963                         rte_errno = errno;
1964                         goto error;
1965                 }
1966                 ind_tbl->queues_n = queues_n;
1967         }
1968         rte_atomic32_inc(&ind_tbl->refcnt);
1969         LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
1970         return ind_tbl;
1971 error:
1972         for (j = 0; j < i; j++)
1973                 mlx5_rxq_release(dev, ind_tbl->queues[j]);
1974         rte_free(ind_tbl);
1975         DEBUG("port %u cannot create indirection table", dev->data->port_id);
1976         return NULL;
1977 }
1978
1979 /**
1980  * Get an indirection table.
1981  *
1982  * @param dev
1983  *   Pointer to Ethernet device.
1984  * @param queues
1985  *   Queues entering in the indirection table.
1986  * @param queues_n
1987  *   Number of queues in the array.
1988  *
1989  * @return
1990  *   An indirection table if found.
1991  */
1992 static struct mlx5_ind_table_obj *
1993 mlx5_ind_table_obj_get(struct rte_eth_dev *dev, const uint16_t *queues,
1994                        uint32_t queues_n)
1995 {
1996         struct mlx5_priv *priv = dev->data->dev_private;
1997         struct mlx5_ind_table_obj *ind_tbl;
1998
1999         LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
2000                 if ((ind_tbl->queues_n == queues_n) &&
2001                     (memcmp(ind_tbl->queues, queues,
2002                             ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
2003                      == 0))
2004                         break;
2005         }
2006         if (ind_tbl) {
2007                 unsigned int i;
2008
2009                 rte_atomic32_inc(&ind_tbl->refcnt);
2010                 for (i = 0; i != ind_tbl->queues_n; ++i)
2011                         mlx5_rxq_get(dev, ind_tbl->queues[i]);
2012         }
2013         return ind_tbl;
2014 }
2015
2016 /**
2017  * Release an indirection table.
2018  *
2019  * @param dev
2020  *   Pointer to Ethernet device.
2021  * @param ind_table
2022  *   Indirection table to release.
2023  *
2024  * @return
2025  *   1 while a reference on it exists, 0 when freed.
2026  */
2027 static int
2028 mlx5_ind_table_obj_release(struct rte_eth_dev *dev,
2029                            struct mlx5_ind_table_obj *ind_tbl)
2030 {
2031         unsigned int i;
2032
2033         if (rte_atomic32_dec_and_test(&ind_tbl->refcnt)) {
2034                 if (ind_tbl->type == MLX5_IND_TBL_TYPE_IBV)
2035                         claim_zero(mlx5_glue->destroy_rwq_ind_table
2036                                                         (ind_tbl->ind_table));
2037                 else if (ind_tbl->type == MLX5_IND_TBL_TYPE_DEVX)
2038                         claim_zero(mlx5_devx_cmd_destroy(ind_tbl->rqt));
2039         }
2040         for (i = 0; i != ind_tbl->queues_n; ++i)
2041                 claim_nonzero(mlx5_rxq_release(dev, ind_tbl->queues[i]));
2042         if (!rte_atomic32_read(&ind_tbl->refcnt)) {
2043                 LIST_REMOVE(ind_tbl, next);
2044                 rte_free(ind_tbl);
2045                 return 0;
2046         }
2047         return 1;
2048 }
2049
2050 /**
2051  * Verify the Rx Queue list is empty
2052  *
2053  * @param dev
2054  *   Pointer to Ethernet device.
2055  *
2056  * @return
2057  *   The number of object not released.
2058  */
2059 int
2060 mlx5_ind_table_obj_verify(struct rte_eth_dev *dev)
2061 {
2062         struct mlx5_priv *priv = dev->data->dev_private;
2063         struct mlx5_ind_table_obj *ind_tbl;
2064         int ret = 0;
2065
2066         LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
2067                 DRV_LOG(DEBUG,
2068                         "port %u indirection table obj %p still referenced",
2069                         dev->data->port_id, (void *)ind_tbl);
2070                 ++ret;
2071         }
2072         return ret;
2073 }
2074
2075 /**
2076  * Create an Rx Hash queue.
2077  *
2078  * @param dev
2079  *   Pointer to Ethernet device.
2080  * @param rss_key
2081  *   RSS key for the Rx hash queue.
2082  * @param rss_key_len
2083  *   RSS key length.
2084  * @param hash_fields
2085  *   Verbs protocol hash field to make the RSS on.
2086  * @param queues
2087  *   Queues entering in hash queue. In case of empty hash_fields only the
2088  *   first queue index will be taken for the indirection table.
2089  * @param queues_n
2090  *   Number of queues.
2091  * @param tunnel
2092  *   Tunnel type.
2093  * @param lro
2094  *   Flow rule is relevant for LRO, i.e. contains IPv4/IPv6 and TCP.
2095  *
2096  * @return
2097  *   The Verbs/DevX object initialised, NULL otherwise and rte_errno is set.
2098  */
2099 struct mlx5_hrxq *
2100 mlx5_hrxq_new(struct rte_eth_dev *dev,
2101               const uint8_t *rss_key, uint32_t rss_key_len,
2102               uint64_t hash_fields,
2103               const uint16_t *queues, uint32_t queues_n,
2104               int tunnel __rte_unused, int lro)
2105 {
2106         struct mlx5_priv *priv = dev->data->dev_private;
2107         struct mlx5_hrxq *hrxq;
2108         struct ibv_qp *qp = NULL;
2109         struct mlx5_ind_table_obj *ind_tbl;
2110         int err;
2111         struct mlx5_devx_obj *tir = NULL;
2112
2113         queues_n = hash_fields ? queues_n : 1;
2114         ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
2115         if (!ind_tbl) {
2116                 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[queues[0]];
2117                 struct mlx5_rxq_ctrl *rxq_ctrl =
2118                         container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
2119                 enum mlx5_ind_tbl_type type;
2120
2121                 type = rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_IBV ?
2122                                 MLX5_IND_TBL_TYPE_IBV : MLX5_IND_TBL_TYPE_DEVX;
2123                 ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n, type);
2124         }
2125         if (!ind_tbl) {
2126                 rte_errno = ENOMEM;
2127                 return NULL;
2128         }
2129         if (ind_tbl->type == MLX5_IND_TBL_TYPE_IBV) {
2130 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
2131                 struct mlx5dv_qp_init_attr qp_init_attr;
2132
2133                 memset(&qp_init_attr, 0, sizeof(qp_init_attr));
2134                 if (tunnel) {
2135                         qp_init_attr.comp_mask =
2136                                 MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
2137                         qp_init_attr.create_flags =
2138                                 MLX5DV_QP_CREATE_TUNNEL_OFFLOADS;
2139                 }
2140 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2141                 if (dev->data->dev_conf.lpbk_mode) {
2142                         /*
2143                          * Allow packet sent from NIC loop back
2144                          * w/o source MAC check.
2145                          */
2146                         qp_init_attr.comp_mask |=
2147                                 MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
2148                         qp_init_attr.create_flags |=
2149                                 MLX5DV_QP_CREATE_TIR_ALLOW_SELF_LOOPBACK_UC;
2150                 }
2151 #endif
2152                 qp = mlx5_glue->dv_create_qp
2153                         (priv->sh->ctx,
2154                          &(struct ibv_qp_init_attr_ex){
2155                                 .qp_type = IBV_QPT_RAW_PACKET,
2156                                 .comp_mask =
2157                                         IBV_QP_INIT_ATTR_PD |
2158                                         IBV_QP_INIT_ATTR_IND_TABLE |
2159                                         IBV_QP_INIT_ATTR_RX_HASH,
2160                                 .rx_hash_conf = (struct ibv_rx_hash_conf){
2161                                         .rx_hash_function =
2162                                                 IBV_RX_HASH_FUNC_TOEPLITZ,
2163                                         .rx_hash_key_len = rss_key_len,
2164                                         .rx_hash_key =
2165                                                 (void *)(uintptr_t)rss_key,
2166                                         .rx_hash_fields_mask = hash_fields,
2167                                 },
2168                                 .rwq_ind_tbl = ind_tbl->ind_table,
2169                                 .pd = priv->sh->pd,
2170                           },
2171                           &qp_init_attr);
2172 #else
2173                 qp = mlx5_glue->create_qp_ex
2174                         (priv->sh->ctx,
2175                          &(struct ibv_qp_init_attr_ex){
2176                                 .qp_type = IBV_QPT_RAW_PACKET,
2177                                 .comp_mask =
2178                                         IBV_QP_INIT_ATTR_PD |
2179                                         IBV_QP_INIT_ATTR_IND_TABLE |
2180                                         IBV_QP_INIT_ATTR_RX_HASH,
2181                                 .rx_hash_conf = (struct ibv_rx_hash_conf){
2182                                         .rx_hash_function =
2183                                                 IBV_RX_HASH_FUNC_TOEPLITZ,
2184                                         .rx_hash_key_len = rss_key_len,
2185                                         .rx_hash_key =
2186                                                 (void *)(uintptr_t)rss_key,
2187                                         .rx_hash_fields_mask = hash_fields,
2188                                 },
2189                                 .rwq_ind_tbl = ind_tbl->ind_table,
2190                                 .pd = priv->sh->pd,
2191                          });
2192 #endif
2193                 if (!qp) {
2194                         rte_errno = errno;
2195                         goto error;
2196                 }
2197         } else { /* ind_tbl->type == MLX5_IND_TBL_TYPE_DEVX */
2198                 struct mlx5_devx_tir_attr tir_attr;
2199
2200                 memset(&tir_attr, 0, sizeof(tir_attr));
2201                 tir_attr.disp_type = MLX5_TIRC_DISP_TYPE_INDIRECT;
2202                 tir_attr.rx_hash_fn = MLX5_RX_HASH_FN_TOEPLITZ;
2203                 memcpy(&tir_attr.rx_hash_field_selector_outer, &hash_fields,
2204                        sizeof(uint64_t));
2205                 tir_attr.transport_domain = priv->sh->tdn;
2206                 memcpy(tir_attr.rx_hash_toeplitz_key, rss_key, rss_key_len);
2207                 tir_attr.indirect_table = ind_tbl->rqt->id;
2208                 if (dev->data->dev_conf.lpbk_mode)
2209                         tir_attr.self_lb_block =
2210                                         MLX5_TIRC_SELF_LB_BLOCK_BLOCK_UNICAST;
2211                 if (lro) {
2212                         tir_attr.lro_timeout_period_usecs =
2213                                         priv->config.lro.timeout;
2214                         tir_attr.lro_max_msg_sz = priv->max_lro_msg_size;
2215                         tir_attr.lro_enable_mask = lro;
2216                 }
2217                 tir = mlx5_devx_cmd_create_tir(priv->sh->ctx, &tir_attr);
2218                 if (!tir) {
2219                         DRV_LOG(ERR, "port %u cannot create DevX TIR",
2220                                 dev->data->port_id);
2221                         rte_errno = errno;
2222                         goto error;
2223                 }
2224         }
2225         hrxq = rte_calloc(__func__, 1, sizeof(*hrxq) + rss_key_len, 0);
2226         if (!hrxq)
2227                 goto error;
2228         hrxq->ind_table = ind_tbl;
2229         if (ind_tbl->type == MLX5_IND_TBL_TYPE_IBV) {
2230                 hrxq->qp = qp;
2231 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2232                 hrxq->action =
2233                         mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
2234                 if (!hrxq->action) {
2235                         rte_errno = errno;
2236                         goto error;
2237                 }
2238 #endif
2239         } else { /* ind_tbl->type == MLX5_IND_TBL_TYPE_DEVX */
2240                 hrxq->tir = tir;
2241 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2242                 hrxq->action = mlx5_glue->dv_create_flow_action_dest_devx_tir
2243                                                         (hrxq->tir->obj);
2244                 if (!hrxq->action) {
2245                         rte_errno = errno;
2246                         goto error;
2247                 }
2248 #endif
2249         }
2250         hrxq->rss_key_len = rss_key_len;
2251         hrxq->hash_fields = hash_fields;
2252         memcpy(hrxq->rss_key, rss_key, rss_key_len);
2253         rte_atomic32_inc(&hrxq->refcnt);
2254         LIST_INSERT_HEAD(&priv->hrxqs, hrxq, next);
2255         return hrxq;
2256 error:
2257         err = rte_errno; /* Save rte_errno before cleanup. */
2258         mlx5_ind_table_obj_release(dev, ind_tbl);
2259         if (qp)
2260                 claim_zero(mlx5_glue->destroy_qp(qp));
2261         else if (tir)
2262                 claim_zero(mlx5_devx_cmd_destroy(tir));
2263         rte_errno = err; /* Restore rte_errno. */
2264         return NULL;
2265 }
2266
2267 /**
2268  * Get an Rx Hash queue.
2269  *
2270  * @param dev
2271  *   Pointer to Ethernet device.
2272  * @param rss_conf
2273  *   RSS configuration for the Rx hash queue.
2274  * @param queues
2275  *   Queues entering in hash queue. In case of empty hash_fields only the
2276  *   first queue index will be taken for the indirection table.
2277  * @param queues_n
2278  *   Number of queues.
2279  *
2280  * @return
2281  *   An hash Rx queue on success.
2282  */
2283 struct mlx5_hrxq *
2284 mlx5_hrxq_get(struct rte_eth_dev *dev,
2285               const uint8_t *rss_key, uint32_t rss_key_len,
2286               uint64_t hash_fields,
2287               const uint16_t *queues, uint32_t queues_n)
2288 {
2289         struct mlx5_priv *priv = dev->data->dev_private;
2290         struct mlx5_hrxq *hrxq;
2291
2292         queues_n = hash_fields ? queues_n : 1;
2293         LIST_FOREACH(hrxq, &priv->hrxqs, next) {
2294                 struct mlx5_ind_table_obj *ind_tbl;
2295
2296                 if (hrxq->rss_key_len != rss_key_len)
2297                         continue;
2298                 if (memcmp(hrxq->rss_key, rss_key, rss_key_len))
2299                         continue;
2300                 if (hrxq->hash_fields != hash_fields)
2301                         continue;
2302                 ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
2303                 if (!ind_tbl)
2304                         continue;
2305                 if (ind_tbl != hrxq->ind_table) {
2306                         mlx5_ind_table_obj_release(dev, ind_tbl);
2307                         continue;
2308                 }
2309                 rte_atomic32_inc(&hrxq->refcnt);
2310                 return hrxq;
2311         }
2312         return NULL;
2313 }
2314
2315 /**
2316  * Release the hash Rx queue.
2317  *
2318  * @param dev
2319  *   Pointer to Ethernet device.
2320  * @param hrxq
2321  *   Pointer to Hash Rx queue to release.
2322  *
2323  * @return
2324  *   1 while a reference on it exists, 0 when freed.
2325  */
2326 int
2327 mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
2328 {
2329         if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
2330 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2331                 mlx5_glue->destroy_flow_action(hrxq->action);
2332 #endif
2333                 if (hrxq->ind_table->type == MLX5_IND_TBL_TYPE_IBV)
2334                         claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
2335                 else /* hrxq->ind_table->type == MLX5_IND_TBL_TYPE_DEVX */
2336                         claim_zero(mlx5_devx_cmd_destroy(hrxq->tir));
2337                 mlx5_ind_table_obj_release(dev, hrxq->ind_table);
2338                 LIST_REMOVE(hrxq, next);
2339                 rte_free(hrxq);
2340                 return 0;
2341         }
2342         claim_nonzero(mlx5_ind_table_obj_release(dev, hrxq->ind_table));
2343         return 1;
2344 }
2345
2346 /**
2347  * Verify the Rx Queue list is empty
2348  *
2349  * @param dev
2350  *   Pointer to Ethernet device.
2351  *
2352  * @return
2353  *   The number of object not released.
2354  */
2355 int
2356 mlx5_hrxq_verify(struct rte_eth_dev *dev)
2357 {
2358         struct mlx5_priv *priv = dev->data->dev_private;
2359         struct mlx5_hrxq *hrxq;
2360         int ret = 0;
2361
2362         LIST_FOREACH(hrxq, &priv->hrxqs, next) {
2363                 DRV_LOG(DEBUG,
2364                         "port %u hash Rx queue %p still referenced",
2365                         dev->data->port_id, (void *)hrxq);
2366                 ++ret;
2367         }
2368         return ret;
2369 }
2370
2371 /**
2372  * Create a drop Rx queue Verbs/DevX object.
2373  *
2374  * @param dev
2375  *   Pointer to Ethernet device.
2376  *
2377  * @return
2378  *   The Verbs/DevX object initialised, NULL otherwise and rte_errno is set.
2379  */
2380 static struct mlx5_rxq_obj *
2381 mlx5_rxq_obj_drop_new(struct rte_eth_dev *dev)
2382 {
2383         struct mlx5_priv *priv = dev->data->dev_private;
2384         struct ibv_context *ctx = priv->sh->ctx;
2385         struct ibv_cq *cq;
2386         struct ibv_wq *wq = NULL;
2387         struct mlx5_rxq_obj *rxq;
2388
2389         if (priv->drop_queue.rxq)
2390                 return priv->drop_queue.rxq;
2391         cq = mlx5_glue->create_cq(ctx, 1, NULL, NULL, 0);
2392         if (!cq) {
2393                 DEBUG("port %u cannot allocate CQ for drop queue",
2394                       dev->data->port_id);
2395                 rte_errno = errno;
2396                 goto error;
2397         }
2398         wq = mlx5_glue->create_wq(ctx,
2399                  &(struct ibv_wq_init_attr){
2400                         .wq_type = IBV_WQT_RQ,
2401                         .max_wr = 1,
2402                         .max_sge = 1,
2403                         .pd = priv->sh->pd,
2404                         .cq = cq,
2405                  });
2406         if (!wq) {
2407                 DEBUG("port %u cannot allocate WQ for drop queue",
2408                       dev->data->port_id);
2409                 rte_errno = errno;
2410                 goto error;
2411         }
2412         rxq = rte_calloc(__func__, 1, sizeof(*rxq), 0);
2413         if (!rxq) {
2414                 DEBUG("port %u cannot allocate drop Rx queue memory",
2415                       dev->data->port_id);
2416                 rte_errno = ENOMEM;
2417                 goto error;
2418         }
2419         rxq->cq = cq;
2420         rxq->wq = wq;
2421         priv->drop_queue.rxq = rxq;
2422         return rxq;
2423 error:
2424         if (wq)
2425                 claim_zero(mlx5_glue->destroy_wq(wq));
2426         if (cq)
2427                 claim_zero(mlx5_glue->destroy_cq(cq));
2428         return NULL;
2429 }
2430
2431 /**
2432  * Release a drop Rx queue Verbs/DevX object.
2433  *
2434  * @param dev
2435  *   Pointer to Ethernet device.
2436  *
2437  * @return
2438  *   The Verbs/DevX object initialised, NULL otherwise and rte_errno is set.
2439  */
2440 static void
2441 mlx5_rxq_obj_drop_release(struct rte_eth_dev *dev)
2442 {
2443         struct mlx5_priv *priv = dev->data->dev_private;
2444         struct mlx5_rxq_obj *rxq = priv->drop_queue.rxq;
2445
2446         if (rxq->wq)
2447                 claim_zero(mlx5_glue->destroy_wq(rxq->wq));
2448         if (rxq->cq)
2449                 claim_zero(mlx5_glue->destroy_cq(rxq->cq));
2450         rte_free(rxq);
2451         priv->drop_queue.rxq = NULL;
2452 }
2453
2454 /**
2455  * Create a drop indirection table.
2456  *
2457  * @param dev
2458  *   Pointer to Ethernet device.
2459  *
2460  * @return
2461  *   The Verbs/DevX object initialised, NULL otherwise and rte_errno is set.
2462  */
2463 static struct mlx5_ind_table_obj *
2464 mlx5_ind_table_obj_drop_new(struct rte_eth_dev *dev)
2465 {
2466         struct mlx5_priv *priv = dev->data->dev_private;
2467         struct mlx5_ind_table_obj *ind_tbl;
2468         struct mlx5_rxq_obj *rxq;
2469         struct mlx5_ind_table_obj tmpl;
2470
2471         rxq = mlx5_rxq_obj_drop_new(dev);
2472         if (!rxq)
2473                 return NULL;
2474         tmpl.ind_table = mlx5_glue->create_rwq_ind_table
2475                 (priv->sh->ctx,
2476                  &(struct ibv_rwq_ind_table_init_attr){
2477                         .log_ind_tbl_size = 0,
2478                         .ind_tbl = &rxq->wq,
2479                         .comp_mask = 0,
2480                  });
2481         if (!tmpl.ind_table) {
2482                 DEBUG("port %u cannot allocate indirection table for drop"
2483                       " queue",
2484                       dev->data->port_id);
2485                 rte_errno = errno;
2486                 goto error;
2487         }
2488         ind_tbl = rte_calloc(__func__, 1, sizeof(*ind_tbl), 0);
2489         if (!ind_tbl) {
2490                 rte_errno = ENOMEM;
2491                 goto error;
2492         }
2493         ind_tbl->ind_table = tmpl.ind_table;
2494         return ind_tbl;
2495 error:
2496         mlx5_rxq_obj_drop_release(dev);
2497         return NULL;
2498 }
2499
2500 /**
2501  * Release a drop indirection table.
2502  *
2503  * @param dev
2504  *   Pointer to Ethernet device.
2505  */
2506 static void
2507 mlx5_ind_table_obj_drop_release(struct rte_eth_dev *dev)
2508 {
2509         struct mlx5_priv *priv = dev->data->dev_private;
2510         struct mlx5_ind_table_obj *ind_tbl = priv->drop_queue.hrxq->ind_table;
2511
2512         claim_zero(mlx5_glue->destroy_rwq_ind_table(ind_tbl->ind_table));
2513         mlx5_rxq_obj_drop_release(dev);
2514         rte_free(ind_tbl);
2515         priv->drop_queue.hrxq->ind_table = NULL;
2516 }
2517
2518 /**
2519  * Create a drop Rx Hash queue.
2520  *
2521  * @param dev
2522  *   Pointer to Ethernet device.
2523  *
2524  * @return
2525  *   The Verbs/DevX object initialised, NULL otherwise and rte_errno is set.
2526  */
2527 struct mlx5_hrxq *
2528 mlx5_hrxq_drop_new(struct rte_eth_dev *dev)
2529 {
2530         struct mlx5_priv *priv = dev->data->dev_private;
2531         struct mlx5_ind_table_obj *ind_tbl;
2532         struct ibv_qp *qp;
2533         struct mlx5_hrxq *hrxq;
2534
2535         if (priv->drop_queue.hrxq) {
2536                 rte_atomic32_inc(&priv->drop_queue.hrxq->refcnt);
2537                 return priv->drop_queue.hrxq;
2538         }
2539         ind_tbl = mlx5_ind_table_obj_drop_new(dev);
2540         if (!ind_tbl)
2541                 return NULL;
2542         qp = mlx5_glue->create_qp_ex(priv->sh->ctx,
2543                  &(struct ibv_qp_init_attr_ex){
2544                         .qp_type = IBV_QPT_RAW_PACKET,
2545                         .comp_mask =
2546                                 IBV_QP_INIT_ATTR_PD |
2547                                 IBV_QP_INIT_ATTR_IND_TABLE |
2548                                 IBV_QP_INIT_ATTR_RX_HASH,
2549                         .rx_hash_conf = (struct ibv_rx_hash_conf){
2550                                 .rx_hash_function =
2551                                         IBV_RX_HASH_FUNC_TOEPLITZ,
2552                                 .rx_hash_key_len = MLX5_RSS_HASH_KEY_LEN,
2553                                 .rx_hash_key = rss_hash_default_key,
2554                                 .rx_hash_fields_mask = 0,
2555                                 },
2556                         .rwq_ind_tbl = ind_tbl->ind_table,
2557                         .pd = priv->sh->pd
2558                  });
2559         if (!qp) {
2560                 DEBUG("port %u cannot allocate QP for drop queue",
2561                       dev->data->port_id);
2562                 rte_errno = errno;
2563                 goto error;
2564         }
2565         hrxq = rte_calloc(__func__, 1, sizeof(*hrxq), 0);
2566         if (!hrxq) {
2567                 DRV_LOG(WARNING,
2568                         "port %u cannot allocate memory for drop queue",
2569                         dev->data->port_id);
2570                 rte_errno = ENOMEM;
2571                 goto error;
2572         }
2573         hrxq->ind_table = ind_tbl;
2574         hrxq->qp = qp;
2575 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2576         hrxq->action = mlx5_glue->dv_create_flow_action_dest_ibv_qp(hrxq->qp);
2577         if (!hrxq->action) {
2578                 rte_errno = errno;
2579                 goto error;
2580         }
2581 #endif
2582         priv->drop_queue.hrxq = hrxq;
2583         rte_atomic32_set(&hrxq->refcnt, 1);
2584         return hrxq;
2585 error:
2586         if (ind_tbl)
2587                 mlx5_ind_table_obj_drop_release(dev);
2588         return NULL;
2589 }
2590
2591 /**
2592  * Release a drop hash Rx queue.
2593  *
2594  * @param dev
2595  *   Pointer to Ethernet device.
2596  */
2597 void
2598 mlx5_hrxq_drop_release(struct rte_eth_dev *dev)
2599 {
2600         struct mlx5_priv *priv = dev->data->dev_private;
2601         struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
2602
2603         if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
2604 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
2605                 mlx5_glue->destroy_flow_action(hrxq->action);
2606 #endif
2607                 claim_zero(mlx5_glue->destroy_qp(hrxq->qp));
2608                 mlx5_ind_table_obj_drop_release(dev);
2609                 rte_free(hrxq);
2610                 priv->drop_queue.hrxq = NULL;
2611         }
2612 }