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
  54 #include "virtio_logs.h"
  55 #include "virtio_ethdev.h"
  56 #include "virtio_pci.h"
  57 #include "virtqueue.h"
  58 #include "virtio_rxtx.h"
  59
  60 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  61 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  62 #else
  63 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  64 #endif
  65
  66
  67 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
  68         ETH_TXQ_FLAGS_NOOFFLOADS)
  69
  70 static int use_simple_rxtx;
  71
  72 static void
  73 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  74 {
  75         struct vring_desc *dp, *dp_tail;
  76         struct vq_desc_extra *dxp;
  77         uint16_t desc_idx_last = desc_idx;
  78
  79         dp  = &vq->vq_ring.desc[desc_idx];
  80         dxp = &vq->vq_descx[desc_idx];
  81         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  82         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  83                 while (dp->flags & VRING_DESC_F_NEXT) {
  84                         desc_idx_last = dp->next;
  85                         dp = &vq->vq_ring.desc[dp->next];
  86                 }
  87         }
  88         dxp->ndescs = 0;
  89
  90         /*
  91          * We must append the existing free chain, if any, to the end of
  92          * newly freed chain. If the virtqueue was completely used, then
  93          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  94          */
  95         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  96                 vq->vq_desc_head_idx = desc_idx;
  97         } else {
  98                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
  99                 dp_tail->next = desc_idx;
 100         }
 101
 102         vq->vq_desc_tail_idx = desc_idx_last;
 103         dp->next = VQ_RING_DESC_CHAIN_END;
 104 }
 105
 106 static uint16_t
 107 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 108                            uint32_t *len, uint16_t num)
 109 {
 110         struct vring_used_elem *uep;
 111         struct rte_mbuf *cookie;
 112         uint16_t used_idx, desc_idx;
 113         uint16_t i;
 114
 115         /*  Caller does the check */
 116         for (i = 0; i < num ; i++) {
 117                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 118                 uep = &vq->vq_ring.used->ring[used_idx];
 119                 desc_idx = (uint16_t) uep->id;
 120                 len[i] = uep->len;
 121                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 122
 123                 if (unlikely(cookie == NULL)) {
 124                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 125                                 vq->vq_used_cons_idx);
 126                         break;
 127                 }
 128
 129                 rte_prefetch0(cookie);
 130                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 131                 rx_pkts[i]  = cookie;
 132                 vq->vq_used_cons_idx++;
 133                 vq_ring_free_chain(vq, desc_idx);
 134                 vq->vq_descx[desc_idx].cookie = NULL;
 135         }
 136
 137         return i;
 138 }
 139
 140 #ifndef DEFAULT_TX_FREE_THRESH
 141 #define DEFAULT_TX_FREE_THRESH 32
 142 #endif
 143
 144 /* Cleanup from completed transmits. */
 145 static void
 146 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 147 {
 148         uint16_t i, used_idx, desc_idx;
 149         for (i = 0; i < num; i++) {
 150                 struct vring_used_elem *uep;
 151                 struct vq_desc_extra *dxp;
 152
 153                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 154                 uep = &vq->vq_ring.used->ring[used_idx];
 155
 156                 desc_idx = (uint16_t) uep->id;
 157                 dxp = &vq->vq_descx[desc_idx];
 158                 vq->vq_used_cons_idx++;
 159                 vq_ring_free_chain(vq, desc_idx);
 160
 161                 if (dxp->cookie != NULL) {
 162                         rte_pktmbuf_free(dxp->cookie);
 163                         dxp->cookie = NULL;
 164                 }
 165         }
 166 }
 167
 168
 169 static inline int
 170 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 171 {
 172         struct vq_desc_extra *dxp;
 173         struct virtio_hw *hw = vq->hw;
 174         struct vring_desc *start_dp;
 175         uint16_t needed = 1;
 176         uint16_t head_idx, idx;
 177
 178         if (unlikely(vq->vq_free_cnt == 0))
 179                 return -ENOSPC;
 180         if (unlikely(vq->vq_free_cnt < needed))
 181                 return -EMSGSIZE;
 182
 183         head_idx = vq->vq_desc_head_idx;
 184         if (unlikely(head_idx >= vq->vq_nentries))
 185                 return -EFAULT;
 186
 187         idx = head_idx;
 188         dxp = &vq->vq_descx[idx];
 189         dxp->cookie = (void *)cookie;
 190         dxp->ndescs = needed;
 191
 192         start_dp = vq->vq_ring.desc;
 193         start_dp[idx].addr =
 194                 (uint64_t)(cookie->buf_physaddr + RTE_PKTMBUF_HEADROOM
 195                 - hw->vtnet_hdr_size);
 196         start_dp[idx].len =
 197                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 198         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 199         idx = start_dp[idx].next;
 200         vq->vq_desc_head_idx = idx;
 201         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 202                 vq->vq_desc_tail_idx = idx;
 203         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 204         vq_update_avail_ring(vq, head_idx);
 205
 206         return 0;
 207 }
 208
 209 static int
 210 virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
 211 {
 212         struct vq_desc_extra *dxp;
 213         struct vring_desc *start_dp;
 214         uint16_t seg_num = cookie->nb_segs;
 215         uint16_t needed = 1 + seg_num;
 216         uint16_t head_idx, idx;
 217         size_t head_size = txvq->hw->vtnet_hdr_size;
 218
 219         if (unlikely(txvq->vq_free_cnt == 0))
 220                 return -ENOSPC;
 221         if (unlikely(txvq->vq_free_cnt < needed))
 222                 return -EMSGSIZE;
 223         head_idx = txvq->vq_desc_head_idx;
 224         if (unlikely(head_idx >= txvq->vq_nentries))
 225                 return -EFAULT;
 226
 227         idx = head_idx;
 228         dxp = &txvq->vq_descx[idx];
 229         dxp->cookie = (void *)cookie;
 230         dxp->ndescs = needed;
 231
 232         start_dp = txvq->vq_ring.desc;
 233         start_dp[idx].addr =
 234                 txvq->virtio_net_hdr_mem + idx * head_size;
 235         start_dp[idx].len = head_size;
 236         start_dp[idx].flags = VRING_DESC_F_NEXT;
 237
 238         for (; ((seg_num > 0) && (cookie != NULL)); seg_num--) {
 239                 idx = start_dp[idx].next;
 240                 start_dp[idx].addr  = RTE_MBUF_DATA_DMA_ADDR(cookie);
 241                 start_dp[idx].len   = cookie->data_len;
 242                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 243                 cookie = cookie->next;
 244         }
 245
 246         start_dp[idx].flags &= ~VRING_DESC_F_NEXT;
 247         idx = start_dp[idx].next;
 248         txvq->vq_desc_head_idx = idx;
 249         if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 250                 txvq->vq_desc_tail_idx = idx;
 251         txvq->vq_free_cnt = (uint16_t)(txvq->vq_free_cnt - needed);
 252         vq_update_avail_ring(txvq, head_idx);
 253
 254         return 0;
 255 }
 256
 257 static inline struct rte_mbuf *
 258 rte_rxmbuf_alloc(struct rte_mempool *mp)
 259 {
 260         struct rte_mbuf *m;
 261
 262         m = __rte_mbuf_raw_alloc(mp);
 263         __rte_mbuf_sanity_check_raw(m, 0);
 264
 265         return m;
 266 }
 267
 268 static void
 269 virtio_dev_vring_start(struct virtqueue *vq, int queue_type)
 270 {
 271         struct rte_mbuf *m;
 272         int i, nbufs, error, size = vq->vq_nentries;
 273         struct vring *vr = &vq->vq_ring;
 274         uint8_t *ring_mem = vq->vq_ring_virt_mem;
 275
 276         PMD_INIT_FUNC_TRACE();
 277
 278         /*
 279          * Reinitialise since virtio port might have been stopped and restarted
 280          */
 281         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
 282         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
 283         vq->vq_used_cons_idx = 0;
 284         vq->vq_desc_head_idx = 0;
 285         vq->vq_avail_idx = 0;
 286         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
 287         vq->vq_free_cnt = vq->vq_nentries;
 288         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 289
 290         /* Chain all the descriptors in the ring with an END */
 291         for (i = 0; i < size - 1; i++)
 292                 vr->desc[i].next = (uint16_t)(i + 1);
 293         vr->desc[i].next = VQ_RING_DESC_CHAIN_END;
 294
 295         /*
 296          * Disable device(host) interrupting guest
 297          */
 298         virtqueue_disable_intr(vq);
 299
 300         /* Only rx virtqueue needs mbufs to be allocated at initialization */
 301         if (queue_type == VTNET_RQ) {
 302                 if (vq->mpool == NULL)
 303                         rte_exit(EXIT_FAILURE,
 304                         "Cannot allocate initial mbufs for rx virtqueue");
 305
 306                 /* Allocate blank mbufs for the each rx descriptor */
 307                 nbufs = 0;
 308                 error = ENOSPC;
 309
 310                 if (use_simple_rxtx)
 311                         for (i = 0; i < vq->vq_nentries; i++) {
 312                                 vq->vq_ring.avail->ring[i] = i;
 313                                 vq->vq_ring.desc[i].flags = VRING_DESC_F_WRITE;
 314                         }
 315
 316                 memset(&vq->fake_mbuf, 0, sizeof(vq->fake_mbuf));
 317                 for (i = 0; i < RTE_PMD_VIRTIO_RX_MAX_BURST; i++)
 318                         vq->sw_ring[vq->vq_nentries + i] = &vq->fake_mbuf;
 319
 320                 while (!virtqueue_full(vq)) {
 321                         m = rte_rxmbuf_alloc(vq->mpool);
 322                         if (m == NULL)
 323                                 break;
 324
 325                         /******************************************
 326                         *         Enqueue allocated buffers        *
 327                         *******************************************/
 328                         if (use_simple_rxtx)
 329                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 330                         else
 331                                 error = virtqueue_enqueue_recv_refill(vq, m);
 332                         if (error) {
 333                                 rte_pktmbuf_free(m);
 334                                 break;
 335                         }
 336                         nbufs++;
 337                 }
 338
 339                 vq_update_avail_idx(vq);
 340
 341                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 342
 343                 VIRTIO_WRITE_REG_2(vq->hw, VIRTIO_PCI_QUEUE_SEL,
 344                         vq->vq_queue_index);
 345                 VIRTIO_WRITE_REG_4(vq->hw, VIRTIO_PCI_QUEUE_PFN,
 346                         vq->mz->phys_addr >> VIRTIO_PCI_QUEUE_ADDR_SHIFT);
 347         } else if (queue_type == VTNET_TQ) {
 348                 if (use_simple_rxtx) {
 349                         int mid_idx  = vq->vq_nentries >> 1;
 350                         for (i = 0; i < mid_idx; i++) {
 351                                 vq->vq_ring.avail->ring[i] = i + mid_idx;
 352                                 vq->vq_ring.desc[i + mid_idx].next = i;
 353                                 vq->vq_ring.desc[i + mid_idx].addr =
 354                                         vq->virtio_net_hdr_mem +
 355                                                 mid_idx * vq->hw->vtnet_hdr_size;
 356                                 vq->vq_ring.desc[i + mid_idx].len =
 357                                         vq->hw->vtnet_hdr_size;
 358                                 vq->vq_ring.desc[i + mid_idx].flags =
 359                                         VRING_DESC_F_NEXT;
 360                                 vq->vq_ring.desc[i].flags = 0;
 361                         }
 362                         for (i = mid_idx; i < vq->vq_nentries; i++)
 363                                 vq->vq_ring.avail->ring[i] = i;
 364                 }
 365
 366                 VIRTIO_WRITE_REG_2(vq->hw, VIRTIO_PCI_QUEUE_SEL,
 367                         vq->vq_queue_index);
 368                 VIRTIO_WRITE_REG_4(vq->hw, VIRTIO_PCI_QUEUE_PFN,
 369                         vq->mz->phys_addr >> VIRTIO_PCI_QUEUE_ADDR_SHIFT);
 370         } else {
 371                 VIRTIO_WRITE_REG_2(vq->hw, VIRTIO_PCI_QUEUE_SEL,
 372                         vq->vq_queue_index);
 373                 VIRTIO_WRITE_REG_4(vq->hw, VIRTIO_PCI_QUEUE_PFN,
 374                         vq->mz->phys_addr >> VIRTIO_PCI_QUEUE_ADDR_SHIFT);
 375         }
 376 }
 377
 378 void
 379 virtio_dev_cq_start(struct rte_eth_dev *dev)
 380 {
 381         struct virtio_hw *hw = dev->data->dev_private;
 382
 383         if (hw->cvq) {
 384                 virtio_dev_vring_start(hw->cvq, VTNET_CQ);
 385                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq);
 386         }
 387 }
 388
 389 void
 390 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 391 {
 392         /*
 393          * Start receive and transmit vrings
 394          * -    Setup vring structure for all queues
 395          * -    Initialize descriptor for the rx vring
 396          * -    Allocate blank mbufs for the each rx descriptor
 397          *
 398          */
 399         int i;
 400
 401         PMD_INIT_FUNC_TRACE();
 402
 403         /* Start rx vring. */
 404         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 405                 virtio_dev_vring_start(dev->data->rx_queues[i], VTNET_RQ);
 406                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->rx_queues[i]);
 407         }
 408
 409         /* Start tx vring. */
 410         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 411                 virtio_dev_vring_start(dev->data->tx_queues[i], VTNET_TQ);
 412                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->tx_queues[i]);
 413         }
 414 }
 415
 416 int
 417 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 418                         uint16_t queue_idx,
 419                         uint16_t nb_desc,
 420                         unsigned int socket_id,
 421                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 422                         struct rte_mempool *mp)
 423 {
 424         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 425         struct virtqueue *vq;
 426         int ret;
 427
 428         PMD_INIT_FUNC_TRACE();
 429         ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
 430                         nb_desc, socket_id, &vq);
 431         if (ret < 0) {
 432                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 433                 return ret;
 434         }
 435
 436         /* Create mempool for rx mbuf allocation */
 437         vq->mpool = mp;
 438
 439         dev->data->rx_queues[queue_idx] = vq;
 440
 441         virtio_rxq_vec_setup(vq);
 442
 443         return 0;
 444 }
 445
 446 void
 447 virtio_dev_rx_queue_release(void *rxq)
 448 {
 449         virtio_dev_queue_release(rxq);
 450 }
 451
 452 /*
 453  * struct rte_eth_dev *dev: Used to update dev
 454  * uint16_t nb_desc: Defaults to values read from config space
 455  * unsigned int socket_id: Used to allocate memzone
 456  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 457  * uint16_t queue_idx: Just used as an index in dev txq list
 458  */
 459 int
 460 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 461                         uint16_t queue_idx,
 462                         uint16_t nb_desc,
 463                         unsigned int socket_id,
 464                         const struct rte_eth_txconf *tx_conf)
 465 {
 466         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 467         struct virtio_hw *hw = dev->data->dev_private;
 468         struct virtqueue *vq;
 469         uint16_t tx_free_thresh;
 470         int ret;
 471
 472         PMD_INIT_FUNC_TRACE();
 473
 474         if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
 475             != ETH_TXQ_FLAGS_NOXSUMS) {
 476                 PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
 477                 return -EINVAL;
 478         }
 479
 480         /* Use simple rx/tx func if single segment and no offloads */
 481         if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 482              !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 483                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 484                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 485                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 486                 use_simple_rxtx = 1;
 487         }
 488
 489         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
 490                         nb_desc, socket_id, &vq);
 491         if (ret < 0) {
 492                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 493                 return ret;
 494         }
 495
 496         tx_free_thresh = tx_conf->tx_free_thresh;
 497         if (tx_free_thresh == 0)
 498                 tx_free_thresh =
 499                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 500
 501         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 502                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 503                         "number of TX entries minus 3 (%u)."
 504                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 505                         vq->vq_nentries - 3,
 506                         tx_free_thresh, dev->data->port_id, queue_idx);
 507                 return -EINVAL;
 508         }
 509
 510         vq->vq_free_thresh = tx_free_thresh;
 511
 512         dev->data->tx_queues[queue_idx] = vq;
 513         return 0;
 514 }
 515
 516 void
 517 virtio_dev_tx_queue_release(void *txq)
 518 {
 519         virtio_dev_queue_release(txq);
 520 }
 521
 522 static void
 523 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 524 {
 525         int error;
 526         /*
 527          * Requeue the discarded mbuf. This should always be
 528          * successful since it was just dequeued.
 529          */
 530         error = virtqueue_enqueue_recv_refill(vq, m);
 531         if (unlikely(error)) {
 532                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 533                 rte_pktmbuf_free(m);
 534         }
 535 }
 536
 537 static void
 538 virtio_update_packet_stats(struct virtqueue *vq, struct rte_mbuf *mbuf)
 539 {
 540         uint32_t s = mbuf->pkt_len;
 541         struct ether_addr *ea;
 542
 543         if (s == 64) {
 544                 vq->size_bins[1]++;
 545         } else if (s > 64 && s < 1024) {
 546                 uint32_t bin;
 547
 548                 /* count zeros, and offset into correct bin */
 549                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 550                 vq->size_bins[bin]++;
 551         } else {
 552                 if (s < 64)
 553                         vq->size_bins[0]++;
 554                 else if (s < 1519)
 555                         vq->size_bins[6]++;
 556                 else if (s >= 1519)
 557                         vq->size_bins[7]++;
 558         }
 559
 560         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 561         vq->multicast += is_multicast_ether_addr(ea);
 562         vq->broadcast += is_broadcast_ether_addr(ea);
 563 }
 564
 565 #define VIRTIO_MBUF_BURST_SZ 64
 566 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 567 uint16_t
 568 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 569 {
 570         struct virtqueue *rxvq = rx_queue;
 571         struct virtio_hw *hw;
 572         struct rte_mbuf *rxm, *new_mbuf;
 573         uint16_t nb_used, num, nb_rx;
 574         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 575         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 576         int error;
 577         uint32_t i, nb_enqueued;
 578         const uint32_t hdr_size = sizeof(struct virtio_net_hdr);
 579
 580         nb_used = VIRTQUEUE_NUSED(rxvq);
 581
 582         virtio_rmb();
 583
 584         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 585         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 586         if (likely(num > DESC_PER_CACHELINE))
 587                 num = num - ((rxvq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 588
 589         if (num == 0)
 590                 return 0;
 591
 592         num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, num);
 593         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 594
 595         hw = rxvq->hw;
 596         nb_rx = 0;
 597         nb_enqueued = 0;
 598
 599         for (i = 0; i < num ; i++) {
 600                 rxm = rcv_pkts[i];
 601
 602                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 603
 604                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 605                         PMD_RX_LOG(ERR, "Packet drop");
 606                         nb_enqueued++;
 607                         virtio_discard_rxbuf(rxvq, rxm);
 608                         rxvq->errors++;
 609                         continue;
 610                 }
 611
 612                 rxm->port = rxvq->port_id;
 613                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 614
 615                 rxm->nb_segs = 1;
 616                 rxm->next = NULL;
 617                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 618                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 619
 620                 if (hw->vlan_strip)
 621                         rte_vlan_strip(rxm);
 622
 623                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 624
 625                 rx_pkts[nb_rx++] = rxm;
 626
 627                 rxvq->bytes += rx_pkts[nb_rx - 1]->pkt_len;
 628                 virtio_update_packet_stats(rxvq, rxm);
 629         }
 630
 631         rxvq->packets += nb_rx;
 632
 633         /* Allocate new mbuf for the used descriptor */
 634         error = ENOSPC;
 635         while (likely(!virtqueue_full(rxvq))) {
 636                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
 637                 if (unlikely(new_mbuf == NULL)) {
 638                         struct rte_eth_dev *dev
 639                                 = &rte_eth_devices[rxvq->port_id];
 640                         dev->data->rx_mbuf_alloc_failed++;
 641                         break;
 642                 }
 643                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 644                 if (unlikely(error)) {
 645                         rte_pktmbuf_free(new_mbuf);
 646                         break;
 647                 }
 648                 nb_enqueued++;
 649         }
 650
 651         if (likely(nb_enqueued)) {
 652                 vq_update_avail_idx(rxvq);
 653
 654                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 655                         virtqueue_notify(rxvq);
 656                         PMD_RX_LOG(DEBUG, "Notified\n");
 657                 }
 658         }
 659
 660         return nb_rx;
 661 }
 662
 663 uint16_t
 664 virtio_recv_mergeable_pkts(void *rx_queue,
 665                         struct rte_mbuf **rx_pkts,
 666                         uint16_t nb_pkts)
 667 {
 668         struct virtqueue *rxvq = rx_queue;
 669         struct virtio_hw *hw;
 670         struct rte_mbuf *rxm, *new_mbuf;
 671         uint16_t nb_used, num, nb_rx;
 672         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 673         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 674         struct rte_mbuf *prev;
 675         int error;
 676         uint32_t i, nb_enqueued;
 677         uint32_t seg_num;
 678         uint16_t extra_idx;
 679         uint32_t seg_res;
 680         const uint32_t hdr_size = sizeof(struct virtio_net_hdr_mrg_rxbuf);
 681
 682         nb_used = VIRTQUEUE_NUSED(rxvq);
 683
 684         virtio_rmb();
 685
 686         if (nb_used == 0)
 687                 return 0;
 688
 689         PMD_RX_LOG(DEBUG, "used:%d\n", nb_used);
 690
 691         hw = rxvq->hw;
 692         nb_rx = 0;
 693         i = 0;
 694         nb_enqueued = 0;
 695         seg_num = 0;
 696         extra_idx = 0;
 697         seg_res = 0;
 698
 699         while (i < nb_used) {
 700                 struct virtio_net_hdr_mrg_rxbuf *header;
 701
 702                 if (nb_rx == nb_pkts)
 703                         break;
 704
 705                 num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, 1);
 706                 if (num != 1)
 707                         continue;
 708
 709                 i++;
 710
 711                 PMD_RX_LOG(DEBUG, "dequeue:%d\n", num);
 712                 PMD_RX_LOG(DEBUG, "packet len:%d\n", len[0]);
 713
 714                 rxm = rcv_pkts[0];
 715
 716                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 717                         PMD_RX_LOG(ERR, "Packet drop\n");
 718                         nb_enqueued++;
 719                         virtio_discard_rxbuf(rxvq, rxm);
 720                         rxvq->errors++;
 721                         continue;
 722                 }
 723
 724                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 725                         RTE_PKTMBUF_HEADROOM - hdr_size);
 726                 seg_num = header->num_buffers;
 727
 728                 if (seg_num == 0)
 729                         seg_num = 1;
 730
 731                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 732                 rxm->nb_segs = seg_num;
 733                 rxm->next = NULL;
 734                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 735                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 736
 737                 rxm->port = rxvq->port_id;
 738                 rx_pkts[nb_rx] = rxm;
 739                 prev = rxm;
 740
 741                 seg_res = seg_num - 1;
 742
 743                 while (seg_res != 0) {
 744                         /*
 745                          * Get extra segments for current uncompleted packet.
 746                          */
 747                         uint16_t  rcv_cnt =
 748                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 749                         if (likely(VIRTQUEUE_NUSED(rxvq) >= rcv_cnt)) {
 750                                 uint32_t rx_num =
 751                                         virtqueue_dequeue_burst_rx(rxvq,
 752                                         rcv_pkts, len, rcv_cnt);
 753                                 i += rx_num;
 754                                 rcv_cnt = rx_num;
 755                         } else {
 756                                 PMD_RX_LOG(ERR,
 757                                         "No enough segments for packet.\n");
 758                                 nb_enqueued++;
 759                                 virtio_discard_rxbuf(rxvq, rxm);
 760                                 rxvq->errors++;
 761                                 break;
 762                         }
 763
 764                         extra_idx = 0;
 765
 766                         while (extra_idx < rcv_cnt) {
 767                                 rxm = rcv_pkts[extra_idx];
 768
 769                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 770                                 rxm->next = NULL;
 771                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 772                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 773
 774                                 if (prev)
 775                                         prev->next = rxm;
 776
 777                                 prev = rxm;
 778                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 779                                 extra_idx++;
 780                         };
 781                         seg_res -= rcv_cnt;
 782                 }
 783
 784                 if (hw->vlan_strip)
 785                         rte_vlan_strip(rx_pkts[nb_rx]);
 786
 787                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 788                         rx_pkts[nb_rx]->data_len);
 789
 790                 rxvq->bytes += rx_pkts[nb_rx]->pkt_len;
 791                 virtio_update_packet_stats(rxvq, rx_pkts[nb_rx]);
 792                 nb_rx++;
 793         }
 794
 795         rxvq->packets += nb_rx;
 796
 797         /* Allocate new mbuf for the used descriptor */
 798         error = ENOSPC;
 799         while (likely(!virtqueue_full(rxvq))) {
 800                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
 801                 if (unlikely(new_mbuf == NULL)) {
 802                         struct rte_eth_dev *dev
 803                                 = &rte_eth_devices[rxvq->port_id];
 804                         dev->data->rx_mbuf_alloc_failed++;
 805                         break;
 806                 }
 807                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 808                 if (unlikely(error)) {
 809                         rte_pktmbuf_free(new_mbuf);
 810                         break;
 811                 }
 812                 nb_enqueued++;
 813         }
 814
 815         if (likely(nb_enqueued)) {
 816                 vq_update_avail_idx(rxvq);
 817
 818                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 819                         virtqueue_notify(rxvq);
 820                         PMD_RX_LOG(DEBUG, "Notified");
 821                 }
 822         }
 823
 824         return nb_rx;
 825 }
 826
 827 uint16_t
 828 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 829 {
 830         struct virtqueue *txvq = tx_queue;
 831         struct rte_mbuf *txm;
 832         uint16_t nb_used, nb_tx;
 833         int error;
 834
 835         if (unlikely(nb_pkts < 1))
 836                 return nb_pkts;
 837
 838         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
 839         nb_used = VIRTQUEUE_NUSED(txvq);
 840
 841         virtio_rmb();
 842         if (likely(nb_used > txvq->vq_nentries - txvq->vq_free_thresh))
 843                 virtio_xmit_cleanup(txvq, nb_used);
 844
 845         nb_tx = 0;
 846
 847         while (nb_tx < nb_pkts) {
 848                 /* Need one more descriptor for virtio header. */
 849                 int need = tx_pkts[nb_tx]->nb_segs - txvq->vq_free_cnt + 1;
 850
 851                 /*Positive value indicates it need free vring descriptors */
 852                 if (unlikely(need > 0)) {
 853                         nb_used = VIRTQUEUE_NUSED(txvq);
 854                         virtio_rmb();
 855                         need = RTE_MIN(need, (int)nb_used);
 856
 857                         virtio_xmit_cleanup(txvq, need);
 858                         need = (int)tx_pkts[nb_tx]->nb_segs -
 859                                 txvq->vq_free_cnt + 1;
 860                 }
 861
 862                 /*
 863                  * Zero or negative value indicates it has enough free
 864                  * descriptors to use for transmitting.
 865                  */
 866                 if (likely(need <= 0)) {
 867                         txm = tx_pkts[nb_tx];
 868
 869                         /* Do VLAN tag insertion */
 870                         if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
 871                                 error = rte_vlan_insert(&txm);
 872                                 if (unlikely(error)) {
 873                                         rte_pktmbuf_free(txm);
 874                                         ++nb_tx;
 875                                         continue;
 876                                 }
 877                         }
 878
 879                         /* Enqueue Packet buffers */
 880                         error = virtqueue_enqueue_xmit(txvq, txm);
 881                         if (unlikely(error)) {
 882                                 if (error == ENOSPC)
 883                                         PMD_TX_LOG(ERR, "virtqueue_enqueue Free count = 0");
 884                                 else if (error == EMSGSIZE)
 885                                         PMD_TX_LOG(ERR, "virtqueue_enqueue Free count < 1");
 886                                 else
 887                                         PMD_TX_LOG(ERR, "virtqueue_enqueue error: %d", error);
 888                                 break;
 889                         }
 890                         nb_tx++;
 891                         txvq->bytes += txm->pkt_len;
 892                         virtio_update_packet_stats(txvq, txm);
 893                 } else {
 894                         PMD_TX_LOG(ERR, "No free tx descriptors to transmit");
 895                         break;
 896                 }
 897         }
 898
 899         txvq->packets += nb_tx;
 900
 901         if (likely(nb_tx)) {
 902                 vq_update_avail_idx(txvq);
 903
 904                 if (unlikely(virtqueue_kick_prepare(txvq))) {
 905                         virtqueue_notify(txvq);
 906                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
 907                 }
 908         }
 909
 910         return nb_tx;
 911 }