drivers/net/virtio/virtio_rxtx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <stdlib.h>
   8 #include <string.h>
   9 #include <errno.h>
  10
  11 #include <rte_cycles.h>
  12 #include <rte_memory.h>
  13 #include <rte_branch_prediction.h>
  14 #include <rte_mempool.h>
  15 #include <rte_malloc.h>
  16 #include <rte_mbuf.h>
  17 #include <rte_ether.h>
  18 #include <rte_ethdev_driver.h>
  19 #include <rte_prefetch.h>
  20 #include <rte_string_fns.h>
  21 #include <rte_errno.h>
  22 #include <rte_byteorder.h>
  23 #include <rte_net.h>
  24 #include <rte_ip.h>
  25 #include <rte_udp.h>
  26 #include <rte_tcp.h>
  27
  28 #include "virtio_logs.h"
  29 #include "virtio_ethdev.h"
  30 #include "virtio_pci.h"
  31 #include "virtqueue.h"
  32 #include "virtio_rxtx.h"
  33 #include "virtio_rxtx_simple.h"
  34
  35 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  36 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  37 #else
  38 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  39 #endif
  40
  41 int
  42 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
  43 {
  44         struct virtnet_rx *rxvq = rxq;
  45         struct virtqueue *vq = rxvq->vq;
  46
  47         return VIRTQUEUE_NUSED(vq) >= offset;
  48 }
  49
  50 void
  51 vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx, uint16_t num)
  52 {
  53         vq->vq_free_cnt += num;
  54         vq->vq_desc_tail_idx = desc_idx & (vq->vq_nentries - 1);
  55 }
  56
  57 void
  58 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  59 {
  60         struct vring_desc *dp, *dp_tail;
  61         struct vq_desc_extra *dxp;
  62         uint16_t desc_idx_last = desc_idx;
  63
  64         dp  = &vq->vq_ring.desc[desc_idx];
  65         dxp = &vq->vq_descx[desc_idx];
  66         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  67         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  68                 while (dp->flags & VRING_DESC_F_NEXT) {
  69                         desc_idx_last = dp->next;
  70                         dp = &vq->vq_ring.desc[dp->next];
  71                 }
  72         }
  73         dxp->ndescs = 0;
  74
  75         /*
  76          * We must append the existing free chain, if any, to the end of
  77          * newly freed chain. If the virtqueue was completely used, then
  78          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  79          */
  80         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  81                 vq->vq_desc_head_idx = desc_idx;
  82         } else {
  83                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
  84                 dp_tail->next = desc_idx;
  85         }
  86
  87         vq->vq_desc_tail_idx = desc_idx_last;
  88         dp->next = VQ_RING_DESC_CHAIN_END;
  89 }
  90
  91 static uint16_t
  92 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
  93                            uint32_t *len, uint16_t num)
  94 {
  95         struct vring_used_elem *uep;
  96         struct rte_mbuf *cookie;
  97         uint16_t used_idx, desc_idx;
  98         uint16_t i;
  99
 100         /*  Caller does the check */
 101         for (i = 0; i < num ; i++) {
 102                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 103                 uep = &vq->vq_ring.used->ring[used_idx];
 104                 desc_idx = (uint16_t) uep->id;
 105                 len[i] = uep->len;
 106                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 107
 108                 if (unlikely(cookie == NULL)) {
 109                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 110                                 vq->vq_used_cons_idx);
 111                         break;
 112                 }
 113
 114                 rte_prefetch0(cookie);
 115                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 116                 rx_pkts[i]  = cookie;
 117                 vq->vq_used_cons_idx++;
 118                 vq_ring_free_chain(vq, desc_idx);
 119                 vq->vq_descx[desc_idx].cookie = NULL;
 120         }
 121
 122         return i;
 123 }
 124
 125 #ifndef DEFAULT_TX_FREE_THRESH
 126 #define DEFAULT_TX_FREE_THRESH 32
 127 #endif
 128
 129 /* Cleanup from completed transmits. */
 130 static void
 131 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 132 {
 133         uint16_t i, used_idx, desc_idx;
 134         for (i = 0; i < num; i++) {
 135                 struct vring_used_elem *uep;
 136                 struct vq_desc_extra *dxp;
 137
 138                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 139                 uep = &vq->vq_ring.used->ring[used_idx];
 140
 141                 desc_idx = (uint16_t) uep->id;
 142                 dxp = &vq->vq_descx[desc_idx];
 143                 vq->vq_used_cons_idx++;
 144                 vq_ring_free_chain(vq, desc_idx);
 145
 146                 if (dxp->cookie != NULL) {
 147                         rte_pktmbuf_free(dxp->cookie);
 148                         dxp->cookie = NULL;
 149                 }
 150         }
 151 }
 152
 153
 154 static inline int
 155 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 156 {
 157         struct vq_desc_extra *dxp;
 158         struct virtio_hw *hw = vq->hw;
 159         struct vring_desc *start_dp;
 160         uint16_t needed = 1;
 161         uint16_t head_idx, idx;
 162
 163         if (unlikely(vq->vq_free_cnt == 0))
 164                 return -ENOSPC;
 165         if (unlikely(vq->vq_free_cnt < needed))
 166                 return -EMSGSIZE;
 167
 168         head_idx = vq->vq_desc_head_idx;
 169         if (unlikely(head_idx >= vq->vq_nentries))
 170                 return -EFAULT;
 171
 172         idx = head_idx;
 173         dxp = &vq->vq_descx[idx];
 174         dxp->cookie = (void *)cookie;
 175         dxp->ndescs = needed;
 176
 177         start_dp = vq->vq_ring.desc;
 178         start_dp[idx].addr =
 179                 VIRTIO_MBUF_ADDR(cookie, vq) +
 180                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 181         start_dp[idx].len =
 182                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 183         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 184         idx = start_dp[idx].next;
 185         vq->vq_desc_head_idx = idx;
 186         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 187                 vq->vq_desc_tail_idx = idx;
 188         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 189         vq_update_avail_ring(vq, head_idx);
 190
 191         return 0;
 192 }
 193
 194 /* When doing TSO, the IP length is not included in the pseudo header
 195  * checksum of the packet given to the PMD, but for virtio it is
 196  * expected.
 197  */
 198 static void
 199 virtio_tso_fix_cksum(struct rte_mbuf *m)
 200 {
 201         /* common case: header is not fragmented */
 202         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 203                         m->l4_len)) {
 204                 struct ipv4_hdr *iph;
 205                 struct ipv6_hdr *ip6h;
 206                 struct tcp_hdr *th;
 207                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 208                 uint32_t tmp;
 209
 210                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 211                 th = RTE_PTR_ADD(iph, m->l3_len);
 212                 if ((iph->version_ihl >> 4) == 4) {
 213                         iph->hdr_checksum = 0;
 214                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 215                         ip_len = iph->total_length;
 216                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 217                                 m->l3_len);
 218                 } else {
 219                         ip6h = (struct ipv6_hdr *)iph;
 220                         ip_paylen = ip6h->payload_len;
 221                 }
 222
 223                 /* calculate the new phdr checksum not including ip_paylen */
 224                 prev_cksum = th->cksum;
 225                 tmp = prev_cksum;
 226                 tmp += ip_paylen;
 227                 tmp = (tmp & 0xffff) + (tmp >> 16);
 228                 new_cksum = tmp;
 229
 230                 /* replace it in the packet */
 231                 th->cksum = new_cksum;
 232         }
 233 }
 234
 235 static inline int
 236 tx_offload_enabled(struct virtio_hw *hw)
 237 {
 238         return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
 239                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
 240                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
 241 }
 242
 243 /* avoid write operation when necessary, to lessen cache issues */
 244 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 245         if ((var) != (val))                     \
 246                 (var) = (val);                  \
 247 } while (0)
 248
 249 static inline void
 250 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 251                        uint16_t needed, int use_indirect, int can_push)
 252 {
 253         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 254         struct vq_desc_extra *dxp;
 255         struct virtqueue *vq = txvq->vq;
 256         struct vring_desc *start_dp;
 257         uint16_t seg_num = cookie->nb_segs;
 258         uint16_t head_idx, idx;
 259         uint16_t head_size = vq->hw->vtnet_hdr_size;
 260         struct virtio_net_hdr *hdr;
 261         int offload;
 262
 263         offload = tx_offload_enabled(vq->hw);
 264         head_idx = vq->vq_desc_head_idx;
 265         idx = head_idx;
 266         dxp = &vq->vq_descx[idx];
 267         dxp->cookie = (void *)cookie;
 268         dxp->ndescs = needed;
 269
 270         start_dp = vq->vq_ring.desc;
 271
 272         if (can_push) {
 273                 /* prepend cannot fail, checked by caller */
 274                 hdr = (struct virtio_net_hdr *)
 275                         rte_pktmbuf_prepend(cookie, head_size);
 276                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 277                  * which is wrong. Below subtract restores correct pkt size.
 278                  */
 279                 cookie->pkt_len -= head_size;
 280                 /* if offload disabled, it is not zeroed below, do it now */
 281                 if (offload == 0) {
 282                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 283                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 284                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 285                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 286                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 287                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 288                 }
 289         } else if (use_indirect) {
 290                 /* setup tx ring slot to point to indirect
 291                  * descriptor list stored in reserved region.
 292                  *
 293                  * the first slot in indirect ring is already preset
 294                  * to point to the header in reserved region
 295                  */
 296                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 297                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 298                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 299                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 300                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 301
 302                 /* loop below will fill in rest of the indirect elements */
 303                 start_dp = txr[idx].tx_indir;
 304                 idx = 1;
 305         } else {
 306                 /* setup first tx ring slot to point to header
 307                  * stored in reserved region.
 308                  */
 309                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 310                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 311                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 312                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 313                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 314
 315                 idx = start_dp[idx].next;
 316         }
 317
 318         /* Checksum Offload / TSO */
 319         if (offload) {
 320                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 321                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 322
 323                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 324                 case PKT_TX_UDP_CKSUM:
 325                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 326                         hdr->csum_offset = offsetof(struct udp_hdr,
 327                                 dgram_cksum);
 328                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 329                         break;
 330
 331                 case PKT_TX_TCP_CKSUM:
 332                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 333                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 334                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 335                         break;
 336
 337                 default:
 338                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 339                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 340                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 341                         break;
 342                 }
 343
 344                 /* TCP Segmentation Offload */
 345                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 346                         virtio_tso_fix_cksum(cookie);
 347                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 348                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 349                                 VIRTIO_NET_HDR_GSO_TCPV4;
 350                         hdr->gso_size = cookie->tso_segsz;
 351                         hdr->hdr_len =
 352                                 cookie->l2_len +
 353                                 cookie->l3_len +
 354                                 cookie->l4_len;
 355                 } else {
 356                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 357                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 358                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 359                 }
 360         }
 361
 362         do {
 363                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 364                 start_dp[idx].len   = cookie->data_len;
 365                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 366                 idx = start_dp[idx].next;
 367         } while ((cookie = cookie->next) != NULL);
 368
 369         if (use_indirect)
 370                 idx = vq->vq_ring.desc[head_idx].next;
 371
 372         vq->vq_desc_head_idx = idx;
 373         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 374                 vq->vq_desc_tail_idx = idx;
 375         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 376         vq_update_avail_ring(vq, head_idx);
 377 }
 378
 379 void
 380 virtio_dev_cq_start(struct rte_eth_dev *dev)
 381 {
 382         struct virtio_hw *hw = dev->data->dev_private;
 383
 384         if (hw->cvq && hw->cvq->vq) {
 385                 rte_spinlock_init(&hw->cvq->lock);
 386                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 387         }
 388 }
 389
 390 int
 391 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 392                         uint16_t queue_idx,
 393                         uint16_t nb_desc,
 394                         unsigned int socket_id __rte_unused,
 395                         const struct rte_eth_rxconf *rx_conf __rte_unused,
 396                         struct rte_mempool *mp)
 397 {
 398         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 399         struct virtio_hw *hw = dev->data->dev_private;
 400         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 401         struct virtnet_rx *rxvq;
 402
 403         PMD_INIT_FUNC_TRACE();
 404
 405         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 406                 nb_desc = vq->vq_nentries;
 407         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 408
 409         rxvq = &vq->rxq;
 410         rxvq->queue_id = queue_idx;
 411         rxvq->mpool = mp;
 412         if (rxvq->mpool == NULL) {
 413                 rte_exit(EXIT_FAILURE,
 414                         "Cannot allocate mbufs for rx virtqueue");
 415         }
 416
 417         dev->data->rx_queues[queue_idx] = rxvq;
 418
 419         return 0;
 420 }
 421
 422 int
 423 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 424 {
 425         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 426         struct virtio_hw *hw = dev->data->dev_private;
 427         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 428         struct virtnet_rx *rxvq = &vq->rxq;
 429         struct rte_mbuf *m;
 430         uint16_t desc_idx;
 431         int error, nbufs;
 432
 433         PMD_INIT_FUNC_TRACE();
 434
 435         /* Allocate blank mbufs for the each rx descriptor */
 436         nbufs = 0;
 437
 438         if (hw->use_simple_rx) {
 439                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 440                      desc_idx++) {
 441                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 442                         vq->vq_ring.desc[desc_idx].flags =
 443                                 VRING_DESC_F_WRITE;
 444                 }
 445
 446                 virtio_rxq_vec_setup(rxvq);
 447         }
 448
 449         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 450         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 451              desc_idx++) {
 452                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 453                         &rxvq->fake_mbuf;
 454         }
 455
 456         if (hw->use_simple_rx) {
 457                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 458                         virtio_rxq_rearm_vec(rxvq);
 459                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 460                 }
 461         } else {
 462                 while (!virtqueue_full(vq)) {
 463                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 464                         if (m == NULL)
 465                                 break;
 466
 467                         /* Enqueue allocated buffers */
 468                         error = virtqueue_enqueue_recv_refill(vq, m);
 469                         if (error) {
 470                                 rte_pktmbuf_free(m);
 471                                 break;
 472                         }
 473                         nbufs++;
 474                 }
 475
 476                 vq_update_avail_idx(vq);
 477         }
 478
 479         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 480
 481         VIRTQUEUE_DUMP(vq);
 482
 483         return 0;
 484 }
 485
 486 /*
 487  * struct rte_eth_dev *dev: Used to update dev
 488  * uint16_t nb_desc: Defaults to values read from config space
 489  * unsigned int socket_id: Used to allocate memzone
 490  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 491  * uint16_t queue_idx: Just used as an index in dev txq list
 492  */
 493 int
 494 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 495                         uint16_t queue_idx,
 496                         uint16_t nb_desc,
 497                         unsigned int socket_id __rte_unused,
 498                         const struct rte_eth_txconf *tx_conf)
 499 {
 500         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 501         struct virtio_hw *hw = dev->data->dev_private;
 502         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 503         struct virtnet_tx *txvq;
 504         uint16_t tx_free_thresh;
 505
 506         PMD_INIT_FUNC_TRACE();
 507
 508         /* cannot use simple rxtx funcs with multisegs or offloads */
 509         if (dev->data->dev_conf.txmode.offloads)
 510                 hw->use_simple_tx = 0;
 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         txvq = &vq->txq;
 517         txvq->queue_id = queue_idx;
 518
 519         tx_free_thresh = tx_conf->tx_free_thresh;
 520         if (tx_free_thresh == 0)
 521                 tx_free_thresh =
 522                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 523
 524         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 525                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 526                         "number of TX entries minus 3 (%u)."
 527                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 528                         vq->vq_nentries - 3,
 529                         tx_free_thresh, dev->data->port_id, queue_idx);
 530                 return -EINVAL;
 531         }
 532
 533         vq->vq_free_thresh = tx_free_thresh;
 534
 535         dev->data->tx_queues[queue_idx] = txvq;
 536         return 0;
 537 }
 538
 539 int
 540 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 541                                 uint16_t queue_idx)
 542 {
 543         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 544         struct virtio_hw *hw = dev->data->dev_private;
 545         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 546         uint16_t mid_idx = vq->vq_nentries >> 1;
 547         struct virtnet_tx *txvq = &vq->txq;
 548         uint16_t desc_idx;
 549
 550         PMD_INIT_FUNC_TRACE();
 551
 552         if (hw->use_simple_tx) {
 553                 for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 554                         vq->vq_ring.avail->ring[desc_idx] =
 555                                 desc_idx + mid_idx;
 556                         vq->vq_ring.desc[desc_idx + mid_idx].next =
 557                                 desc_idx;
 558                         vq->vq_ring.desc[desc_idx + mid_idx].addr =
 559                                 txvq->virtio_net_hdr_mem +
 560                                 offsetof(struct virtio_tx_region, tx_hdr);
 561                         vq->vq_ring.desc[desc_idx + mid_idx].len =
 562                                 vq->hw->vtnet_hdr_size;
 563                         vq->vq_ring.desc[desc_idx + mid_idx].flags =
 564                                 VRING_DESC_F_NEXT;
 565                         vq->vq_ring.desc[desc_idx].flags = 0;
 566                 }
 567                 for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 568                      desc_idx++)
 569                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 570         }
 571
 572         VIRTQUEUE_DUMP(vq);
 573
 574         return 0;
 575 }
 576
 577 static void
 578 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 579 {
 580         int error;
 581         /*
 582          * Requeue the discarded mbuf. This should always be
 583          * successful since it was just dequeued.
 584          */
 585         error = virtqueue_enqueue_recv_refill(vq, m);
 586         if (unlikely(error)) {
 587                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 588                 rte_pktmbuf_free(m);
 589         }
 590 }
 591
 592 static void
 593 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 594 {
 595         uint32_t s = mbuf->pkt_len;
 596         struct ether_addr *ea;
 597
 598         if (s == 64) {
 599                 stats->size_bins[1]++;
 600         } else if (s > 64 && s < 1024) {
 601                 uint32_t bin;
 602
 603                 /* count zeros, and offset into correct bin */
 604                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 605                 stats->size_bins[bin]++;
 606         } else {
 607                 if (s < 64)
 608                         stats->size_bins[0]++;
 609                 else if (s < 1519)
 610                         stats->size_bins[6]++;
 611                 else if (s >= 1519)
 612                         stats->size_bins[7]++;
 613         }
 614
 615         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 616         if (is_multicast_ether_addr(ea)) {
 617                 if (is_broadcast_ether_addr(ea))
 618                         stats->broadcast++;
 619                 else
 620                         stats->multicast++;
 621         }
 622 }
 623
 624 /* Optionally fill offload information in structure */
 625 static int
 626 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 627 {
 628         struct rte_net_hdr_lens hdr_lens;
 629         uint32_t hdrlen, ptype;
 630         int l4_supported = 0;
 631
 632         /* nothing to do */
 633         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 634                 return 0;
 635
 636         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 637
 638         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 639         m->packet_type = ptype;
 640         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 641             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 642             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 643                 l4_supported = 1;
 644
 645         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 646                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 647                 if (hdr->csum_start <= hdrlen && l4_supported) {
 648                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 649                 } else {
 650                         /* Unknown proto or tunnel, do sw cksum. We can assume
 651                          * the cksum field is in the first segment since the
 652                          * buffers we provided to the host are large enough.
 653                          * In case of SCTP, this will be wrong since it's a CRC
 654                          * but there's nothing we can do.
 655                          */
 656                         uint16_t csum = 0, off;
 657
 658                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 659                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 660                                 &csum);
 661                         if (likely(csum != 0xffff))
 662                                 csum = ~csum;
 663                         off = hdr->csum_offset + hdr->csum_start;
 664                         if (rte_pktmbuf_data_len(m) >= off + 1)
 665                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 666                                         off) = csum;
 667                 }
 668         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 669                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 670         }
 671
 672         /* GSO request, save required information in mbuf */
 673         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 674                 /* Check unsupported modes */
 675                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 676                     (hdr->gso_size == 0)) {
 677                         return -EINVAL;
 678                 }
 679
 680                 /* Update mss lengthes in mbuf */
 681                 m->tso_segsz = hdr->gso_size;
 682                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 683                         case VIRTIO_NET_HDR_GSO_TCPV4:
 684                         case VIRTIO_NET_HDR_GSO_TCPV6:
 685                                 m->ol_flags |= PKT_RX_LRO | \
 686                                         PKT_RX_L4_CKSUM_NONE;
 687                                 break;
 688                         default:
 689                                 return -EINVAL;
 690                 }
 691         }
 692
 693         return 0;
 694 }
 695
 696 static inline int
 697 rx_offload_enabled(struct virtio_hw *hw)
 698 {
 699         return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
 700                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
 701                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
 702 }
 703
 704 #define VIRTIO_MBUF_BURST_SZ 64
 705 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 706 uint16_t
 707 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 708 {
 709         struct virtnet_rx *rxvq = rx_queue;
 710         struct virtqueue *vq = rxvq->vq;
 711         struct virtio_hw *hw = vq->hw;
 712         struct rte_mbuf *rxm, *new_mbuf;
 713         uint16_t nb_used, num, nb_rx;
 714         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 715         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 716         int error;
 717         uint32_t i, nb_enqueued;
 718         uint32_t hdr_size;
 719         int offload;
 720         struct virtio_net_hdr *hdr;
 721
 722         nb_rx = 0;
 723         if (unlikely(hw->started == 0))
 724                 return nb_rx;
 725
 726         nb_used = VIRTQUEUE_NUSED(vq);
 727
 728         virtio_rmb();
 729
 730         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
 731         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
 732                 num = VIRTIO_MBUF_BURST_SZ;
 733         if (likely(num > DESC_PER_CACHELINE))
 734                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 735
 736         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 737         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 738
 739         nb_enqueued = 0;
 740         hdr_size = hw->vtnet_hdr_size;
 741         offload = rx_offload_enabled(hw);
 742
 743         for (i = 0; i < num ; i++) {
 744                 rxm = rcv_pkts[i];
 745
 746                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 747
 748                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 749                         PMD_RX_LOG(ERR, "Packet drop");
 750                         nb_enqueued++;
 751                         virtio_discard_rxbuf(vq, rxm);
 752                         rxvq->stats.errors++;
 753                         continue;
 754                 }
 755
 756                 rxm->port = rxvq->port_id;
 757                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 758                 rxm->ol_flags = 0;
 759                 rxm->vlan_tci = 0;
 760
 761                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 762                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 763
 764                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 765                         RTE_PKTMBUF_HEADROOM - hdr_size);
 766
 767                 if (hw->vlan_strip)
 768                         rte_vlan_strip(rxm);
 769
 770                 if (offload && virtio_rx_offload(rxm, hdr) < 0) {
 771                         virtio_discard_rxbuf(vq, rxm);
 772                         rxvq->stats.errors++;
 773                         continue;
 774                 }
 775
 776                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 777
 778                 rx_pkts[nb_rx++] = rxm;
 779
 780                 rxvq->stats.bytes += rxm->pkt_len;
 781                 virtio_update_packet_stats(&rxvq->stats, rxm);
 782         }
 783
 784         rxvq->stats.packets += nb_rx;
 785
 786         /* Allocate new mbuf for the used descriptor */
 787         error = ENOSPC;
 788         while (likely(!virtqueue_full(vq))) {
 789                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 790                 if (unlikely(new_mbuf == NULL)) {
 791                         struct rte_eth_dev *dev
 792                                 = &rte_eth_devices[rxvq->port_id];
 793                         dev->data->rx_mbuf_alloc_failed++;
 794                         break;
 795                 }
 796                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 797                 if (unlikely(error)) {
 798                         rte_pktmbuf_free(new_mbuf);
 799                         break;
 800                 }
 801                 nb_enqueued++;
 802         }
 803
 804         if (likely(nb_enqueued)) {
 805                 vq_update_avail_idx(vq);
 806
 807                 if (unlikely(virtqueue_kick_prepare(vq))) {
 808                         virtqueue_notify(vq);
 809                         PMD_RX_LOG(DEBUG, "Notified");
 810                 }
 811         }
 812
 813         return nb_rx;
 814 }
 815
 816 uint16_t
 817 virtio_recv_mergeable_pkts(void *rx_queue,
 818                         struct rte_mbuf **rx_pkts,
 819                         uint16_t nb_pkts)
 820 {
 821         struct virtnet_rx *rxvq = rx_queue;
 822         struct virtqueue *vq = rxvq->vq;
 823         struct virtio_hw *hw = vq->hw;
 824         struct rte_mbuf *rxm, *new_mbuf;
 825         uint16_t nb_used, num, nb_rx;
 826         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 827         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 828         struct rte_mbuf *prev;
 829         int error;
 830         uint32_t i, nb_enqueued;
 831         uint32_t seg_num;
 832         uint16_t extra_idx;
 833         uint32_t seg_res;
 834         uint32_t hdr_size;
 835         int offload;
 836
 837         nb_rx = 0;
 838         if (unlikely(hw->started == 0))
 839                 return nb_rx;
 840
 841         nb_used = VIRTQUEUE_NUSED(vq);
 842
 843         virtio_rmb();
 844
 845         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 846
 847         i = 0;
 848         nb_enqueued = 0;
 849         seg_num = 0;
 850         extra_idx = 0;
 851         seg_res = 0;
 852         hdr_size = hw->vtnet_hdr_size;
 853         offload = rx_offload_enabled(hw);
 854
 855         while (i < nb_used) {
 856                 struct virtio_net_hdr_mrg_rxbuf *header;
 857
 858                 if (nb_rx == nb_pkts)
 859                         break;
 860
 861                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
 862                 if (num != 1)
 863                         continue;
 864
 865                 i++;
 866
 867                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 868                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 869
 870                 rxm = rcv_pkts[0];
 871
 872                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 873                         PMD_RX_LOG(ERR, "Packet drop");
 874                         nb_enqueued++;
 875                         virtio_discard_rxbuf(vq, rxm);
 876                         rxvq->stats.errors++;
 877                         continue;
 878                 }
 879
 880                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 881                         RTE_PKTMBUF_HEADROOM - hdr_size);
 882                 seg_num = header->num_buffers;
 883
 884                 if (seg_num == 0)
 885                         seg_num = 1;
 886
 887                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 888                 rxm->nb_segs = seg_num;
 889                 rxm->ol_flags = 0;
 890                 rxm->vlan_tci = 0;
 891                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 892                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 893
 894                 rxm->port = rxvq->port_id;
 895                 rx_pkts[nb_rx] = rxm;
 896                 prev = rxm;
 897
 898                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
 899                         virtio_discard_rxbuf(vq, rxm);
 900                         rxvq->stats.errors++;
 901                         continue;
 902                 }
 903
 904                 seg_res = seg_num - 1;
 905
 906                 while (seg_res != 0) {
 907                         /*
 908                          * Get extra segments for current uncompleted packet.
 909                          */
 910                         uint16_t  rcv_cnt =
 911                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 912                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
 913                                 uint32_t rx_num =
 914                                         virtqueue_dequeue_burst_rx(vq,
 915                                         rcv_pkts, len, rcv_cnt);
 916                                 i += rx_num;
 917                                 rcv_cnt = rx_num;
 918                         } else {
 919                                 PMD_RX_LOG(ERR,
 920                                            "No enough segments for packet.");
 921                                 nb_enqueued++;
 922                                 virtio_discard_rxbuf(vq, rxm);
 923                                 rxvq->stats.errors++;
 924                                 break;
 925                         }
 926
 927                         extra_idx = 0;
 928
 929                         while (extra_idx < rcv_cnt) {
 930                                 rxm = rcv_pkts[extra_idx];
 931
 932                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 933                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 934                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 935
 936                                 if (prev)
 937                                         prev->next = rxm;
 938
 939                                 prev = rxm;
 940                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 941                                 extra_idx++;
 942                         };
 943                         seg_res -= rcv_cnt;
 944                 }
 945
 946                 if (hw->vlan_strip)
 947                         rte_vlan_strip(rx_pkts[nb_rx]);
 948
 949                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 950                         rx_pkts[nb_rx]->data_len);
 951
 952                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
 953                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
 954                 nb_rx++;
 955         }
 956
 957         rxvq->stats.packets += nb_rx;
 958
 959         /* Allocate new mbuf for the used descriptor */
 960         error = ENOSPC;
 961         while (likely(!virtqueue_full(vq))) {
 962                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 963                 if (unlikely(new_mbuf == NULL)) {
 964                         struct rte_eth_dev *dev
 965                                 = &rte_eth_devices[rxvq->port_id];
 966                         dev->data->rx_mbuf_alloc_failed++;
 967                         break;
 968                 }
 969                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 970                 if (unlikely(error)) {
 971                         rte_pktmbuf_free(new_mbuf);
 972                         break;
 973                 }
 974                 nb_enqueued++;
 975         }
 976
 977         if (likely(nb_enqueued)) {
 978                 vq_update_avail_idx(vq);
 979
 980                 if (unlikely(virtqueue_kick_prepare(vq))) {
 981                         virtqueue_notify(vq);
 982                         PMD_RX_LOG(DEBUG, "Notified");
 983                 }
 984         }
 985
 986         return nb_rx;
 987 }
 988
 989 uint16_t
 990 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 991 {
 992         struct virtnet_tx *txvq = tx_queue;
 993         struct virtqueue *vq = txvq->vq;
 994         struct virtio_hw *hw = vq->hw;
 995         uint16_t hdr_size = hw->vtnet_hdr_size;
 996         uint16_t nb_used, nb_tx = 0;
 997         int error;
 998
 999         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1000                 return nb_tx;
1001
1002         if (unlikely(nb_pkts < 1))
1003                 return nb_pkts;
1004
1005         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1006         nb_used = VIRTQUEUE_NUSED(vq);
1007
1008         virtio_rmb();
1009         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1010                 virtio_xmit_cleanup(vq, nb_used);
1011
1012         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1013                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1014                 int can_push = 0, use_indirect = 0, slots, need;
1015
1016                 /* Do VLAN tag insertion */
1017                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1018                         error = rte_vlan_insert(&txm);
1019                         if (unlikely(error)) {
1020                                 rte_pktmbuf_free(txm);
1021                                 continue;
1022                         }
1023                 }
1024
1025                 /* optimize ring usage */
1026                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1027                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1028                     rte_mbuf_refcnt_read(txm) == 1 &&
1029                     RTE_MBUF_DIRECT(txm) &&
1030                     txm->nb_segs == 1 &&
1031                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1032                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1033                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1034                         can_push = 1;
1035                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1036                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1037                         use_indirect = 1;
1038
1039                 /* How many main ring entries are needed to this Tx?
1040                  * any_layout => number of segments
1041                  * indirect   => 1
1042                  * default    => number of segments + 1
1043                  */
1044                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1045                 need = slots - vq->vq_free_cnt;
1046
1047                 /* Positive value indicates it need free vring descriptors */
1048                 if (unlikely(need > 0)) {
1049                         nb_used = VIRTQUEUE_NUSED(vq);
1050                         virtio_rmb();
1051                         need = RTE_MIN(need, (int)nb_used);
1052
1053                         virtio_xmit_cleanup(vq, need);
1054                         need = slots - vq->vq_free_cnt;
1055                         if (unlikely(need > 0)) {
1056                                 PMD_TX_LOG(ERR,
1057                                            "No free tx descriptors to transmit");
1058                                 break;
1059                         }
1060                 }
1061
1062                 /* Enqueue Packet buffers */
1063                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1064
1065                 txvq->stats.bytes += txm->pkt_len;
1066                 virtio_update_packet_stats(&txvq->stats, txm);
1067         }
1068
1069         txvq->stats.packets += nb_tx;
1070
1071         if (likely(nb_tx)) {
1072                 vq_update_avail_idx(vq);
1073
1074                 if (unlikely(virtqueue_kick_prepare(vq))) {
1075                         virtqueue_notify(vq);
1076                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1077                 }
1078         }
1079
1080         return nb_tx;
1081 }