lib/librte_pmd_e1000/igb_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2013 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  *
  33  */
  34
  35 #include <sys/queue.h>
  36
  37 #include <endian.h>
  38 #include <stdio.h>
  39 #include <stdlib.h>
  40 #include <string.h>
  41 #include <errno.h>
  42 #include <stdint.h>
  43 #include <stdarg.h>
  44 #include <inttypes.h>
  45
  46 #include <rte_interrupts.h>
  47 #include <rte_byteorder.h>
  48 #include <rte_common.h>
  49 #include <rte_log.h>
  50 #include <rte_debug.h>
  51 #include <rte_pci.h>
  52 #include <rte_memory.h>
  53 #include <rte_memcpy.h>
  54 #include <rte_memzone.h>
  55 #include <rte_launch.h>
  56 #include <rte_tailq.h>
  57 #include <rte_eal.h>
  58 #include <rte_per_lcore.h>
  59 #include <rte_lcore.h>
  60 #include <rte_atomic.h>
  61 #include <rte_branch_prediction.h>
  62 #include <rte_ring.h>
  63 #include <rte_mempool.h>
  64 #include <rte_malloc.h>
  65 #include <rte_mbuf.h>
  66 #include <rte_ether.h>
  67 #include <rte_ethdev.h>
  68 #include <rte_prefetch.h>
  69 #include <rte_udp.h>
  70 #include <rte_tcp.h>
  71 #include <rte_sctp.h>
  72 #include <rte_string_fns.h>
  73
  74 #include "e1000_logs.h"
  75 #include "e1000/e1000_api.h"
  76 #include "e1000_ethdev.h"
  77
  78 static inline struct rte_mbuf *
  79 rte_rxmbuf_alloc(struct rte_mempool *mp)
  80 {
  81         struct rte_mbuf *m;
  82
  83         m = __rte_mbuf_raw_alloc(mp);
  84         __rte_mbuf_sanity_check_raw(m, RTE_MBUF_PKT, 0);
  85         return (m);
  86 }
  87
  88 #define RTE_MBUF_DATA_DMA_ADDR(mb) \
  89         (uint64_t) ((mb)->buf_physaddr +                   \
  90                         (uint64_t) ((char *)((mb)->pkt.data) -     \
  91                                 (char *)(mb)->buf_addr))
  92
  93 #define RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mb) \
  94         (uint64_t) ((mb)->buf_physaddr + RTE_PKTMBUF_HEADROOM)
  95
  96 /**
  97  * Structure associated with each descriptor of the RX ring of a RX queue.
  98  */
  99 struct igb_rx_entry {
 100         struct rte_mbuf *mbuf; /**< mbuf associated with RX descriptor. */
 101 };
 102
 103 /**
 104  * Structure associated with each descriptor of the TX ring of a TX queue.
 105  */
 106 struct igb_tx_entry {
 107         struct rte_mbuf *mbuf; /**< mbuf associated with TX desc, if any. */
 108         uint16_t next_id; /**< Index of next descriptor in ring. */
 109         uint16_t last_id; /**< Index of last scattered descriptor. */
 110 };
 111
 112 /**
 113  * Structure associated with each RX queue.
 114  */
 115 struct igb_rx_queue {
 116         struct rte_mempool  *mb_pool;   /**< mbuf pool to populate RX ring. */
 117         volatile union e1000_adv_rx_desc *rx_ring; /**< RX ring virtual address. */
 118         uint64_t            rx_ring_phys_addr; /**< RX ring DMA address. */
 119         volatile uint32_t   *rdt_reg_addr; /**< RDT register address. */
 120         volatile uint32_t   *rdh_reg_addr; /**< RDH register address. */
 121         struct igb_rx_entry *sw_ring;   /**< address of RX software ring. */
 122         struct rte_mbuf *pkt_first_seg; /**< First segment of current packet. */
 123         struct rte_mbuf *pkt_last_seg;  /**< Last segment of current packet. */
 124         uint16_t            nb_rx_desc; /**< number of RX descriptors. */
 125         uint16_t            rx_tail;    /**< current value of RDT register. */
 126         uint16_t            nb_rx_hold; /**< number of held free RX desc. */
 127         uint16_t            rx_free_thresh; /**< max free RX desc to hold. */
 128         uint16_t            queue_id;   /**< RX queue index. */
 129         uint16_t            reg_idx;    /**< RX queue register index. */
 130         uint8_t             port_id;    /**< Device port identifier. */
 131         uint8_t             pthresh;    /**< Prefetch threshold register. */
 132         uint8_t             hthresh;    /**< Host threshold register. */
 133         uint8_t             wthresh;    /**< Write-back threshold register. */
 134         uint8_t             crc_len;    /**< 0 if CRC stripped, 4 otherwise. */
 135         uint8_t             drop_en;  /**< If not 0, set SRRCTL.Drop_En. */
 136 };
 137
 138 /**
 139  * Hardware context number
 140  */
 141 enum igb_advctx_num {
 142         IGB_CTX_0    = 0, /**< CTX0    */
 143         IGB_CTX_1    = 1, /**< CTX1    */
 144         IGB_CTX_NUM  = 2, /**< CTX_NUM */
 145 };
 146
 147 /**
 148  * Strucutre to check if new context need be built
 149  */
 150 struct igb_advctx_info {
 151         uint16_t flags;           /**< ol_flags related to context build. */
 152         uint32_t cmp_mask;        /**< compare mask for vlan_macip_lens */
 153         union rte_vlan_macip vlan_macip_lens; /**< vlan, mac & ip length. */
 154 };
 155
 156 /**
 157  * Structure associated with each TX queue.
 158  */
 159 struct igb_tx_queue {
 160         volatile union e1000_adv_tx_desc *tx_ring; /**< TX ring address */
 161         uint64_t               tx_ring_phys_addr; /**< TX ring DMA address. */
 162         struct igb_tx_entry    *sw_ring; /**< virtual address of SW ring. */
 163         volatile uint32_t      *tdt_reg_addr; /**< Address of TDT register. */
 164         uint32_t               txd_type;      /**< Device-specific TXD type */
 165         uint16_t               nb_tx_desc;    /**< number of TX descriptors. */
 166         uint16_t               tx_tail; /**< Current value of TDT register. */
 167         uint16_t               tx_head;
 168         /**< Index of first used TX descriptor. */
 169         uint16_t               queue_id; /**< TX queue index. */
 170         uint16_t               reg_idx;  /**< TX queue register index. */
 171         uint8_t                port_id;  /**< Device port identifier. */
 172         uint8_t                pthresh;  /**< Prefetch threshold register. */
 173         uint8_t                hthresh;  /**< Host threshold register. */
 174         uint8_t                wthresh;  /**< Write-back threshold register. */
 175         uint32_t               ctx_curr;
 176         /**< Current used hardware descriptor. */
 177         uint32_t               ctx_start;
 178         /**< Start context position for transmit queue. */
 179         struct igb_advctx_info ctx_cache[IGB_CTX_NUM];
 180         /**< Hardware context history.*/
 181 };
 182
 183 #if 1
 184 #define RTE_PMD_USE_PREFETCH
 185 #endif
 186
 187 #ifdef RTE_PMD_USE_PREFETCH
 188 #define rte_igb_prefetch(p)     rte_prefetch0(p)
 189 #else
 190 #define rte_igb_prefetch(p)     do {} while(0)
 191 #endif
 192
 193 #ifdef RTE_PMD_PACKET_PREFETCH
 194 #define rte_packet_prefetch(p) rte_prefetch1(p)
 195 #else
 196 #define rte_packet_prefetch(p)  do {} while(0)
 197 #endif
 198
 199 /*********************************************************************
 200  *
 201  *  TX function
 202  *
 203  **********************************************************************/
 204
 205 /*
 206  * Advanced context descriptor are almost same between igb/ixgbe
 207  * This is a separate function, looking for optimization opportunity here
 208  * Rework required to go with the pre-defined values.
 209  */
 210
 211 static inline void
 212 igbe_set_xmit_ctx(struct igb_tx_queue* txq,
 213                 volatile struct e1000_adv_tx_context_desc *ctx_txd,
 214                 uint16_t ol_flags, uint32_t vlan_macip_lens)
 215 {
 216         uint32_t type_tucmd_mlhl;
 217         uint32_t mss_l4len_idx;
 218         uint32_t ctx_idx, ctx_curr;
 219         uint32_t cmp_mask;
 220
 221         ctx_curr = txq->ctx_curr;
 222         ctx_idx = ctx_curr + txq->ctx_start;
 223
 224         cmp_mask = 0;
 225         type_tucmd_mlhl = 0;
 226
 227         if (ol_flags & PKT_TX_VLAN_PKT) {
 228                 cmp_mask |= TX_VLAN_CMP_MASK;
 229         }
 230
 231         if (ol_flags & PKT_TX_IP_CKSUM) {
 232                 type_tucmd_mlhl = E1000_ADVTXD_TUCMD_IPV4;
 233                 cmp_mask |= TX_MAC_LEN_CMP_MASK;
 234         }
 235
 236         /* Specify which HW CTX to upload. */
 237         mss_l4len_idx = (ctx_idx << E1000_ADVTXD_IDX_SHIFT);
 238         switch (ol_flags & PKT_TX_L4_MASK) {
 239         case PKT_TX_UDP_CKSUM:
 240                 type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_UDP |
 241                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
 242                 mss_l4len_idx |= sizeof(struct udp_hdr) << E1000_ADVTXD_L4LEN_SHIFT;
 243                 cmp_mask |= TX_MACIP_LEN_CMP_MASK;
 244                 break;
 245         case PKT_TX_TCP_CKSUM:
 246                 type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_TCP |
 247                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
 248                 mss_l4len_idx |= sizeof(struct tcp_hdr) << E1000_ADVTXD_L4LEN_SHIFT;
 249                 cmp_mask |= TX_MACIP_LEN_CMP_MASK;
 250                 break;
 251         case PKT_TX_SCTP_CKSUM:
 252                 type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_SCTP |
 253                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
 254                 mss_l4len_idx |= sizeof(struct sctp_hdr) << E1000_ADVTXD_L4LEN_SHIFT;
 255                 cmp_mask |= TX_MACIP_LEN_CMP_MASK;
 256                 break;
 257         default:
 258                 type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_RSV |
 259                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
 260                 break;
 261         }
 262
 263         txq->ctx_cache[ctx_curr].flags           = ol_flags;
 264         txq->ctx_cache[ctx_curr].cmp_mask        = cmp_mask;
 265         txq->ctx_cache[ctx_curr].vlan_macip_lens.data =
 266                 vlan_macip_lens & cmp_mask;
 267
 268         ctx_txd->type_tucmd_mlhl = rte_cpu_to_le_32(type_tucmd_mlhl);
 269         ctx_txd->vlan_macip_lens = rte_cpu_to_le_32(vlan_macip_lens);
 270         ctx_txd->mss_l4len_idx   = rte_cpu_to_le_32(mss_l4len_idx);
 271         ctx_txd->seqnum_seed     = 0;
 272 }
 273
 274 /*
 275  * Check which hardware context can be used. Use the existing match
 276  * or create a new context descriptor.
 277  */
 278 static inline uint32_t
 279 what_advctx_update(struct igb_tx_queue *txq, uint16_t flags,
 280                 uint32_t vlan_macip_lens)
 281 {
 282         /* If match with the current context */
 283         if (likely((txq->ctx_cache[txq->ctx_curr].flags == flags) &&
 284                 (txq->ctx_cache[txq->ctx_curr].vlan_macip_lens.data ==
 285                 (txq->ctx_cache[txq->ctx_curr].cmp_mask & vlan_macip_lens)))) {
 286                         return txq->ctx_curr;
 287         }
 288
 289         /* If match with the second context */
 290         txq->ctx_curr ^= 1;
 291         if (likely((txq->ctx_cache[txq->ctx_curr].flags == flags) &&
 292                 (txq->ctx_cache[txq->ctx_curr].vlan_macip_lens.data ==
 293                 (txq->ctx_cache[txq->ctx_curr].cmp_mask & vlan_macip_lens)))) {
 294                         return txq->ctx_curr;
 295         }
 296
 297         /* Mismatch, use the previous context */
 298         return (IGB_CTX_NUM);
 299 }
 300
 301 static inline uint32_t
 302 tx_desc_cksum_flags_to_olinfo(uint16_t ol_flags)
 303 {
 304         static const uint32_t l4_olinfo[2] = {0, E1000_ADVTXD_POPTS_TXSM};
 305         static const uint32_t l3_olinfo[2] = {0, E1000_ADVTXD_POPTS_IXSM};
 306         uint32_t tmp;
 307
 308         tmp  = l4_olinfo[(ol_flags & PKT_TX_L4_MASK)  != PKT_TX_L4_NO_CKSUM];
 309         tmp |= l3_olinfo[(ol_flags & PKT_TX_IP_CKSUM) != 0];
 310         return tmp;
 311 }
 312
 313 static inline uint32_t
 314 tx_desc_vlan_flags_to_cmdtype(uint16_t ol_flags)
 315 {
 316         static uint32_t vlan_cmd[2] = {0, E1000_ADVTXD_DCMD_VLE};
 317         return vlan_cmd[(ol_flags & PKT_TX_VLAN_PKT) != 0];
 318 }
 319
 320 uint16_t
 321 eth_igb_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 322                uint16_t nb_pkts)
 323 {
 324         struct igb_tx_queue *txq;
 325         struct igb_tx_entry *sw_ring;
 326         struct igb_tx_entry *txe, *txn;
 327         volatile union e1000_adv_tx_desc *txr;
 328         volatile union e1000_adv_tx_desc *txd;
 329         struct rte_mbuf     *tx_pkt;
 330         struct rte_mbuf     *m_seg;
 331         uint64_t buf_dma_addr;
 332         uint32_t olinfo_status;
 333         uint32_t cmd_type_len;
 334         uint32_t pkt_len;
 335         uint16_t slen;
 336         uint16_t ol_flags;
 337         uint16_t tx_end;
 338         uint16_t tx_id;
 339         uint16_t tx_last;
 340         uint16_t nb_tx;
 341         uint16_t tx_ol_req;
 342         uint32_t new_ctx = 0;
 343         uint32_t ctx = 0;
 344         uint32_t vlan_macip_lens;
 345
 346         txq = tx_queue;
 347         sw_ring = txq->sw_ring;
 348         txr     = txq->tx_ring;
 349         tx_id   = txq->tx_tail;
 350         txe = &sw_ring[tx_id];
 351
 352         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 353                 tx_pkt = *tx_pkts++;
 354                 pkt_len = tx_pkt->pkt.pkt_len;
 355
 356                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
 357
 358                 /*
 359                  * The number of descriptors that must be allocated for a
 360                  * packet is the number of segments of that packet, plus 1
 361                  * Context Descriptor for the VLAN Tag Identifier, if any.
 362                  * Determine the last TX descriptor to allocate in the TX ring
 363                  * for the packet, starting from the current position (tx_id)
 364                  * in the ring.
 365                  */
 366                 tx_last = (uint16_t) (tx_id + tx_pkt->pkt.nb_segs - 1);
 367
 368                 ol_flags = tx_pkt->ol_flags;
 369                 vlan_macip_lens = tx_pkt->pkt.vlan_macip.data;
 370                 tx_ol_req = (uint16_t)(ol_flags & PKT_TX_OFFLOAD_MASK);
 371
 372                 /* If a Context Descriptor need be built . */
 373                 if (tx_ol_req) {
 374                         ctx = what_advctx_update(txq, tx_ol_req,
 375                                 vlan_macip_lens);
 376                         /* Only allocate context descriptor if required*/
 377                         new_ctx = (ctx == IGB_CTX_NUM);
 378                         ctx = txq->ctx_curr;
 379                         tx_last = (uint16_t) (tx_last + new_ctx);
 380                 }
 381                 if (tx_last >= txq->nb_tx_desc)
 382                         tx_last = (uint16_t) (tx_last - txq->nb_tx_desc);
 383
 384                 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u pktlen=%u"
 385                            " tx_first=%u tx_last=%u\n",
 386                            (unsigned) txq->port_id,
 387                            (unsigned) txq->queue_id,
 388                            (unsigned) pkt_len,
 389                            (unsigned) tx_id,
 390                            (unsigned) tx_last);
 391
 392                 /*
 393                  * Check if there are enough free descriptors in the TX ring
 394                  * to transmit the next packet.
 395                  * This operation is based on the two following rules:
 396                  *
 397                  *   1- Only check that the last needed TX descriptor can be
 398                  *      allocated (by construction, if that descriptor is free,
 399                  *      all intermediate ones are also free).
 400                  *
 401                  *      For this purpose, the index of the last TX descriptor
 402                  *      used for a packet (the "last descriptor" of a packet)
 403                  *      is recorded in the TX entries (the last one included)
 404                  *      that are associated with all TX descriptors allocated
 405                  *      for that packet.
 406                  *
 407                  *   2- Avoid to allocate the last free TX descriptor of the
 408                  *      ring, in order to never set the TDT register with the
 409                  *      same value stored in parallel by the NIC in the TDH
 410                  *      register, which makes the TX engine of the NIC enter
 411                  *      in a deadlock situation.
 412                  *
 413                  *      By extension, avoid to allocate a free descriptor that
 414                  *      belongs to the last set of free descriptors allocated
 415                  *      to the same packet previously transmitted.
 416                  */
 417
 418                 /*
 419                  * The "last descriptor" of the previously sent packet, if any,
 420                  * which used the last descriptor to allocate.
 421                  */
 422                 tx_end = sw_ring[tx_last].last_id;
 423
 424                 /*
 425                  * The next descriptor following that "last descriptor" in the
 426                  * ring.
 427                  */
 428                 tx_end = sw_ring[tx_end].next_id;
 429
 430                 /*
 431                  * The "last descriptor" associated with that next descriptor.
 432                  */
 433                 tx_end = sw_ring[tx_end].last_id;
 434
 435                 /*
 436                  * Check that this descriptor is free.
 437                  */
 438                 if (! (txr[tx_end].wb.status & E1000_TXD_STAT_DD)) {
 439                         if (nb_tx == 0)
 440                                 return (0);
 441                         goto end_of_tx;
 442                 }
 443
 444                 /*
 445                  * Set common flags of all TX Data Descriptors.
 446                  *
 447                  * The following bits must be set in all Data Descriptors:
 448                  *   - E1000_ADVTXD_DTYP_DATA
 449                  *   - E1000_ADVTXD_DCMD_DEXT
 450                  *
 451                  * The following bits must be set in the first Data Descriptor
 452                  * and are ignored in the other ones:
 453                  *   - E1000_ADVTXD_DCMD_IFCS
 454                  *   - E1000_ADVTXD_MAC_1588
 455                  *   - E1000_ADVTXD_DCMD_VLE
 456                  *
 457                  * The following bits must only be set in the last Data
 458                  * Descriptor:
 459                  *   - E1000_TXD_CMD_EOP
 460                  *
 461                  * The following bits can be set in any Data Descriptor, but
 462                  * are only set in the last Data Descriptor:
 463                  *   - E1000_TXD_CMD_RS
 464                  */
 465                 cmd_type_len = txq->txd_type |
 466                         E1000_ADVTXD_DCMD_IFCS | E1000_ADVTXD_DCMD_DEXT;
 467                 olinfo_status = (pkt_len << E1000_ADVTXD_PAYLEN_SHIFT);
 468 #if defined(RTE_LIBRTE_IEEE1588)
 469                 if (ol_flags & PKT_TX_IEEE1588_TMST)
 470                         cmd_type_len |= E1000_ADVTXD_MAC_TSTAMP;
 471 #endif
 472                 if (tx_ol_req) {
 473                         /* Setup TX Advanced context descriptor if required */
 474                         if (new_ctx) {
 475                                 volatile struct e1000_adv_tx_context_desc *
 476                                     ctx_txd;
 477
 478                                 ctx_txd = (volatile struct
 479                                     e1000_adv_tx_context_desc *)
 480                                     &txr[tx_id];
 481
 482                                 txn = &sw_ring[txe->next_id];
 483                                 RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
 484
 485                                 if (txe->mbuf != NULL) {
 486                                         rte_pktmbuf_free_seg(txe->mbuf);
 487                                         txe->mbuf = NULL;
 488                                 }
 489
 490                                 igbe_set_xmit_ctx(txq, ctx_txd, tx_ol_req,
 491                                     vlan_macip_lens);
 492
 493                                 txe->last_id = tx_last;
 494                                 tx_id = txe->next_id;
 495                                 txe = txn;
 496                         }
 497
 498                         /* Setup the TX Advanced Data Descriptor */
 499                         cmd_type_len  |= tx_desc_vlan_flags_to_cmdtype(ol_flags);
 500                         olinfo_status |= tx_desc_cksum_flags_to_olinfo(ol_flags);
 501                         olinfo_status |= (ctx << E1000_ADVTXD_IDX_SHIFT);
 502                 }
 503
 504                 m_seg = tx_pkt;
 505                 do {
 506                         txn = &sw_ring[txe->next_id];
 507                         txd = &txr[tx_id];
 508
 509                         if (txe->mbuf != NULL)
 510                                 rte_pktmbuf_free_seg(txe->mbuf);
 511                         txe->mbuf = m_seg;
 512
 513                         /*
 514                          * Set up transmit descriptor.
 515                          */
 516                         slen = (uint16_t) m_seg->pkt.data_len;
 517                         buf_dma_addr = RTE_MBUF_DATA_DMA_ADDR(m_seg);
 518                         txd->read.buffer_addr =
 519                                 rte_cpu_to_le_64(buf_dma_addr);
 520                         txd->read.cmd_type_len =
 521                                 rte_cpu_to_le_32(cmd_type_len | slen);
 522                         txd->read.olinfo_status =
 523                                 rte_cpu_to_le_32(olinfo_status);
 524                         txe->last_id = tx_last;
 525                         tx_id = txe->next_id;
 526                         txe = txn;
 527                         m_seg = m_seg->pkt.next;
 528                 } while (m_seg != NULL);
 529
 530                 /*
 531                  * The last packet data descriptor needs End Of Packet (EOP)
 532                  * and Report Status (RS).
 533                  */
 534                 txd->read.cmd_type_len |=
 535                         rte_cpu_to_le_32(E1000_TXD_CMD_EOP | E1000_TXD_CMD_RS);
 536         }
 537  end_of_tx:
 538         rte_wmb();
 539
 540         /*
 541          * Set the Transmit Descriptor Tail (TDT).
 542          */
 543         E1000_PCI_REG_WRITE(txq->tdt_reg_addr, tx_id);
 544         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
 545                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
 546                    (unsigned) tx_id, (unsigned) nb_tx);
 547         txq->tx_tail = tx_id;
 548
 549         return (nb_tx);
 550 }
 551
 552 /*********************************************************************
 553  *
 554  *  RX functions
 555  *
 556  **********************************************************************/
 557 static inline uint16_t
 558 rx_desc_hlen_type_rss_to_pkt_flags(uint32_t hl_tp_rs)
 559 {
 560         uint16_t pkt_flags;
 561
 562         static uint16_t ip_pkt_types_map[16] = {
 563                 0, PKT_RX_IPV4_HDR, PKT_RX_IPV4_HDR_EXT, PKT_RX_IPV4_HDR_EXT,
 564                 PKT_RX_IPV6_HDR, 0, 0, 0,
 565                 PKT_RX_IPV6_HDR_EXT, 0, 0, 0,
 566                 PKT_RX_IPV6_HDR_EXT, 0, 0, 0,
 567         };
 568
 569 #if defined(RTE_LIBRTE_IEEE1588)
 570         static uint32_t ip_pkt_etqf_map[8] = {
 571                 0, 0, 0, PKT_RX_IEEE1588_PTP,
 572                 0, 0, 0, 0,
 573         };
 574
 575         pkt_flags = (uint16_t)((hl_tp_rs & E1000_RXDADV_PKTTYPE_ETQF) ?
 576                                 ip_pkt_etqf_map[(hl_tp_rs >> 4) & 0x07] :
 577                                 ip_pkt_types_map[(hl_tp_rs >> 4) & 0x0F]);
 578 #else
 579         pkt_flags = (uint16_t)((hl_tp_rs & E1000_RXDADV_PKTTYPE_ETQF) ? 0 :
 580                                 ip_pkt_types_map[(hl_tp_rs >> 4) & 0x0F]);
 581 #endif
 582         return (uint16_t)(pkt_flags | (((hl_tp_rs & 0x0F) == 0) ?
 583                                                 0 : PKT_RX_RSS_HASH));
 584 }
 585
 586 static inline uint16_t
 587 rx_desc_status_to_pkt_flags(uint32_t rx_status)
 588 {
 589         uint16_t pkt_flags;
 590
 591         /* Check if VLAN present */
 592         pkt_flags = (uint16_t)((rx_status & E1000_RXD_STAT_VP) ?
 593                                                 PKT_RX_VLAN_PKT : 0);
 594
 595 #if defined(RTE_LIBRTE_IEEE1588)
 596         if (rx_status & E1000_RXD_STAT_TMST)
 597                 pkt_flags = (uint16_t)(pkt_flags | PKT_RX_IEEE1588_TMST);
 598 #endif
 599         return pkt_flags;
 600 }
 601
 602 static inline uint16_t
 603 rx_desc_error_to_pkt_flags(uint32_t rx_status)
 604 {
 605         /*
 606          * Bit 30: IPE, IPv4 checksum error
 607          * Bit 29: L4I, L4I integrity error
 608          */
 609
 610         static uint16_t error_to_pkt_flags_map[4] = {
 611                 0,  PKT_RX_L4_CKSUM_BAD, PKT_RX_IP_CKSUM_BAD,
 612                 PKT_RX_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD
 613         };
 614         return error_to_pkt_flags_map[(rx_status >>
 615                 E1000_RXD_ERR_CKSUM_BIT) & E1000_RXD_ERR_CKSUM_MSK];
 616 }
 617
 618 uint16_t
 619 eth_igb_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 620                uint16_t nb_pkts)
 621 {
 622         struct igb_rx_queue *rxq;
 623         volatile union e1000_adv_rx_desc *rx_ring;
 624         volatile union e1000_adv_rx_desc *rxdp;
 625         struct igb_rx_entry *sw_ring;
 626         struct igb_rx_entry *rxe;
 627         struct rte_mbuf *rxm;
 628         struct rte_mbuf *nmb;
 629         union e1000_adv_rx_desc rxd;
 630         uint64_t dma_addr;
 631         uint32_t staterr;
 632         uint32_t hlen_type_rss;
 633         uint16_t pkt_len;
 634         uint16_t rx_id;
 635         uint16_t nb_rx;
 636         uint16_t nb_hold;
 637         uint16_t pkt_flags;
 638
 639         nb_rx = 0;
 640         nb_hold = 0;
 641         rxq = rx_queue;
 642         rx_id = rxq->rx_tail;
 643         rx_ring = rxq->rx_ring;
 644         sw_ring = rxq->sw_ring;
 645         while (nb_rx < nb_pkts) {
 646                 /*
 647                  * The order of operations here is important as the DD status
 648                  * bit must not be read after any other descriptor fields.
 649                  * rx_ring and rxdp are pointing to volatile data so the order
 650                  * of accesses cannot be reordered by the compiler. If they were
 651                  * not volatile, they could be reordered which could lead to
 652                  * using invalid descriptor fields when read from rxd.
 653                  */
 654                 rxdp = &rx_ring[rx_id];
 655                 staterr = rxdp->wb.upper.status_error;
 656                 if (! (staterr & rte_cpu_to_le_32(E1000_RXD_STAT_DD)))
 657                         break;
 658                 rxd = *rxdp;
 659
 660                 /*
 661                  * End of packet.
 662                  *
 663                  * If the E1000_RXD_STAT_EOP flag is not set, the RX packet is
 664                  * likely to be invalid and to be dropped by the various
 665                  * validation checks performed by the network stack.
 666                  *
 667                  * Allocate a new mbuf to replenish the RX ring descriptor.
 668                  * If the allocation fails:
 669                  *    - arrange for that RX descriptor to be the first one
 670                  *      being parsed the next time the receive function is
 671                  *      invoked [on the same queue].
 672                  *
 673                  *    - Stop parsing the RX ring and return immediately.
 674                  *
 675                  * This policy do not drop the packet received in the RX
 676                  * descriptor for which the allocation of a new mbuf failed.
 677                  * Thus, it allows that packet to be later retrieved if
 678                  * mbuf have been freed in the mean time.
 679                  * As a side effect, holding RX descriptors instead of
 680                  * systematically giving them back to the NIC may lead to
 681                  * RX ring exhaustion situations.
 682                  * However, the NIC can gracefully prevent such situations
 683                  * to happen by sending specific "back-pressure" flow control
 684                  * frames to its peer(s).
 685                  */
 686                 PMD_RX_LOG(DEBUG, "\nport_id=%u queue_id=%u rx_id=%u "
 687                            "staterr=0x%x pkt_len=%u\n",
 688                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
 689                            (unsigned) rx_id, (unsigned) staterr,
 690                            (unsigned) rte_le_to_cpu_16(rxd.wb.upper.length));
 691
 692                 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
 693                 if (nmb == NULL) {
 694                         PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
 695                                    "queue_id=%u\n", (unsigned) rxq->port_id,
 696                                    (unsigned) rxq->queue_id);
 697                         rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
 698                         break;
 699                 }
 700
 701                 nb_hold++;
 702                 rxe = &sw_ring[rx_id];
 703                 rx_id++;
 704                 if (rx_id == rxq->nb_rx_desc)
 705                         rx_id = 0;
 706
 707                 /* Prefetch next mbuf while processing current one. */
 708                 rte_igb_prefetch(sw_ring[rx_id].mbuf);
 709
 710                 /*
 711                  * When next RX descriptor is on a cache-line boundary,
 712                  * prefetch the next 4 RX descriptors and the next 8 pointers
 713                  * to mbufs.
 714                  */
 715                 if ((rx_id & 0x3) == 0) {
 716                         rte_igb_prefetch(&rx_ring[rx_id]);
 717                         rte_igb_prefetch(&sw_ring[rx_id]);
 718                 }
 719
 720                 rxm = rxe->mbuf;
 721                 rxe->mbuf = nmb;
 722                 dma_addr =
 723                         rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
 724                 rxdp->read.hdr_addr = dma_addr;
 725                 rxdp->read.pkt_addr = dma_addr;
 726
 727                 /*
 728                  * Initialize the returned mbuf.
 729                  * 1) setup generic mbuf fields:
 730                  *    - number of segments,
 731                  *    - next segment,
 732                  *    - packet length,
 733                  *    - RX port identifier.
 734                  * 2) integrate hardware offload data, if any:
 735                  *    - RSS flag & hash,
 736                  *    - IP checksum flag,
 737                  *    - VLAN TCI, if any,
 738                  *    - error flags.
 739                  */
 740                 pkt_len = (uint16_t) (rte_le_to_cpu_16(rxd.wb.upper.length) -
 741                                       rxq->crc_len);
 742                 rxm->pkt.data = (char*) rxm->buf_addr + RTE_PKTMBUF_HEADROOM;
 743                 rte_packet_prefetch(rxm->pkt.data);
 744                 rxm->pkt.nb_segs = 1;
 745                 rxm->pkt.next = NULL;
 746                 rxm->pkt.pkt_len = pkt_len;
 747                 rxm->pkt.data_len = pkt_len;
 748                 rxm->pkt.in_port = rxq->port_id;
 749
 750                 rxm->pkt.hash.rss = rxd.wb.lower.hi_dword.rss;
 751                 hlen_type_rss = rte_le_to_cpu_32(rxd.wb.lower.lo_dword.data);
 752                 /* Only valid if PKT_RX_VLAN_PKT set in pkt_flags */
 753                 rxm->pkt.vlan_macip.f.vlan_tci =
 754                         rte_le_to_cpu_16(rxd.wb.upper.vlan);
 755
 756                 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(hlen_type_rss);
 757                 pkt_flags = (uint16_t)(pkt_flags |
 758                                 rx_desc_status_to_pkt_flags(staterr));
 759                 pkt_flags = (uint16_t)(pkt_flags |
 760                                 rx_desc_error_to_pkt_flags(staterr));
 761                 rxm->ol_flags = pkt_flags;
 762
 763                 /*
 764                  * Store the mbuf address into the next entry of the array
 765                  * of returned packets.
 766                  */
 767                 rx_pkts[nb_rx++] = rxm;
 768         }
 769         rxq->rx_tail = rx_id;
 770
 771         /*
 772          * If the number of free RX descriptors is greater than the RX free
 773          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
 774          * register.
 775          * Update the RDT with the value of the last processed RX descriptor
 776          * minus 1, to guarantee that the RDT register is never equal to the
 777          * RDH register, which creates a "full" ring situtation from the
 778          * hardware point of view...
 779          */
 780         nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
 781         if (nb_hold > rxq->rx_free_thresh) {
 782                 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
 783                            "nb_hold=%u nb_rx=%u\n",
 784                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
 785                            (unsigned) rx_id, (unsigned) nb_hold,
 786                            (unsigned) nb_rx);
 787                 rx_id = (uint16_t) ((rx_id == 0) ?
 788                                      (rxq->nb_rx_desc - 1) : (rx_id - 1));
 789                 E1000_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
 790                 nb_hold = 0;
 791         }
 792         rxq->nb_rx_hold = nb_hold;
 793         return (nb_rx);
 794 }
 795
 796 uint16_t
 797 eth_igb_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 798                          uint16_t nb_pkts)
 799 {
 800         struct igb_rx_queue *rxq;
 801         volatile union e1000_adv_rx_desc *rx_ring;
 802         volatile union e1000_adv_rx_desc *rxdp;
 803         struct igb_rx_entry *sw_ring;
 804         struct igb_rx_entry *rxe;
 805         struct rte_mbuf *first_seg;
 806         struct rte_mbuf *last_seg;
 807         struct rte_mbuf *rxm;
 808         struct rte_mbuf *nmb;
 809         union e1000_adv_rx_desc rxd;
 810         uint64_t dma; /* Physical address of mbuf data buffer */
 811         uint32_t staterr;
 812         uint32_t hlen_type_rss;
 813         uint16_t rx_id;
 814         uint16_t nb_rx;
 815         uint16_t nb_hold;
 816         uint16_t data_len;
 817         uint16_t pkt_flags;
 818
 819         nb_rx = 0;
 820         nb_hold = 0;
 821         rxq = rx_queue;
 822         rx_id = rxq->rx_tail;
 823         rx_ring = rxq->rx_ring;
 824         sw_ring = rxq->sw_ring;
 825
 826         /*
 827          * Retrieve RX context of current packet, if any.
 828          */
 829         first_seg = rxq->pkt_first_seg;
 830         last_seg = rxq->pkt_last_seg;
 831
 832         while (nb_rx < nb_pkts) {
 833         next_desc:
 834                 /*
 835                  * The order of operations here is important as the DD status
 836                  * bit must not be read after any other descriptor fields.
 837                  * rx_ring and rxdp are pointing to volatile data so the order
 838                  * of accesses cannot be reordered by the compiler. If they were
 839                  * not volatile, they could be reordered which could lead to
 840                  * using invalid descriptor fields when read from rxd.
 841                  */
 842                 rxdp = &rx_ring[rx_id];
 843                 staterr = rxdp->wb.upper.status_error;
 844                 if (! (staterr & rte_cpu_to_le_32(E1000_RXD_STAT_DD)))
 845                         break;
 846                 rxd = *rxdp;
 847
 848                 /*
 849                  * Descriptor done.
 850                  *
 851                  * Allocate a new mbuf to replenish the RX ring descriptor.
 852                  * If the allocation fails:
 853                  *    - arrange for that RX descriptor to be the first one
 854                  *      being parsed the next time the receive function is
 855                  *      invoked [on the same queue].
 856                  *
 857                  *    - Stop parsing the RX ring and return immediately.
 858                  *
 859                  * This policy does not drop the packet received in the RX
 860                  * descriptor for which the allocation of a new mbuf failed.
 861                  * Thus, it allows that packet to be later retrieved if
 862                  * mbuf have been freed in the mean time.
 863                  * As a side effect, holding RX descriptors instead of
 864                  * systematically giving them back to the NIC may lead to
 865                  * RX ring exhaustion situations.
 866                  * However, the NIC can gracefully prevent such situations
 867                  * to happen by sending specific "back-pressure" flow control
 868                  * frames to its peer(s).
 869                  */
 870                 PMD_RX_LOG(DEBUG, "\nport_id=%u queue_id=%u rx_id=%u "
 871                            "staterr=0x%x data_len=%u\n",
 872                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
 873                            (unsigned) rx_id, (unsigned) staterr,
 874                            (unsigned) rte_le_to_cpu_16(rxd.wb.upper.length));
 875
 876                 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
 877                 if (nmb == NULL) {
 878                         PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
 879                                    "queue_id=%u\n", (unsigned) rxq->port_id,
 880                                    (unsigned) rxq->queue_id);
 881                         rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
 882                         break;
 883                 }
 884
 885                 nb_hold++;
 886                 rxe = &sw_ring[rx_id];
 887                 rx_id++;
 888                 if (rx_id == rxq->nb_rx_desc)
 889                         rx_id = 0;
 890
 891                 /* Prefetch next mbuf while processing current one. */
 892                 rte_igb_prefetch(sw_ring[rx_id].mbuf);
 893
 894                 /*
 895                  * When next RX descriptor is on a cache-line boundary,
 896                  * prefetch the next 4 RX descriptors and the next 8 pointers
 897                  * to mbufs.
 898                  */
 899                 if ((rx_id & 0x3) == 0) {
 900                         rte_igb_prefetch(&rx_ring[rx_id]);
 901                         rte_igb_prefetch(&sw_ring[rx_id]);
 902                 }
 903
 904                 /*
 905                  * Update RX descriptor with the physical address of the new
 906                  * data buffer of the new allocated mbuf.
 907                  */
 908                 rxm = rxe->mbuf;
 909                 rxe->mbuf = nmb;
 910                 dma = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
 911                 rxdp->read.pkt_addr = dma;
 912                 rxdp->read.hdr_addr = dma;
 913
 914                 /*
 915                  * Set data length & data buffer address of mbuf.
 916                  */
 917                 data_len = rte_le_to_cpu_16(rxd.wb.upper.length);
 918                 rxm->pkt.data_len = data_len;
 919                 rxm->pkt.data = (char*) rxm->buf_addr + RTE_PKTMBUF_HEADROOM;
 920
 921                 /*
 922                  * If this is the first buffer of the received packet,
 923                  * set the pointer to the first mbuf of the packet and
 924                  * initialize its context.
 925                  * Otherwise, update the total length and the number of segments
 926                  * of the current scattered packet, and update the pointer to
 927                  * the last mbuf of the current packet.
 928                  */
 929                 if (first_seg == NULL) {
 930                         first_seg = rxm;
 931                         first_seg->pkt.pkt_len = data_len;
 932                         first_seg->pkt.nb_segs = 1;
 933                 } else {
 934                         first_seg->pkt.pkt_len += data_len;
 935                         first_seg->pkt.nb_segs++;
 936                         last_seg->pkt.next = rxm;
 937                 }
 938
 939                 /*
 940                  * If this is not the last buffer of the received packet,
 941                  * update the pointer to the last mbuf of the current scattered
 942                  * packet and continue to parse the RX ring.
 943                  */
 944                 if (! (staterr & E1000_RXD_STAT_EOP)) {
 945                         last_seg = rxm;
 946                         goto next_desc;
 947                 }
 948
 949                 /*
 950                  * This is the last buffer of the received packet.
 951                  * If the CRC is not stripped by the hardware:
 952                  *   - Subtract the CRC length from the total packet length.
 953                  *   - If the last buffer only contains the whole CRC or a part
 954                  *     of it, free the mbuf associated to the last buffer.
 955                  *     If part of the CRC is also contained in the previous
 956                  *     mbuf, subtract the length of that CRC part from the
 957                  *     data length of the previous mbuf.
 958                  */
 959                 rxm->pkt.next = NULL;
 960                 if (unlikely(rxq->crc_len > 0)) {
 961                         first_seg->pkt.pkt_len -= ETHER_CRC_LEN;
 962                         if (data_len <= ETHER_CRC_LEN) {
 963                                 rte_pktmbuf_free_seg(rxm);
 964                                 first_seg->pkt.nb_segs--;
 965                                 last_seg->pkt.data_len = (uint16_t)
 966                                         (last_seg->pkt.data_len -
 967                                          (ETHER_CRC_LEN - data_len));
 968                                 last_seg->pkt.next = NULL;
 969                         } else
 970                                 rxm->pkt.data_len =
 971                                         (uint16_t) (data_len - ETHER_CRC_LEN);
 972                 }
 973
 974                 /*
 975                  * Initialize the first mbuf of the returned packet:
 976                  *    - RX port identifier,
 977                  *    - hardware offload data, if any:
 978                  *      - RSS flag & hash,
 979                  *      - IP checksum flag,
 980                  *      - VLAN TCI, if any,
 981                  *      - error flags.
 982                  */
 983                 first_seg->pkt.in_port = rxq->port_id;
 984                 first_seg->pkt.hash.rss = rxd.wb.lower.hi_dword.rss;
 985
 986                 /*
 987                  * The vlan_tci field is only valid when PKT_RX_VLAN_PKT is
 988                  * set in the pkt_flags field.
 989                  */
 990                 first_seg->pkt.vlan_macip.f.vlan_tci =
 991                         rte_le_to_cpu_16(rxd.wb.upper.vlan);
 992                 hlen_type_rss = rte_le_to_cpu_32(rxd.wb.lower.lo_dword.data);
 993                 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(hlen_type_rss);
 994                 pkt_flags = (uint16_t)(pkt_flags |
 995                                 rx_desc_status_to_pkt_flags(staterr));
 996                 pkt_flags = (uint16_t)(pkt_flags |
 997                                 rx_desc_error_to_pkt_flags(staterr));
 998                 first_seg->ol_flags = pkt_flags;
 999
1000                 /* Prefetch data of first segment, if configured to do so. */
1001                 rte_packet_prefetch(first_seg->pkt.data);
1002
1003                 /*
1004                  * Store the mbuf address into the next entry of the array
1005                  * of returned packets.
1006                  */
1007                 rx_pkts[nb_rx++] = first_seg;
1008
1009                 /*
1010                  * Setup receipt context for a new packet.
1011                  */
1012                 first_seg = NULL;
1013         }
1014
1015         /*
1016          * Record index of the next RX descriptor to probe.
1017          */
1018         rxq->rx_tail = rx_id;
1019
1020         /*
1021          * Save receive context.
1022          */
1023         rxq->pkt_first_seg = first_seg;
1024         rxq->pkt_last_seg = last_seg;
1025
1026         /*
1027          * If the number of free RX descriptors is greater than the RX free
1028          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1029          * register.
1030          * Update the RDT with the value of the last processed RX descriptor
1031          * minus 1, to guarantee that the RDT register is never equal to the
1032          * RDH register, which creates a "full" ring situtation from the
1033          * hardware point of view...
1034          */
1035         nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
1036         if (nb_hold > rxq->rx_free_thresh) {
1037                 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1038                            "nb_hold=%u nb_rx=%u\n",
1039                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1040                            (unsigned) rx_id, (unsigned) nb_hold,
1041                            (unsigned) nb_rx);
1042                 rx_id = (uint16_t) ((rx_id == 0) ?
1043                                      (rxq->nb_rx_desc - 1) : (rx_id - 1));
1044                 E1000_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
1045                 nb_hold = 0;
1046         }
1047         rxq->nb_rx_hold = nb_hold;
1048         return (nb_rx);
1049 }
1050
1051 /*
1052  * Rings setup and release.
1053  *
1054  * TDBA/RDBA should be aligned on 16 byte boundary. But TDLEN/RDLEN should be
1055  * multiple of 128 bytes. So we align TDBA/RDBA on 128 byte boundary.
1056  * This will also optimize cache line size effect.
1057  * H/W supports up to cache line size 128.
1058  */
1059 #define IGB_ALIGN 128
1060
1061 /*
1062  * Maximum number of Ring Descriptors.
1063  *
1064  * Since RDLEN/TDLEN should be multiple of 128bytes, the number of ring
1065  * desscriptors should meet the following condition:
1066  *      (num_ring_desc * sizeof(struct e1000_rx/tx_desc)) % 128 == 0
1067  */
1068 #define IGB_MIN_RING_DESC 32
1069 #define IGB_MAX_RING_DESC 4096
1070
1071 static const struct rte_memzone *
1072 ring_dma_zone_reserve(struct rte_eth_dev *dev, const char *ring_name,
1073                       uint16_t queue_id, uint32_t ring_size, int socket_id)
1074 {
1075         char z_name[RTE_MEMZONE_NAMESIZE];
1076         const struct rte_memzone *mz;
1077
1078         rte_snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
1079                         dev->driver->pci_drv.name, ring_name,
1080                                 dev->data->port_id, queue_id);
1081         mz = rte_memzone_lookup(z_name);
1082         if (mz)
1083                 return mz;
1084
1085         return rte_memzone_reserve_aligned(z_name, ring_size,
1086                         socket_id, 0, IGB_ALIGN);
1087 }
1088
1089 static void
1090 igb_tx_queue_release_mbufs(struct igb_tx_queue *txq)
1091 {
1092         unsigned i;
1093
1094         if (txq->sw_ring != NULL) {
1095                 for (i = 0; i < txq->nb_tx_desc; i++) {
1096                         if (txq->sw_ring[i].mbuf != NULL) {
1097                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
1098                                 txq->sw_ring[i].mbuf = NULL;
1099                         }
1100                 }
1101         }
1102 }
1103
1104 static void
1105 igb_tx_queue_release(struct igb_tx_queue *txq)
1106 {
1107         if (txq != NULL) {
1108                 igb_tx_queue_release_mbufs(txq);
1109                 rte_free(txq->sw_ring);
1110                 rte_free(txq);
1111         }
1112 }
1113
1114 void
1115 eth_igb_tx_queue_release(void *txq)
1116 {
1117         igb_tx_queue_release(txq);
1118 }
1119
1120 static void
1121 igb_reset_tx_queue_stat(struct igb_tx_queue *txq)
1122 {
1123         txq->tx_head = 0;
1124         txq->tx_tail = 0;
1125         txq->ctx_curr = 0;
1126         memset((void*)&txq->ctx_cache, 0,
1127                 IGB_CTX_NUM * sizeof(struct igb_advctx_info));
1128 }
1129
1130 static void
1131 igb_reset_tx_queue(struct igb_tx_queue *txq, struct rte_eth_dev *dev)
1132 {
1133         struct igb_tx_entry *txe = txq->sw_ring;
1134         uint32_t size;
1135         uint16_t i, prev;
1136         struct e1000_hw *hw;
1137
1138         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1139         size = sizeof(union e1000_adv_tx_desc) * txq->nb_tx_desc;
1140         /* Zero out HW ring memory */
1141         for (i = 0; i < size; i++) {
1142                 ((volatile char *)txq->tx_ring)[i] = 0;
1143         }
1144
1145         /* Initialize ring entries */
1146         prev = (uint16_t)(txq->nb_tx_desc - 1);
1147         for (i = 0; i < txq->nb_tx_desc; i++) {
1148                 volatile union e1000_adv_tx_desc *txd = &(txq->tx_ring[i]);
1149
1150                 txd->wb.status = E1000_TXD_STAT_DD;
1151                 txe[i].mbuf = NULL;
1152                 txe[i].last_id = i;
1153                 txe[prev].next_id = i;
1154                 prev = i;
1155         }
1156
1157         txq->txd_type = E1000_ADVTXD_DTYP_DATA;
1158         /* 82575 specific, each tx queue will use 2 hw contexts */
1159         if (hw->mac.type == e1000_82575)
1160                 txq->ctx_start = txq->queue_id * IGB_CTX_NUM;
1161
1162         igb_reset_tx_queue_stat(txq);
1163 }
1164
1165 int
1166 eth_igb_tx_queue_setup(struct rte_eth_dev *dev,
1167                          uint16_t queue_idx,
1168                          uint16_t nb_desc,
1169                          unsigned int socket_id,
1170                          const struct rte_eth_txconf *tx_conf)
1171 {
1172         const struct rte_memzone *tz;
1173         struct igb_tx_queue *txq;
1174         struct e1000_hw     *hw;
1175         uint32_t size;
1176
1177         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1178
1179         /*
1180          * Validate number of transmit descriptors.
1181          * It must not exceed hardware maximum, and must be multiple
1182          * of IGB_ALIGN.
1183          */
1184         if (((nb_desc * sizeof(union e1000_adv_tx_desc)) % IGB_ALIGN) != 0 ||
1185             (nb_desc > IGB_MAX_RING_DESC) || (nb_desc < IGB_MIN_RING_DESC)) {
1186                 return -EINVAL;
1187         }
1188
1189         /*
1190          * The tx_free_thresh and tx_rs_thresh values are not used in the 1G
1191          * driver.
1192          */
1193         if (tx_conf->tx_free_thresh != 0)
1194                 RTE_LOG(WARNING, PMD,
1195                         "The tx_free_thresh parameter is not "
1196                         "used for the 1G driver.\n");
1197         if (tx_conf->tx_rs_thresh != 0)
1198                 RTE_LOG(WARNING, PMD,
1199                         "The tx_rs_thresh parameter is not "
1200                         "used for the 1G driver.\n");
1201         if (tx_conf->tx_thresh.wthresh == 0)
1202                 RTE_LOG(WARNING, PMD,
1203                         "To improve 1G driver performance, consider setting "
1204                         "the TX WTHRESH value to 4, 8, or 16.\n");
1205
1206         /* Free memory prior to re-allocation if needed */
1207         if (dev->data->tx_queues[queue_idx] != NULL)
1208                 igb_tx_queue_release(dev->data->tx_queues[queue_idx]);
1209
1210         /* First allocate the tx queue data structure */
1211         txq = rte_zmalloc("ethdev TX queue", sizeof(struct igb_tx_queue),
1212                                                         CACHE_LINE_SIZE);
1213         if (txq == NULL)
1214                 return (-ENOMEM);
1215
1216         /*
1217          * Allocate TX ring hardware descriptors. A memzone large enough to
1218          * handle the maximum ring size is allocated in order to allow for
1219          * resizing in later calls to the queue setup function.
1220          */
1221         size = sizeof(union e1000_adv_tx_desc) * IGB_MAX_RING_DESC;
1222         tz = ring_dma_zone_reserve(dev, "tx_ring", queue_idx,
1223                                         size, socket_id);
1224         if (tz == NULL) {
1225                 igb_tx_queue_release(txq);
1226                 return (-ENOMEM);
1227         }
1228
1229         txq->nb_tx_desc = nb_desc;
1230         txq->pthresh = tx_conf->tx_thresh.pthresh;
1231         txq->hthresh = tx_conf->tx_thresh.hthresh;
1232         txq->wthresh = tx_conf->tx_thresh.wthresh;
1233         txq->queue_id = queue_idx;
1234         txq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
1235                 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
1236         txq->port_id = dev->data->port_id;
1237
1238         txq->tdt_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_TDT(txq->reg_idx));
1239         txq->tx_ring_phys_addr = (uint64_t) tz->phys_addr;
1240         txq->tx_ring = (union e1000_adv_tx_desc *) tz->addr;
1241
1242         /* Allocate software ring */
1243         txq->sw_ring = rte_zmalloc("txq->sw_ring",
1244                                    sizeof(struct igb_tx_entry) * nb_desc,
1245                                    CACHE_LINE_SIZE);
1246         if (txq->sw_ring == NULL) {
1247                 igb_tx_queue_release(txq);
1248                 return (-ENOMEM);
1249         }
1250         PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64"\n",
1251                      txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr);
1252
1253         igb_reset_tx_queue(txq, dev);
1254         dev->tx_pkt_burst = eth_igb_xmit_pkts;
1255         dev->data->tx_queues[queue_idx] = txq;
1256
1257         return (0);
1258 }
1259
1260 static void
1261 igb_rx_queue_release_mbufs(struct igb_rx_queue *rxq)
1262 {
1263         unsigned i;
1264
1265         if (rxq->sw_ring != NULL) {
1266                 for (i = 0; i < rxq->nb_rx_desc; i++) {
1267                         if (rxq->sw_ring[i].mbuf != NULL) {
1268                                 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
1269                                 rxq->sw_ring[i].mbuf = NULL;
1270                         }
1271                 }
1272         }
1273 }
1274
1275 static void
1276 igb_rx_queue_release(struct igb_rx_queue *rxq)
1277 {
1278         if (rxq != NULL) {
1279                 igb_rx_queue_release_mbufs(rxq);
1280                 rte_free(rxq->sw_ring);
1281                 rte_free(rxq);
1282         }
1283 }
1284
1285 void
1286 eth_igb_rx_queue_release(void *rxq)
1287 {
1288         igb_rx_queue_release(rxq);
1289 }
1290
1291 static void
1292 igb_reset_rx_queue(struct igb_rx_queue *rxq)
1293 {
1294         unsigned size;
1295         unsigned i;
1296
1297         /* Zero out HW ring memory */
1298         size = sizeof(union e1000_adv_rx_desc) * rxq->nb_rx_desc;
1299         for (i = 0; i < size; i++) {
1300                 ((volatile char *)rxq->rx_ring)[i] = 0;
1301         }
1302
1303         rxq->rx_tail = 0;
1304         rxq->pkt_first_seg = NULL;
1305         rxq->pkt_last_seg = NULL;
1306 }
1307
1308 int
1309 eth_igb_rx_queue_setup(struct rte_eth_dev *dev,
1310                          uint16_t queue_idx,
1311                          uint16_t nb_desc,
1312                          unsigned int socket_id,
1313                          const struct rte_eth_rxconf *rx_conf,
1314                          struct rte_mempool *mp)
1315 {
1316         const struct rte_memzone *rz;
1317         struct igb_rx_queue *rxq;
1318         struct e1000_hw     *hw;
1319         unsigned int size;
1320
1321         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1322
1323         /*
1324          * Validate number of receive descriptors.
1325          * It must not exceed hardware maximum, and must be multiple
1326          * of IGB_ALIGN.
1327          */
1328         if (((nb_desc * sizeof(union e1000_adv_rx_desc)) % IGB_ALIGN) != 0 ||
1329             (nb_desc > IGB_MAX_RING_DESC) || (nb_desc < IGB_MIN_RING_DESC)) {
1330                 return (-EINVAL);
1331         }
1332
1333         /* Free memory prior to re-allocation if needed */
1334         if (dev->data->rx_queues[queue_idx] != NULL) {
1335                 igb_rx_queue_release(dev->data->rx_queues[queue_idx]);
1336                 dev->data->rx_queues[queue_idx] = NULL;
1337         }
1338
1339         /* First allocate the RX queue data structure. */
1340         rxq = rte_zmalloc("ethdev RX queue", sizeof(struct igb_rx_queue),
1341                           CACHE_LINE_SIZE);
1342         if (rxq == NULL)
1343                 return (-ENOMEM);
1344         rxq->mb_pool = mp;
1345         rxq->nb_rx_desc = nb_desc;
1346         rxq->pthresh = rx_conf->rx_thresh.pthresh;
1347         rxq->hthresh = rx_conf->rx_thresh.hthresh;
1348         rxq->wthresh = rx_conf->rx_thresh.wthresh;
1349         rxq->drop_en = rx_conf->rx_drop_en;
1350         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1351         rxq->queue_id = queue_idx;
1352         rxq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
1353                 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
1354         rxq->port_id = dev->data->port_id;
1355         rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ? 0 :
1356                                   ETHER_CRC_LEN);
1357
1358         /*
1359          *  Allocate RX ring hardware descriptors. A memzone large enough to
1360          *  handle the maximum ring size is allocated in order to allow for
1361          *  resizing in later calls to the queue setup function.
1362          */
1363         size = sizeof(union e1000_adv_rx_desc) * IGB_MAX_RING_DESC;
1364         rz = ring_dma_zone_reserve(dev, "rx_ring", queue_idx, size, socket_id);
1365         if (rz == NULL) {
1366                 igb_rx_queue_release(rxq);
1367                 return (-ENOMEM);
1368         }
1369         rxq->rdt_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_RDT(rxq->reg_idx));
1370         rxq->rdh_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_RDH(rxq->reg_idx));
1371         rxq->rx_ring_phys_addr = (uint64_t) rz->phys_addr;
1372         rxq->rx_ring = (union e1000_adv_rx_desc *) rz->addr;
1373
1374         /* Allocate software ring. */
1375         rxq->sw_ring = rte_zmalloc("rxq->sw_ring",
1376                                    sizeof(struct igb_rx_entry) * nb_desc,
1377                                    CACHE_LINE_SIZE);
1378         if (rxq->sw_ring == NULL) {
1379                 igb_rx_queue_release(rxq);
1380                 return (-ENOMEM);
1381         }
1382         PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64"\n",
1383                      rxq->sw_ring, rxq->rx_ring, rxq->rx_ring_phys_addr);
1384
1385         dev->data->rx_queues[queue_idx] = rxq;
1386         igb_reset_rx_queue(rxq);
1387
1388         return 0;
1389 }
1390
1391 uint32_t
1392 eth_igb_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1393 {
1394 #define IGB_RXQ_SCAN_INTERVAL 4
1395         volatile union e1000_adv_rx_desc *rxdp;
1396         struct igb_rx_queue *rxq;
1397         uint32_t desc = 0;
1398
1399         if (rx_queue_id >= dev->data->nb_rx_queues) {
1400                 PMD_RX_LOG(ERR, "Invalid RX queue id=%d\n", rx_queue_id);
1401                 return 0;
1402         }
1403
1404         rxq = dev->data->rx_queues[rx_queue_id];
1405         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
1406
1407         while ((desc < rxq->nb_rx_desc) &&
1408                 (rxdp->wb.upper.status_error & E1000_RXD_STAT_DD)) {
1409                 desc += IGB_RXQ_SCAN_INTERVAL;
1410                 rxdp += IGB_RXQ_SCAN_INTERVAL;
1411                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
1412                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
1413                                 desc - rxq->nb_rx_desc]);
1414         }
1415
1416         return 0;
1417 }
1418
1419 void
1420 igb_dev_clear_queues(struct rte_eth_dev *dev)
1421 {
1422         uint16_t i;
1423         struct igb_tx_queue *txq;
1424         struct igb_rx_queue *rxq;
1425
1426         for (i = 0; i < dev->data->nb_tx_queues; i++) {
1427                 txq = dev->data->tx_queues[i];
1428                 if (txq != NULL) {
1429                         igb_tx_queue_release_mbufs(txq);
1430                         igb_reset_tx_queue(txq, dev);
1431                 }
1432         }
1433
1434         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1435                 rxq = dev->data->rx_queues[i];
1436                 if (rxq != NULL) {
1437                         igb_rx_queue_release_mbufs(rxq);
1438                         igb_reset_rx_queue(rxq);
1439                 }
1440         }
1441 }
1442
1443 /**
1444  * Receive Side Scaling (RSS).
1445  * See section 7.1.1.7 in the following document:
1446  *     "Intel 82576 GbE Controller Datasheet" - Revision 2.45 October 2009
1447  *
1448  * Principles:
1449  * The source and destination IP addresses of the IP header and the source and
1450  * destination ports of TCP/UDP headers, if any, of received packets are hashed
1451  * against a configurable random key to compute a 32-bit RSS hash result.
1452  * The seven (7) LSBs of the 32-bit hash result are used as an index into a
1453  * 128-entry redirection table (RETA).  Each entry of the RETA provides a 3-bit
1454  * RSS output index which is used as the RX queue index where to store the
1455  * received packets.
1456  * The following output is supplied in the RX write-back descriptor:
1457  *     - 32-bit result of the Microsoft RSS hash function,
1458  *     - 4-bit RSS type field.
1459  */
1460
1461 /*
1462  * RSS random key supplied in section 7.1.1.7.3 of the Intel 82576 datasheet.
1463  * Used as the default key.
1464  */
1465 static uint8_t rss_intel_key[40] = {
1466         0x6D, 0x5A, 0x56, 0xDA, 0x25, 0x5B, 0x0E, 0xC2,
1467         0x41, 0x67, 0x25, 0x3D, 0x43, 0xA3, 0x8F, 0xB0,
1468         0xD0, 0xCA, 0x2B, 0xCB, 0xAE, 0x7B, 0x30, 0xB4,
1469         0x77, 0xCB, 0x2D, 0xA3, 0x80, 0x30, 0xF2, 0x0C,
1470         0x6A, 0x42, 0xB7, 0x3B, 0xBE, 0xAC, 0x01, 0xFA,
1471 };
1472
1473 static void
1474 igb_rss_disable(struct rte_eth_dev *dev)
1475 {
1476         struct e1000_hw *hw;
1477         uint32_t mrqc;
1478
1479         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1480         mrqc = E1000_READ_REG(hw, E1000_MRQC);
1481         mrqc &= ~E1000_MRQC_ENABLE_MASK;
1482         E1000_WRITE_REG(hw, E1000_MRQC, mrqc);
1483 }
1484
1485 static void
1486 igb_rss_configure(struct rte_eth_dev *dev)
1487 {
1488         struct e1000_hw *hw;
1489         uint8_t *hash_key;
1490         uint32_t rss_key;
1491         uint32_t mrqc;
1492         uint32_t shift;
1493         uint16_t rss_hf;
1494         uint16_t i;
1495
1496         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1497
1498         rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
1499         if (rss_hf == 0) /* Disable RSS. */ {
1500                 igb_rss_disable(dev);
1501                 return;
1502         }
1503         hash_key = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
1504         if (hash_key == NULL)
1505                 hash_key = rss_intel_key; /* Default hash key. */
1506
1507         /* Fill in RSS hash key. */
1508         for (i = 0; i < 10; i++) {
1509                 rss_key  = hash_key[(i * 4)];
1510                 rss_key |= hash_key[(i * 4) + 1] << 8;
1511                 rss_key |= hash_key[(i * 4) + 2] << 16;
1512                 rss_key |= hash_key[(i * 4) + 3] << 24;
1513                 E1000_WRITE_REG_ARRAY(hw, E1000_RSSRK(0), i, rss_key);
1514         }
1515
1516         /* Fill in redirection table. */
1517         shift = (hw->mac.type == e1000_82575) ? 6 : 0;
1518         for (i = 0; i < 128; i++) {
1519                 union e1000_reta {
1520                         uint32_t dword;
1521                         uint8_t  bytes[4];
1522                 } reta;
1523                 uint8_t q_idx;
1524
1525                 q_idx = (uint8_t) ((dev->data->nb_rx_queues > 1) ?
1526                                    i % dev->data->nb_rx_queues : 0);
1527                 reta.bytes[i & 3] = (uint8_t) (q_idx << shift);
1528                 if ((i & 3) == 3)
1529                         E1000_WRITE_REG(hw, E1000_RETA(i >> 2), reta.dword);
1530         }
1531
1532         /* Set configured hashing functions in MRQC register. */
1533         mrqc = E1000_MRQC_ENABLE_RSS_4Q; /* RSS enabled. */
1534         if (rss_hf & ETH_RSS_IPV4)
1535                 mrqc |= E1000_MRQC_RSS_FIELD_IPV4;
1536         if (rss_hf & ETH_RSS_IPV4_TCP)
1537                 mrqc |= E1000_MRQC_RSS_FIELD_IPV4_TCP;
1538         if (rss_hf & ETH_RSS_IPV6)
1539                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6;
1540         if (rss_hf & ETH_RSS_IPV6_EX)
1541                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_EX;
1542         if (rss_hf & ETH_RSS_IPV6_TCP)
1543                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_TCP;
1544         if (rss_hf & ETH_RSS_IPV6_TCP_EX)
1545                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_TCP_EX;
1546         if (rss_hf & ETH_RSS_IPV4_UDP)
1547                 mrqc |= E1000_MRQC_RSS_FIELD_IPV4_UDP;
1548         if (rss_hf & ETH_RSS_IPV6_UDP)
1549                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_UDP;
1550         if (rss_hf & ETH_RSS_IPV6_UDP_EX)
1551                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_UDP_EX;
1552         E1000_WRITE_REG(hw, E1000_MRQC, mrqc);
1553 }
1554
1555 /*********************************************************************
1556  *
1557  *  Enable receive unit.
1558  *
1559  **********************************************************************/
1560
1561 static int
1562 igb_alloc_rx_queue_mbufs(struct igb_rx_queue *rxq)
1563 {
1564         struct igb_rx_entry *rxe = rxq->sw_ring;
1565         uint64_t dma_addr;
1566         unsigned i;
1567
1568         /* Initialize software ring entries. */
1569         for (i = 0; i < rxq->nb_rx_desc; i++) {
1570                 volatile union e1000_adv_rx_desc *rxd;
1571                 struct rte_mbuf *mbuf = rte_rxmbuf_alloc(rxq->mb_pool);
1572
1573                 if (mbuf == NULL) {
1574                         PMD_INIT_LOG(ERR, "RX mbuf alloc failed "
1575                                 "queue_id=%hu\n", rxq->queue_id);
1576                         igb_rx_queue_release(rxq);
1577                         return (-ENOMEM);
1578                 }
1579                 dma_addr =
1580                         rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mbuf));
1581                 rxd = &rxq->rx_ring[i];
1582                 rxd->read.hdr_addr = dma_addr;
1583                 rxd->read.pkt_addr = dma_addr;
1584                 rxe[i].mbuf = mbuf;
1585         }
1586
1587         return 0;
1588 }
1589
1590 #define E1000_MRQC_DEF_Q_SHIFT               (3)
1591 static int
1592 igb_dev_mq_rx_configure(struct rte_eth_dev *dev)
1593 {
1594         struct e1000_hw *hw =
1595                 E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1596         uint32_t mrqc;
1597
1598         if (RTE_ETH_DEV_SRIOV(dev).active == ETH_8_POOLS) {
1599                 /*
1600                 * SRIOV active scheme
1601                 * FIXME if support RSS together with VMDq & SRIOV
1602                 */
1603                 mrqc = E1000_MRQC_ENABLE_VMDQ;
1604                 /* 011b Def_Q ignore, according to VT_CTL.DEF_PL */
1605                 mrqc |= 0x3 << E1000_MRQC_DEF_Q_SHIFT;
1606                 E1000_WRITE_REG(hw, E1000_MRQC, mrqc);
1607         } else if(RTE_ETH_DEV_SRIOV(dev).active == 0) {
1608                 /*
1609                 * SRIOV inactive scheme
1610                 */
1611                 if (dev->data->nb_rx_queues > 1)
1612                         igb_rss_configure(dev);
1613                 else
1614                         igb_rss_disable(dev);
1615         }
1616
1617         return 0;
1618 }
1619
1620 int
1621 eth_igb_rx_init(struct rte_eth_dev *dev)
1622 {
1623         struct e1000_hw     *hw;
1624         struct igb_rx_queue *rxq;
1625         struct rte_pktmbuf_pool_private *mbp_priv;
1626         uint32_t rctl;
1627         uint32_t rxcsum;
1628         uint32_t srrctl;
1629         uint16_t buf_size;
1630         uint16_t rctl_bsize;
1631         uint16_t i;
1632         int ret;
1633
1634         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1635         srrctl = 0;
1636
1637         /*
1638          * Make sure receives are disabled while setting
1639          * up the descriptor ring.
1640          */
1641         rctl = E1000_READ_REG(hw, E1000_RCTL);
1642         E1000_WRITE_REG(hw, E1000_RCTL, rctl & ~E1000_RCTL_EN);
1643
1644         /*
1645          * Configure support of jumbo frames, if any.
1646          */
1647         if (dev->data->dev_conf.rxmode.jumbo_frame == 1) {
1648                 rctl |= E1000_RCTL_LPE;
1649
1650                 /*
1651                  * Set maximum packet length by default, and might be updated
1652                  * together with enabling/disabling dual VLAN.
1653                  */
1654                 E1000_WRITE_REG(hw, E1000_RLPML,
1655                         dev->data->dev_conf.rxmode.max_rx_pkt_len +
1656                                                 VLAN_TAG_SIZE);
1657         } else
1658                 rctl &= ~E1000_RCTL_LPE;
1659
1660         /* Configure and enable each RX queue. */
1661         rctl_bsize = 0;
1662         dev->rx_pkt_burst = eth_igb_recv_pkts;
1663         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1664                 uint64_t bus_addr;
1665                 uint32_t rxdctl;
1666
1667                 rxq = dev->data->rx_queues[i];
1668
1669                 /* Allocate buffers for descriptor rings and set up queue */
1670                 ret = igb_alloc_rx_queue_mbufs(rxq);
1671                 if (ret)
1672                         return ret;
1673
1674                 /*
1675                  * Reset crc_len in case it was changed after queue setup by a
1676                  *  call to configure
1677                  */
1678                 rxq->crc_len =
1679                         (uint8_t)(dev->data->dev_conf.rxmode.hw_strip_crc ?
1680                                                         0 : ETHER_CRC_LEN);
1681
1682                 bus_addr = rxq->rx_ring_phys_addr;
1683                 E1000_WRITE_REG(hw, E1000_RDLEN(rxq->reg_idx),
1684                                 rxq->nb_rx_desc *
1685                                 sizeof(union e1000_adv_rx_desc));
1686                 E1000_WRITE_REG(hw, E1000_RDBAH(rxq->reg_idx),
1687                                 (uint32_t)(bus_addr >> 32));
1688                 E1000_WRITE_REG(hw, E1000_RDBAL(rxq->reg_idx), (uint32_t)bus_addr);
1689
1690                 srrctl = E1000_SRRCTL_DESCTYPE_ADV_ONEBUF;
1691
1692                 /*
1693                  * Configure RX buffer size.
1694                  */
1695                 mbp_priv = (struct rte_pktmbuf_pool_private *)
1696                         ((char *)rxq->mb_pool + sizeof(struct rte_mempool));
1697                 buf_size = (uint16_t) (mbp_priv->mbuf_data_room_size -
1698                                        RTE_PKTMBUF_HEADROOM);
1699                 if (buf_size >= 1024) {
1700                         /*
1701                          * Configure the BSIZEPACKET field of the SRRCTL
1702                          * register of the queue.
1703                          * Value is in 1 KB resolution, from 1 KB to 127 KB.
1704                          * If this field is equal to 0b, then RCTL.BSIZE
1705                          * determines the RX packet buffer size.
1706                          */
1707                         srrctl |= ((buf_size >> E1000_SRRCTL_BSIZEPKT_SHIFT) &
1708                                    E1000_SRRCTL_BSIZEPKT_MASK);
1709                         buf_size = (uint16_t) ((srrctl &
1710                                                 E1000_SRRCTL_BSIZEPKT_MASK) <<
1711                                                E1000_SRRCTL_BSIZEPKT_SHIFT);
1712
1713                         /* It adds dual VLAN length for supporting dual VLAN */
1714                         if ((dev->data->dev_conf.rxmode.max_rx_pkt_len +
1715                                                 2 * VLAN_TAG_SIZE) > buf_size){
1716                                 dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
1717                                 dev->data->scattered_rx = 1;
1718                         }
1719                 } else {
1720                         /*
1721                          * Use BSIZE field of the device RCTL register.
1722                          */
1723                         if ((rctl_bsize == 0) || (rctl_bsize > buf_size))
1724                                 rctl_bsize = buf_size;
1725                         dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
1726                         dev->data->scattered_rx = 1;
1727                 }
1728
1729                 /* Set if packets are dropped when no descriptors available */
1730                 if (rxq->drop_en)
1731                         srrctl |= E1000_SRRCTL_DROP_EN;
1732
1733                 E1000_WRITE_REG(hw, E1000_SRRCTL(rxq->reg_idx), srrctl);
1734
1735                 /* Enable this RX queue. */
1736                 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(rxq->reg_idx));
1737                 rxdctl |= E1000_RXDCTL_QUEUE_ENABLE;
1738                 rxdctl &= 0xFFF00000;
1739                 rxdctl |= (rxq->pthresh & 0x1F);
1740                 rxdctl |= ((rxq->hthresh & 0x1F) << 8);
1741                 rxdctl |= ((rxq->wthresh & 0x1F) << 16);
1742                 E1000_WRITE_REG(hw, E1000_RXDCTL(rxq->reg_idx), rxdctl);
1743         }
1744
1745         /*
1746          * Setup BSIZE field of RCTL register, if needed.
1747          * Buffer sizes >= 1024 are not [supposed to be] setup in the RCTL
1748          * register, since the code above configures the SRRCTL register of
1749          * the RX queue in such a case.
1750          * All configurable sizes are:
1751          * 16384: rctl |= (E1000_RCTL_SZ_16384 | E1000_RCTL_BSEX);
1752          *  8192: rctl |= (E1000_RCTL_SZ_8192  | E1000_RCTL_BSEX);
1753          *  4096: rctl |= (E1000_RCTL_SZ_4096  | E1000_RCTL_BSEX);
1754          *  2048: rctl |= E1000_RCTL_SZ_2048;
1755          *  1024: rctl |= E1000_RCTL_SZ_1024;
1756          *   512: rctl |= E1000_RCTL_SZ_512;
1757          *   256: rctl |= E1000_RCTL_SZ_256;
1758          */
1759         if (rctl_bsize > 0) {
1760                 if (rctl_bsize >= 512) /* 512 <= buf_size < 1024 - use 512 */
1761                         rctl |= E1000_RCTL_SZ_512;
1762                 else /* 256 <= buf_size < 512 - use 256 */
1763                         rctl |= E1000_RCTL_SZ_256;
1764         }
1765
1766         /*
1767          * Configure RSS if device configured with multiple RX queues.
1768          */
1769         igb_dev_mq_rx_configure(dev);
1770
1771         /*
1772          * Setup the Checksum Register.
1773          * Receive Full-Packet Checksum Offload is mutually exclusive with RSS.
1774          */
1775         rxcsum = E1000_READ_REG(hw, E1000_RXCSUM);
1776         rxcsum |= E1000_RXCSUM_PCSD;
1777
1778         /* Enable both L3/L4 rx checksum offload */
1779         if (dev->data->dev_conf.rxmode.hw_ip_checksum)
1780                 rxcsum |= (E1000_RXCSUM_IPOFL  | E1000_RXCSUM_TUOFL);
1781         else
1782                 rxcsum &= ~(E1000_RXCSUM_IPOFL | E1000_RXCSUM_TUOFL);
1783         E1000_WRITE_REG(hw, E1000_RXCSUM, rxcsum);
1784
1785         /* Setup the Receive Control Register. */
1786         if (dev->data->dev_conf.rxmode.hw_strip_crc) {
1787                 rctl |= E1000_RCTL_SECRC; /* Strip Ethernet CRC. */
1788
1789                 /* set STRCRC bit in all queues for Powerville/Springville */
1790                 if (hw->mac.type == e1000_i350 || hw->mac.type == e1000_i210) {
1791                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1792                                 rxq = dev->data->rx_queues[i];
1793                                 uint32_t dvmolr = E1000_READ_REG(hw,
1794                                         E1000_DVMOLR(rxq->reg_idx));
1795                                 dvmolr |= E1000_DVMOLR_STRCRC;
1796                                 E1000_WRITE_REG(hw, E1000_DVMOLR(rxq->reg_idx), dvmolr);
1797                         }
1798                 }
1799         } else {
1800                 rctl &= ~E1000_RCTL_SECRC; /* Do not Strip Ethernet CRC. */
1801
1802                 /* clear STRCRC bit in all queues for Powerville/Springville */
1803                 if (hw->mac.type == e1000_i350 || hw->mac.type == e1000_i210) {
1804                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1805                                 rxq = dev->data->rx_queues[i];
1806                                 uint32_t dvmolr = E1000_READ_REG(hw,
1807                                         E1000_DVMOLR(rxq->reg_idx));
1808                                 dvmolr &= ~E1000_DVMOLR_STRCRC;
1809                                 E1000_WRITE_REG(hw, E1000_DVMOLR(rxq->reg_idx), dvmolr);
1810                         }
1811                 }
1812         }
1813
1814         rctl &= ~(3 << E1000_RCTL_MO_SHIFT);
1815         rctl |= E1000_RCTL_EN | E1000_RCTL_BAM | E1000_RCTL_LBM_NO |
1816                 E1000_RCTL_RDMTS_HALF |
1817                 (hw->mac.mc_filter_type << E1000_RCTL_MO_SHIFT);
1818
1819         /* Make sure VLAN Filters are off. */
1820         rctl &= ~E1000_RCTL_VFE;
1821         /* Don't store bad packets. */
1822         rctl &= ~E1000_RCTL_SBP;
1823
1824         /* Enable Receives. */
1825         E1000_WRITE_REG(hw, E1000_RCTL, rctl);
1826
1827         /*
1828          * Setup the HW Rx Head and Tail Descriptor Pointers.
1829          * This needs to be done after enable.
1830          */
1831         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1832                 rxq = dev->data->rx_queues[i];
1833                 E1000_WRITE_REG(hw, E1000_RDH(rxq->reg_idx), 0);
1834                 E1000_WRITE_REG(hw, E1000_RDT(rxq->reg_idx), rxq->nb_rx_desc - 1);
1835         }
1836
1837         return 0;
1838 }
1839
1840 /*********************************************************************
1841  *
1842  *  Enable transmit unit.
1843  *
1844  **********************************************************************/
1845 void
1846 eth_igb_tx_init(struct rte_eth_dev *dev)
1847 {
1848         struct e1000_hw     *hw;
1849         struct igb_tx_queue *txq;
1850         uint32_t tctl;
1851         uint32_t txdctl;
1852         uint16_t i;
1853
1854         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1855
1856         /* Setup the Base and Length of the Tx Descriptor Rings. */
1857         for (i = 0; i < dev->data->nb_tx_queues; i++) {
1858                 uint64_t bus_addr;
1859                 txq = dev->data->tx_queues[i];
1860                 bus_addr = txq->tx_ring_phys_addr;
1861
1862                 E1000_WRITE_REG(hw, E1000_TDLEN(txq->reg_idx),
1863                                 txq->nb_tx_desc *
1864                                 sizeof(union e1000_adv_tx_desc));
1865                 E1000_WRITE_REG(hw, E1000_TDBAH(txq->reg_idx),
1866                                 (uint32_t)(bus_addr >> 32));
1867                 E1000_WRITE_REG(hw, E1000_TDBAL(txq->reg_idx), (uint32_t)bus_addr);
1868
1869                 /* Setup the HW Tx Head and Tail descriptor pointers. */
1870                 E1000_WRITE_REG(hw, E1000_TDT(txq->reg_idx), 0);
1871                 E1000_WRITE_REG(hw, E1000_TDH(txq->reg_idx), 0);
1872
1873                 /* Setup Transmit threshold registers. */
1874                 txdctl = E1000_READ_REG(hw, E1000_TXDCTL(txq->reg_idx));
1875                 txdctl |= txq->pthresh & 0x1F;
1876                 txdctl |= ((txq->hthresh & 0x1F) << 8);
1877                 txdctl |= ((txq->wthresh & 0x1F) << 16);
1878                 txdctl |= E1000_TXDCTL_QUEUE_ENABLE;
1879                 E1000_WRITE_REG(hw, E1000_TXDCTL(txq->reg_idx), txdctl);
1880         }
1881
1882         /* Program the Transmit Control Register. */
1883         tctl = E1000_READ_REG(hw, E1000_TCTL);
1884         tctl &= ~E1000_TCTL_CT;
1885         tctl |= (E1000_TCTL_PSP | E1000_TCTL_RTLC | E1000_TCTL_EN |
1886                  (E1000_COLLISION_THRESHOLD << E1000_CT_SHIFT));
1887
1888         e1000_config_collision_dist(hw);
1889
1890         /* This write will effectively turn on the transmit unit. */
1891         E1000_WRITE_REG(hw, E1000_TCTL, tctl);
1892 }
1893
1894 /*********************************************************************
1895  *
1896  *  Enable VF receive unit.
1897  *
1898  **********************************************************************/
1899 int
1900 eth_igbvf_rx_init(struct rte_eth_dev *dev)
1901 {
1902         struct e1000_hw     *hw;
1903         struct igb_rx_queue *rxq;
1904         struct rte_pktmbuf_pool_private *mbp_priv;
1905         uint32_t srrctl;
1906         uint16_t buf_size;
1907         uint16_t rctl_bsize;
1908         uint16_t i;
1909         int ret;
1910
1911         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1912
1913         /* Configure and enable each RX queue. */
1914         rctl_bsize = 0;
1915         dev->rx_pkt_burst = eth_igb_recv_pkts;
1916         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1917                 uint64_t bus_addr;
1918                 uint32_t rxdctl;
1919
1920                 rxq = dev->data->rx_queues[i];
1921
1922                 /* Allocate buffers for descriptor rings and set up queue */
1923                 ret = igb_alloc_rx_queue_mbufs(rxq);
1924                 if (ret)
1925                         return ret;
1926
1927                 bus_addr = rxq->rx_ring_phys_addr;
1928                 E1000_WRITE_REG(hw, E1000_RDLEN(i),
1929                                 rxq->nb_rx_desc *
1930                                 sizeof(union e1000_adv_rx_desc));
1931                 E1000_WRITE_REG(hw, E1000_RDBAH(i),
1932                                 (uint32_t)(bus_addr >> 32));
1933                 E1000_WRITE_REG(hw, E1000_RDBAL(i), (uint32_t)bus_addr);
1934
1935                 srrctl = E1000_SRRCTL_DESCTYPE_ADV_ONEBUF;
1936
1937                 /*
1938                  * Configure RX buffer size.
1939                  */
1940                 mbp_priv = (struct rte_pktmbuf_pool_private *)
1941                         ((char *)rxq->mb_pool + sizeof(struct rte_mempool));
1942                 buf_size = (uint16_t) (mbp_priv->mbuf_data_room_size -
1943                                        RTE_PKTMBUF_HEADROOM);
1944                 if (buf_size >= 1024) {
1945                         /*
1946                          * Configure the BSIZEPACKET field of the SRRCTL
1947                          * register of the queue.
1948                          * Value is in 1 KB resolution, from 1 KB to 127 KB.
1949                          * If this field is equal to 0b, then RCTL.BSIZE
1950                          * determines the RX packet buffer size.
1951                          */
1952                         srrctl |= ((buf_size >> E1000_SRRCTL_BSIZEPKT_SHIFT) &
1953                                    E1000_SRRCTL_BSIZEPKT_MASK);
1954                         buf_size = (uint16_t) ((srrctl &
1955                                                 E1000_SRRCTL_BSIZEPKT_MASK) <<
1956                                                E1000_SRRCTL_BSIZEPKT_SHIFT);
1957
1958                         /* It adds dual VLAN length for supporting dual VLAN */
1959                         if ((dev->data->dev_conf.rxmode.max_rx_pkt_len +
1960                                                 2 * VLAN_TAG_SIZE) > buf_size){
1961                                 dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
1962                                 dev->data->scattered_rx = 1;
1963                         }
1964                 } else {
1965                         /*
1966                          * Use BSIZE field of the device RCTL register.
1967                          */
1968                         if ((rctl_bsize == 0) || (rctl_bsize > buf_size))
1969                                 rctl_bsize = buf_size;
1970                         dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
1971                         dev->data->scattered_rx = 1;
1972                 }
1973
1974                 /* Set if packets are dropped when no descriptors available */
1975                 if (rxq->drop_en)
1976                         srrctl |= E1000_SRRCTL_DROP_EN;
1977
1978                 E1000_WRITE_REG(hw, E1000_SRRCTL(i), srrctl);
1979
1980                 /* Enable this RX queue. */
1981                 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(i));
1982                 rxdctl |= E1000_RXDCTL_QUEUE_ENABLE;
1983                 rxdctl &= 0xFFF00000;
1984                 rxdctl |= (rxq->pthresh & 0x1F);
1985                 rxdctl |= ((rxq->hthresh & 0x1F) << 8);
1986                 if (hw->mac.type == e1000_82576) {
1987                         /*
1988                          * Workaround of 82576 VF Erratum
1989                          * force set WTHRESH to 1
1990                          * to avoid Write-Back not triggered sometimes
1991                          */
1992                         rxdctl |= 0x10000;
1993                         PMD_INIT_LOG(DEBUG, "Force set RX WTHRESH to 1 !\n");
1994                 }
1995                 else
1996                         rxdctl |= ((rxq->wthresh & 0x1F) << 16);
1997                 E1000_WRITE_REG(hw, E1000_RXDCTL(i), rxdctl);
1998         }
1999
2000         /*
2001          * Setup the HW Rx Head and Tail Descriptor Pointers.
2002          * This needs to be done after enable.
2003          */
2004         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2005                 rxq = dev->data->rx_queues[i];
2006                 E1000_WRITE_REG(hw, E1000_RDH(i), 0);
2007                 E1000_WRITE_REG(hw, E1000_RDT(i), rxq->nb_rx_desc - 1);
2008         }
2009
2010         return 0;
2011 }
2012
2013 /*********************************************************************
2014  *
2015  *  Enable VF transmit unit.
2016  *
2017  **********************************************************************/
2018 void
2019 eth_igbvf_tx_init(struct rte_eth_dev *dev)
2020 {
2021         struct e1000_hw     *hw;
2022         struct igb_tx_queue *txq;
2023         uint32_t txdctl;
2024         uint16_t i;
2025
2026         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2027
2028         /* Setup the Base and Length of the Tx Descriptor Rings. */
2029         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2030                 uint64_t bus_addr;
2031
2032                 txq = dev->data->tx_queues[i];
2033                 bus_addr = txq->tx_ring_phys_addr;
2034                 E1000_WRITE_REG(hw, E1000_TDLEN(i),
2035                                 txq->nb_tx_desc *
2036                                 sizeof(union e1000_adv_tx_desc));
2037                 E1000_WRITE_REG(hw, E1000_TDBAH(i),
2038                                 (uint32_t)(bus_addr >> 32));
2039                 E1000_WRITE_REG(hw, E1000_TDBAL(i), (uint32_t)bus_addr);
2040
2041                 /* Setup the HW Tx Head and Tail descriptor pointers. */
2042                 E1000_WRITE_REG(hw, E1000_TDT(i), 0);
2043                 E1000_WRITE_REG(hw, E1000_TDH(i), 0);
2044
2045                 /* Setup Transmit threshold registers. */
2046                 txdctl = E1000_READ_REG(hw, E1000_TXDCTL(i));
2047                 txdctl |= txq->pthresh & 0x1F;
2048                 txdctl |= ((txq->hthresh & 0x1F) << 8);
2049                 if (hw->mac.type == e1000_82576) {
2050                         /*
2051                          * Workaround of 82576 VF Erratum
2052                          * force set WTHRESH to 1
2053                          * to avoid Write-Back not triggered sometimes
2054                          */
2055                         txdctl |= 0x10000;
2056                         PMD_INIT_LOG(DEBUG, "Force set TX WTHRESH to 1 !\n");
2057                 }
2058                 else
2059                         txdctl |= ((txq->wthresh & 0x1F) << 16);
2060                 txdctl |= E1000_TXDCTL_QUEUE_ENABLE;
2061                 E1000_WRITE_REG(hw, E1000_TXDCTL(i), txdctl);
2062         }
2063
2064 }
2065