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 int
  76 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
  77 {
  78         struct virtnet_rx *rxvq = rxq;
  79         struct virtqueue *vq = rxvq->vq;
  80
  81         return VIRTQUEUE_NUSED(vq) >= offset;
  82 }
  83
  84 static void
  85 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  86 {
  87         struct vring_desc *dp, *dp_tail;
  88         struct vq_desc_extra *dxp;
  89         uint16_t desc_idx_last = desc_idx;
  90
  91         dp  = &vq->vq_ring.desc[desc_idx];
  92         dxp = &vq->vq_descx[desc_idx];
  93         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  94         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  95                 while (dp->flags & VRING_DESC_F_NEXT) {
  96                         desc_idx_last = dp->next;
  97                         dp = &vq->vq_ring.desc[dp->next];
  98                 }
  99         }
 100         dxp->ndescs = 0;
 101
 102         /*
 103          * We must append the existing free chain, if any, to the end of
 104          * newly freed chain. If the virtqueue was completely used, then
 105          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
 106          */
 107         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
 108                 vq->vq_desc_head_idx = desc_idx;
 109         } else {
 110                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
 111                 dp_tail->next = desc_idx;
 112         }
 113
 114         vq->vq_desc_tail_idx = desc_idx_last;
 115         dp->next = VQ_RING_DESC_CHAIN_END;
 116 }
 117
 118 static uint16_t
 119 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 120                            uint32_t *len, uint16_t num)
 121 {
 122         struct vring_used_elem *uep;
 123         struct rte_mbuf *cookie;
 124         uint16_t used_idx, desc_idx;
 125         uint16_t i;
 126
 127         /*  Caller does the check */
 128         for (i = 0; i < num ; i++) {
 129                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 130                 uep = &vq->vq_ring.used->ring[used_idx];
 131                 desc_idx = (uint16_t) uep->id;
 132                 len[i] = uep->len;
 133                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 134
 135                 if (unlikely(cookie == NULL)) {
 136                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 137                                 vq->vq_used_cons_idx);
 138                         break;
 139                 }
 140
 141                 rte_prefetch0(cookie);
 142                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 143                 rx_pkts[i]  = cookie;
 144                 vq->vq_used_cons_idx++;
 145                 vq_ring_free_chain(vq, desc_idx);
 146                 vq->vq_descx[desc_idx].cookie = NULL;
 147         }
 148
 149         return i;
 150 }
 151
 152 #ifndef DEFAULT_TX_FREE_THRESH
 153 #define DEFAULT_TX_FREE_THRESH 32
 154 #endif
 155
 156 /* Cleanup from completed transmits. */
 157 static void
 158 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 159 {
 160         uint16_t i, used_idx, desc_idx;
 161         for (i = 0; i < num; i++) {
 162                 struct vring_used_elem *uep;
 163                 struct vq_desc_extra *dxp;
 164
 165                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 166                 uep = &vq->vq_ring.used->ring[used_idx];
 167
 168                 desc_idx = (uint16_t) uep->id;
 169                 dxp = &vq->vq_descx[desc_idx];
 170                 vq->vq_used_cons_idx++;
 171                 vq_ring_free_chain(vq, desc_idx);
 172
 173                 if (dxp->cookie != NULL) {
 174                         rte_pktmbuf_free(dxp->cookie);
 175                         dxp->cookie = NULL;
 176                 }
 177         }
 178 }
 179
 180
 181 static inline int
 182 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 183 {
 184         struct vq_desc_extra *dxp;
 185         struct virtio_hw *hw = vq->hw;
 186         struct vring_desc *start_dp;
 187         uint16_t needed = 1;
 188         uint16_t head_idx, idx;
 189
 190         if (unlikely(vq->vq_free_cnt == 0))
 191                 return -ENOSPC;
 192         if (unlikely(vq->vq_free_cnt < needed))
 193                 return -EMSGSIZE;
 194
 195         head_idx = vq->vq_desc_head_idx;
 196         if (unlikely(head_idx >= vq->vq_nentries))
 197                 return -EFAULT;
 198
 199         idx = head_idx;
 200         dxp = &vq->vq_descx[idx];
 201         dxp->cookie = (void *)cookie;
 202         dxp->ndescs = needed;
 203
 204         start_dp = vq->vq_ring.desc;
 205         start_dp[idx].addr =
 206                 VIRTIO_MBUF_ADDR(cookie, vq) +
 207                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 208         start_dp[idx].len =
 209                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 210         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 211         idx = start_dp[idx].next;
 212         vq->vq_desc_head_idx = idx;
 213         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 214                 vq->vq_desc_tail_idx = idx;
 215         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 216         vq_update_avail_ring(vq, head_idx);
 217
 218         return 0;
 219 }
 220
 221 /* When doing TSO, the IP length is not included in the pseudo header
 222  * checksum of the packet given to the PMD, but for virtio it is
 223  * expected.
 224  */
 225 static void
 226 virtio_tso_fix_cksum(struct rte_mbuf *m)
 227 {
 228         /* common case: header is not fragmented */
 229         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 230                         m->l4_len)) {
 231                 struct ipv4_hdr *iph;
 232                 struct ipv6_hdr *ip6h;
 233                 struct tcp_hdr *th;
 234                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 235                 uint32_t tmp;
 236
 237                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 238                 th = RTE_PTR_ADD(iph, m->l3_len);
 239                 if ((iph->version_ihl >> 4) == 4) {
 240                         iph->hdr_checksum = 0;
 241                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 242                         ip_len = iph->total_length;
 243                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 244                                 m->l3_len);
 245                 } else {
 246                         ip6h = (struct ipv6_hdr *)iph;
 247                         ip_paylen = ip6h->payload_len;
 248                 }
 249
 250                 /* calculate the new phdr checksum not including ip_paylen */
 251                 prev_cksum = th->cksum;
 252                 tmp = prev_cksum;
 253                 tmp += ip_paylen;
 254                 tmp = (tmp & 0xffff) + (tmp >> 16);
 255                 new_cksum = tmp;
 256
 257                 /* replace it in the packet */
 258                 th->cksum = new_cksum;
 259         }
 260 }
 261
 262 static inline int
 263 tx_offload_enabled(struct virtio_hw *hw)
 264 {
 265         return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
 266                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
 267                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
 268 }
 269
 270 static inline void
 271 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 272                        uint16_t needed, int use_indirect, int can_push)
 273 {
 274         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 275         struct vq_desc_extra *dxp;
 276         struct virtqueue *vq = txvq->vq;
 277         struct vring_desc *start_dp;
 278         uint16_t seg_num = cookie->nb_segs;
 279         uint16_t head_idx, idx;
 280         uint16_t head_size = vq->hw->vtnet_hdr_size;
 281         struct virtio_net_hdr *hdr;
 282         int offload;
 283
 284         offload = tx_offload_enabled(vq->hw);
 285         head_idx = vq->vq_desc_head_idx;
 286         idx = head_idx;
 287         dxp = &vq->vq_descx[idx];
 288         dxp->cookie = (void *)cookie;
 289         dxp->ndescs = needed;
 290
 291         start_dp = vq->vq_ring.desc;
 292
 293         if (can_push) {
 294                 /* prepend cannot fail, checked by caller */
 295                 hdr = (struct virtio_net_hdr *)
 296                         rte_pktmbuf_prepend(cookie, head_size);
 297                 /* if offload disabled, it is not zeroed below, do it now */
 298                 if (offload == 0)
 299                         memset(hdr, 0, head_size);
 300         } else if (use_indirect) {
 301                 /* setup tx ring slot to point to indirect
 302                  * descriptor list stored in reserved region.
 303                  *
 304                  * the first slot in indirect ring is already preset
 305                  * to point to the header in reserved region
 306                  */
 307                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 308                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 309                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 310                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 311                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 312
 313                 /* loop below will fill in rest of the indirect elements */
 314                 start_dp = txr[idx].tx_indir;
 315                 idx = 1;
 316         } else {
 317                 /* setup first tx ring slot to point to header
 318                  * stored in reserved region.
 319                  */
 320                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 321                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 322                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 323                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 324                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 325
 326                 idx = start_dp[idx].next;
 327         }
 328
 329         /* Checksum Offload / TSO */
 330         if (offload) {
 331                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 332                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 333
 334                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 335                 case PKT_TX_UDP_CKSUM:
 336                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 337                         hdr->csum_offset = offsetof(struct udp_hdr,
 338                                 dgram_cksum);
 339                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 340                         break;
 341
 342                 case PKT_TX_TCP_CKSUM:
 343                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 344                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 345                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 346                         break;
 347
 348                 default:
 349                         hdr->csum_start = 0;
 350                         hdr->csum_offset = 0;
 351                         hdr->flags = 0;
 352                         break;
 353                 }
 354
 355                 /* TCP Segmentation Offload */
 356                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 357                         virtio_tso_fix_cksum(cookie);
 358                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 359                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 360                                 VIRTIO_NET_HDR_GSO_TCPV4;
 361                         hdr->gso_size = cookie->tso_segsz;
 362                         hdr->hdr_len =
 363                                 cookie->l2_len +
 364                                 cookie->l3_len +
 365                                 cookie->l4_len;
 366                 } else {
 367                         hdr->gso_type = 0;
 368                         hdr->gso_size = 0;
 369                         hdr->hdr_len = 0;
 370                 }
 371         }
 372
 373         do {
 374                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 375                 start_dp[idx].len   = cookie->data_len;
 376                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 377                 idx = start_dp[idx].next;
 378         } while ((cookie = cookie->next) != NULL);
 379
 380         if (use_indirect)
 381                 idx = vq->vq_ring.desc[head_idx].next;
 382
 383         vq->vq_desc_head_idx = idx;
 384         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 385                 vq->vq_desc_tail_idx = idx;
 386         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 387         vq_update_avail_ring(vq, head_idx);
 388 }
 389
 390 void
 391 virtio_dev_cq_start(struct rte_eth_dev *dev)
 392 {
 393         struct virtio_hw *hw = dev->data->dev_private;
 394
 395         if (hw->cvq && hw->cvq->vq) {
 396                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 397         }
 398 }
 399
 400 int
 401 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 402                         uint16_t queue_idx,
 403                         uint16_t nb_desc,
 404                         unsigned int socket_id __rte_unused,
 405                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 406                         struct rte_mempool *mp)
 407 {
 408         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 409         struct virtio_hw *hw = dev->data->dev_private;
 410         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 411         struct virtnet_rx *rxvq;
 412         int error, nbufs;
 413         struct rte_mbuf *m;
 414         uint16_t desc_idx;
 415
 416         PMD_INIT_FUNC_TRACE();
 417
 418         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 419                 nb_desc = vq->vq_nentries;
 420         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 421
 422         rxvq = &vq->rxq;
 423         rxvq->queue_id = queue_idx;
 424         rxvq->mpool = mp;
 425         if (rxvq->mpool == NULL) {
 426                 rte_exit(EXIT_FAILURE,
 427                         "Cannot allocate mbufs for rx virtqueue");
 428         }
 429         dev->data->rx_queues[queue_idx] = rxvq;
 430
 431
 432         /* Allocate blank mbufs for the each rx descriptor */
 433         nbufs = 0;
 434         error = ENOSPC;
 435
 436         if (hw->use_simple_rxtx) {
 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
 445         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 446         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 447              desc_idx++) {
 448                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 449                         &rxvq->fake_mbuf;
 450         }
 451
 452         while (!virtqueue_full(vq)) {
 453                 m = rte_mbuf_raw_alloc(rxvq->mpool);
 454                 if (m == NULL)
 455                         break;
 456
 457                 /* Enqueue allocated buffers */
 458                 if (hw->use_simple_rxtx)
 459                         error = virtqueue_enqueue_recv_refill_simple(vq, m);
 460                 else
 461                         error = virtqueue_enqueue_recv_refill(vq, m);
 462
 463                 if (error) {
 464                         rte_pktmbuf_free(m);
 465                         break;
 466                 }
 467                 nbufs++;
 468         }
 469
 470         vq_update_avail_idx(vq);
 471
 472         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 473
 474         virtio_rxq_vec_setup(rxvq);
 475
 476         VIRTQUEUE_DUMP(vq);
 477
 478         return 0;
 479 }
 480
 481 static void
 482 virtio_update_rxtx_handler(struct rte_eth_dev *dev,
 483                            const struct rte_eth_txconf *tx_conf)
 484 {
 485         uint8_t use_simple_rxtx = 0;
 486         struct virtio_hw *hw = dev->data->dev_private;
 487
 488 #if defined RTE_ARCH_X86
 489         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE3))
 490                 use_simple_rxtx = 1;
 491 #elif defined RTE_ARCH_ARM64 || defined CONFIG_RTE_ARCH_ARM
 492         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
 493                 use_simple_rxtx = 1;
 494 #endif
 495         /* Use simple rx/tx func if single segment and no offloads */
 496         if (use_simple_rxtx &&
 497             (tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 498             !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 499                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 500                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 501                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 502                 hw->use_simple_rxtx = use_simple_rxtx;
 503         }
 504 }
 505
 506 /*
 507  * struct rte_eth_dev *dev: Used to update dev
 508  * uint16_t nb_desc: Defaults to values read from config space
 509  * unsigned int socket_id: Used to allocate memzone
 510  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 511  * uint16_t queue_idx: Just used as an index in dev txq list
 512  */
 513 int
 514 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 515                         uint16_t queue_idx,
 516                         uint16_t nb_desc,
 517                         unsigned int socket_id __rte_unused,
 518                         const struct rte_eth_txconf *tx_conf)
 519 {
 520         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 521         struct virtio_hw *hw = dev->data->dev_private;
 522         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 523         struct virtnet_tx *txvq;
 524         uint16_t tx_free_thresh;
 525         uint16_t desc_idx;
 526
 527         PMD_INIT_FUNC_TRACE();
 528
 529         virtio_update_rxtx_handler(dev, tx_conf);
 530
 531         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 532                 nb_desc = vq->vq_nentries;
 533         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 534
 535         txvq = &vq->txq;
 536         txvq->queue_id = queue_idx;
 537
 538         tx_free_thresh = tx_conf->tx_free_thresh;
 539         if (tx_free_thresh == 0)
 540                 tx_free_thresh =
 541                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 542
 543         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 544                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 545                         "number of TX entries minus 3 (%u)."
 546                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 547                         vq->vq_nentries - 3,
 548                         tx_free_thresh, dev->data->port_id, queue_idx);
 549                 return -EINVAL;
 550         }
 551
 552         vq->vq_free_thresh = tx_free_thresh;
 553
 554         if (hw->use_simple_rxtx) {
 555                 uint16_t mid_idx  = vq->vq_nentries >> 1;
 556
 557                 for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 558                         vq->vq_ring.avail->ring[desc_idx] =
 559                                 desc_idx + mid_idx;
 560                         vq->vq_ring.desc[desc_idx + mid_idx].next =
 561                                 desc_idx;
 562                         vq->vq_ring.desc[desc_idx + mid_idx].addr =
 563                                 txvq->virtio_net_hdr_mem +
 564                                 offsetof(struct virtio_tx_region, tx_hdr);
 565                         vq->vq_ring.desc[desc_idx + mid_idx].len =
 566                                 vq->hw->vtnet_hdr_size;
 567                         vq->vq_ring.desc[desc_idx + mid_idx].flags =
 568                                 VRING_DESC_F_NEXT;
 569                         vq->vq_ring.desc[desc_idx].flags = 0;
 570                 }
 571                 for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 572                      desc_idx++)
 573                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 574         }
 575
 576         VIRTQUEUE_DUMP(vq);
 577
 578         dev->data->tx_queues[queue_idx] = txvq;
 579         return 0;
 580 }
 581
 582 static void
 583 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 584 {
 585         int error;
 586         /*
 587          * Requeue the discarded mbuf. This should always be
 588          * successful since it was just dequeued.
 589          */
 590         error = virtqueue_enqueue_recv_refill(vq, m);
 591         if (unlikely(error)) {
 592                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 593                 rte_pktmbuf_free(m);
 594         }
 595 }
 596
 597 static void
 598 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 599 {
 600         uint32_t s = mbuf->pkt_len;
 601         struct ether_addr *ea;
 602
 603         if (s == 64) {
 604                 stats->size_bins[1]++;
 605         } else if (s > 64 && s < 1024) {
 606                 uint32_t bin;
 607
 608                 /* count zeros, and offset into correct bin */
 609                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 610                 stats->size_bins[bin]++;
 611         } else {
 612                 if (s < 64)
 613                         stats->size_bins[0]++;
 614                 else if (s < 1519)
 615                         stats->size_bins[6]++;
 616                 else if (s >= 1519)
 617                         stats->size_bins[7]++;
 618         }
 619
 620         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 621         if (is_multicast_ether_addr(ea)) {
 622                 if (is_broadcast_ether_addr(ea))
 623                         stats->broadcast++;
 624                 else
 625                         stats->multicast++;
 626         }
 627 }
 628
 629 /* Optionally fill offload information in structure */
 630 static int
 631 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 632 {
 633         struct rte_net_hdr_lens hdr_lens;
 634         uint32_t hdrlen, ptype;
 635         int l4_supported = 0;
 636
 637         /* nothing to do */
 638         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 639                 return 0;
 640
 641         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 642
 643         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 644         m->packet_type = ptype;
 645         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 646             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 647             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 648                 l4_supported = 1;
 649
 650         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 651                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 652                 if (hdr->csum_start <= hdrlen && l4_supported) {
 653                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 654                 } else {
 655                         /* Unknown proto or tunnel, do sw cksum. We can assume
 656                          * the cksum field is in the first segment since the
 657                          * buffers we provided to the host are large enough.
 658                          * In case of SCTP, this will be wrong since it's a CRC
 659                          * but there's nothing we can do.
 660                          */
 661                         uint16_t csum, off;
 662
 663                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 664                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 665                                 &csum);
 666                         if (likely(csum != 0xffff))
 667                                 csum = ~csum;
 668                         off = hdr->csum_offset + hdr->csum_start;
 669                         if (rte_pktmbuf_data_len(m) >= off + 1)
 670                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 671                                         off) = csum;
 672                 }
 673         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 674                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 675         }
 676
 677         /* GSO request, save required information in mbuf */
 678         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 679                 /* Check unsupported modes */
 680                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 681                     (hdr->gso_size == 0)) {
 682                         return -EINVAL;
 683                 }
 684
 685                 /* Update mss lengthes in mbuf */
 686                 m->tso_segsz = hdr->gso_size;
 687                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 688                         case VIRTIO_NET_HDR_GSO_TCPV4:
 689                         case VIRTIO_NET_HDR_GSO_TCPV6:
 690                                 m->ol_flags |= PKT_RX_LRO | \
 691                                         PKT_RX_L4_CKSUM_NONE;
 692                                 break;
 693                         default:
 694                                 return -EINVAL;
 695                 }
 696         }
 697
 698         return 0;
 699 }
 700
 701 static inline int
 702 rx_offload_enabled(struct virtio_hw *hw)
 703 {
 704         return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
 705                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
 706                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
 707 }
 708
 709 #define VIRTIO_MBUF_BURST_SZ 64
 710 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 711 uint16_t
 712 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 713 {
 714         struct virtnet_rx *rxvq = rx_queue;
 715         struct virtqueue *vq = rxvq->vq;
 716         struct virtio_hw *hw;
 717         struct rte_mbuf *rxm, *new_mbuf;
 718         uint16_t nb_used, num, nb_rx;
 719         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 720         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 721         int error;
 722         uint32_t i, nb_enqueued;
 723         uint32_t hdr_size;
 724         int offload;
 725         struct virtio_net_hdr *hdr;
 726
 727         nb_used = VIRTQUEUE_NUSED(vq);
 728
 729         virtio_rmb();
 730
 731         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 732         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? 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         hw = vq->hw;
 740         nb_rx = 0;
 741         nb_enqueued = 0;
 742         hdr_size = hw->vtnet_hdr_size;
 743         offload = rx_offload_enabled(hw);
 744
 745         for (i = 0; i < num ; i++) {
 746                 rxm = rcv_pkts[i];
 747
 748                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 749
 750                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 751                         PMD_RX_LOG(ERR, "Packet drop");
 752                         nb_enqueued++;
 753                         virtio_discard_rxbuf(vq, rxm);
 754                         rxvq->stats.errors++;
 755                         continue;
 756                 }
 757
 758                 rxm->port = rxvq->port_id;
 759                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 760                 rxm->ol_flags = 0;
 761                 rxm->vlan_tci = 0;
 762
 763                 rxm->nb_segs = 1;
 764                 rxm->next = NULL;
 765                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 766                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 767
 768                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 769                         RTE_PKTMBUF_HEADROOM - hdr_size);
 770
 771                 if (hw->vlan_strip)
 772                         rte_vlan_strip(rxm);
 773
 774                 if (offload && virtio_rx_offload(rxm, hdr) < 0) {
 775                         virtio_discard_rxbuf(vq, rxm);
 776                         rxvq->stats.errors++;
 777                         continue;
 778                 }
 779
 780                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 781
 782                 rx_pkts[nb_rx++] = rxm;
 783
 784                 rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
 785                 virtio_update_packet_stats(&rxvq->stats, rxm);
 786         }
 787
 788         rxvq->stats.packets += nb_rx;
 789
 790         /* Allocate new mbuf for the used descriptor */
 791         error = ENOSPC;
 792         while (likely(!virtqueue_full(vq))) {
 793                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 794                 if (unlikely(new_mbuf == NULL)) {
 795                         struct rte_eth_dev *dev
 796                                 = &rte_eth_devices[rxvq->port_id];
 797                         dev->data->rx_mbuf_alloc_failed++;
 798                         break;
 799                 }
 800                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 801                 if (unlikely(error)) {
 802                         rte_pktmbuf_free(new_mbuf);
 803                         break;
 804                 }
 805                 nb_enqueued++;
 806         }
 807
 808         if (likely(nb_enqueued)) {
 809                 vq_update_avail_idx(vq);
 810
 811                 if (unlikely(virtqueue_kick_prepare(vq))) {
 812                         virtqueue_notify(vq);
 813                         PMD_RX_LOG(DEBUG, "Notified");
 814                 }
 815         }
 816
 817         return nb_rx;
 818 }
 819
 820 uint16_t
 821 virtio_recv_mergeable_pkts(void *rx_queue,
 822                         struct rte_mbuf **rx_pkts,
 823                         uint16_t nb_pkts)
 824 {
 825         struct virtnet_rx *rxvq = rx_queue;
 826         struct virtqueue *vq = rxvq->vq;
 827         struct virtio_hw *hw;
 828         struct rte_mbuf *rxm, *new_mbuf;
 829         uint16_t nb_used, num, nb_rx;
 830         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 831         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 832         struct rte_mbuf *prev;
 833         int error;
 834         uint32_t i, nb_enqueued;
 835         uint32_t seg_num;
 836         uint16_t extra_idx;
 837         uint32_t seg_res;
 838         uint32_t hdr_size;
 839         int offload;
 840
 841         nb_used = VIRTQUEUE_NUSED(vq);
 842
 843         virtio_rmb();
 844
 845         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 846
 847         hw = vq->hw;
 848         nb_rx = 0;
 849         i = 0;
 850         nb_enqueued = 0;
 851         seg_num = 0;
 852         extra_idx = 0;
 853         seg_res = 0;
 854         hdr_size = hw->vtnet_hdr_size;
 855         offload = rx_offload_enabled(hw);
 856
 857         while (i < nb_used) {
 858                 struct virtio_net_hdr_mrg_rxbuf *header;
 859
 860                 if (nb_rx == nb_pkts)
 861                         break;
 862
 863                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
 864                 if (num != 1)
 865                         continue;
 866
 867                 i++;
 868
 869                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 870                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 871
 872                 rxm = rcv_pkts[0];
 873
 874                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 875                         PMD_RX_LOG(ERR, "Packet drop");
 876                         nb_enqueued++;
 877                         virtio_discard_rxbuf(vq, rxm);
 878                         rxvq->stats.errors++;
 879                         continue;
 880                 }
 881
 882                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 883                         RTE_PKTMBUF_HEADROOM - hdr_size);
 884                 seg_num = header->num_buffers;
 885
 886                 if (seg_num == 0)
 887                         seg_num = 1;
 888
 889                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 890                 rxm->nb_segs = seg_num;
 891                 rxm->next = NULL;
 892                 rxm->ol_flags = 0;
 893                 rxm->vlan_tci = 0;
 894                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 895                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 896
 897                 rxm->port = rxvq->port_id;
 898                 rx_pkts[nb_rx] = rxm;
 899                 prev = rxm;
 900
 901                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
 902                         virtio_discard_rxbuf(vq, rxm);
 903                         rxvq->stats.errors++;
 904                         continue;
 905                 }
 906
 907                 seg_res = seg_num - 1;
 908
 909                 while (seg_res != 0) {
 910                         /*
 911                          * Get extra segments for current uncompleted packet.
 912                          */
 913                         uint16_t  rcv_cnt =
 914                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 915                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
 916                                 uint32_t rx_num =
 917                                         virtqueue_dequeue_burst_rx(vq,
 918                                         rcv_pkts, len, rcv_cnt);
 919                                 i += rx_num;
 920                                 rcv_cnt = rx_num;
 921                         } else {
 922                                 PMD_RX_LOG(ERR,
 923                                            "No enough segments for packet.");
 924                                 nb_enqueued++;
 925                                 virtio_discard_rxbuf(vq, rxm);
 926                                 rxvq->stats.errors++;
 927                                 break;
 928                         }
 929
 930                         extra_idx = 0;
 931
 932                         while (extra_idx < rcv_cnt) {
 933                                 rxm = rcv_pkts[extra_idx];
 934
 935                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 936                                 rxm->next = NULL;
 937                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 938                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 939
 940                                 if (prev)
 941                                         prev->next = rxm;
 942
 943                                 prev = rxm;
 944                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 945                                 extra_idx++;
 946                         };
 947                         seg_res -= rcv_cnt;
 948                 }
 949
 950                 if (hw->vlan_strip)
 951                         rte_vlan_strip(rx_pkts[nb_rx]);
 952
 953                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 954                         rx_pkts[nb_rx]->data_len);
 955
 956                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
 957                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
 958                 nb_rx++;
 959         }
 960
 961         rxvq->stats.packets += nb_rx;
 962
 963         /* Allocate new mbuf for the used descriptor */
 964         error = ENOSPC;
 965         while (likely(!virtqueue_full(vq))) {
 966                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 967                 if (unlikely(new_mbuf == NULL)) {
 968                         struct rte_eth_dev *dev
 969                                 = &rte_eth_devices[rxvq->port_id];
 970                         dev->data->rx_mbuf_alloc_failed++;
 971                         break;
 972                 }
 973                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 974                 if (unlikely(error)) {
 975                         rte_pktmbuf_free(new_mbuf);
 976                         break;
 977                 }
 978                 nb_enqueued++;
 979         }
 980
 981         if (likely(nb_enqueued)) {
 982                 vq_update_avail_idx(vq);
 983
 984                 if (unlikely(virtqueue_kick_prepare(vq))) {
 985                         virtqueue_notify(vq);
 986                         PMD_RX_LOG(DEBUG, "Notified");
 987                 }
 988         }
 989
 990         return nb_rx;
 991 }
 992
 993 uint16_t
 994 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 995 {
 996         struct virtnet_tx *txvq = tx_queue;
 997         struct virtqueue *vq = txvq->vq;
 998         struct virtio_hw *hw = vq->hw;
 999         uint16_t hdr_size = hw->vtnet_hdr_size;
1000         uint16_t nb_used, nb_tx;
1001         int error;
1002
1003         if (unlikely(nb_pkts < 1))
1004                 return nb_pkts;
1005
1006         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1007         nb_used = VIRTQUEUE_NUSED(vq);
1008
1009         virtio_rmb();
1010         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1011                 virtio_xmit_cleanup(vq, nb_used);
1012
1013         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1014                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1015                 int can_push = 0, use_indirect = 0, slots, need;
1016
1017                 /* Do VLAN tag insertion */
1018                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1019                         error = rte_vlan_insert(&txm);
1020                         if (unlikely(error)) {
1021                                 rte_pktmbuf_free(txm);
1022                                 continue;
1023                         }
1024                 }
1025
1026                 /* optimize ring usage */
1027                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1028                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1029                     rte_mbuf_refcnt_read(txm) == 1 &&
1030                     RTE_MBUF_DIRECT(txm) &&
1031                     txm->nb_segs == 1 &&
1032                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1033                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1034                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1035                         can_push = 1;
1036                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1037                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1038                         use_indirect = 1;
1039
1040                 /* How many main ring entries are needed to this Tx?
1041                  * any_layout => number of segments
1042                  * indirect   => 1
1043                  * default    => number of segments + 1
1044                  */
1045                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1046                 need = slots - vq->vq_free_cnt;
1047
1048                 /* Positive value indicates it need free vring descriptors */
1049                 if (unlikely(need > 0)) {
1050                         nb_used = VIRTQUEUE_NUSED(vq);
1051                         virtio_rmb();
1052                         need = RTE_MIN(need, (int)nb_used);
1053
1054                         virtio_xmit_cleanup(vq, need);
1055                         need = slots - vq->vq_free_cnt;
1056                         if (unlikely(need > 0)) {
1057                                 PMD_TX_LOG(ERR,
1058                                            "No free tx descriptors to transmit");
1059                                 break;
1060                         }
1061                 }
1062
1063                 /* Enqueue Packet buffers */
1064                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1065
1066                 txvq->stats.bytes += txm->pkt_len;
1067                 virtio_update_packet_stats(&txvq->stats, txm);
1068         }
1069
1070         txvq->stats.packets += nb_tx;
1071
1072         if (likely(nb_tx)) {
1073                 vq_update_avail_idx(vq);
1074
1075                 if (unlikely(virtqueue_kick_prepare(vq))) {
1076                         virtqueue_notify(vq);
1077                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1078                 }
1079         }
1080
1081         return nb_tx;
1082 }