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