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_chain(struct virtqueue *vq, uint16_t desc_idx)
  52 {
  53         struct vring_desc *dp, *dp_tail;
  54         struct vq_desc_extra *dxp;
  55         uint16_t desc_idx_last = desc_idx;
  56
  57         dp  = &vq->vq_ring.desc[desc_idx];
  58         dxp = &vq->vq_descx[desc_idx];
  59         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  60         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  61                 while (dp->flags & VRING_DESC_F_NEXT) {
  62                         desc_idx_last = dp->next;
  63                         dp = &vq->vq_ring.desc[dp->next];
  64                 }
  65         }
  66         dxp->ndescs = 0;
  67
  68         /*
  69          * We must append the existing free chain, if any, to the end of
  70          * newly freed chain. If the virtqueue was completely used, then
  71          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  72          */
  73         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  74                 vq->vq_desc_head_idx = desc_idx;
  75         } else {
  76                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
  77                 dp_tail->next = desc_idx;
  78         }
  79
  80         vq->vq_desc_tail_idx = desc_idx_last;
  81         dp->next = VQ_RING_DESC_CHAIN_END;
  82 }
  83
  84 static uint16_t
  85 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
  86                            uint32_t *len, uint16_t num)
  87 {
  88         struct vring_used_elem *uep;
  89         struct rte_mbuf *cookie;
  90         uint16_t used_idx, desc_idx;
  91         uint16_t i;
  92
  93         /*  Caller does the check */
  94         for (i = 0; i < num ; i++) {
  95                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
  96                 uep = &vq->vq_ring.used->ring[used_idx];
  97                 desc_idx = (uint16_t) uep->id;
  98                 len[i] = uep->len;
  99                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 100
 101                 if (unlikely(cookie == NULL)) {
 102                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 103                                 vq->vq_used_cons_idx);
 104                         break;
 105                 }
 106
 107                 rte_prefetch0(cookie);
 108                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 109                 rx_pkts[i]  = cookie;
 110                 vq->vq_used_cons_idx++;
 111                 vq_ring_free_chain(vq, desc_idx);
 112                 vq->vq_descx[desc_idx].cookie = NULL;
 113         }
 114
 115         return i;
 116 }
 117
 118 #ifndef DEFAULT_TX_FREE_THRESH
 119 #define DEFAULT_TX_FREE_THRESH 32
 120 #endif
 121
 122 /* Cleanup from completed transmits. */
 123 static void
 124 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 125 {
 126         uint16_t i, used_idx, desc_idx;
 127         for (i = 0; i < num; i++) {
 128                 struct vring_used_elem *uep;
 129                 struct vq_desc_extra *dxp;
 130
 131                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 132                 uep = &vq->vq_ring.used->ring[used_idx];
 133
 134                 desc_idx = (uint16_t) uep->id;
 135                 dxp = &vq->vq_descx[desc_idx];
 136                 vq->vq_used_cons_idx++;
 137                 vq_ring_free_chain(vq, desc_idx);
 138
 139                 if (dxp->cookie != NULL) {
 140                         rte_pktmbuf_free(dxp->cookie);
 141                         dxp->cookie = NULL;
 142                 }
 143         }
 144 }
 145
 146
 147 static inline int
 148 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 149 {
 150         struct vq_desc_extra *dxp;
 151         struct virtio_hw *hw = vq->hw;
 152         struct vring_desc *start_dp;
 153         uint16_t needed = 1;
 154         uint16_t head_idx, idx;
 155
 156         if (unlikely(vq->vq_free_cnt == 0))
 157                 return -ENOSPC;
 158         if (unlikely(vq->vq_free_cnt < needed))
 159                 return -EMSGSIZE;
 160
 161         head_idx = vq->vq_desc_head_idx;
 162         if (unlikely(head_idx >= vq->vq_nentries))
 163                 return -EFAULT;
 164
 165         idx = head_idx;
 166         dxp = &vq->vq_descx[idx];
 167         dxp->cookie = (void *)cookie;
 168         dxp->ndescs = needed;
 169
 170         start_dp = vq->vq_ring.desc;
 171         start_dp[idx].addr =
 172                 VIRTIO_MBUF_ADDR(cookie, vq) +
 173                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 174         start_dp[idx].len =
 175                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 176         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 177         idx = start_dp[idx].next;
 178         vq->vq_desc_head_idx = idx;
 179         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 180                 vq->vq_desc_tail_idx = idx;
 181         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 182         vq_update_avail_ring(vq, head_idx);
 183
 184         return 0;
 185 }
 186
 187 /* When doing TSO, the IP length is not included in the pseudo header
 188  * checksum of the packet given to the PMD, but for virtio it is
 189  * expected.
 190  */
 191 static void
 192 virtio_tso_fix_cksum(struct rte_mbuf *m)
 193 {
 194         /* common case: header is not fragmented */
 195         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 196                         m->l4_len)) {
 197                 struct ipv4_hdr *iph;
 198                 struct ipv6_hdr *ip6h;
 199                 struct tcp_hdr *th;
 200                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 201                 uint32_t tmp;
 202
 203                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 204                 th = RTE_PTR_ADD(iph, m->l3_len);
 205                 if ((iph->version_ihl >> 4) == 4) {
 206                         iph->hdr_checksum = 0;
 207                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 208                         ip_len = iph->total_length;
 209                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 210                                 m->l3_len);
 211                 } else {
 212                         ip6h = (struct ipv6_hdr *)iph;
 213                         ip_paylen = ip6h->payload_len;
 214                 }
 215
 216                 /* calculate the new phdr checksum not including ip_paylen */
 217                 prev_cksum = th->cksum;
 218                 tmp = prev_cksum;
 219                 tmp += ip_paylen;
 220                 tmp = (tmp & 0xffff) + (tmp >> 16);
 221                 new_cksum = tmp;
 222
 223                 /* replace it in the packet */
 224                 th->cksum = new_cksum;
 225         }
 226 }
 227
 228 static inline int
 229 tx_offload_enabled(struct virtio_hw *hw)
 230 {
 231         return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
 232                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
 233                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
 234 }
 235
 236 /* avoid write operation when necessary, to lessen cache issues */
 237 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 238         if ((var) != (val))                     \
 239                 (var) = (val);                  \
 240 } while (0)
 241
 242 static inline void
 243 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 244                        uint16_t needed, int use_indirect, int can_push)
 245 {
 246         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 247         struct vq_desc_extra *dxp;
 248         struct virtqueue *vq = txvq->vq;
 249         struct vring_desc *start_dp;
 250         uint16_t seg_num = cookie->nb_segs;
 251         uint16_t head_idx, idx;
 252         uint16_t head_size = vq->hw->vtnet_hdr_size;
 253         struct virtio_net_hdr *hdr;
 254         int offload;
 255
 256         offload = tx_offload_enabled(vq->hw);
 257         head_idx = vq->vq_desc_head_idx;
 258         idx = head_idx;
 259         dxp = &vq->vq_descx[idx];
 260         dxp->cookie = (void *)cookie;
 261         dxp->ndescs = needed;
 262
 263         start_dp = vq->vq_ring.desc;
 264
 265         if (can_push) {
 266                 /* prepend cannot fail, checked by caller */
 267                 hdr = (struct virtio_net_hdr *)
 268                         rte_pktmbuf_prepend(cookie, head_size);
 269                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 270                  * which is wrong. Below subtract restores correct pkt size.
 271                  */
 272                 cookie->pkt_len -= head_size;
 273                 /* if offload disabled, it is not zeroed below, do it now */
 274                 if (offload == 0) {
 275                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 276                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 277                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 278                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 279                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 280                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 281                 }
 282         } else if (use_indirect) {
 283                 /* setup tx ring slot to point to indirect
 284                  * descriptor list stored in reserved region.
 285                  *
 286                  * the first slot in indirect ring is already preset
 287                  * to point to the header in reserved region
 288                  */
 289                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 290                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 291                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 292                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 293                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 294
 295                 /* loop below will fill in rest of the indirect elements */
 296                 start_dp = txr[idx].tx_indir;
 297                 idx = 1;
 298         } else {
 299                 /* setup first tx ring slot to point to header
 300                  * stored in reserved region.
 301                  */
 302                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 303                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 304                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 305                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 306                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 307
 308                 idx = start_dp[idx].next;
 309         }
 310
 311         /* Checksum Offload / TSO */
 312         if (offload) {
 313                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 314                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 315
 316                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 317                 case PKT_TX_UDP_CKSUM:
 318                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 319                         hdr->csum_offset = offsetof(struct udp_hdr,
 320                                 dgram_cksum);
 321                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 322                         break;
 323
 324                 case PKT_TX_TCP_CKSUM:
 325                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 326                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 327                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 328                         break;
 329
 330                 default:
 331                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 332                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 333                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 334                         break;
 335                 }
 336
 337                 /* TCP Segmentation Offload */
 338                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 339                         virtio_tso_fix_cksum(cookie);
 340                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 341                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 342                                 VIRTIO_NET_HDR_GSO_TCPV4;
 343                         hdr->gso_size = cookie->tso_segsz;
 344                         hdr->hdr_len =
 345                                 cookie->l2_len +
 346                                 cookie->l3_len +
 347                                 cookie->l4_len;
 348                 } else {
 349                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 350                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 351                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 352                 }
 353         }
 354
 355         do {
 356                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 357                 start_dp[idx].len   = cookie->data_len;
 358                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 359                 idx = start_dp[idx].next;
 360         } while ((cookie = cookie->next) != NULL);
 361
 362         if (use_indirect)
 363                 idx = vq->vq_ring.desc[head_idx].next;
 364
 365         vq->vq_desc_head_idx = idx;
 366         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 367                 vq->vq_desc_tail_idx = idx;
 368         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 369         vq_update_avail_ring(vq, head_idx);
 370 }
 371
 372 void
 373 virtio_dev_cq_start(struct rte_eth_dev *dev)
 374 {
 375         struct virtio_hw *hw = dev->data->dev_private;
 376
 377         if (hw->cvq && hw->cvq->vq) {
 378                 rte_spinlock_init(&hw->cvq->lock);
 379                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 380         }
 381 }
 382
 383 int
 384 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 385                         uint16_t queue_idx,
 386                         uint16_t nb_desc,
 387                         unsigned int socket_id __rte_unused,
 388                         const struct rte_eth_rxconf *rx_conf,
 389                         struct rte_mempool *mp)
 390 {
 391         const struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
 392         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 393         struct virtio_hw *hw = dev->data->dev_private;
 394         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 395         struct virtnet_rx *rxvq;
 396
 397         PMD_INIT_FUNC_TRACE();
 398
 399         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 400                 nb_desc = vq->vq_nentries;
 401         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 402
 403         rxvq = &vq->rxq;
 404         rxvq->queue_id = queue_idx;
 405         rxvq->mpool = mp;
 406         if (rxvq->mpool == NULL) {
 407                 rte_exit(EXIT_FAILURE,
 408                         "Cannot allocate mbufs for rx virtqueue");
 409         }
 410
 411         if ((rx_conf->offloads ^ rxmode->offloads) &
 412             VIRTIO_PMD_PER_DEVICE_RX_OFFLOADS)
 413                 return -EINVAL;
 414
 415         dev->data->rx_queues[queue_idx] = rxvq;
 416
 417         return 0;
 418 }
 419
 420 int
 421 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 422 {
 423         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 424         struct virtio_hw *hw = dev->data->dev_private;
 425         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 426         struct virtnet_rx *rxvq = &vq->rxq;
 427         struct rte_mbuf *m;
 428         uint16_t desc_idx;
 429         int error, nbufs;
 430
 431         PMD_INIT_FUNC_TRACE();
 432
 433         /* Allocate blank mbufs for the each rx descriptor */
 434         nbufs = 0;
 435
 436         if (hw->use_simple_rx) {
 437                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 438                      desc_idx++) {
 439                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 440                         vq->vq_ring.desc[desc_idx].flags =
 441                                 VRING_DESC_F_WRITE;
 442                 }
 443
 444                 virtio_rxq_vec_setup(rxvq);
 445         }
 446
 447         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 448         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 449              desc_idx++) {
 450                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 451                         &rxvq->fake_mbuf;
 452         }
 453
 454         if (hw->use_simple_rx) {
 455                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 456                         virtio_rxq_rearm_vec(rxvq);
 457                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 458                 }
 459         } else {
 460                 while (!virtqueue_full(vq)) {
 461                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 462                         if (m == NULL)
 463                                 break;
 464
 465                         /* Enqueue allocated buffers */
 466                         error = virtqueue_enqueue_recv_refill(vq, m);
 467                         if (error) {
 468                                 rte_pktmbuf_free(m);
 469                                 break;
 470                         }
 471                         nbufs++;
 472                 }
 473
 474                 vq_update_avail_idx(vq);
 475         }
 476
 477         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 478
 479         VIRTQUEUE_DUMP(vq);
 480
 481         return 0;
 482 }
 483
 484 /*
 485  * struct rte_eth_dev *dev: Used to update dev
 486  * uint16_t nb_desc: Defaults to values read from config space
 487  * unsigned int socket_id: Used to allocate memzone
 488  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 489  * uint16_t queue_idx: Just used as an index in dev txq list
 490  */
 491 int
 492 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 493                         uint16_t queue_idx,
 494                         uint16_t nb_desc,
 495                         unsigned int socket_id __rte_unused,
 496                         const struct rte_eth_txconf *tx_conf)
 497 {
 498         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 499         struct virtio_hw *hw = dev->data->dev_private;
 500         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 501         struct virtnet_tx *txvq;
 502         uint16_t tx_free_thresh;
 503
 504         PMD_INIT_FUNC_TRACE();
 505
 506         /* cannot use simple rxtx funcs with multisegs or offloads */
 507         if (tx_conf->offloads)
 508                 hw->use_simple_tx = 0;
 509
 510         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 511                 nb_desc = vq->vq_nentries;
 512         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 513
 514         txvq = &vq->txq;
 515         txvq->queue_id = queue_idx;
 516
 517         tx_free_thresh = tx_conf->tx_free_thresh;
 518         if (tx_free_thresh == 0)
 519                 tx_free_thresh =
 520                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 521
 522         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 523                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 524                         "number of TX entries minus 3 (%u)."
 525                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 526                         vq->vq_nentries - 3,
 527                         tx_free_thresh, dev->data->port_id, queue_idx);
 528                 return -EINVAL;
 529         }
 530
 531         vq->vq_free_thresh = tx_free_thresh;
 532
 533         dev->data->tx_queues[queue_idx] = txvq;
 534         return 0;
 535 }
 536
 537 int
 538 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 539                                 uint16_t queue_idx)
 540 {
 541         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 542         struct virtio_hw *hw = dev->data->dev_private;
 543         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 544         uint16_t mid_idx = vq->vq_nentries >> 1;
 545         struct virtnet_tx *txvq = &vq->txq;
 546         uint16_t desc_idx;
 547
 548         PMD_INIT_FUNC_TRACE();
 549
 550         if (hw->use_simple_tx) {
 551                 for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 552                         vq->vq_ring.avail->ring[desc_idx] =
 553                                 desc_idx + mid_idx;
 554                         vq->vq_ring.desc[desc_idx + mid_idx].next =
 555                                 desc_idx;
 556                         vq->vq_ring.desc[desc_idx + mid_idx].addr =
 557                                 txvq->virtio_net_hdr_mem +
 558                                 offsetof(struct virtio_tx_region, tx_hdr);
 559                         vq->vq_ring.desc[desc_idx + mid_idx].len =
 560                                 vq->hw->vtnet_hdr_size;
 561                         vq->vq_ring.desc[desc_idx + mid_idx].flags =
 562                                 VRING_DESC_F_NEXT;
 563                         vq->vq_ring.desc[desc_idx].flags = 0;
 564                 }
 565                 for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 566                      desc_idx++)
 567                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 568         }
 569
 570         VIRTQUEUE_DUMP(vq);
 571
 572         return 0;
 573 }
 574
 575 static void
 576 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 577 {
 578         int error;
 579         /*
 580          * Requeue the discarded mbuf. This should always be
 581          * successful since it was just dequeued.
 582          */
 583         error = virtqueue_enqueue_recv_refill(vq, m);
 584         if (unlikely(error)) {
 585                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 586                 rte_pktmbuf_free(m);
 587         }
 588 }
 589
 590 static void
 591 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 592 {
 593         uint32_t s = mbuf->pkt_len;
 594         struct ether_addr *ea;
 595
 596         if (s == 64) {
 597                 stats->size_bins[1]++;
 598         } else if (s > 64 && s < 1024) {
 599                 uint32_t bin;
 600
 601                 /* count zeros, and offset into correct bin */
 602                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 603                 stats->size_bins[bin]++;
 604         } else {
 605                 if (s < 64)
 606                         stats->size_bins[0]++;
 607                 else if (s < 1519)
 608                         stats->size_bins[6]++;
 609                 else if (s >= 1519)
 610                         stats->size_bins[7]++;
 611         }
 612
 613         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 614         if (is_multicast_ether_addr(ea)) {
 615                 if (is_broadcast_ether_addr(ea))
 616                         stats->broadcast++;
 617                 else
 618                         stats->multicast++;
 619         }
 620 }
 621
 622 /* Optionally fill offload information in structure */
 623 static int
 624 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 625 {
 626         struct rte_net_hdr_lens hdr_lens;
 627         uint32_t hdrlen, ptype;
 628         int l4_supported = 0;
 629
 630         /* nothing to do */
 631         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 632                 return 0;
 633
 634         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 635
 636         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 637         m->packet_type = ptype;
 638         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 639             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 640             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 641                 l4_supported = 1;
 642
 643         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 644                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 645                 if (hdr->csum_start <= hdrlen && l4_supported) {
 646                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 647                 } else {
 648                         /* Unknown proto or tunnel, do sw cksum. We can assume
 649                          * the cksum field is in the first segment since the
 650                          * buffers we provided to the host are large enough.
 651                          * In case of SCTP, this will be wrong since it's a CRC
 652                          * but there's nothing we can do.
 653                          */
 654                         uint16_t csum = 0, off;
 655
 656                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 657                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 658                                 &csum);
 659                         if (likely(csum != 0xffff))
 660                                 csum = ~csum;
 661                         off = hdr->csum_offset + hdr->csum_start;
 662                         if (rte_pktmbuf_data_len(m) >= off + 1)
 663                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 664                                         off) = csum;
 665                 }
 666         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 667                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 668         }
 669
 670         /* GSO request, save required information in mbuf */
 671         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 672                 /* Check unsupported modes */
 673                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 674                     (hdr->gso_size == 0)) {
 675                         return -EINVAL;
 676                 }
 677
 678                 /* Update mss lengthes in mbuf */
 679                 m->tso_segsz = hdr->gso_size;
 680                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 681                         case VIRTIO_NET_HDR_GSO_TCPV4:
 682                         case VIRTIO_NET_HDR_GSO_TCPV6:
 683                                 m->ol_flags |= PKT_RX_LRO | \
 684                                         PKT_RX_L4_CKSUM_NONE;
 685                                 break;
 686                         default:
 687                                 return -EINVAL;
 688                 }
 689         }
 690
 691         return 0;
 692 }
 693
 694 static inline int
 695 rx_offload_enabled(struct virtio_hw *hw)
 696 {
 697         return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
 698                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
 699                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
 700 }
 701
 702 #define VIRTIO_MBUF_BURST_SZ 64
 703 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 704 uint16_t
 705 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 706 {
 707         struct virtnet_rx *rxvq = rx_queue;
 708         struct virtqueue *vq = rxvq->vq;
 709         struct virtio_hw *hw = vq->hw;
 710         struct rte_mbuf *rxm, *new_mbuf;
 711         uint16_t nb_used, num, nb_rx;
 712         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 713         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 714         int error;
 715         uint32_t i, nb_enqueued;
 716         uint32_t hdr_size;
 717         int offload;
 718         struct virtio_net_hdr *hdr;
 719
 720         nb_rx = 0;
 721         if (unlikely(hw->started == 0))
 722                 return nb_rx;
 723
 724         nb_used = VIRTQUEUE_NUSED(vq);
 725
 726         virtio_rmb();
 727
 728         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
 729         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
 730                 num = VIRTIO_MBUF_BURST_SZ;
 731         if (likely(num > DESC_PER_CACHELINE))
 732                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 733
 734         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 735         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 736
 737         nb_enqueued = 0;
 738         hdr_size = hw->vtnet_hdr_size;
 739         offload = rx_offload_enabled(hw);
 740
 741         for (i = 0; i < num ; i++) {
 742                 rxm = rcv_pkts[i];
 743
 744                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 745
 746                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 747                         PMD_RX_LOG(ERR, "Packet drop");
 748                         nb_enqueued++;
 749                         virtio_discard_rxbuf(vq, rxm);
 750                         rxvq->stats.errors++;
 751                         continue;
 752                 }
 753
 754                 rxm->port = rxvq->port_id;
 755                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 756                 rxm->ol_flags = 0;
 757                 rxm->vlan_tci = 0;
 758
 759                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 760                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 761
 762                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 763                         RTE_PKTMBUF_HEADROOM - hdr_size);
 764
 765                 if (hw->vlan_strip)
 766                         rte_vlan_strip(rxm);
 767
 768                 if (offload && virtio_rx_offload(rxm, hdr) < 0) {
 769                         virtio_discard_rxbuf(vq, rxm);
 770                         rxvq->stats.errors++;
 771                         continue;
 772                 }
 773
 774                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 775
 776                 rx_pkts[nb_rx++] = rxm;
 777
 778                 rxvq->stats.bytes += rxm->pkt_len;
 779                 virtio_update_packet_stats(&rxvq->stats, rxm);
 780         }
 781
 782         rxvq->stats.packets += nb_rx;
 783
 784         /* Allocate new mbuf for the used descriptor */
 785         error = ENOSPC;
 786         while (likely(!virtqueue_full(vq))) {
 787                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 788                 if (unlikely(new_mbuf == NULL)) {
 789                         struct rte_eth_dev *dev
 790                                 = &rte_eth_devices[rxvq->port_id];
 791                         dev->data->rx_mbuf_alloc_failed++;
 792                         break;
 793                 }
 794                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 795                 if (unlikely(error)) {
 796                         rte_pktmbuf_free(new_mbuf);
 797                         break;
 798                 }
 799                 nb_enqueued++;
 800         }
 801
 802         if (likely(nb_enqueued)) {
 803                 vq_update_avail_idx(vq);
 804
 805                 if (unlikely(virtqueue_kick_prepare(vq))) {
 806                         virtqueue_notify(vq);
 807                         PMD_RX_LOG(DEBUG, "Notified");
 808                 }
 809         }
 810
 811         return nb_rx;
 812 }
 813
 814 uint16_t
 815 virtio_recv_mergeable_pkts(void *rx_queue,
 816                         struct rte_mbuf **rx_pkts,
 817                         uint16_t nb_pkts)
 818 {
 819         struct virtnet_rx *rxvq = rx_queue;
 820         struct virtqueue *vq = rxvq->vq;
 821         struct virtio_hw *hw = vq->hw;
 822         struct rte_mbuf *rxm, *new_mbuf;
 823         uint16_t nb_used, num, nb_rx;
 824         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 825         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 826         struct rte_mbuf *prev;
 827         int error;
 828         uint32_t i, nb_enqueued;
 829         uint32_t seg_num;
 830         uint16_t extra_idx;
 831         uint32_t seg_res;
 832         uint32_t hdr_size;
 833         int offload;
 834
 835         nb_rx = 0;
 836         if (unlikely(hw->started == 0))
 837                 return nb_rx;
 838
 839         nb_used = VIRTQUEUE_NUSED(vq);
 840
 841         virtio_rmb();
 842
 843         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 844
 845         i = 0;
 846         nb_enqueued = 0;
 847         seg_num = 0;
 848         extra_idx = 0;
 849         seg_res = 0;
 850         hdr_size = hw->vtnet_hdr_size;
 851         offload = rx_offload_enabled(hw);
 852
 853         while (i < nb_used) {
 854                 struct virtio_net_hdr_mrg_rxbuf *header;
 855
 856                 if (nb_rx == nb_pkts)
 857                         break;
 858
 859                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
 860                 if (num != 1)
 861                         continue;
 862
 863                 i++;
 864
 865                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 866                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 867
 868                 rxm = rcv_pkts[0];
 869
 870                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 871                         PMD_RX_LOG(ERR, "Packet drop");
 872                         nb_enqueued++;
 873                         virtio_discard_rxbuf(vq, rxm);
 874                         rxvq->stats.errors++;
 875                         continue;
 876                 }
 877
 878                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 879                         RTE_PKTMBUF_HEADROOM - hdr_size);
 880                 seg_num = header->num_buffers;
 881
 882                 if (seg_num == 0)
 883                         seg_num = 1;
 884
 885                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 886                 rxm->nb_segs = seg_num;
 887                 rxm->ol_flags = 0;
 888                 rxm->vlan_tci = 0;
 889                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 890                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 891
 892                 rxm->port = rxvq->port_id;
 893                 rx_pkts[nb_rx] = rxm;
 894                 prev = rxm;
 895
 896                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
 897                         virtio_discard_rxbuf(vq, rxm);
 898                         rxvq->stats.errors++;
 899                         continue;
 900                 }
 901
 902                 seg_res = seg_num - 1;
 903
 904                 while (seg_res != 0) {
 905                         /*
 906                          * Get extra segments for current uncompleted packet.
 907                          */
 908                         uint16_t  rcv_cnt =
 909                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 910                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
 911                                 uint32_t rx_num =
 912                                         virtqueue_dequeue_burst_rx(vq,
 913                                         rcv_pkts, len, rcv_cnt);
 914                                 i += rx_num;
 915                                 rcv_cnt = rx_num;
 916                         } else {
 917                                 PMD_RX_LOG(ERR,
 918                                            "No enough segments for packet.");
 919                                 nb_enqueued++;
 920                                 virtio_discard_rxbuf(vq, rxm);
 921                                 rxvq->stats.errors++;
 922                                 break;
 923                         }
 924
 925                         extra_idx = 0;
 926
 927                         while (extra_idx < rcv_cnt) {
 928                                 rxm = rcv_pkts[extra_idx];
 929
 930                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 931                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 932                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 933
 934                                 if (prev)
 935                                         prev->next = rxm;
 936
 937                                 prev = rxm;
 938                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 939                                 extra_idx++;
 940                         };
 941                         seg_res -= rcv_cnt;
 942                 }
 943
 944                 if (hw->vlan_strip)
 945                         rte_vlan_strip(rx_pkts[nb_rx]);
 946
 947                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 948                         rx_pkts[nb_rx]->data_len);
 949
 950                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
 951                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
 952                 nb_rx++;
 953         }
 954
 955         rxvq->stats.packets += nb_rx;
 956
 957         /* Allocate new mbuf for the used descriptor */
 958         error = ENOSPC;
 959         while (likely(!virtqueue_full(vq))) {
 960                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 961                 if (unlikely(new_mbuf == NULL)) {
 962                         struct rte_eth_dev *dev
 963                                 = &rte_eth_devices[rxvq->port_id];
 964                         dev->data->rx_mbuf_alloc_failed++;
 965                         break;
 966                 }
 967                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 968                 if (unlikely(error)) {
 969                         rte_pktmbuf_free(new_mbuf);
 970                         break;
 971                 }
 972                 nb_enqueued++;
 973         }
 974
 975         if (likely(nb_enqueued)) {
 976                 vq_update_avail_idx(vq);
 977
 978                 if (unlikely(virtqueue_kick_prepare(vq))) {
 979                         virtqueue_notify(vq);
 980                         PMD_RX_LOG(DEBUG, "Notified");
 981                 }
 982         }
 983
 984         return nb_rx;
 985 }
 986
 987 uint16_t
 988 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 989 {
 990         struct virtnet_tx *txvq = tx_queue;
 991         struct virtqueue *vq = txvq->vq;
 992         struct virtio_hw *hw = vq->hw;
 993         uint16_t hdr_size = hw->vtnet_hdr_size;
 994         uint16_t nb_used, nb_tx = 0;
 995         int error;
 996
 997         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
 998                 return nb_tx;
 999
1000         if (unlikely(nb_pkts < 1))
1001                 return nb_pkts;
1002
1003         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1004         nb_used = VIRTQUEUE_NUSED(vq);
1005
1006         virtio_rmb();
1007         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1008                 virtio_xmit_cleanup(vq, nb_used);
1009
1010         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1011                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1012                 int can_push = 0, use_indirect = 0, slots, need;
1013
1014                 /* Do VLAN tag insertion */
1015                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1016                         error = rte_vlan_insert(&txm);
1017                         if (unlikely(error)) {
1018                                 rte_pktmbuf_free(txm);
1019                                 continue;
1020                         }
1021                 }
1022
1023                 /* optimize ring usage */
1024                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1025                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1026                     rte_mbuf_refcnt_read(txm) == 1 &&
1027                     RTE_MBUF_DIRECT(txm) &&
1028                     txm->nb_segs == 1 &&
1029                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1030                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1031                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1032                         can_push = 1;
1033                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1034                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1035                         use_indirect = 1;
1036
1037                 /* How many main ring entries are needed to this Tx?
1038                  * any_layout => number of segments
1039                  * indirect   => 1
1040                  * default    => number of segments + 1
1041                  */
1042                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1043                 need = slots - vq->vq_free_cnt;
1044
1045                 /* Positive value indicates it need free vring descriptors */
1046                 if (unlikely(need > 0)) {
1047                         nb_used = VIRTQUEUE_NUSED(vq);
1048                         virtio_rmb();
1049                         need = RTE_MIN(need, (int)nb_used);
1050
1051                         virtio_xmit_cleanup(vq, need);
1052                         need = slots - vq->vq_free_cnt;
1053                         if (unlikely(need > 0)) {
1054                                 PMD_TX_LOG(ERR,
1055                                            "No free tx descriptors to transmit");
1056                                 break;
1057                         }
1058                 }
1059
1060                 /* Enqueue Packet buffers */
1061                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1062
1063                 txvq->stats.bytes += txm->pkt_len;
1064                 virtio_update_packet_stats(&txvq->stats, txm);
1065         }
1066
1067         txvq->stats.packets += nb_tx;
1068
1069         if (likely(nb_tx)) {
1070                 vq_update_avail_idx(vq);
1071
1072                 if (unlikely(virtqueue_kick_prepare(vq))) {
1073                         virtqueue_notify(vq);
1074                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1075                 }
1076         }
1077
1078         return nb_tx;
1079 }