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 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
  71 static int use_simple_rxtx;
  72 #endif
  73
  74 static void
  75 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  76 {
  77         struct vring_desc *dp, *dp_tail;
  78         struct vq_desc_extra *dxp;
  79         uint16_t desc_idx_last = desc_idx;
  80
  81         dp  = &vq->vq_ring.desc[desc_idx];
  82         dxp = &vq->vq_descx[desc_idx];
  83         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  84         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  85                 while (dp->flags & VRING_DESC_F_NEXT) {
  86                         desc_idx_last = dp->next;
  87                         dp = &vq->vq_ring.desc[dp->next];
  88                 }
  89         }
  90         dxp->ndescs = 0;
  91
  92         /*
  93          * We must append the existing free chain, if any, to the end of
  94          * newly freed chain. If the virtqueue was completely used, then
  95          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  96          */
  97         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  98                 vq->vq_desc_head_idx = desc_idx;
  99         } else {
 100                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
 101                 dp_tail->next = desc_idx;
 102         }
 103
 104         vq->vq_desc_tail_idx = desc_idx_last;
 105         dp->next = VQ_RING_DESC_CHAIN_END;
 106 }
 107
 108 static uint16_t
 109 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 110                            uint32_t *len, uint16_t num)
 111 {
 112         struct vring_used_elem *uep;
 113         struct rte_mbuf *cookie;
 114         uint16_t used_idx, desc_idx;
 115         uint16_t i;
 116
 117         /*  Caller does the check */
 118         for (i = 0; i < num ; i++) {
 119                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 120                 uep = &vq->vq_ring.used->ring[used_idx];
 121                 desc_idx = (uint16_t) uep->id;
 122                 len[i] = uep->len;
 123                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 124
 125                 if (unlikely(cookie == NULL)) {
 126                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 127                                 vq->vq_used_cons_idx);
 128                         break;
 129                 }
 130
 131                 rte_prefetch0(cookie);
 132                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 133                 rx_pkts[i]  = cookie;
 134                 vq->vq_used_cons_idx++;
 135                 vq_ring_free_chain(vq, desc_idx);
 136                 vq->vq_descx[desc_idx].cookie = NULL;
 137         }
 138
 139         return i;
 140 }
 141
 142 #ifndef DEFAULT_TX_FREE_THRESH
 143 #define DEFAULT_TX_FREE_THRESH 32
 144 #endif
 145
 146 /* Cleanup from completed transmits. */
 147 static void
 148 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 149 {
 150         uint16_t i, used_idx, desc_idx;
 151         for (i = 0; i < num; i++) {
 152                 struct vring_used_elem *uep;
 153                 struct vq_desc_extra *dxp;
 154
 155                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 156                 uep = &vq->vq_ring.used->ring[used_idx];
 157
 158                 desc_idx = (uint16_t) uep->id;
 159                 dxp = &vq->vq_descx[desc_idx];
 160                 vq->vq_used_cons_idx++;
 161                 vq_ring_free_chain(vq, desc_idx);
 162
 163                 if (dxp->cookie != NULL) {
 164                         rte_pktmbuf_free(dxp->cookie);
 165                         dxp->cookie = NULL;
 166                 }
 167         }
 168 }
 169
 170
 171 static inline int
 172 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 173 {
 174         struct vq_desc_extra *dxp;
 175         struct virtio_hw *hw = vq->hw;
 176         struct vring_desc *start_dp;
 177         uint16_t needed = 1;
 178         uint16_t head_idx, idx;
 179
 180         if (unlikely(vq->vq_free_cnt == 0))
 181                 return -ENOSPC;
 182         if (unlikely(vq->vq_free_cnt < needed))
 183                 return -EMSGSIZE;
 184
 185         head_idx = vq->vq_desc_head_idx;
 186         if (unlikely(head_idx >= vq->vq_nentries))
 187                 return -EFAULT;
 188
 189         idx = head_idx;
 190         dxp = &vq->vq_descx[idx];
 191         dxp->cookie = (void *)cookie;
 192         dxp->ndescs = needed;
 193
 194         start_dp = vq->vq_ring.desc;
 195         start_dp[idx].addr =
 196                 (uint64_t)(cookie->buf_physaddr + RTE_PKTMBUF_HEADROOM
 197                 - hw->vtnet_hdr_size);
 198         start_dp[idx].len =
 199                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 200         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 201         idx = start_dp[idx].next;
 202         vq->vq_desc_head_idx = idx;
 203         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 204                 vq->vq_desc_tail_idx = idx;
 205         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 206         vq_update_avail_ring(vq, head_idx);
 207
 208         return 0;
 209 }
 210
 211 static inline void
 212 virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie,
 213                        uint16_t needed, int use_indirect, int can_push)
 214 {
 215         struct vq_desc_extra *dxp;
 216         struct vring_desc *start_dp;
 217         uint16_t seg_num = cookie->nb_segs;
 218         uint16_t head_idx, idx;
 219         uint16_t head_size = txvq->hw->vtnet_hdr_size;
 220         unsigned long offs;
 221
 222         head_idx = txvq->vq_desc_head_idx;
 223         idx = head_idx;
 224         dxp = &txvq->vq_descx[idx];
 225         dxp->cookie = (void *)cookie;
 226         dxp->ndescs = needed;
 227
 228         start_dp = txvq->vq_ring.desc;
 229
 230         if (can_push) {
 231                 /* put on zero'd transmit header (no offloads) */
 232                 void *hdr = rte_pktmbuf_prepend(cookie, head_size);
 233
 234                 memset(hdr, 0, head_size);
 235         } else if (use_indirect) {
 236                 /* setup tx ring slot to point to indirect
 237                  * descriptor list stored in reserved region.
 238                  *
 239                  * the first slot in indirect ring is already preset
 240                  * to point to the header in reserved region
 241                  */
 242                 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 243
 244                 offs = idx * sizeof(struct virtio_tx_region)
 245                         + offsetof(struct virtio_tx_region, tx_indir);
 246
 247                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 248                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 249                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 250
 251                 /* loop below will fill in rest of the indirect elements */
 252                 start_dp = txr[idx].tx_indir;
 253                 idx = 1;
 254         } else {
 255                 /* setup first tx ring slot to point to header
 256                  * stored in reserved region.
 257                  */
 258                 offs = idx * sizeof(struct virtio_tx_region)
 259                         + offsetof(struct virtio_tx_region, tx_hdr);
 260
 261                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 262                 start_dp[idx].len   = txvq->hw->vtnet_hdr_size;
 263                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 264                 idx = start_dp[idx].next;
 265         }
 266
 267         do {
 268                 start_dp[idx].addr  = rte_mbuf_data_dma_addr(cookie);
 269                 start_dp[idx].len   = cookie->data_len;
 270                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 271                 idx = start_dp[idx].next;
 272         } while ((cookie = cookie->next) != NULL);
 273
 274         if (use_indirect)
 275                 idx = txvq->vq_ring.desc[head_idx].next;
 276
 277         txvq->vq_desc_head_idx = idx;
 278         if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 279                 txvq->vq_desc_tail_idx = idx;
 280         txvq->vq_free_cnt = (uint16_t)(txvq->vq_free_cnt - needed);
 281         vq_update_avail_ring(txvq, head_idx);
 282 }
 283
 284 static void
 285 virtio_dev_vring_start(struct virtqueue *vq, int queue_type)
 286 {
 287         struct rte_mbuf *m;
 288         int i, nbufs, error, size = vq->vq_nentries;
 289         struct vring *vr = &vq->vq_ring;
 290         uint8_t *ring_mem = vq->vq_ring_virt_mem;
 291
 292         PMD_INIT_FUNC_TRACE();
 293
 294         /*
 295          * Reinitialise since virtio port might have been stopped and restarted
 296          */
 297         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
 298         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
 299         vq->vq_used_cons_idx = 0;
 300         vq->vq_desc_head_idx = 0;
 301         vq->vq_avail_idx = 0;
 302         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
 303         vq->vq_free_cnt = vq->vq_nentries;
 304         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 305
 306         vring_desc_init(vr->desc, size);
 307
 308         /*
 309          * Disable device(host) interrupting guest
 310          */
 311         virtqueue_disable_intr(vq);
 312
 313         /* Only rx virtqueue needs mbufs to be allocated at initialization */
 314         if (queue_type == VTNET_RQ) {
 315                 if (vq->mpool == NULL)
 316                         rte_exit(EXIT_FAILURE,
 317                         "Cannot allocate initial mbufs for rx virtqueue");
 318
 319                 /* Allocate blank mbufs for the each rx descriptor */
 320                 nbufs = 0;
 321                 error = ENOSPC;
 322
 323 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 324                 if (use_simple_rxtx)
 325                         for (i = 0; i < vq->vq_nentries; i++) {
 326                                 vq->vq_ring.avail->ring[i] = i;
 327                                 vq->vq_ring.desc[i].flags = VRING_DESC_F_WRITE;
 328                         }
 329 #endif
 330                 memset(&vq->fake_mbuf, 0, sizeof(vq->fake_mbuf));
 331                 for (i = 0; i < RTE_PMD_VIRTIO_RX_MAX_BURST; i++)
 332                         vq->sw_ring[vq->vq_nentries + i] = &vq->fake_mbuf;
 333
 334                 while (!virtqueue_full(vq)) {
 335                         m = rte_mbuf_raw_alloc(vq->mpool);
 336                         if (m == NULL)
 337                                 break;
 338
 339                         /******************************************
 340                         *         Enqueue allocated buffers        *
 341                         *******************************************/
 342 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 343                         if (use_simple_rxtx)
 344                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 345                         else
 346 #endif
 347                                 error = virtqueue_enqueue_recv_refill(vq, m);
 348                         if (error) {
 349                                 rte_pktmbuf_free(m);
 350                                 break;
 351                         }
 352                         nbufs++;
 353                 }
 354
 355                 vq_update_avail_idx(vq);
 356
 357                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 358         } else if (queue_type == VTNET_TQ) {
 359 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 360                 if (use_simple_rxtx) {
 361                         int mid_idx  = vq->vq_nentries >> 1;
 362                         for (i = 0; i < mid_idx; i++) {
 363                                 vq->vq_ring.avail->ring[i] = i + mid_idx;
 364                                 vq->vq_ring.desc[i + mid_idx].next = i;
 365                                 vq->vq_ring.desc[i + mid_idx].addr =
 366                                         vq->virtio_net_hdr_mem +
 367                                         offsetof(struct virtio_tx_region, tx_hdr);
 368                                 vq->vq_ring.desc[i + mid_idx].len =
 369                                         vq->hw->vtnet_hdr_size;
 370                                 vq->vq_ring.desc[i + mid_idx].flags =
 371                                         VRING_DESC_F_NEXT;
 372                                 vq->vq_ring.desc[i].flags = 0;
 373                         }
 374                         for (i = mid_idx; i < vq->vq_nentries; i++)
 375                                 vq->vq_ring.avail->ring[i] = i;
 376                 }
 377 #endif
 378         }
 379 }
 380
 381 void
 382 virtio_dev_cq_start(struct rte_eth_dev *dev)
 383 {
 384         struct virtio_hw *hw = dev->data->dev_private;
 385
 386         if (hw->cvq) {
 387                 virtio_dev_vring_start(hw->cvq, VTNET_CQ);
 388                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq);
 389         }
 390 }
 391
 392 void
 393 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 394 {
 395         /*
 396          * Start receive and transmit vrings
 397          * -    Setup vring structure for all queues
 398          * -    Initialize descriptor for the rx vring
 399          * -    Allocate blank mbufs for the each rx descriptor
 400          *
 401          */
 402         int i;
 403
 404         PMD_INIT_FUNC_TRACE();
 405
 406         /* Start rx vring. */
 407         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 408                 virtio_dev_vring_start(dev->data->rx_queues[i], VTNET_RQ);
 409                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->rx_queues[i]);
 410         }
 411
 412         /* Start tx vring. */
 413         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 414                 virtio_dev_vring_start(dev->data->tx_queues[i], VTNET_TQ);
 415                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->tx_queues[i]);
 416         }
 417 }
 418
 419 int
 420 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 421                         uint16_t queue_idx,
 422                         uint16_t nb_desc,
 423                         unsigned int socket_id,
 424                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 425                         struct rte_mempool *mp)
 426 {
 427         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 428         struct virtqueue *vq;
 429         int ret;
 430
 431         PMD_INIT_FUNC_TRACE();
 432         ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
 433                         nb_desc, socket_id, &vq);
 434         if (ret < 0) {
 435                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 436                 return ret;
 437         }
 438
 439         /* Create mempool for rx mbuf allocation */
 440         vq->mpool = mp;
 441
 442         dev->data->rx_queues[queue_idx] = vq;
 443
 444 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 445         virtio_rxq_vec_setup(vq);
 446 #endif
 447
 448         return 0;
 449 }
 450
 451 void
 452 virtio_dev_rx_queue_release(void *rxq)
 453 {
 454         virtio_dev_queue_release(rxq);
 455 }
 456
 457 /*
 458  * struct rte_eth_dev *dev: Used to update dev
 459  * uint16_t nb_desc: Defaults to values read from config space
 460  * unsigned int socket_id: Used to allocate memzone
 461  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 462  * uint16_t queue_idx: Just used as an index in dev txq list
 463  */
 464 int
 465 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 466                         uint16_t queue_idx,
 467                         uint16_t nb_desc,
 468                         unsigned int socket_id,
 469                         const struct rte_eth_txconf *tx_conf)
 470 {
 471         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 472
 473 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 474         struct virtio_hw *hw = dev->data->dev_private;
 475 #endif
 476         struct virtqueue *vq;
 477         uint16_t tx_free_thresh;
 478         int ret;
 479
 480         PMD_INIT_FUNC_TRACE();
 481
 482         if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
 483             != ETH_TXQ_FLAGS_NOXSUMS) {
 484                 PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
 485                 return -EINVAL;
 486         }
 487
 488 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 489         /* Use simple rx/tx func if single segment and no offloads */
 490         if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 491              !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 492                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 493                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 494                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 495                 use_simple_rxtx = 1;
 496         }
 497 #endif
 498
 499         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
 500                         nb_desc, socket_id, &vq);
 501         if (ret < 0) {
 502                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 503                 return ret;
 504         }
 505
 506         tx_free_thresh = tx_conf->tx_free_thresh;
 507         if (tx_free_thresh == 0)
 508                 tx_free_thresh =
 509                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 510
 511         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 512                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 513                         "number of TX entries minus 3 (%u)."
 514                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 515                         vq->vq_nentries - 3,
 516                         tx_free_thresh, dev->data->port_id, queue_idx);
 517                 return -EINVAL;
 518         }
 519
 520         vq->vq_free_thresh = tx_free_thresh;
 521
 522         dev->data->tx_queues[queue_idx] = vq;
 523         return 0;
 524 }
 525
 526 void
 527 virtio_dev_tx_queue_release(void *txq)
 528 {
 529         virtio_dev_queue_release(txq);
 530 }
 531
 532 static void
 533 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 534 {
 535         int error;
 536         /*
 537          * Requeue the discarded mbuf. This should always be
 538          * successful since it was just dequeued.
 539          */
 540         error = virtqueue_enqueue_recv_refill(vq, m);
 541         if (unlikely(error)) {
 542                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 543                 rte_pktmbuf_free(m);
 544         }
 545 }
 546
 547 static void
 548 virtio_update_packet_stats(struct virtqueue *vq, struct rte_mbuf *mbuf)
 549 {
 550         uint32_t s = mbuf->pkt_len;
 551         struct ether_addr *ea;
 552
 553         if (s == 64) {
 554                 vq->size_bins[1]++;
 555         } else if (s > 64 && s < 1024) {
 556                 uint32_t bin;
 557
 558                 /* count zeros, and offset into correct bin */
 559                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 560                 vq->size_bins[bin]++;
 561         } else {
 562                 if (s < 64)
 563                         vq->size_bins[0]++;
 564                 else if (s < 1519)
 565                         vq->size_bins[6]++;
 566                 else if (s >= 1519)
 567                         vq->size_bins[7]++;
 568         }
 569
 570         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 571         if (is_multicast_ether_addr(ea)) {
 572                 if (is_broadcast_ether_addr(ea))
 573                         vq->broadcast++;
 574                 else
 575                         vq->multicast++;
 576         }
 577 }
 578
 579 #define VIRTIO_MBUF_BURST_SZ 64
 580 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 581 uint16_t
 582 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 583 {
 584         struct virtqueue *rxvq = rx_queue;
 585         struct virtio_hw *hw;
 586         struct rte_mbuf *rxm, *new_mbuf;
 587         uint16_t nb_used, num, nb_rx;
 588         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 589         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 590         int error;
 591         uint32_t i, nb_enqueued;
 592         uint32_t hdr_size;
 593
 594         nb_used = VIRTQUEUE_NUSED(rxvq);
 595
 596         virtio_rmb();
 597
 598         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 599         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 600         if (likely(num > DESC_PER_CACHELINE))
 601                 num = num - ((rxvq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 602
 603         num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, num);
 604         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 605
 606         hw = rxvq->hw;
 607         nb_rx = 0;
 608         nb_enqueued = 0;
 609         hdr_size = hw->vtnet_hdr_size;
 610
 611         for (i = 0; i < num ; i++) {
 612                 rxm = rcv_pkts[i];
 613
 614                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 615
 616                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 617                         PMD_RX_LOG(ERR, "Packet drop");
 618                         nb_enqueued++;
 619                         virtio_discard_rxbuf(rxvq, rxm);
 620                         rxvq->errors++;
 621                         continue;
 622                 }
 623
 624                 rxm->port = rxvq->port_id;
 625                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 626                 rxm->ol_flags = 0;
 627                 rxm->vlan_tci = 0;
 628
 629                 rxm->nb_segs = 1;
 630                 rxm->next = NULL;
 631                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 632                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 633
 634                 if (hw->vlan_strip)
 635                         rte_vlan_strip(rxm);
 636
 637                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 638
 639                 rx_pkts[nb_rx++] = rxm;
 640
 641                 rxvq->bytes += rx_pkts[nb_rx - 1]->pkt_len;
 642                 virtio_update_packet_stats(rxvq, rxm);
 643         }
 644
 645         rxvq->packets += nb_rx;
 646
 647         /* Allocate new mbuf for the used descriptor */
 648         error = ENOSPC;
 649         while (likely(!virtqueue_full(rxvq))) {
 650                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 651                 if (unlikely(new_mbuf == NULL)) {
 652                         struct rte_eth_dev *dev
 653                                 = &rte_eth_devices[rxvq->port_id];
 654                         dev->data->rx_mbuf_alloc_failed++;
 655                         break;
 656                 }
 657                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 658                 if (unlikely(error)) {
 659                         rte_pktmbuf_free(new_mbuf);
 660                         break;
 661                 }
 662                 nb_enqueued++;
 663         }
 664
 665         if (likely(nb_enqueued)) {
 666                 vq_update_avail_idx(rxvq);
 667
 668                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 669                         virtqueue_notify(rxvq);
 670                         PMD_RX_LOG(DEBUG, "Notified");
 671                 }
 672         }
 673
 674         return nb_rx;
 675 }
 676
 677 uint16_t
 678 virtio_recv_mergeable_pkts(void *rx_queue,
 679                         struct rte_mbuf **rx_pkts,
 680                         uint16_t nb_pkts)
 681 {
 682         struct virtqueue *rxvq = rx_queue;
 683         struct virtio_hw *hw;
 684         struct rte_mbuf *rxm, *new_mbuf;
 685         uint16_t nb_used, num, nb_rx;
 686         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 687         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 688         struct rte_mbuf *prev;
 689         int error;
 690         uint32_t i, nb_enqueued;
 691         uint32_t seg_num;
 692         uint16_t extra_idx;
 693         uint32_t seg_res;
 694         uint32_t hdr_size;
 695
 696         nb_used = VIRTQUEUE_NUSED(rxvq);
 697
 698         virtio_rmb();
 699
 700         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 701
 702         hw = rxvq->hw;
 703         nb_rx = 0;
 704         i = 0;
 705         nb_enqueued = 0;
 706         seg_num = 0;
 707         extra_idx = 0;
 708         seg_res = 0;
 709         hdr_size = hw->vtnet_hdr_size;
 710
 711         while (i < nb_used) {
 712                 struct virtio_net_hdr_mrg_rxbuf *header;
 713
 714                 if (nb_rx == nb_pkts)
 715                         break;
 716
 717                 num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, 1);
 718                 if (num != 1)
 719                         continue;
 720
 721                 i++;
 722
 723                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 724                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 725
 726                 rxm = rcv_pkts[0];
 727
 728                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 729                         PMD_RX_LOG(ERR, "Packet drop");
 730                         nb_enqueued++;
 731                         virtio_discard_rxbuf(rxvq, rxm);
 732                         rxvq->errors++;
 733                         continue;
 734                 }
 735
 736                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 737                         RTE_PKTMBUF_HEADROOM - hdr_size);
 738                 seg_num = header->num_buffers;
 739
 740                 if (seg_num == 0)
 741                         seg_num = 1;
 742
 743                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 744                 rxm->nb_segs = seg_num;
 745                 rxm->next = NULL;
 746                 rxm->ol_flags = 0;
 747                 rxm->vlan_tci = 0;
 748                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 749                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 750
 751                 rxm->port = rxvq->port_id;
 752                 rx_pkts[nb_rx] = rxm;
 753                 prev = rxm;
 754
 755                 seg_res = seg_num - 1;
 756
 757                 while (seg_res != 0) {
 758                         /*
 759                          * Get extra segments for current uncompleted packet.
 760                          */
 761                         uint16_t  rcv_cnt =
 762                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 763                         if (likely(VIRTQUEUE_NUSED(rxvq) >= rcv_cnt)) {
 764                                 uint32_t rx_num =
 765                                         virtqueue_dequeue_burst_rx(rxvq,
 766                                         rcv_pkts, len, rcv_cnt);
 767                                 i += rx_num;
 768                                 rcv_cnt = rx_num;
 769                         } else {
 770                                 PMD_RX_LOG(ERR,
 771                                            "No enough segments for packet.");
 772                                 nb_enqueued++;
 773                                 virtio_discard_rxbuf(rxvq, rxm);
 774                                 rxvq->errors++;
 775                                 break;
 776                         }
 777
 778                         extra_idx = 0;
 779
 780                         while (extra_idx < rcv_cnt) {
 781                                 rxm = rcv_pkts[extra_idx];
 782
 783                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 784                                 rxm->next = NULL;
 785                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 786                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 787
 788                                 if (prev)
 789                                         prev->next = rxm;
 790
 791                                 prev = rxm;
 792                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 793                                 extra_idx++;
 794                         };
 795                         seg_res -= rcv_cnt;
 796                 }
 797
 798                 if (hw->vlan_strip)
 799                         rte_vlan_strip(rx_pkts[nb_rx]);
 800
 801                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 802                         rx_pkts[nb_rx]->data_len);
 803
 804                 rxvq->bytes += rx_pkts[nb_rx]->pkt_len;
 805                 virtio_update_packet_stats(rxvq, rx_pkts[nb_rx]);
 806                 nb_rx++;
 807         }
 808
 809         rxvq->packets += nb_rx;
 810
 811         /* Allocate new mbuf for the used descriptor */
 812         error = ENOSPC;
 813         while (likely(!virtqueue_full(rxvq))) {
 814                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 815                 if (unlikely(new_mbuf == NULL)) {
 816                         struct rte_eth_dev *dev
 817                                 = &rte_eth_devices[rxvq->port_id];
 818                         dev->data->rx_mbuf_alloc_failed++;
 819                         break;
 820                 }
 821                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 822                 if (unlikely(error)) {
 823                         rte_pktmbuf_free(new_mbuf);
 824                         break;
 825                 }
 826                 nb_enqueued++;
 827         }
 828
 829         if (likely(nb_enqueued)) {
 830                 vq_update_avail_idx(rxvq);
 831
 832                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 833                         virtqueue_notify(rxvq);
 834                         PMD_RX_LOG(DEBUG, "Notified");
 835                 }
 836         }
 837
 838         return nb_rx;
 839 }
 840
 841 uint16_t
 842 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 843 {
 844         struct virtqueue *txvq = tx_queue;
 845         struct virtio_hw *hw = txvq->hw;
 846         uint16_t hdr_size = hw->vtnet_hdr_size;
 847         uint16_t nb_used, nb_tx;
 848         int error;
 849
 850         if (unlikely(nb_pkts < 1))
 851                 return nb_pkts;
 852
 853         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
 854         nb_used = VIRTQUEUE_NUSED(txvq);
 855
 856         virtio_rmb();
 857         if (likely(nb_used > txvq->vq_nentries - txvq->vq_free_thresh))
 858                 virtio_xmit_cleanup(txvq, nb_used);
 859
 860         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 861                 struct rte_mbuf *txm = tx_pkts[nb_tx];
 862                 int can_push = 0, use_indirect = 0, slots, need;
 863
 864                 /* Do VLAN tag insertion */
 865                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
 866                         error = rte_vlan_insert(&txm);
 867                         if (unlikely(error)) {
 868                                 rte_pktmbuf_free(txm);
 869                                 continue;
 870                         }
 871                 }
 872
 873                 /* optimize ring usage */
 874                 if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
 875                     rte_mbuf_refcnt_read(txm) == 1 &&
 876                     txm->nb_segs == 1 &&
 877                     rte_pktmbuf_headroom(txm) >= hdr_size &&
 878                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
 879                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
 880                         can_push = 1;
 881                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
 882                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
 883                         use_indirect = 1;
 884
 885                 /* How many main ring entries are needed to this Tx?
 886                  * any_layout => number of segments
 887                  * indirect   => 1
 888                  * default    => number of segments + 1
 889                  */
 890                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
 891                 need = slots - txvq->vq_free_cnt;
 892
 893                 /* Positive value indicates it need free vring descriptors */
 894                 if (unlikely(need > 0)) {
 895                         nb_used = VIRTQUEUE_NUSED(txvq);
 896                         virtio_rmb();
 897                         need = RTE_MIN(need, (int)nb_used);
 898
 899                         virtio_xmit_cleanup(txvq, need);
 900                         need = slots - txvq->vq_free_cnt;
 901                         if (unlikely(need > 0)) {
 902                                 PMD_TX_LOG(ERR,
 903                                            "No free tx descriptors to transmit");
 904                                 break;
 905                         }
 906                 }
 907
 908                 /* Enqueue Packet buffers */
 909                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
 910
 911                 txvq->bytes += txm->pkt_len;
 912                 virtio_update_packet_stats(txvq, txm);
 913         }
 914
 915         txvq->packets += nb_tx;
 916
 917         if (likely(nb_tx)) {
 918                 vq_update_avail_idx(txvq);
 919
 920                 if (unlikely(virtqueue_kick_prepare(txvq))) {
 921                         virtqueue_notify(txvq);
 922                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
 923                 }
 924         }
 925
 926         return nb_tx;
 927 }