ceph/src/spdk/dpdk/drivers/net/mlx4/mlx4_rxq.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2017 6WIND S.A.
   3  * Copyright 2017 Mellanox Technologies, Ltd
   4  */
   5
   6 /**
   7  * @file
   8  * Rx queues configuration for mlx4 driver.
   9  */
  10
  11 #include <assert.h>
  12 #include <errno.h>
  13 #include <stddef.h>
  14 #include <stdint.h>
  15 #include <string.h>
  16
  17 /* Verbs headers do not support -pedantic. */
  18 #ifdef PEDANTIC
  19 #pragma GCC diagnostic ignored "-Wpedantic"
  20 #endif
  21 #include <infiniband/mlx4dv.h>
  22 #include <infiniband/verbs.h>
  23 #ifdef PEDANTIC
  24 #pragma GCC diagnostic error "-Wpedantic"
  25 #endif
  26
  27 #include <rte_byteorder.h>
  28 #include <rte_common.h>
  29 #include <rte_errno.h>
  30 #include <rte_ethdev_driver.h>
  31 #include <rte_flow.h>
  32 #include <rte_malloc.h>
  33 #include <rte_mbuf.h>
  34 #include <rte_mempool.h>
  35
  36 #include "mlx4.h"
  37 #include "mlx4_glue.h"
  38 #include "mlx4_flow.h"
  39 #include "mlx4_rxtx.h"
  40 #include "mlx4_utils.h"
  41
  42 /**
  43  * Historical RSS hash key.
  44  *
  45  * This used to be the default for mlx4 in Linux before v3.19 switched to
  46  * generating random hash keys through netdev_rss_key_fill().
  47  *
  48  * It is used in this PMD for consistency with past DPDK releases but can
  49  * now be overridden through user configuration.
  50  *
  51  * Note: this is not const to work around API quirks.
  52  */
  53 uint8_t
  54 mlx4_rss_hash_key_default[MLX4_RSS_HASH_KEY_SIZE] = {
  55         0x2c, 0xc6, 0x81, 0xd1,
  56         0x5b, 0xdb, 0xf4, 0xf7,
  57         0xfc, 0xa2, 0x83, 0x19,
  58         0xdb, 0x1a, 0x3e, 0x94,
  59         0x6b, 0x9e, 0x38, 0xd9,
  60         0x2c, 0x9c, 0x03, 0xd1,
  61         0xad, 0x99, 0x44, 0xa7,
  62         0xd9, 0x56, 0x3d, 0x59,
  63         0x06, 0x3c, 0x25, 0xf3,
  64         0xfc, 0x1f, 0xdc, 0x2a,
  65 };
  66
  67 /**
  68  * Obtain a RSS context with specified properties.
  69  *
  70  * Used when creating a flow rule targeting one or several Rx queues.
  71  *
  72  * If a matching RSS context already exists, it is returned with its
  73  * reference count incremented.
  74  *
  75  * @param priv
  76  *   Pointer to private structure.
  77  * @param fields
  78  *   Fields for RSS processing (Verbs format).
  79  * @param[in] key
  80  *   Hash key to use (whose size is exactly MLX4_RSS_HASH_KEY_SIZE).
  81  * @param queues
  82  *   Number of target queues.
  83  * @param[in] queue_id
  84  *   Target queues.
  85  *
  86  * @return
  87  *   Pointer to RSS context on success, NULL otherwise and rte_errno is set.
  88  */
  89 struct mlx4_rss *
  90 mlx4_rss_get(struct priv *priv, uint64_t fields,
  91              const uint8_t key[MLX4_RSS_HASH_KEY_SIZE],
  92              uint16_t queues, const uint16_t queue_id[])
  93 {
  94         struct mlx4_rss *rss;
  95         size_t queue_id_size = sizeof(queue_id[0]) * queues;
  96
  97         LIST_FOREACH(rss, &priv->rss, next)
  98                 if (fields == rss->fields &&
  99                     queues == rss->queues &&
 100                     !memcmp(key, rss->key, MLX4_RSS_HASH_KEY_SIZE) &&
 101                     !memcmp(queue_id, rss->queue_id, queue_id_size)) {
 102                         ++rss->refcnt;
 103                         return rss;
 104                 }
 105         rss = rte_malloc(__func__, offsetof(struct mlx4_rss, queue_id) +
 106                          queue_id_size, 0);
 107         if (!rss)
 108                 goto error;
 109         *rss = (struct mlx4_rss){
 110                 .priv = priv,
 111                 .refcnt = 1,
 112                 .usecnt = 0,
 113                 .qp = NULL,
 114                 .ind = NULL,
 115                 .fields = fields,
 116                 .queues = queues,
 117         };
 118         memcpy(rss->key, key, MLX4_RSS_HASH_KEY_SIZE);
 119         memcpy(rss->queue_id, queue_id, queue_id_size);
 120         LIST_INSERT_HEAD(&priv->rss, rss, next);
 121         return rss;
 122 error:
 123         rte_errno = ENOMEM;
 124         return NULL;
 125 }
 126
 127 /**
 128  * Release a RSS context instance.
 129  *
 130  * Used when destroying a flow rule targeting one or several Rx queues.
 131  *
 132  * This function decrements the reference count of the context and destroys
 133  * it after reaching 0. The context must have no users at this point; all
 134  * prior calls to mlx4_rss_attach() must have been followed by matching
 135  * calls to mlx4_rss_detach().
 136  *
 137  * @param rss
 138  *   RSS context to release.
 139  */
 140 void
 141 mlx4_rss_put(struct mlx4_rss *rss)
 142 {
 143         assert(rss->refcnt);
 144         if (--rss->refcnt)
 145                 return;
 146         assert(!rss->usecnt);
 147         assert(!rss->qp);
 148         assert(!rss->ind);
 149         LIST_REMOVE(rss, next);
 150         rte_free(rss);
 151 }
 152
 153 /**
 154  * Attach a user to a RSS context instance.
 155  *
 156  * Used when the RSS QP and indirection table objects must be instantiated,
 157  * that is, when a flow rule must be enabled.
 158  *
 159  * This function increments the usage count of the context.
 160  *
 161  * @param rss
 162  *   RSS context to attach to.
 163  *
 164  * @return
 165  *   0 on success, a negative errno value otherwise and rte_errno is set.
 166  */
 167 int
 168 mlx4_rss_attach(struct mlx4_rss *rss)
 169 {
 170         assert(rss->refcnt);
 171         if (rss->usecnt++) {
 172                 assert(rss->qp);
 173                 assert(rss->ind);
 174                 return 0;
 175         }
 176
 177         struct ibv_wq *ind_tbl[rss->queues];
 178         struct priv *priv = rss->priv;
 179         const char *msg;
 180         unsigned int i = 0;
 181         int ret;
 182
 183         if (!rte_is_power_of_2(RTE_DIM(ind_tbl))) {
 184                 ret = EINVAL;
 185                 msg = "number of RSS queues must be a power of two";
 186                 goto error;
 187         }
 188         for (i = 0; i != RTE_DIM(ind_tbl); ++i) {
 189                 uint16_t id = rss->queue_id[i];
 190                 struct rxq *rxq = NULL;
 191
 192                 if (id < priv->dev->data->nb_rx_queues)
 193                         rxq = priv->dev->data->rx_queues[id];
 194                 if (!rxq) {
 195                         ret = EINVAL;
 196                         msg = "RSS target queue is not configured";
 197                         goto error;
 198                 }
 199                 ret = mlx4_rxq_attach(rxq);
 200                 if (ret) {
 201                         ret = -ret;
 202                         msg = "unable to attach RSS target queue";
 203                         goto error;
 204                 }
 205                 ind_tbl[i] = rxq->wq;
 206         }
 207         rss->ind = mlx4_glue->create_rwq_ind_table
 208                 (priv->ctx,
 209                  &(struct ibv_rwq_ind_table_init_attr){
 210                         .log_ind_tbl_size = rte_log2_u32(RTE_DIM(ind_tbl)),
 211                         .ind_tbl = ind_tbl,
 212                         .comp_mask = 0,
 213                  });
 214         if (!rss->ind) {
 215                 ret = errno ? errno : EINVAL;
 216                 msg = "RSS indirection table creation failure";
 217                 goto error;
 218         }
 219         rss->qp = mlx4_glue->create_qp_ex
 220                 (priv->ctx,
 221                  &(struct ibv_qp_init_attr_ex){
 222                         .comp_mask = (IBV_QP_INIT_ATTR_PD |
 223                                       IBV_QP_INIT_ATTR_RX_HASH |
 224                                       IBV_QP_INIT_ATTR_IND_TABLE),
 225                         .qp_type = IBV_QPT_RAW_PACKET,
 226                         .pd = priv->pd,
 227                         .rwq_ind_tbl = rss->ind,
 228                         .rx_hash_conf = {
 229                                 .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
 230                                 .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
 231                                 .rx_hash_key = rss->key,
 232                                 .rx_hash_fields_mask = rss->fields,
 233                         },
 234                  });
 235         if (!rss->qp) {
 236                 ret = errno ? errno : EINVAL;
 237                 msg = "RSS hash QP creation failure";
 238                 goto error;
 239         }
 240         ret = mlx4_glue->modify_qp
 241                 (rss->qp,
 242                  &(struct ibv_qp_attr){
 243                         .qp_state = IBV_QPS_INIT,
 244                         .port_num = priv->port,
 245                  },
 246                  IBV_QP_STATE | IBV_QP_PORT);
 247         if (ret) {
 248                 msg = "failed to switch RSS hash QP to INIT state";
 249                 goto error;
 250         }
 251         ret = mlx4_glue->modify_qp
 252                 (rss->qp,
 253                  &(struct ibv_qp_attr){
 254                         .qp_state = IBV_QPS_RTR,
 255                  },
 256                  IBV_QP_STATE);
 257         if (ret) {
 258                 msg = "failed to switch RSS hash QP to RTR state";
 259                 goto error;
 260         }
 261         return 0;
 262 error:
 263         if (rss->qp) {
 264                 claim_zero(mlx4_glue->destroy_qp(rss->qp));
 265                 rss->qp = NULL;
 266         }
 267         if (rss->ind) {
 268                 claim_zero(mlx4_glue->destroy_rwq_ind_table(rss->ind));
 269                 rss->ind = NULL;
 270         }
 271         while (i--)
 272                 mlx4_rxq_detach(priv->dev->data->rx_queues[rss->queue_id[i]]);
 273         ERROR("mlx4: %s", msg);
 274         --rss->usecnt;
 275         rte_errno = ret;
 276         return -ret;
 277 }
 278
 279 /**
 280  * Detach a user from a RSS context instance.
 281  *
 282  * Used when disabling (not destroying) a flow rule.
 283  *
 284  * This function decrements the usage count of the context and destroys
 285  * usage resources after reaching 0.
 286  *
 287  * @param rss
 288  *   RSS context to detach from.
 289  */
 290 void
 291 mlx4_rss_detach(struct mlx4_rss *rss)
 292 {
 293         struct priv *priv = rss->priv;
 294         unsigned int i;
 295
 296         assert(rss->refcnt);
 297         assert(rss->qp);
 298         assert(rss->ind);
 299         if (--rss->usecnt)
 300                 return;
 301         claim_zero(mlx4_glue->destroy_qp(rss->qp));
 302         rss->qp = NULL;
 303         claim_zero(mlx4_glue->destroy_rwq_ind_table(rss->ind));
 304         rss->ind = NULL;
 305         for (i = 0; i != rss->queues; ++i)
 306                 mlx4_rxq_detach(priv->dev->data->rx_queues[rss->queue_id[i]]);
 307 }
 308
 309 /**
 310  * Initialize common RSS context resources.
 311  *
 312  * Because ConnectX-3 hardware limitations require a fixed order in the
 313  * indirection table, WQs must be allocated sequentially to be part of a
 314  * common RSS context.
 315  *
 316  * Since a newly created WQ cannot be moved to a different context, this
 317  * function allocates them all at once, one for each configured Rx queue,
 318  * as well as all related resources (CQs and mbufs).
 319  *
 320  * This must therefore be done before creating any Rx flow rules relying on
 321  * indirection tables.
 322  *
 323  * @param priv
 324  *   Pointer to private structure.
 325  *
 326  * @return
 327  *   0 on success, a negative errno value otherwise and rte_errno is set.
 328  */
 329 int
 330 mlx4_rss_init(struct priv *priv)
 331 {
 332         struct rte_eth_dev *dev = priv->dev;
 333         uint8_t log2_range = rte_log2_u32(dev->data->nb_rx_queues);
 334         uint32_t wq_num_prev = 0;
 335         const char *msg;
 336         unsigned int i;
 337         int ret;
 338
 339         if (priv->rss_init)
 340                 return 0;
 341         if (priv->dev->data->nb_rx_queues > priv->hw_rss_max_qps) {
 342                 ERROR("RSS does not support more than %d queues",
 343                       priv->hw_rss_max_qps);
 344                 rte_errno = EINVAL;
 345                 return -rte_errno;
 346         }
 347         /* Prepare range for RSS contexts before creating the first WQ. */
 348         ret = mlx4_glue->dv_set_context_attr
 349                 (priv->ctx,
 350                  MLX4DV_SET_CTX_ATTR_LOG_WQS_RANGE_SZ,
 351                  &log2_range);
 352         if (ret) {
 353                 ERROR("cannot set up range size for RSS context to %u"
 354                       " (for %u Rx queues), error: %s",
 355                       1 << log2_range, dev->data->nb_rx_queues, strerror(ret));
 356                 rte_errno = ret;
 357                 return -ret;
 358         }
 359         for (i = 0; i != priv->dev->data->nb_rx_queues; ++i) {
 360                 struct rxq *rxq = priv->dev->data->rx_queues[i];
 361                 struct ibv_cq *cq;
 362                 struct ibv_wq *wq;
 363                 uint32_t wq_num;
 364
 365                 /* Attach the configured Rx queues. */
 366                 if (rxq) {
 367                         assert(!rxq->usecnt);
 368                         ret = mlx4_rxq_attach(rxq);
 369                         if (!ret) {
 370                                 wq_num = rxq->wq->wq_num;
 371                                 goto wq_num_check;
 372                         }
 373                         ret = -ret;
 374                         msg = "unable to create Rx queue resources";
 375                         goto error;
 376                 }
 377                 /*
 378                  * WQs are temporarily allocated for unconfigured Rx queues
 379                  * to maintain proper index alignment in indirection table
 380                  * by skipping unused WQ numbers.
 381                  *
 382                  * The reason this works at all even though these WQs are
 383                  * immediately destroyed is that WQNs are allocated
 384                  * sequentially and are guaranteed to never be reused in the
 385                  * same context by the underlying implementation.
 386                  */
 387                 cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
 388                 if (!cq) {
 389                         ret = ENOMEM;
 390                         msg = "placeholder CQ creation failure";
 391                         goto error;
 392                 }
 393                 wq = mlx4_glue->create_wq
 394                         (priv->ctx,
 395                          &(struct ibv_wq_init_attr){
 396                                 .wq_type = IBV_WQT_RQ,
 397                                 .max_wr = 1,
 398                                 .max_sge = 1,
 399                                 .pd = priv->pd,
 400                                 .cq = cq,
 401                          });
 402                 if (wq) {
 403                         wq_num = wq->wq_num;
 404                         claim_zero(mlx4_glue->destroy_wq(wq));
 405                 } else {
 406                         wq_num = 0; /* Shut up GCC 4.8 warnings. */
 407                 }
 408                 claim_zero(mlx4_glue->destroy_cq(cq));
 409                 if (!wq) {
 410                         ret = ENOMEM;
 411                         msg = "placeholder WQ creation failure";
 412                         goto error;
 413                 }
 414 wq_num_check:
 415                 /*
 416                  * While guaranteed by the implementation, make sure WQ
 417                  * numbers are really sequential (as the saying goes,
 418                  * trust, but verify).
 419                  */
 420                 if (i && wq_num - wq_num_prev != 1) {
 421                         if (rxq)
 422                                 mlx4_rxq_detach(rxq);
 423                         ret = ERANGE;
 424                         msg = "WQ numbers are not sequential";
 425                         goto error;
 426                 }
 427                 wq_num_prev = wq_num;
 428         }
 429         priv->rss_init = 1;
 430         return 0;
 431 error:
 432         ERROR("cannot initialize common RSS resources (queue %u): %s: %s",
 433               i, msg, strerror(ret));
 434         while (i--) {
 435                 struct rxq *rxq = priv->dev->data->rx_queues[i];
 436
 437                 if (rxq)
 438                         mlx4_rxq_detach(rxq);
 439         }
 440         rte_errno = ret;
 441         return -ret;
 442 }
 443
 444 /**
 445  * Release common RSS context resources.
 446  *
 447  * As the reverse of mlx4_rss_init(), this must be done after removing all
 448  * flow rules relying on indirection tables.
 449  *
 450  * @param priv
 451  *   Pointer to private structure.
 452  */
 453 void
 454 mlx4_rss_deinit(struct priv *priv)
 455 {
 456         unsigned int i;
 457
 458         if (!priv->rss_init)
 459                 return;
 460         for (i = 0; i != priv->dev->data->nb_rx_queues; ++i) {
 461                 struct rxq *rxq = priv->dev->data->rx_queues[i];
 462
 463                 if (rxq) {
 464                         assert(rxq->usecnt == 1);
 465                         mlx4_rxq_detach(rxq);
 466                 }
 467         }
 468         priv->rss_init = 0;
 469 }
 470
 471 /**
 472  * Attach a user to a Rx queue.
 473  *
 474  * Used when the resources of an Rx queue must be instantiated for it to
 475  * become in a usable state.
 476  *
 477  * This function increments the usage count of the Rx queue.
 478  *
 479  * @param rxq
 480  *   Pointer to Rx queue structure.
 481  *
 482  * @return
 483  *   0 on success, negative errno value otherwise and rte_errno is set.
 484  */
 485 int
 486 mlx4_rxq_attach(struct rxq *rxq)
 487 {
 488         if (rxq->usecnt++) {
 489                 assert(rxq->cq);
 490                 assert(rxq->wq);
 491                 assert(rxq->wqes);
 492                 assert(rxq->rq_db);
 493                 return 0;
 494         }
 495
 496         struct priv *priv = rxq->priv;
 497         struct rte_eth_dev *dev = priv->dev;
 498         const uint32_t elts_n = 1 << rxq->elts_n;
 499         const uint32_t sges_n = 1 << rxq->sges_n;
 500         struct rte_mbuf *(*elts)[elts_n] = rxq->elts;
 501         struct mlx4dv_obj mlxdv;
 502         struct mlx4dv_rwq dv_rwq;
 503         struct mlx4dv_cq dv_cq = { .comp_mask = MLX4DV_CQ_MASK_UAR, };
 504         const char *msg;
 505         struct ibv_cq *cq = NULL;
 506         struct ibv_wq *wq = NULL;
 507         uint32_t create_flags = 0;
 508         uint32_t comp_mask = 0;
 509         volatile struct mlx4_wqe_data_seg (*wqes)[];
 510         unsigned int i;
 511         int ret;
 512
 513         assert(rte_is_power_of_2(elts_n));
 514         cq = mlx4_glue->create_cq(priv->ctx, elts_n / sges_n, NULL,
 515                                   rxq->channel, 0);
 516         if (!cq) {
 517                 ret = ENOMEM;
 518                 msg = "CQ creation failure";
 519                 goto error;
 520         }
 521         /* By default, FCS (CRC) is stripped by hardware. */
 522         if (rxq->crc_present) {
 523                 create_flags |= IBV_WQ_FLAGS_SCATTER_FCS;
 524                 comp_mask |= IBV_WQ_INIT_ATTR_FLAGS;
 525         }
 526         wq = mlx4_glue->create_wq
 527                 (priv->ctx,
 528                  &(struct ibv_wq_init_attr){
 529                         .wq_type = IBV_WQT_RQ,
 530                         .max_wr = elts_n / sges_n,
 531                         .max_sge = sges_n,
 532                         .pd = priv->pd,
 533                         .cq = cq,
 534                         .comp_mask = comp_mask,
 535                         .create_flags = create_flags,
 536                  });
 537         if (!wq) {
 538                 ret = errno ? errno : EINVAL;
 539                 msg = "WQ creation failure";
 540                 goto error;
 541         }
 542         ret = mlx4_glue->modify_wq
 543                 (wq,
 544                  &(struct ibv_wq_attr){
 545                         .attr_mask = IBV_WQ_ATTR_STATE,
 546                         .wq_state = IBV_WQS_RDY,
 547                  });
 548         if (ret) {
 549                 msg = "WQ state change to IBV_WQS_RDY failed";
 550                 goto error;
 551         }
 552         /* Retrieve device queue information. */
 553         mlxdv.cq.in = cq;
 554         mlxdv.cq.out = &dv_cq;
 555         mlxdv.rwq.in = wq;
 556         mlxdv.rwq.out = &dv_rwq;
 557         ret = mlx4_glue->dv_init_obj(&mlxdv, MLX4DV_OBJ_RWQ | MLX4DV_OBJ_CQ);
 558         if (ret) {
 559                 msg = "failed to obtain device information from WQ/CQ objects";
 560                 goto error;
 561         }
 562         /* Pre-register Rx mempool. */
 563         DEBUG("port %u Rx queue %u registering mp %s having %u chunks",
 564               priv->dev->data->port_id, rxq->stats.idx,
 565               rxq->mp->name, rxq->mp->nb_mem_chunks);
 566         mlx4_mr_update_mp(dev, &rxq->mr_ctrl, rxq->mp);
 567         wqes = (volatile struct mlx4_wqe_data_seg (*)[])
 568                 ((uintptr_t)dv_rwq.buf.buf + dv_rwq.rq.offset);
 569         for (i = 0; i != RTE_DIM(*elts); ++i) {
 570                 volatile struct mlx4_wqe_data_seg *scat = &(*wqes)[i];
 571                 struct rte_mbuf *buf = rte_pktmbuf_alloc(rxq->mp);
 572
 573                 if (buf == NULL) {
 574                         while (i--) {
 575                                 rte_pktmbuf_free_seg((*elts)[i]);
 576                                 (*elts)[i] = NULL;
 577                         }
 578                         ret = ENOMEM;
 579                         msg = "cannot allocate mbuf";
 580                         goto error;
 581                 }
 582                 /* Headroom is reserved by rte_pktmbuf_alloc(). */
 583                 assert(buf->data_off == RTE_PKTMBUF_HEADROOM);
 584                 /* Buffer is supposed to be empty. */
 585                 assert(rte_pktmbuf_data_len(buf) == 0);
 586                 assert(rte_pktmbuf_pkt_len(buf) == 0);
 587                 /* Only the first segment keeps headroom. */
 588                 if (i % sges_n)
 589                         buf->data_off = 0;
 590                 buf->port = rxq->port_id;
 591                 buf->data_len = rte_pktmbuf_tailroom(buf);
 592                 buf->pkt_len = rte_pktmbuf_tailroom(buf);
 593                 buf->nb_segs = 1;
 594                 *scat = (struct mlx4_wqe_data_seg){
 595                         .addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf,
 596                                                                   uintptr_t)),
 597                         .byte_count = rte_cpu_to_be_32(buf->data_len),
 598                         .lkey = mlx4_rx_mb2mr(rxq, buf),
 599                 };
 600                 (*elts)[i] = buf;
 601         }
 602         DEBUG("%p: allocated and configured %u segments (max %u packets)",
 603               (void *)rxq, elts_n, elts_n / sges_n);
 604         rxq->cq = cq;
 605         rxq->wq = wq;
 606         rxq->wqes = wqes;
 607         rxq->rq_db = dv_rwq.rdb;
 608         rxq->mcq.buf = dv_cq.buf.buf;
 609         rxq->mcq.cqe_cnt = dv_cq.cqe_cnt;
 610         rxq->mcq.set_ci_db = dv_cq.set_ci_db;
 611         rxq->mcq.cqe_64 = (dv_cq.cqe_size & 64) ? 1 : 0;
 612         rxq->mcq.arm_db = dv_cq.arm_db;
 613         rxq->mcq.arm_sn = dv_cq.arm_sn;
 614         rxq->mcq.cqn = dv_cq.cqn;
 615         rxq->mcq.cq_uar = dv_cq.cq_uar;
 616         rxq->mcq.cq_db_reg = (uint8_t *)dv_cq.cq_uar + MLX4_CQ_DOORBELL;
 617         /* Update doorbell counter. */
 618         rxq->rq_ci = elts_n / sges_n;
 619         rte_wmb();
 620         *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
 621         return 0;
 622 error:
 623         if (wq)
 624                 claim_zero(mlx4_glue->destroy_wq(wq));
 625         if (cq)
 626                 claim_zero(mlx4_glue->destroy_cq(cq));
 627         --rxq->usecnt;
 628         rte_errno = ret;
 629         ERROR("error while attaching Rx queue %p: %s: %s",
 630               (void *)rxq, msg, strerror(ret));
 631         return -ret;
 632 }
 633
 634 /**
 635  * Detach a user from a Rx queue.
 636  *
 637  * This function decrements the usage count of the Rx queue and destroys
 638  * usage resources after reaching 0.
 639  *
 640  * @param rxq
 641  *   Pointer to Rx queue structure.
 642  */
 643 void
 644 mlx4_rxq_detach(struct rxq *rxq)
 645 {
 646         unsigned int i;
 647         struct rte_mbuf *(*elts)[1 << rxq->elts_n] = rxq->elts;
 648
 649         if (--rxq->usecnt)
 650                 return;
 651         rxq->rq_ci = 0;
 652         memset(&rxq->mcq, 0, sizeof(rxq->mcq));
 653         rxq->rq_db = NULL;
 654         rxq->wqes = NULL;
 655         claim_zero(mlx4_glue->destroy_wq(rxq->wq));
 656         rxq->wq = NULL;
 657         claim_zero(mlx4_glue->destroy_cq(rxq->cq));
 658         rxq->cq = NULL;
 659         DEBUG("%p: freeing Rx queue elements", (void *)rxq);
 660         for (i = 0; (i != RTE_DIM(*elts)); ++i) {
 661                 if (!(*elts)[i])
 662                         continue;
 663                 rte_pktmbuf_free_seg((*elts)[i]);
 664                 (*elts)[i] = NULL;
 665         }
 666 }
 667
 668 /**
 669  * Returns the per-queue supported offloads.
 670  *
 671  * @param priv
 672  *   Pointer to private structure.
 673  *
 674  * @return
 675  *   Supported Tx offloads.
 676  */
 677 uint64_t
 678 mlx4_get_rx_queue_offloads(struct priv *priv)
 679 {
 680         uint64_t offloads = DEV_RX_OFFLOAD_SCATTER |
 681                             DEV_RX_OFFLOAD_CRC_STRIP |
 682                             DEV_RX_OFFLOAD_KEEP_CRC |
 683                             DEV_RX_OFFLOAD_JUMBO_FRAME;
 684
 685         if (priv->hw_csum)
 686                 offloads |= DEV_RX_OFFLOAD_CHECKSUM;
 687         return offloads;
 688 }
 689
 690 /**
 691  * Returns the per-port supported offloads.
 692  *
 693  * @param priv
 694  *   Pointer to private structure.
 695  *
 696  * @return
 697  *   Supported Rx offloads.
 698  */
 699 uint64_t
 700 mlx4_get_rx_port_offloads(struct priv *priv)
 701 {
 702         uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
 703
 704         (void)priv;
 705         return offloads;
 706 }
 707
 708 /**
 709  * DPDK callback to configure a Rx queue.
 710  *
 711  * @param dev
 712  *   Pointer to Ethernet device structure.
 713  * @param idx
 714  *   Rx queue index.
 715  * @param desc
 716  *   Number of descriptors to configure in queue.
 717  * @param socket
 718  *   NUMA socket on which memory must be allocated.
 719  * @param[in] conf
 720  *   Thresholds parameters.
 721  * @param mp
 722  *   Memory pool for buffer allocations.
 723  *
 724  * @return
 725  *   0 on success, negative errno value otherwise and rte_errno is set.
 726  */
 727 int
 728 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 729                     unsigned int socket, const struct rte_eth_rxconf *conf,
 730                     struct rte_mempool *mp)
 731 {
 732         struct priv *priv = dev->data->dev_private;
 733         uint32_t mb_len = rte_pktmbuf_data_room_size(mp);
 734         struct rte_mbuf *(*elts)[rte_align32pow2(desc)];
 735         struct rxq *rxq;
 736         struct mlx4_malloc_vec vec[] = {
 737                 {
 738                         .align = RTE_CACHE_LINE_SIZE,
 739                         .size = sizeof(*rxq),
 740                         .addr = (void **)&rxq,
 741                 },
 742                 {
 743                         .align = RTE_CACHE_LINE_SIZE,
 744                         .size = sizeof(*elts),
 745                         .addr = (void **)&elts,
 746                 },
 747         };
 748         int ret;
 749         uint32_t crc_present;
 750         uint64_t offloads;
 751
 752         offloads = conf->offloads | dev->data->dev_conf.rxmode.offloads;
 753
 754         DEBUG("%p: configuring queue %u for %u descriptors",
 755               (void *)dev, idx, desc);
 756
 757         if (idx >= dev->data->nb_rx_queues) {
 758                 rte_errno = EOVERFLOW;
 759                 ERROR("%p: queue index out of range (%u >= %u)",
 760                       (void *)dev, idx, dev->data->nb_rx_queues);
 761                 return -rte_errno;
 762         }
 763         rxq = dev->data->rx_queues[idx];
 764         if (rxq) {
 765                 rte_errno = EEXIST;
 766                 ERROR("%p: Rx queue %u already configured, release it first",
 767                       (void *)dev, idx);
 768                 return -rte_errno;
 769         }
 770         if (!desc) {
 771                 rte_errno = EINVAL;
 772                 ERROR("%p: invalid number of Rx descriptors", (void *)dev);
 773                 return -rte_errno;
 774         }
 775         if (desc != RTE_DIM(*elts)) {
 776                 desc = RTE_DIM(*elts);
 777                 WARN("%p: increased number of descriptors in Rx queue %u"
 778                      " to the next power of two (%u)",
 779                      (void *)dev, idx, desc);
 780         }
 781         /* By default, FCS (CRC) is stripped by hardware. */
 782         crc_present = 0;
 783         if (rte_eth_dev_must_keep_crc(offloads)) {
 784                 if (priv->hw_fcs_strip) {
 785                         crc_present = 1;
 786                 } else {
 787                         WARN("%p: CRC stripping has been disabled but will still"
 788                              " be performed by hardware, make sure MLNX_OFED and"
 789                              " firmware are up to date",
 790                              (void *)dev);
 791                 }
 792         }
 793         DEBUG("%p: CRC stripping is %s, %u bytes will be subtracted from"
 794               " incoming frames to hide it",
 795               (void *)dev,
 796               crc_present ? "disabled" : "enabled",
 797               crc_present << 2);
 798         /* Allocate and initialize Rx queue. */
 799         mlx4_zmallocv_socket("RXQ", vec, RTE_DIM(vec), socket);
 800         if (!rxq) {
 801                 ERROR("%p: unable to allocate queue index %u",
 802                       (void *)dev, idx);
 803                 return -rte_errno;
 804         }
 805         *rxq = (struct rxq){
 806                 .priv = priv,
 807                 .mp = mp,
 808                 .port_id = dev->data->port_id,
 809                 .sges_n = 0,
 810                 .elts_n = rte_log2_u32(desc),
 811                 .elts = elts,
 812                 /* Toggle Rx checksum offload if hardware supports it. */
 813                 .csum = priv->hw_csum &&
 814                         (offloads & DEV_RX_OFFLOAD_CHECKSUM),
 815                 .csum_l2tun = priv->hw_csum_l2tun &&
 816                               (offloads & DEV_RX_OFFLOAD_CHECKSUM),
 817                 .crc_present = crc_present,
 818                 .l2tun_offload = priv->hw_csum_l2tun,
 819                 .stats = {
 820                         .idx = idx,
 821                 },
 822                 .socket = socket,
 823         };
 824         /* Enable scattered packets support for this queue if necessary. */
 825         assert(mb_len >= RTE_PKTMBUF_HEADROOM);
 826         if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
 827             (mb_len - RTE_PKTMBUF_HEADROOM)) {
 828                 ;
 829         } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
 830                 uint32_t size =
 831                         RTE_PKTMBUF_HEADROOM +
 832                         dev->data->dev_conf.rxmode.max_rx_pkt_len;
 833                 uint32_t sges_n;
 834
 835                 /*
 836                  * Determine the number of SGEs needed for a full packet
 837                  * and round it to the next power of two.
 838                  */
 839                 sges_n = rte_log2_u32((size / mb_len) + !!(size % mb_len));
 840                 rxq->sges_n = sges_n;
 841                 /* Make sure sges_n did not overflow. */
 842                 size = mb_len * (1 << rxq->sges_n);
 843                 size -= RTE_PKTMBUF_HEADROOM;
 844                 if (size < dev->data->dev_conf.rxmode.max_rx_pkt_len) {
 845                         rte_errno = EOVERFLOW;
 846                         ERROR("%p: too many SGEs (%u) needed to handle"
 847                               " requested maximum packet size %u",
 848                               (void *)dev,
 849                               1 << sges_n,
 850                               dev->data->dev_conf.rxmode.max_rx_pkt_len);
 851                         goto error;
 852                 }
 853         } else {
 854                 WARN("%p: the requested maximum Rx packet size (%u) is"
 855                      " larger than a single mbuf (%u) and scattered"
 856                      " mode has not been requested",
 857                      (void *)dev,
 858                      dev->data->dev_conf.rxmode.max_rx_pkt_len,
 859                      mb_len - RTE_PKTMBUF_HEADROOM);
 860         }
 861         DEBUG("%p: maximum number of segments per packet: %u",
 862               (void *)dev, 1 << rxq->sges_n);
 863         if (desc % (1 << rxq->sges_n)) {
 864                 rte_errno = EINVAL;
 865                 ERROR("%p: number of Rx queue descriptors (%u) is not a"
 866                       " multiple of maximum segments per packet (%u)",
 867                       (void *)dev,
 868                       desc,
 869                       1 << rxq->sges_n);
 870                 goto error;
 871         }
 872         if (mlx4_mr_btree_init(&rxq->mr_ctrl.cache_bh,
 873                                MLX4_MR_BTREE_CACHE_N, socket)) {
 874                 /* rte_errno is already set. */
 875                 goto error;
 876         }
 877         if (dev->data->dev_conf.intr_conf.rxq) {
 878                 rxq->channel = mlx4_glue->create_comp_channel(priv->ctx);
 879                 if (rxq->channel == NULL) {
 880                         rte_errno = ENOMEM;
 881                         ERROR("%p: Rx interrupt completion channel creation"
 882                               " failure: %s",
 883                               (void *)dev, strerror(rte_errno));
 884                         goto error;
 885                 }
 886                 if (mlx4_fd_set_non_blocking(rxq->channel->fd) < 0) {
 887                         ERROR("%p: unable to make Rx interrupt completion"
 888                               " channel non-blocking: %s",
 889                               (void *)dev, strerror(rte_errno));
 890                         goto error;
 891                 }
 892         }
 893         DEBUG("%p: adding Rx queue %p to list", (void *)dev, (void *)rxq);
 894         dev->data->rx_queues[idx] = rxq;
 895         return 0;
 896 error:
 897         dev->data->rx_queues[idx] = NULL;
 898         ret = rte_errno;
 899         mlx4_rx_queue_release(rxq);
 900         rte_errno = ret;
 901         assert(rte_errno > 0);
 902         return -rte_errno;
 903 }
 904
 905 /**
 906  * DPDK callback to release a Rx queue.
 907  *
 908  * @param dpdk_rxq
 909  *   Generic Rx queue pointer.
 910  */
 911 void
 912 mlx4_rx_queue_release(void *dpdk_rxq)
 913 {
 914         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 915         struct priv *priv;
 916         unsigned int i;
 917
 918         if (rxq == NULL)
 919                 return;
 920         priv = rxq->priv;
 921         for (i = 0; i != priv->dev->data->nb_rx_queues; ++i)
 922                 if (priv->dev->data->rx_queues[i] == rxq) {
 923                         DEBUG("%p: removing Rx queue %p from list",
 924                               (void *)priv->dev, (void *)rxq);
 925                         priv->dev->data->rx_queues[i] = NULL;
 926                         break;
 927                 }
 928         assert(!rxq->cq);
 929         assert(!rxq->wq);
 930         assert(!rxq->wqes);
 931         assert(!rxq->rq_db);
 932         if (rxq->channel)
 933                 claim_zero(mlx4_glue->destroy_comp_channel(rxq->channel));
 934         mlx4_mr_btree_free(&rxq->mr_ctrl.cache_bh);
 935         rte_free(rxq);
 936 }