drivers/net/virtio/virtio_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 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 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38 #include <errno.h>
  39
  40 #include <rte_cycles.h>
  41 #include <rte_memory.h>
  42 #include <rte_memzone.h>
  43 #include <rte_branch_prediction.h>
  44 #include <rte_mempool.h>
  45 #include <rte_malloc.h>
  46 #include <rte_mbuf.h>
  47 #include <rte_ether.h>
  48 #include <rte_ethdev.h>
  49 #include <rte_prefetch.h>
  50 #include <rte_string_fns.h>
  51 #include <rte_errno.h>
  52 #include <rte_byteorder.h>
  53 #include <rte_cpuflags.h>
  54 #include <rte_net.h>
  55 #include <rte_ip.h>
  56 #include <rte_udp.h>
  57 #include <rte_tcp.h>
  58
  59 #include "virtio_logs.h"
  60 #include "virtio_ethdev.h"
  61 #include "virtio_pci.h"
  62 #include "virtqueue.h"
  63 #include "virtio_rxtx.h"
  64
  65 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  66 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  67 #else
  68 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  69 #endif
  70
  71
  72 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
  73         ETH_TXQ_FLAGS_NOOFFLOADS)
  74
  75 static void
  76 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  77 {
  78         struct vring_desc *dp, *dp_tail;
  79         struct vq_desc_extra *dxp;
  80         uint16_t desc_idx_last = desc_idx;
  81
  82         dp  = &vq->vq_ring.desc[desc_idx];
  83         dxp = &vq->vq_descx[desc_idx];
  84         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  85         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  86                 while (dp->flags & VRING_DESC_F_NEXT) {
  87                         desc_idx_last = dp->next;
  88                         dp = &vq->vq_ring.desc[dp->next];
  89                 }
  90         }
  91         dxp->ndescs = 0;
  92
  93         /*
  94          * We must append the existing free chain, if any, to the end of
  95          * newly freed chain. If the virtqueue was completely used, then
  96          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  97          */
  98         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  99                 vq->vq_desc_head_idx = desc_idx;
 100         } else {
 101                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
 102                 dp_tail->next = desc_idx;
 103         }
 104
 105         vq->vq_desc_tail_idx = desc_idx_last;
 106         dp->next = VQ_RING_DESC_CHAIN_END;
 107 }
 108
 109 static uint16_t
 110 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 111                            uint32_t *len, uint16_t num)
 112 {
 113         struct vring_used_elem *uep;
 114         struct rte_mbuf *cookie;
 115         uint16_t used_idx, desc_idx;
 116         uint16_t i;
 117
 118         /*  Caller does the check */
 119         for (i = 0; i < num ; i++) {
 120                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 121                 uep = &vq->vq_ring.used->ring[used_idx];
 122                 desc_idx = (uint16_t) uep->id;
 123                 len[i] = uep->len;
 124                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 125
 126                 if (unlikely(cookie == NULL)) {
 127                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 128                                 vq->vq_used_cons_idx);
 129                         break;
 130                 }
 131
 132                 rte_prefetch0(cookie);
 133                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 134                 rx_pkts[i]  = cookie;
 135                 vq->vq_used_cons_idx++;
 136                 vq_ring_free_chain(vq, desc_idx);
 137                 vq->vq_descx[desc_idx].cookie = NULL;
 138         }
 139
 140         return i;
 141 }
 142
 143 #ifndef DEFAULT_TX_FREE_THRESH
 144 #define DEFAULT_TX_FREE_THRESH 32
 145 #endif
 146
 147 /* Cleanup from completed transmits. */
 148 static void
 149 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 150 {
 151         uint16_t i, used_idx, desc_idx;
 152         for (i = 0; i < num; i++) {
 153                 struct vring_used_elem *uep;
 154                 struct vq_desc_extra *dxp;
 155
 156                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 157                 uep = &vq->vq_ring.used->ring[used_idx];
 158
 159                 desc_idx = (uint16_t) uep->id;
 160                 dxp = &vq->vq_descx[desc_idx];
 161                 vq->vq_used_cons_idx++;
 162                 vq_ring_free_chain(vq, desc_idx);
 163
 164                 if (dxp->cookie != NULL) {
 165                         rte_pktmbuf_free(dxp->cookie);
 166                         dxp->cookie = NULL;
 167                 }
 168         }
 169 }
 170
 171
 172 static inline int
 173 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 174 {
 175         struct vq_desc_extra *dxp;
 176         struct virtio_hw *hw = vq->hw;
 177         struct vring_desc *start_dp;
 178         uint16_t needed = 1;
 179         uint16_t head_idx, idx;
 180
 181         if (unlikely(vq->vq_free_cnt == 0))
 182                 return -ENOSPC;
 183         if (unlikely(vq->vq_free_cnt < needed))
 184                 return -EMSGSIZE;
 185
 186         head_idx = vq->vq_desc_head_idx;
 187         if (unlikely(head_idx >= vq->vq_nentries))
 188                 return -EFAULT;
 189
 190         idx = head_idx;
 191         dxp = &vq->vq_descx[idx];
 192         dxp->cookie = (void *)cookie;
 193         dxp->ndescs = needed;
 194
 195         start_dp = vq->vq_ring.desc;
 196         start_dp[idx].addr =
 197                 VIRTIO_MBUF_ADDR(cookie, vq) +
 198                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 199         start_dp[idx].len =
 200                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 201         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 202         idx = start_dp[idx].next;
 203         vq->vq_desc_head_idx = idx;
 204         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 205                 vq->vq_desc_tail_idx = idx;
 206         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 207         vq_update_avail_ring(vq, head_idx);
 208
 209         return 0;
 210 }
 211
 212 /* When doing TSO, the IP length is not included in the pseudo header
 213  * checksum of the packet given to the PMD, but for virtio it is
 214  * expected.
 215  */
 216 static void
 217 virtio_tso_fix_cksum(struct rte_mbuf *m)
 218 {
 219         /* common case: header is not fragmented */
 220         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 221                         m->l4_len)) {
 222                 struct ipv4_hdr *iph;
 223                 struct ipv6_hdr *ip6h;
 224                 struct tcp_hdr *th;
 225                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 226                 uint32_t tmp;
 227
 228                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 229                 th = RTE_PTR_ADD(iph, m->l3_len);
 230                 if ((iph->version_ihl >> 4) == 4) {
 231                         iph->hdr_checksum = 0;
 232                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 233                         ip_len = iph->total_length;
 234                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 235                                 m->l3_len);
 236                 } else {
 237                         ip6h = (struct ipv6_hdr *)iph;
 238                         ip_paylen = ip6h->payload_len;
 239                 }
 240
 241                 /* calculate the new phdr checksum not including ip_paylen */
 242                 prev_cksum = th->cksum;
 243                 tmp = prev_cksum;
 244                 tmp += ip_paylen;
 245                 tmp = (tmp & 0xffff) + (tmp >> 16);
 246                 new_cksum = tmp;
 247
 248                 /* replace it in the packet */
 249                 th->cksum = new_cksum;
 250         }
 251 }
 252
 253 static inline int
 254 tx_offload_enabled(struct virtio_hw *hw)
 255 {
 256         return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
 257                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
 258                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
 259 }
 260
 261 static inline void
 262 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 263                        uint16_t needed, int use_indirect, int can_push)
 264 {
 265         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 266         struct vq_desc_extra *dxp;
 267         struct virtqueue *vq = txvq->vq;
 268         struct vring_desc *start_dp;
 269         uint16_t seg_num = cookie->nb_segs;
 270         uint16_t head_idx, idx;
 271         uint16_t head_size = vq->hw->vtnet_hdr_size;
 272         struct virtio_net_hdr *hdr;
 273         int offload;
 274
 275         offload = tx_offload_enabled(vq->hw);
 276         head_idx = vq->vq_desc_head_idx;
 277         idx = head_idx;
 278         dxp = &vq->vq_descx[idx];
 279         dxp->cookie = (void *)cookie;
 280         dxp->ndescs = needed;
 281
 282         start_dp = vq->vq_ring.desc;
 283
 284         if (can_push) {
 285                 /* prepend cannot fail, checked by caller */
 286                 hdr = (struct virtio_net_hdr *)
 287                         rte_pktmbuf_prepend(cookie, head_size);
 288                 /* if offload disabled, it is not zeroed below, do it now */
 289                 if (offload == 0)
 290                         memset(hdr, 0, head_size);
 291         } else if (use_indirect) {
 292                 /* setup tx ring slot to point to indirect
 293                  * descriptor list stored in reserved region.
 294                  *
 295                  * the first slot in indirect ring is already preset
 296                  * to point to the header in reserved region
 297                  */
 298                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 299                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 300                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 301                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 302                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 303
 304                 /* loop below will fill in rest of the indirect elements */
 305                 start_dp = txr[idx].tx_indir;
 306                 idx = 1;
 307         } else {
 308                 /* setup first tx ring slot to point to header
 309                  * stored in reserved region.
 310                  */
 311                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 312                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 313                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 314                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 315                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 316
 317                 idx = start_dp[idx].next;
 318         }
 319
 320         /* Checksum Offload / TSO */
 321         if (offload) {
 322                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 323                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 324
 325                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 326                 case PKT_TX_UDP_CKSUM:
 327                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 328                         hdr->csum_offset = offsetof(struct udp_hdr,
 329                                 dgram_cksum);
 330                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 331                         break;
 332
 333                 case PKT_TX_TCP_CKSUM:
 334                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 335                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 336                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 337                         break;
 338
 339                 default:
 340                         hdr->csum_start = 0;
 341                         hdr->csum_offset = 0;
 342                         hdr->flags = 0;
 343                         break;
 344                 }
 345
 346                 /* TCP Segmentation Offload */
 347                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 348                         virtio_tso_fix_cksum(cookie);
 349                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 350                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 351                                 VIRTIO_NET_HDR_GSO_TCPV4;
 352                         hdr->gso_size = cookie->tso_segsz;
 353                         hdr->hdr_len =
 354                                 cookie->l2_len +
 355                                 cookie->l3_len +
 356                                 cookie->l4_len;
 357                 } else {
 358                         hdr->gso_type = 0;
 359                         hdr->gso_size = 0;
 360                         hdr->hdr_len = 0;
 361                 }
 362         }
 363
 364         do {
 365                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 366                 start_dp[idx].len   = cookie->data_len;
 367                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 368                 idx = start_dp[idx].next;
 369         } while ((cookie = cookie->next) != NULL);
 370
 371         if (use_indirect)
 372                 idx = vq->vq_ring.desc[head_idx].next;
 373
 374         vq->vq_desc_head_idx = idx;
 375         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 376                 vq->vq_desc_tail_idx = idx;
 377         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 378         vq_update_avail_ring(vq, head_idx);
 379 }
 380
 381 void
 382 virtio_dev_cq_start(struct rte_eth_dev *dev)
 383 {
 384         struct virtio_hw *hw = dev->data->dev_private;
 385
 386         if (hw->cvq && hw->cvq->vq) {
 387                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 388         }
 389 }
 390
 391 void
 392 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 393 {
 394         /*
 395          * Start receive and transmit vrings
 396          * -    Setup vring structure for all queues
 397          * -    Initialize descriptor for the rx vring
 398          * -    Allocate blank mbufs for the each rx descriptor
 399          *
 400          */
 401         uint16_t i;
 402         uint16_t desc_idx;
 403         struct virtio_hw *hw = dev->data->dev_private;
 404
 405         PMD_INIT_FUNC_TRACE();
 406
 407         /* Start rx vring. */
 408         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 409                 struct virtnet_rx *rxvq = dev->data->rx_queues[i];
 410                 struct virtqueue *vq = rxvq->vq;
 411                 int error, nbufs;
 412                 struct rte_mbuf *m;
 413
 414                 if (rxvq->mpool == NULL) {
 415                         rte_exit(EXIT_FAILURE,
 416                                 "Cannot allocate mbufs for rx virtqueue");
 417                 }
 418
 419                 /* Allocate blank mbufs for the each rx descriptor */
 420                 nbufs = 0;
 421                 error = ENOSPC;
 422
 423                 if (hw->use_simple_rxtx) {
 424                         for (desc_idx = 0; desc_idx < vq->vq_nentries;
 425                              desc_idx++) {
 426                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 427                                 vq->vq_ring.desc[desc_idx].flags =
 428                                         VRING_DESC_F_WRITE;
 429                         }
 430                 }
 431
 432                 memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 433                 for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 434                      desc_idx++) {
 435                         vq->sw_ring[vq->vq_nentries + desc_idx] =
 436                                 &rxvq->fake_mbuf;
 437                 }
 438
 439                 while (!virtqueue_full(vq)) {
 440                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 441                         if (m == NULL)
 442                                 break;
 443
 444                         /******************************************
 445                         *         Enqueue allocated buffers        *
 446                         *******************************************/
 447                         if (hw->use_simple_rxtx)
 448                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 449                         else
 450                                 error = virtqueue_enqueue_recv_refill(vq, m);
 451
 452                         if (error) {
 453                                 rte_pktmbuf_free(m);
 454                                 break;
 455                         }
 456                         nbufs++;
 457                 }
 458
 459                 vq_update_avail_idx(vq);
 460
 461                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 462
 463                 VIRTQUEUE_DUMP(vq);
 464         }
 465
 466         /* Start tx vring. */
 467         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 468                 struct virtnet_tx *txvq = dev->data->tx_queues[i];
 469                 struct virtqueue *vq = txvq->vq;
 470
 471                 if (hw->use_simple_rxtx) {
 472                         uint16_t mid_idx  = vq->vq_nentries >> 1;
 473
 474                         for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 475                                 vq->vq_ring.avail->ring[desc_idx] =
 476                                         desc_idx + mid_idx;
 477                                 vq->vq_ring.desc[desc_idx + mid_idx].next =
 478                                         desc_idx;
 479                                 vq->vq_ring.desc[desc_idx + mid_idx].addr =
 480                                         txvq->virtio_net_hdr_mem +
 481                                         offsetof(struct virtio_tx_region, tx_hdr);
 482                                 vq->vq_ring.desc[desc_idx + mid_idx].len =
 483                                         vq->hw->vtnet_hdr_size;
 484                                 vq->vq_ring.desc[desc_idx + mid_idx].flags =
 485                                         VRING_DESC_F_NEXT;
 486                                 vq->vq_ring.desc[desc_idx].flags = 0;
 487                         }
 488                         for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 489                              desc_idx++)
 490                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 491                 }
 492
 493                 VIRTQUEUE_DUMP(vq);
 494         }
 495 }
 496
 497 int
 498 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 499                         uint16_t queue_idx,
 500                         uint16_t nb_desc,
 501                         unsigned int socket_id __rte_unused,
 502                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 503                         struct rte_mempool *mp)
 504 {
 505         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 506         struct virtio_hw *hw = dev->data->dev_private;
 507         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 508         struct virtnet_rx *rxvq;
 509
 510         PMD_INIT_FUNC_TRACE();
 511
 512         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 513                 nb_desc = vq->vq_nentries;
 514         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 515
 516         rxvq = &vq->rxq;
 517         rxvq->mpool = mp;
 518         rxvq->queue_id = queue_idx;
 519
 520         dev->data->rx_queues[queue_idx] = rxvq;
 521
 522         virtio_rxq_vec_setup(rxvq);
 523
 524         return 0;
 525 }
 526
 527 static void
 528 virtio_update_rxtx_handler(struct rte_eth_dev *dev,
 529                            const struct rte_eth_txconf *tx_conf)
 530 {
 531         uint8_t use_simple_rxtx = 0;
 532         struct virtio_hw *hw = dev->data->dev_private;
 533
 534 #if defined RTE_ARCH_X86
 535         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE3))
 536                 use_simple_rxtx = 1;
 537 #elif defined RTE_ARCH_ARM64 || defined CONFIG_RTE_ARCH_ARM
 538         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
 539                 use_simple_rxtx = 1;
 540 #endif
 541         /* Use simple rx/tx func if single segment and no offloads */
 542         if (use_simple_rxtx &&
 543             (tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 544             !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 545                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 546                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 547                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 548                 hw->use_simple_rxtx = use_simple_rxtx;
 549         }
 550 }
 551
 552 /*
 553  * struct rte_eth_dev *dev: Used to update dev
 554  * uint16_t nb_desc: Defaults to values read from config space
 555  * unsigned int socket_id: Used to allocate memzone
 556  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 557  * uint16_t queue_idx: Just used as an index in dev txq list
 558  */
 559 int
 560 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 561                         uint16_t queue_idx,
 562                         uint16_t nb_desc,
 563                         unsigned int socket_id __rte_unused,
 564                         const struct rte_eth_txconf *tx_conf)
 565 {
 566         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 567         struct virtio_hw *hw = dev->data->dev_private;
 568         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 569         struct virtnet_tx *txvq;
 570         uint16_t tx_free_thresh;
 571
 572         PMD_INIT_FUNC_TRACE();
 573
 574         virtio_update_rxtx_handler(dev, tx_conf);
 575
 576         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 577                 nb_desc = vq->vq_nentries;
 578         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 579
 580         txvq = &vq->txq;
 581         txvq->queue_id = queue_idx;
 582
 583         tx_free_thresh = tx_conf->tx_free_thresh;
 584         if (tx_free_thresh == 0)
 585                 tx_free_thresh =
 586                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 587
 588         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 589                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 590                         "number of TX entries minus 3 (%u)."
 591                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 592                         vq->vq_nentries - 3,
 593                         tx_free_thresh, dev->data->port_id, queue_idx);
 594                 return -EINVAL;
 595         }
 596
 597         vq->vq_free_thresh = tx_free_thresh;
 598
 599         dev->data->tx_queues[queue_idx] = txvq;
 600         return 0;
 601 }
 602
 603 static void
 604 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 605 {
 606         int error;
 607         /*
 608          * Requeue the discarded mbuf. This should always be
 609          * successful since it was just dequeued.
 610          */
 611         error = virtqueue_enqueue_recv_refill(vq, m);
 612         if (unlikely(error)) {
 613                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 614                 rte_pktmbuf_free(m);
 615         }
 616 }
 617
 618 static void
 619 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 620 {
 621         uint32_t s = mbuf->pkt_len;
 622         struct ether_addr *ea;
 623
 624         if (s == 64) {
 625                 stats->size_bins[1]++;
 626         } else if (s > 64 && s < 1024) {
 627                 uint32_t bin;
 628
 629                 /* count zeros, and offset into correct bin */
 630                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 631                 stats->size_bins[bin]++;
 632         } else {
 633                 if (s < 64)
 634                         stats->size_bins[0]++;
 635                 else if (s < 1519)
 636                         stats->size_bins[6]++;
 637                 else if (s >= 1519)
 638                         stats->size_bins[7]++;
 639         }
 640
 641         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 642         if (is_multicast_ether_addr(ea)) {
 643                 if (is_broadcast_ether_addr(ea))
 644                         stats->broadcast++;
 645                 else
 646                         stats->multicast++;
 647         }
 648 }
 649
 650 /* Optionally fill offload information in structure */
 651 static int
 652 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 653 {
 654         struct rte_net_hdr_lens hdr_lens;
 655         uint32_t hdrlen, ptype;
 656         int l4_supported = 0;
 657
 658         /* nothing to do */
 659         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 660                 return 0;
 661
 662         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 663
 664         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 665         m->packet_type = ptype;
 666         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 667             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 668             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 669                 l4_supported = 1;
 670
 671         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 672                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 673                 if (hdr->csum_start <= hdrlen && l4_supported) {
 674                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 675                 } else {
 676                         /* Unknown proto or tunnel, do sw cksum. We can assume
 677                          * the cksum field is in the first segment since the
 678                          * buffers we provided to the host are large enough.
 679                          * In case of SCTP, this will be wrong since it's a CRC
 680                          * but there's nothing we can do.
 681                          */
 682                         uint16_t csum, off;
 683
 684                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 685                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 686                                 &csum);
 687                         if (likely(csum != 0xffff))
 688                                 csum = ~csum;
 689                         off = hdr->csum_offset + hdr->csum_start;
 690                         if (rte_pktmbuf_data_len(m) >= off + 1)
 691                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 692                                         off) = csum;
 693                 }
 694         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 695                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 696         }
 697
 698         /* GSO request, save required information in mbuf */
 699         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 700                 /* Check unsupported modes */
 701                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 702                     (hdr->gso_size == 0)) {
 703                         return -EINVAL;
 704                 }
 705
 706                 /* Update mss lengthes in mbuf */
 707                 m->tso_segsz = hdr->gso_size;
 708                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 709                         case VIRTIO_NET_HDR_GSO_TCPV4:
 710                         case VIRTIO_NET_HDR_GSO_TCPV6:
 711                                 m->ol_flags |= PKT_RX_LRO | \
 712                                         PKT_RX_L4_CKSUM_NONE;
 713                                 break;
 714                         default:
 715                                 return -EINVAL;
 716                 }
 717         }
 718
 719         return 0;
 720 }
 721
 722 static inline int
 723 rx_offload_enabled(struct virtio_hw *hw)
 724 {
 725         return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
 726                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
 727                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
 728 }
 729
 730 #define VIRTIO_MBUF_BURST_SZ 64
 731 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 732 uint16_t
 733 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 734 {
 735         struct virtnet_rx *rxvq = rx_queue;
 736         struct virtqueue *vq = rxvq->vq;
 737         struct virtio_hw *hw;
 738         struct rte_mbuf *rxm, *new_mbuf;
 739         uint16_t nb_used, num, nb_rx;
 740         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 741         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 742         int error;
 743         uint32_t i, nb_enqueued;
 744         uint32_t hdr_size;
 745         int offload;
 746         struct virtio_net_hdr *hdr;
 747
 748         nb_used = VIRTQUEUE_NUSED(vq);
 749
 750         virtio_rmb();
 751
 752         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 753         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 754         if (likely(num > DESC_PER_CACHELINE))
 755                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 756
 757         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 758         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 759
 760         hw = vq->hw;
 761         nb_rx = 0;
 762         nb_enqueued = 0;
 763         hdr_size = hw->vtnet_hdr_size;
 764         offload = rx_offload_enabled(hw);
 765
 766         for (i = 0; i < num ; i++) {
 767                 rxm = rcv_pkts[i];
 768
 769                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 770
 771                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 772                         PMD_RX_LOG(ERR, "Packet drop");
 773                         nb_enqueued++;
 774                         virtio_discard_rxbuf(vq, rxm);
 775                         rxvq->stats.errors++;
 776                         continue;
 777                 }
 778
 779                 rxm->port = rxvq->port_id;
 780                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 781                 rxm->ol_flags = 0;
 782                 rxm->vlan_tci = 0;
 783
 784                 rxm->nb_segs = 1;
 785                 rxm->next = NULL;
 786                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 787                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 788
 789                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 790                         RTE_PKTMBUF_HEADROOM - hdr_size);
 791
 792                 if (hw->vlan_strip)
 793                         rte_vlan_strip(rxm);
 794
 795                 if (offload && virtio_rx_offload(rxm, hdr) < 0) {
 796                         virtio_discard_rxbuf(vq, rxm);
 797                         rxvq->stats.errors++;
 798                         continue;
 799                 }
 800
 801                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 802
 803                 rx_pkts[nb_rx++] = rxm;
 804
 805                 rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
 806                 virtio_update_packet_stats(&rxvq->stats, rxm);
 807         }
 808
 809         rxvq->stats.packets += nb_rx;
 810
 811         /* Allocate new mbuf for the used descriptor */
 812         error = ENOSPC;
 813         while (likely(!virtqueue_full(vq))) {
 814                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 815                 if (unlikely(new_mbuf == NULL)) {
 816                         struct rte_eth_dev *dev
 817                                 = &rte_eth_devices[rxvq->port_id];
 818                         dev->data->rx_mbuf_alloc_failed++;
 819                         break;
 820                 }
 821                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 822                 if (unlikely(error)) {
 823                         rte_pktmbuf_free(new_mbuf);
 824                         break;
 825                 }
 826                 nb_enqueued++;
 827         }
 828
 829         if (likely(nb_enqueued)) {
 830                 vq_update_avail_idx(vq);
 831
 832                 if (unlikely(virtqueue_kick_prepare(vq))) {
 833                         virtqueue_notify(vq);
 834                         PMD_RX_LOG(DEBUG, "Notified");
 835                 }
 836         }
 837
 838         return nb_rx;
 839 }
 840
 841 uint16_t
 842 virtio_recv_mergeable_pkts(void *rx_queue,
 843                         struct rte_mbuf **rx_pkts,
 844                         uint16_t nb_pkts)
 845 {
 846         struct virtnet_rx *rxvq = rx_queue;
 847         struct virtqueue *vq = rxvq->vq;
 848         struct virtio_hw *hw;
 849         struct rte_mbuf *rxm, *new_mbuf;
 850         uint16_t nb_used, num, nb_rx;
 851         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 852         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 853         struct rte_mbuf *prev;
 854         int error;
 855         uint32_t i, nb_enqueued;
 856         uint32_t seg_num;
 857         uint16_t extra_idx;
 858         uint32_t seg_res;
 859         uint32_t hdr_size;
 860         int offload;
 861
 862         nb_used = VIRTQUEUE_NUSED(vq);
 863
 864         virtio_rmb();
 865
 866         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 867
 868         hw = vq->hw;
 869         nb_rx = 0;
 870         i = 0;
 871         nb_enqueued = 0;
 872         seg_num = 0;
 873         extra_idx = 0;
 874         seg_res = 0;
 875         hdr_size = hw->vtnet_hdr_size;
 876         offload = rx_offload_enabled(hw);
 877
 878         while (i < nb_used) {
 879                 struct virtio_net_hdr_mrg_rxbuf *header;
 880
 881                 if (nb_rx == nb_pkts)
 882                         break;
 883
 884                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
 885                 if (num != 1)
 886                         continue;
 887
 888                 i++;
 889
 890                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 891                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 892
 893                 rxm = rcv_pkts[0];
 894
 895                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 896                         PMD_RX_LOG(ERR, "Packet drop");
 897                         nb_enqueued++;
 898                         virtio_discard_rxbuf(vq, rxm);
 899                         rxvq->stats.errors++;
 900                         continue;
 901                 }
 902
 903                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 904                         RTE_PKTMBUF_HEADROOM - hdr_size);
 905                 seg_num = header->num_buffers;
 906
 907                 if (seg_num == 0)
 908                         seg_num = 1;
 909
 910                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 911                 rxm->nb_segs = seg_num;
 912                 rxm->next = NULL;
 913                 rxm->ol_flags = 0;
 914                 rxm->vlan_tci = 0;
 915                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 916                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 917
 918                 rxm->port = rxvq->port_id;
 919                 rx_pkts[nb_rx] = rxm;
 920                 prev = rxm;
 921
 922                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
 923                         virtio_discard_rxbuf(vq, rxm);
 924                         rxvq->stats.errors++;
 925                         continue;
 926                 }
 927
 928                 seg_res = seg_num - 1;
 929
 930                 while (seg_res != 0) {
 931                         /*
 932                          * Get extra segments for current uncompleted packet.
 933                          */
 934                         uint16_t  rcv_cnt =
 935                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 936                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
 937                                 uint32_t rx_num =
 938                                         virtqueue_dequeue_burst_rx(vq,
 939                                         rcv_pkts, len, rcv_cnt);
 940                                 i += rx_num;
 941                                 rcv_cnt = rx_num;
 942                         } else {
 943                                 PMD_RX_LOG(ERR,
 944                                            "No enough segments for packet.");
 945                                 nb_enqueued++;
 946                                 virtio_discard_rxbuf(vq, rxm);
 947                                 rxvq->stats.errors++;
 948                                 break;
 949                         }
 950
 951                         extra_idx = 0;
 952
 953                         while (extra_idx < rcv_cnt) {
 954                                 rxm = rcv_pkts[extra_idx];
 955
 956                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 957                                 rxm->next = NULL;
 958                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 959                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 960
 961                                 if (prev)
 962                                         prev->next = rxm;
 963
 964                                 prev = rxm;
 965                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 966                                 extra_idx++;
 967                         };
 968                         seg_res -= rcv_cnt;
 969                 }
 970
 971                 if (hw->vlan_strip)
 972                         rte_vlan_strip(rx_pkts[nb_rx]);
 973
 974                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 975                         rx_pkts[nb_rx]->data_len);
 976
 977                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
 978                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
 979                 nb_rx++;
 980         }
 981
 982         rxvq->stats.packets += nb_rx;
 983
 984         /* Allocate new mbuf for the used descriptor */
 985         error = ENOSPC;
 986         while (likely(!virtqueue_full(vq))) {
 987                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 988                 if (unlikely(new_mbuf == NULL)) {
 989                         struct rte_eth_dev *dev
 990                                 = &rte_eth_devices[rxvq->port_id];
 991                         dev->data->rx_mbuf_alloc_failed++;
 992                         break;
 993                 }
 994                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 995                 if (unlikely(error)) {
 996                         rte_pktmbuf_free(new_mbuf);
 997                         break;
 998                 }
 999                 nb_enqueued++;
1000         }
1001
1002         if (likely(nb_enqueued)) {
1003                 vq_update_avail_idx(vq);
1004
1005                 if (unlikely(virtqueue_kick_prepare(vq))) {
1006                         virtqueue_notify(vq);
1007                         PMD_RX_LOG(DEBUG, "Notified");
1008                 }
1009         }
1010
1011         return nb_rx;
1012 }
1013
1014 uint16_t
1015 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1016 {
1017         struct virtnet_tx *txvq = tx_queue;
1018         struct virtqueue *vq = txvq->vq;
1019         struct virtio_hw *hw = vq->hw;
1020         uint16_t hdr_size = hw->vtnet_hdr_size;
1021         uint16_t nb_used, nb_tx;
1022         int error;
1023
1024         if (unlikely(nb_pkts < 1))
1025                 return nb_pkts;
1026
1027         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1028         nb_used = VIRTQUEUE_NUSED(vq);
1029
1030         virtio_rmb();
1031         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1032                 virtio_xmit_cleanup(vq, nb_used);
1033
1034         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1035                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1036                 int can_push = 0, use_indirect = 0, slots, need;
1037
1038                 /* Do VLAN tag insertion */
1039                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1040                         error = rte_vlan_insert(&txm);
1041                         if (unlikely(error)) {
1042                                 rte_pktmbuf_free(txm);
1043                                 continue;
1044                         }
1045                 }
1046
1047                 /* optimize ring usage */
1048                 if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
1049                     rte_mbuf_refcnt_read(txm) == 1 &&
1050                     RTE_MBUF_DIRECT(txm) &&
1051                     txm->nb_segs == 1 &&
1052                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1053                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1054                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1055                         can_push = 1;
1056                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1057                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1058                         use_indirect = 1;
1059
1060                 /* How many main ring entries are needed to this Tx?
1061                  * any_layout => number of segments
1062                  * indirect   => 1
1063                  * default    => number of segments + 1
1064                  */
1065                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1066                 need = slots - vq->vq_free_cnt;
1067
1068                 /* Positive value indicates it need free vring descriptors */
1069                 if (unlikely(need > 0)) {
1070                         nb_used = VIRTQUEUE_NUSED(vq);
1071                         virtio_rmb();
1072                         need = RTE_MIN(need, (int)nb_used);
1073
1074                         virtio_xmit_cleanup(vq, need);
1075                         need = slots - vq->vq_free_cnt;
1076                         if (unlikely(need > 0)) {
1077                                 PMD_TX_LOG(ERR,
1078                                            "No free tx descriptors to transmit");
1079                                 break;
1080                         }
1081                 }
1082
1083                 /* Enqueue Packet buffers */
1084                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1085
1086                 txvq->stats.bytes += txm->pkt_len;
1087                 virtio_update_packet_stats(&txvq->stats, txm);
1088         }
1089
1090         txvq->stats.packets += nb_tx;
1091
1092         if (likely(nb_tx)) {
1093                 vq_update_avail_idx(vq);
1094
1095                 if (unlikely(virtqueue_kick_prepare(vq))) {
1096                         virtqueue_notify(vq);
1097                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1098                 }
1099         }
1100
1101         return nb_tx;
1102 }