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 int
 212 virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie,
 213                        int use_indirect)
 214 {
 215         struct vq_desc_extra *dxp;
 216         struct vring_desc *start_dp;
 217         uint16_t seg_num = cookie->nb_segs;
 218         uint16_t needed = use_indirect ? 1 : 1 + seg_num;
 219         uint16_t head_idx, idx;
 220         unsigned long offs;
 221
 222         if (unlikely(txvq->vq_free_cnt == 0))
 223                 return -ENOSPC;
 224         if (unlikely(txvq->vq_free_cnt < needed))
 225                 return -EMSGSIZE;
 226         head_idx = txvq->vq_desc_head_idx;
 227         if (unlikely(head_idx >= txvq->vq_nentries))
 228                 return -EFAULT;
 229
 230         idx = head_idx;
 231         dxp = &txvq->vq_descx[idx];
 232         dxp->cookie = (void *)cookie;
 233         dxp->ndescs = needed;
 234
 235         start_dp = txvq->vq_ring.desc;
 236
 237         if (use_indirect) {
 238                 /* setup tx ring slot to point to indirect
 239                  * descriptor list stored in reserved region.
 240                  *
 241                  * the first slot in indirect ring is already preset
 242                  * to point to the header in reserved region
 243                  */
 244                 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 245
 246                 offs = idx * sizeof(struct virtio_tx_region)
 247                         + offsetof(struct virtio_tx_region, tx_indir);
 248
 249                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 250                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 251                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 252
 253                 /* loop below will fill in rest of the indirect elements */
 254                 start_dp = txr[idx].tx_indir;
 255                 idx = 0;
 256         } else {
 257                 /* setup first tx ring slot to point to header
 258                  * stored in reserved region.
 259                  */
 260                 offs = idx * sizeof(struct virtio_tx_region)
 261                         + offsetof(struct virtio_tx_region, tx_hdr);
 262
 263                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 264                 start_dp[idx].len   = txvq->hw->vtnet_hdr_size;
 265                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 266         }
 267
 268         for (; ((seg_num > 0) && (cookie != NULL)); seg_num--) {
 269                 idx = start_dp[idx].next;
 270                 start_dp[idx].addr  = rte_mbuf_data_dma_addr(cookie);
 271                 start_dp[idx].len   = cookie->data_len;
 272                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 273                 cookie = cookie->next;
 274         }
 275
 276         start_dp[idx].flags &= ~VRING_DESC_F_NEXT;
 277
 278         if (use_indirect)
 279                 idx = txvq->vq_ring.desc[head_idx].next;
 280         else
 281                 idx = start_dp[idx].next;
 282
 283         txvq->vq_desc_head_idx = idx;
 284         if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 285                 txvq->vq_desc_tail_idx = idx;
 286         txvq->vq_free_cnt = (uint16_t)(txvq->vq_free_cnt - needed);
 287         vq_update_avail_ring(txvq, head_idx);
 288
 289         return 0;
 290 }
 291
 292 static inline struct rte_mbuf *
 293 rte_rxmbuf_alloc(struct rte_mempool *mp)
 294 {
 295         struct rte_mbuf *m;
 296
 297         m = __rte_mbuf_raw_alloc(mp);
 298         __rte_mbuf_sanity_check_raw(m, 0);
 299
 300         return m;
 301 }
 302
 303 static void
 304 virtio_dev_vring_start(struct virtqueue *vq, int queue_type)
 305 {
 306         struct rte_mbuf *m;
 307         int i, nbufs, error, size = vq->vq_nentries;
 308         struct vring *vr = &vq->vq_ring;
 309         uint8_t *ring_mem = vq->vq_ring_virt_mem;
 310
 311         PMD_INIT_FUNC_TRACE();
 312
 313         /*
 314          * Reinitialise since virtio port might have been stopped and restarted
 315          */
 316         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
 317         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
 318         vq->vq_used_cons_idx = 0;
 319         vq->vq_desc_head_idx = 0;
 320         vq->vq_avail_idx = 0;
 321         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
 322         vq->vq_free_cnt = vq->vq_nentries;
 323         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 324
 325         vring_desc_init(vr->desc, size);
 326
 327         /*
 328          * Disable device(host) interrupting guest
 329          */
 330         virtqueue_disable_intr(vq);
 331
 332         /* Only rx virtqueue needs mbufs to be allocated at initialization */
 333         if (queue_type == VTNET_RQ) {
 334                 if (vq->mpool == NULL)
 335                         rte_exit(EXIT_FAILURE,
 336                         "Cannot allocate initial mbufs for rx virtqueue");
 337
 338                 /* Allocate blank mbufs for the each rx descriptor */
 339                 nbufs = 0;
 340                 error = ENOSPC;
 341
 342 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 343                 if (use_simple_rxtx)
 344                         for (i = 0; i < vq->vq_nentries; i++) {
 345                                 vq->vq_ring.avail->ring[i] = i;
 346                                 vq->vq_ring.desc[i].flags = VRING_DESC_F_WRITE;
 347                         }
 348 #endif
 349                 memset(&vq->fake_mbuf, 0, sizeof(vq->fake_mbuf));
 350                 for (i = 0; i < RTE_PMD_VIRTIO_RX_MAX_BURST; i++)
 351                         vq->sw_ring[vq->vq_nentries + i] = &vq->fake_mbuf;
 352
 353                 while (!virtqueue_full(vq)) {
 354                         m = rte_rxmbuf_alloc(vq->mpool);
 355                         if (m == NULL)
 356                                 break;
 357
 358                         /******************************************
 359                         *         Enqueue allocated buffers        *
 360                         *******************************************/
 361 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 362                         if (use_simple_rxtx)
 363                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 364                         else
 365 #endif
 366                                 error = virtqueue_enqueue_recv_refill(vq, m);
 367                         if (error) {
 368                                 rte_pktmbuf_free(m);
 369                                 break;
 370                         }
 371                         nbufs++;
 372                 }
 373
 374                 vq_update_avail_idx(vq);
 375
 376                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 377         } else if (queue_type == VTNET_TQ) {
 378 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 379                 if (use_simple_rxtx) {
 380                         int mid_idx  = vq->vq_nentries >> 1;
 381                         for (i = 0; i < mid_idx; i++) {
 382                                 vq->vq_ring.avail->ring[i] = i + mid_idx;
 383                                 vq->vq_ring.desc[i + mid_idx].next = i;
 384                                 vq->vq_ring.desc[i + mid_idx].addr =
 385                                         vq->virtio_net_hdr_mem +
 386                                                 i * vq->hw->vtnet_hdr_size;
 387                                 vq->vq_ring.desc[i + mid_idx].len =
 388                                         vq->hw->vtnet_hdr_size;
 389                                 vq->vq_ring.desc[i + mid_idx].flags =
 390                                         VRING_DESC_F_NEXT;
 391                                 vq->vq_ring.desc[i].flags = 0;
 392                         }
 393                         for (i = mid_idx; i < vq->vq_nentries; i++)
 394                                 vq->vq_ring.avail->ring[i] = i;
 395                 }
 396 #endif
 397         }
 398 }
 399
 400 void
 401 virtio_dev_cq_start(struct rte_eth_dev *dev)
 402 {
 403         struct virtio_hw *hw = dev->data->dev_private;
 404
 405         if (hw->cvq) {
 406                 virtio_dev_vring_start(hw->cvq, VTNET_CQ);
 407                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq);
 408         }
 409 }
 410
 411 void
 412 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 413 {
 414         /*
 415          * Start receive and transmit vrings
 416          * -    Setup vring structure for all queues
 417          * -    Initialize descriptor for the rx vring
 418          * -    Allocate blank mbufs for the each rx descriptor
 419          *
 420          */
 421         int i;
 422
 423         PMD_INIT_FUNC_TRACE();
 424
 425         /* Start rx vring. */
 426         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 427                 virtio_dev_vring_start(dev->data->rx_queues[i], VTNET_RQ);
 428                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->rx_queues[i]);
 429         }
 430
 431         /* Start tx vring. */
 432         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 433                 virtio_dev_vring_start(dev->data->tx_queues[i], VTNET_TQ);
 434                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->tx_queues[i]);
 435         }
 436 }
 437
 438 int
 439 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 440                         uint16_t queue_idx,
 441                         uint16_t nb_desc,
 442                         unsigned int socket_id,
 443                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 444                         struct rte_mempool *mp)
 445 {
 446         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 447         struct virtqueue *vq;
 448         int ret;
 449
 450         PMD_INIT_FUNC_TRACE();
 451         ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
 452                         nb_desc, socket_id, &vq);
 453         if (ret < 0) {
 454                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 455                 return ret;
 456         }
 457
 458         /* Create mempool for rx mbuf allocation */
 459         vq->mpool = mp;
 460
 461         dev->data->rx_queues[queue_idx] = vq;
 462
 463 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 464         virtio_rxq_vec_setup(vq);
 465 #endif
 466
 467         return 0;
 468 }
 469
 470 void
 471 virtio_dev_rx_queue_release(void *rxq)
 472 {
 473         virtio_dev_queue_release(rxq);
 474 }
 475
 476 /*
 477  * struct rte_eth_dev *dev: Used to update dev
 478  * uint16_t nb_desc: Defaults to values read from config space
 479  * unsigned int socket_id: Used to allocate memzone
 480  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 481  * uint16_t queue_idx: Just used as an index in dev txq list
 482  */
 483 int
 484 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 485                         uint16_t queue_idx,
 486                         uint16_t nb_desc,
 487                         unsigned int socket_id,
 488                         const struct rte_eth_txconf *tx_conf)
 489 {
 490         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 491
 492 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 493         struct virtio_hw *hw = dev->data->dev_private;
 494 #endif
 495         struct virtqueue *vq;
 496         uint16_t tx_free_thresh;
 497         int ret;
 498
 499         PMD_INIT_FUNC_TRACE();
 500
 501         if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
 502             != ETH_TXQ_FLAGS_NOXSUMS) {
 503                 PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
 504                 return -EINVAL;
 505         }
 506
 507 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 508         /* Use simple rx/tx func if single segment and no offloads */
 509         if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 510              !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 511                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 512                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 513                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 514                 use_simple_rxtx = 1;
 515         }
 516 #endif
 517
 518         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
 519                         nb_desc, socket_id, &vq);
 520         if (ret < 0) {
 521                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 522                 return ret;
 523         }
 524
 525         tx_free_thresh = tx_conf->tx_free_thresh;
 526         if (tx_free_thresh == 0)
 527                 tx_free_thresh =
 528                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 529
 530         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 531                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 532                         "number of TX entries minus 3 (%u)."
 533                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 534                         vq->vq_nentries - 3,
 535                         tx_free_thresh, dev->data->port_id, queue_idx);
 536                 return -EINVAL;
 537         }
 538
 539         vq->vq_free_thresh = tx_free_thresh;
 540
 541         dev->data->tx_queues[queue_idx] = vq;
 542         return 0;
 543 }
 544
 545 void
 546 virtio_dev_tx_queue_release(void *txq)
 547 {
 548         virtio_dev_queue_release(txq);
 549 }
 550
 551 static void
 552 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 553 {
 554         int error;
 555         /*
 556          * Requeue the discarded mbuf. This should always be
 557          * successful since it was just dequeued.
 558          */
 559         error = virtqueue_enqueue_recv_refill(vq, m);
 560         if (unlikely(error)) {
 561                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 562                 rte_pktmbuf_free(m);
 563         }
 564 }
 565
 566 static void
 567 virtio_update_packet_stats(struct virtqueue *vq, struct rte_mbuf *mbuf)
 568 {
 569         uint32_t s = mbuf->pkt_len;
 570         struct ether_addr *ea;
 571
 572         if (s == 64) {
 573                 vq->size_bins[1]++;
 574         } else if (s > 64 && s < 1024) {
 575                 uint32_t bin;
 576
 577                 /* count zeros, and offset into correct bin */
 578                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 579                 vq->size_bins[bin]++;
 580         } else {
 581                 if (s < 64)
 582                         vq->size_bins[0]++;
 583                 else if (s < 1519)
 584                         vq->size_bins[6]++;
 585                 else if (s >= 1519)
 586                         vq->size_bins[7]++;
 587         }
 588
 589         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 590         if (is_multicast_ether_addr(ea)) {
 591                 if (is_broadcast_ether_addr(ea))
 592                         vq->broadcast++;
 593                 else
 594                         vq->multicast++;
 595         }
 596 }
 597
 598 #define VIRTIO_MBUF_BURST_SZ 64
 599 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 600 uint16_t
 601 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 602 {
 603         struct virtqueue *rxvq = rx_queue;
 604         struct virtio_hw *hw;
 605         struct rte_mbuf *rxm, *new_mbuf;
 606         uint16_t nb_used, num, nb_rx;
 607         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 608         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 609         int error;
 610         uint32_t i, nb_enqueued;
 611         uint32_t hdr_size;
 612
 613         nb_used = VIRTQUEUE_NUSED(rxvq);
 614
 615         virtio_rmb();
 616
 617         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 618         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 619         if (likely(num > DESC_PER_CACHELINE))
 620                 num = num - ((rxvq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 621
 622         if (num == 0)
 623                 return 0;
 624
 625         num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, num);
 626         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 627
 628         hw = rxvq->hw;
 629         nb_rx = 0;
 630         nb_enqueued = 0;
 631         hdr_size = hw->vtnet_hdr_size;
 632
 633         for (i = 0; i < num ; i++) {
 634                 rxm = rcv_pkts[i];
 635
 636                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 637
 638                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 639                         PMD_RX_LOG(ERR, "Packet drop");
 640                         nb_enqueued++;
 641                         virtio_discard_rxbuf(rxvq, rxm);
 642                         rxvq->errors++;
 643                         continue;
 644                 }
 645
 646                 rxm->port = rxvq->port_id;
 647                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 648                 rxm->ol_flags = 0;
 649                 rxm->vlan_tci = 0;
 650
 651                 rxm->nb_segs = 1;
 652                 rxm->next = NULL;
 653                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 654                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 655
 656                 if (hw->vlan_strip)
 657                         rte_vlan_strip(rxm);
 658
 659                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 660
 661                 rx_pkts[nb_rx++] = rxm;
 662
 663                 rxvq->bytes += rx_pkts[nb_rx - 1]->pkt_len;
 664                 virtio_update_packet_stats(rxvq, rxm);
 665         }
 666
 667         rxvq->packets += nb_rx;
 668
 669         /* Allocate new mbuf for the used descriptor */
 670         error = ENOSPC;
 671         while (likely(!virtqueue_full(rxvq))) {
 672                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
 673                 if (unlikely(new_mbuf == NULL)) {
 674                         struct rte_eth_dev *dev
 675                                 = &rte_eth_devices[rxvq->port_id];
 676                         dev->data->rx_mbuf_alloc_failed++;
 677                         break;
 678                 }
 679                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 680                 if (unlikely(error)) {
 681                         rte_pktmbuf_free(new_mbuf);
 682                         break;
 683                 }
 684                 nb_enqueued++;
 685         }
 686
 687         if (likely(nb_enqueued)) {
 688                 vq_update_avail_idx(rxvq);
 689
 690                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 691                         virtqueue_notify(rxvq);
 692                         PMD_RX_LOG(DEBUG, "Notified\n");
 693                 }
 694         }
 695
 696         return nb_rx;
 697 }
 698
 699 uint16_t
 700 virtio_recv_mergeable_pkts(void *rx_queue,
 701                         struct rte_mbuf **rx_pkts,
 702                         uint16_t nb_pkts)
 703 {
 704         struct virtqueue *rxvq = rx_queue;
 705         struct virtio_hw *hw;
 706         struct rte_mbuf *rxm, *new_mbuf;
 707         uint16_t nb_used, num, nb_rx;
 708         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 709         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 710         struct rte_mbuf *prev;
 711         int error;
 712         uint32_t i, nb_enqueued;
 713         uint32_t seg_num;
 714         uint16_t extra_idx;
 715         uint32_t seg_res;
 716         uint32_t hdr_size;
 717
 718         nb_used = VIRTQUEUE_NUSED(rxvq);
 719
 720         virtio_rmb();
 721
 722         if (nb_used == 0)
 723                 return 0;
 724
 725         PMD_RX_LOG(DEBUG, "used:%d\n", nb_used);
 726
 727         hw = rxvq->hw;
 728         nb_rx = 0;
 729         i = 0;
 730         nb_enqueued = 0;
 731         seg_num = 0;
 732         extra_idx = 0;
 733         seg_res = 0;
 734         hdr_size = hw->vtnet_hdr_size;
 735
 736         while (i < nb_used) {
 737                 struct virtio_net_hdr_mrg_rxbuf *header;
 738
 739                 if (nb_rx == nb_pkts)
 740                         break;
 741
 742                 num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, 1);
 743                 if (num != 1)
 744                         continue;
 745
 746                 i++;
 747
 748                 PMD_RX_LOG(DEBUG, "dequeue:%d\n", num);
 749                 PMD_RX_LOG(DEBUG, "packet len:%d\n", len[0]);
 750
 751                 rxm = rcv_pkts[0];
 752
 753                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 754                         PMD_RX_LOG(ERR, "Packet drop\n");
 755                         nb_enqueued++;
 756                         virtio_discard_rxbuf(rxvq, rxm);
 757                         rxvq->errors++;
 758                         continue;
 759                 }
 760
 761                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 762                         RTE_PKTMBUF_HEADROOM - hdr_size);
 763                 seg_num = header->num_buffers;
 764
 765                 if (seg_num == 0)
 766                         seg_num = 1;
 767
 768                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 769                 rxm->nb_segs = seg_num;
 770                 rxm->next = NULL;
 771                 rxm->ol_flags = 0;
 772                 rxm->vlan_tci = 0;
 773                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 774                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 775
 776                 rxm->port = rxvq->port_id;
 777                 rx_pkts[nb_rx] = rxm;
 778                 prev = rxm;
 779
 780                 seg_res = seg_num - 1;
 781
 782                 while (seg_res != 0) {
 783                         /*
 784                          * Get extra segments for current uncompleted packet.
 785                          */
 786                         uint16_t  rcv_cnt =
 787                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 788                         if (likely(VIRTQUEUE_NUSED(rxvq) >= rcv_cnt)) {
 789                                 uint32_t rx_num =
 790                                         virtqueue_dequeue_burst_rx(rxvq,
 791                                         rcv_pkts, len, rcv_cnt);
 792                                 i += rx_num;
 793                                 rcv_cnt = rx_num;
 794                         } else {
 795                                 PMD_RX_LOG(ERR,
 796                                         "No enough segments for packet.\n");
 797                                 nb_enqueued++;
 798                                 virtio_discard_rxbuf(rxvq, rxm);
 799                                 rxvq->errors++;
 800                                 break;
 801                         }
 802
 803                         extra_idx = 0;
 804
 805                         while (extra_idx < rcv_cnt) {
 806                                 rxm = rcv_pkts[extra_idx];
 807
 808                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 809                                 rxm->next = NULL;
 810                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 811                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 812
 813                                 if (prev)
 814                                         prev->next = rxm;
 815
 816                                 prev = rxm;
 817                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 818                                 extra_idx++;
 819                         };
 820                         seg_res -= rcv_cnt;
 821                 }
 822
 823                 if (hw->vlan_strip)
 824                         rte_vlan_strip(rx_pkts[nb_rx]);
 825
 826                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 827                         rx_pkts[nb_rx]->data_len);
 828
 829                 rxvq->bytes += rx_pkts[nb_rx]->pkt_len;
 830                 virtio_update_packet_stats(rxvq, rx_pkts[nb_rx]);
 831                 nb_rx++;
 832         }
 833
 834         rxvq->packets += nb_rx;
 835
 836         /* Allocate new mbuf for the used descriptor */
 837         error = ENOSPC;
 838         while (likely(!virtqueue_full(rxvq))) {
 839                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
 840                 if (unlikely(new_mbuf == NULL)) {
 841                         struct rte_eth_dev *dev
 842                                 = &rte_eth_devices[rxvq->port_id];
 843                         dev->data->rx_mbuf_alloc_failed++;
 844                         break;
 845                 }
 846                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 847                 if (unlikely(error)) {
 848                         rte_pktmbuf_free(new_mbuf);
 849                         break;
 850                 }
 851                 nb_enqueued++;
 852         }
 853
 854         if (likely(nb_enqueued)) {
 855                 vq_update_avail_idx(rxvq);
 856
 857                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 858                         virtqueue_notify(rxvq);
 859                         PMD_RX_LOG(DEBUG, "Notified");
 860                 }
 861         }
 862
 863         return nb_rx;
 864 }
 865
 866 uint16_t
 867 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 868 {
 869         struct virtqueue *txvq = tx_queue;
 870         uint16_t nb_used, nb_tx;
 871         int error;
 872
 873         if (unlikely(nb_pkts < 1))
 874                 return nb_pkts;
 875
 876         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
 877         nb_used = VIRTQUEUE_NUSED(txvq);
 878
 879         virtio_rmb();
 880         if (likely(nb_used > txvq->vq_nentries - txvq->vq_free_thresh))
 881                 virtio_xmit_cleanup(txvq, nb_used);
 882
 883         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 884                 struct rte_mbuf *txm = tx_pkts[nb_tx];
 885                 int use_indirect, slots, need;
 886
 887                 use_indirect = vtpci_with_feature(txvq->hw,
 888                                                   VIRTIO_RING_F_INDIRECT_DESC)
 889                         && (txm->nb_segs < VIRTIO_MAX_TX_INDIRECT);
 890
 891                 /* How many main ring entries are needed to this Tx? */
 892                 slots = use_indirect ? 1 : 1 + txm->nb_segs;
 893                 need = slots - txvq->vq_free_cnt;
 894
 895                 /* Positive value indicates it need free vring descriptors */
 896                 if (unlikely(need > 0)) {
 897                         nb_used = VIRTQUEUE_NUSED(txvq);
 898                         virtio_rmb();
 899                         need = RTE_MIN(need, (int)nb_used);
 900
 901                         virtio_xmit_cleanup(txvq, need);
 902                         need = slots - txvq->vq_free_cnt;
 903                         if (unlikely(need > 0)) {
 904                                 PMD_TX_LOG(ERR,
 905                                            "No free tx descriptors to transmit");
 906                                 break;
 907                         }
 908                 }
 909
 910                 /* Do VLAN tag insertion */
 911                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
 912                         error = rte_vlan_insert(&txm);
 913                         if (unlikely(error)) {
 914                                 rte_pktmbuf_free(txm);
 915                                 continue;
 916                         }
 917                 }
 918
 919                 /* Enqueue Packet buffers */
 920                 error = virtqueue_enqueue_xmit(txvq, txm, use_indirect);
 921                 if (unlikely(error)) {
 922                         if (error == ENOSPC)
 923                                 PMD_TX_LOG(ERR, "virtqueue_enqueue Free count = 0");
 924                         else if (error == EMSGSIZE)
 925                                 PMD_TX_LOG(ERR, "virtqueue_enqueue Free count < 1");
 926                         else
 927                                 PMD_TX_LOG(ERR, "virtqueue_enqueue error: %d", error);
 928                         break;
 929                 }
 930
 931                 txvq->bytes += txm->pkt_len;
 932                 virtio_update_packet_stats(txvq, txm);
 933         }
 934
 935         txvq->packets += nb_tx;
 936
 937         if (likely(nb_tx)) {
 938                 vq_update_avail_idx(txvq);
 939
 940                 if (unlikely(virtqueue_kick_prepare(txvq))) {
 941                         virtqueue_notify(txvq);
 942                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
 943                 }
 944         }
 945
 946         return nb_tx;
 947 }