drivers/net/virtio/virtio_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38 #include <errno.h>
  39
  40 #include <rte_cycles.h>
  41 #include <rte_memory.h>
  42 #include <rte_memzone.h>
  43 #include <rte_branch_prediction.h>
  44 #include <rte_mempool.h>
  45 #include <rte_malloc.h>
  46 #include <rte_mbuf.h>
  47 #include <rte_ether.h>
  48 #include <rte_ethdev.h>
  49 #include <rte_prefetch.h>
  50 #include <rte_string_fns.h>
  51 #include <rte_errno.h>
  52 #include <rte_byteorder.h>
  53 #include <rte_cpuflags.h>
  54 #include <rte_net.h>
  55 #include <rte_ip.h>
  56 #include <rte_udp.h>
  57 #include <rte_tcp.h>
  58
  59 #include "virtio_logs.h"
  60 #include "virtio_ethdev.h"
  61 #include "virtio_pci.h"
  62 #include "virtqueue.h"
  63 #include "virtio_rxtx.h"
  64
  65 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  66 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  67 #else
  68 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  69 #endif
  70
  71
  72 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
  73         ETH_TXQ_FLAGS_NOOFFLOADS)
  74
  75 static void
  76 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  77 {
  78         struct vring_desc *dp, *dp_tail;
  79         struct vq_desc_extra *dxp;
  80         uint16_t desc_idx_last = desc_idx;
  81
  82         dp  = &vq->vq_ring.desc[desc_idx];
  83         dxp = &vq->vq_descx[desc_idx];
  84         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  85         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  86                 while (dp->flags & VRING_DESC_F_NEXT) {
  87                         desc_idx_last = dp->next;
  88                         dp = &vq->vq_ring.desc[dp->next];
  89                 }
  90         }
  91         dxp->ndescs = 0;
  92
  93         /*
  94          * We must append the existing free chain, if any, to the end of
  95          * newly freed chain. If the virtqueue was completely used, then
  96          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  97          */
  98         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  99                 vq->vq_desc_head_idx = desc_idx;
 100         } else {
 101                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
 102                 dp_tail->next = desc_idx;
 103         }
 104
 105         vq->vq_desc_tail_idx = desc_idx_last;
 106         dp->next = VQ_RING_DESC_CHAIN_END;
 107 }
 108
 109 static uint16_t
 110 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 111                            uint32_t *len, uint16_t num)
 112 {
 113         struct vring_used_elem *uep;
 114         struct rte_mbuf *cookie;
 115         uint16_t used_idx, desc_idx;
 116         uint16_t i;
 117
 118         /*  Caller does the check */
 119         for (i = 0; i < num ; i++) {
 120                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 121                 uep = &vq->vq_ring.used->ring[used_idx];
 122                 desc_idx = (uint16_t) uep->id;
 123                 len[i] = uep->len;
 124                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 125
 126                 if (unlikely(cookie == NULL)) {
 127                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 128                                 vq->vq_used_cons_idx);
 129                         break;
 130                 }
 131
 132                 rte_prefetch0(cookie);
 133                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 134                 rx_pkts[i]  = cookie;
 135                 vq->vq_used_cons_idx++;
 136                 vq_ring_free_chain(vq, desc_idx);
 137                 vq->vq_descx[desc_idx].cookie = NULL;
 138         }
 139
 140         return i;
 141 }
 142
 143 #ifndef DEFAULT_TX_FREE_THRESH
 144 #define DEFAULT_TX_FREE_THRESH 32
 145 #endif
 146
 147 /* Cleanup from completed transmits. */
 148 static void
 149 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 150 {
 151         uint16_t i, used_idx, desc_idx;
 152         for (i = 0; i < num; i++) {
 153                 struct vring_used_elem *uep;
 154                 struct vq_desc_extra *dxp;
 155
 156                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 157                 uep = &vq->vq_ring.used->ring[used_idx];
 158
 159                 desc_idx = (uint16_t) uep->id;
 160                 dxp = &vq->vq_descx[desc_idx];
 161                 vq->vq_used_cons_idx++;
 162                 vq_ring_free_chain(vq, desc_idx);
 163
 164                 if (dxp->cookie != NULL) {
 165                         rte_pktmbuf_free(dxp->cookie);
 166                         dxp->cookie = NULL;
 167                 }
 168         }
 169 }
 170
 171
 172 static inline int
 173 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 174 {
 175         struct vq_desc_extra *dxp;
 176         struct virtio_hw *hw = vq->hw;
 177         struct vring_desc *start_dp;
 178         uint16_t needed = 1;
 179         uint16_t head_idx, idx;
 180
 181         if (unlikely(vq->vq_free_cnt == 0))
 182                 return -ENOSPC;
 183         if (unlikely(vq->vq_free_cnt < needed))
 184                 return -EMSGSIZE;
 185
 186         head_idx = vq->vq_desc_head_idx;
 187         if (unlikely(head_idx >= vq->vq_nentries))
 188                 return -EFAULT;
 189
 190         idx = head_idx;
 191         dxp = &vq->vq_descx[idx];
 192         dxp->cookie = (void *)cookie;
 193         dxp->ndescs = needed;
 194
 195         start_dp = vq->vq_ring.desc;
 196         start_dp[idx].addr =
 197                 VIRTIO_MBUF_ADDR(cookie, vq) +
 198                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 199         start_dp[idx].len =
 200                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 201         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 202         idx = start_dp[idx].next;
 203         vq->vq_desc_head_idx = idx;
 204         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 205                 vq->vq_desc_tail_idx = idx;
 206         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 207         vq_update_avail_ring(vq, head_idx);
 208
 209         return 0;
 210 }
 211
 212 static inline int
 213 tx_offload_enabled(struct virtio_hw *hw)
 214 {
 215         return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM);
 216 }
 217
 218 static inline void
 219 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 220                        uint16_t needed, int use_indirect, int can_push)
 221 {
 222         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 223         struct vq_desc_extra *dxp;
 224         struct virtqueue *vq = txvq->vq;
 225         struct vring_desc *start_dp;
 226         uint16_t seg_num = cookie->nb_segs;
 227         uint16_t head_idx, idx;
 228         uint16_t head_size = vq->hw->vtnet_hdr_size;
 229         struct virtio_net_hdr *hdr;
 230         int offload;
 231
 232         offload = tx_offload_enabled(vq->hw);
 233         head_idx = vq->vq_desc_head_idx;
 234         idx = head_idx;
 235         dxp = &vq->vq_descx[idx];
 236         dxp->cookie = (void *)cookie;
 237         dxp->ndescs = needed;
 238
 239         start_dp = vq->vq_ring.desc;
 240
 241         if (can_push) {
 242                 /* prepend cannot fail, checked by caller */
 243                 hdr = (struct virtio_net_hdr *)
 244                         rte_pktmbuf_prepend(cookie, head_size);
 245                 /* if offload disabled, it is not zeroed below, do it now */
 246                 if (offload == 0)
 247                         memset(hdr, 0, head_size);
 248         } else if (use_indirect) {
 249                 /* setup tx ring slot to point to indirect
 250                  * descriptor list stored in reserved region.
 251                  *
 252                  * the first slot in indirect ring is already preset
 253                  * to point to the header in reserved region
 254                  */
 255                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 256                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 257                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 258                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 259                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 260
 261                 /* loop below will fill in rest of the indirect elements */
 262                 start_dp = txr[idx].tx_indir;
 263                 idx = 1;
 264         } else {
 265                 /* setup first tx ring slot to point to header
 266                  * stored in reserved region.
 267                  */
 268                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 269                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 270                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 271                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 272                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 273
 274                 idx = start_dp[idx].next;
 275         }
 276
 277         if (offload) {
 278                 /* Checksum Offload */
 279                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 280                 case PKT_TX_UDP_CKSUM:
 281                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 282                         hdr->csum_offset = offsetof(struct udp_hdr,
 283                                 dgram_cksum);
 284                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 285                         break;
 286
 287                 case PKT_TX_TCP_CKSUM:
 288                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 289                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 290                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 291                         break;
 292
 293                 default:
 294                         hdr->csum_start = 0;
 295                         hdr->csum_offset = 0;
 296                         hdr->flags = 0;
 297                         break;
 298                 }
 299
 300                 hdr->gso_type = 0;
 301                 hdr->gso_size = 0;
 302                 hdr->hdr_len = 0;
 303         }
 304
 305         do {
 306                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 307                 start_dp[idx].len   = cookie->data_len;
 308                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 309                 idx = start_dp[idx].next;
 310         } while ((cookie = cookie->next) != NULL);
 311
 312         if (use_indirect)
 313                 idx = vq->vq_ring.desc[head_idx].next;
 314
 315         vq->vq_desc_head_idx = idx;
 316         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 317                 vq->vq_desc_tail_idx = idx;
 318         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 319         vq_update_avail_ring(vq, head_idx);
 320 }
 321
 322 static void
 323 virtio_dev_vring_start(struct virtqueue *vq)
 324 {
 325         int size = vq->vq_nentries;
 326         struct vring *vr = &vq->vq_ring;
 327         uint8_t *ring_mem = vq->vq_ring_virt_mem;
 328
 329         PMD_INIT_FUNC_TRACE();
 330
 331         /*
 332          * Reinitialise since virtio port might have been stopped and restarted
 333          */
 334         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
 335         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
 336         vq->vq_used_cons_idx = 0;
 337         vq->vq_desc_head_idx = 0;
 338         vq->vq_avail_idx = 0;
 339         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
 340         vq->vq_free_cnt = vq->vq_nentries;
 341         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 342
 343         vring_desc_init(vr->desc, size);
 344
 345         /*
 346          * Disable device(host) interrupting guest
 347          */
 348         virtqueue_disable_intr(vq);
 349 }
 350
 351 void
 352 virtio_dev_cq_start(struct rte_eth_dev *dev)
 353 {
 354         struct virtio_hw *hw = dev->data->dev_private;
 355
 356         if (hw->cvq && hw->cvq->vq) {
 357                 virtio_dev_vring_start(hw->cvq->vq);
 358                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 359         }
 360 }
 361
 362 void
 363 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 364 {
 365         /*
 366          * Start receive and transmit vrings
 367          * -    Setup vring structure for all queues
 368          * -    Initialize descriptor for the rx vring
 369          * -    Allocate blank mbufs for the each rx descriptor
 370          *
 371          */
 372         uint16_t i;
 373         uint16_t desc_idx;
 374         struct virtio_hw *hw = dev->data->dev_private;
 375
 376         PMD_INIT_FUNC_TRACE();
 377
 378         /* Start rx vring. */
 379         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 380                 struct virtnet_rx *rxvq = dev->data->rx_queues[i];
 381                 struct virtqueue *vq = rxvq->vq;
 382                 int error, nbufs;
 383                 struct rte_mbuf *m;
 384
 385                 virtio_dev_vring_start(vq);
 386                 if (rxvq->mpool == NULL) {
 387                         rte_exit(EXIT_FAILURE,
 388                                 "Cannot allocate mbufs for rx virtqueue");
 389                 }
 390
 391                 /* Allocate blank mbufs for the each rx descriptor */
 392                 nbufs = 0;
 393                 error = ENOSPC;
 394
 395                 if (hw->use_simple_rxtx) {
 396                         for (desc_idx = 0; desc_idx < vq->vq_nentries;
 397                              desc_idx++) {
 398                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 399                                 vq->vq_ring.desc[desc_idx].flags =
 400                                         VRING_DESC_F_WRITE;
 401                         }
 402                 }
 403
 404                 memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 405                 for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 406                      desc_idx++) {
 407                         vq->sw_ring[vq->vq_nentries + desc_idx] =
 408                                 &rxvq->fake_mbuf;
 409                 }
 410
 411                 while (!virtqueue_full(vq)) {
 412                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 413                         if (m == NULL)
 414                                 break;
 415
 416                         /******************************************
 417                         *         Enqueue allocated buffers        *
 418                         *******************************************/
 419                         if (hw->use_simple_rxtx)
 420                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 421                         else
 422                                 error = virtqueue_enqueue_recv_refill(vq, m);
 423
 424                         if (error) {
 425                                 rte_pktmbuf_free(m);
 426                                 break;
 427                         }
 428                         nbufs++;
 429                 }
 430
 431                 vq_update_avail_idx(vq);
 432
 433                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 434
 435                 VIRTQUEUE_DUMP(vq);
 436         }
 437
 438         /* Start tx vring. */
 439         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 440                 struct virtnet_tx *txvq = dev->data->tx_queues[i];
 441                 struct virtqueue *vq = txvq->vq;
 442
 443                 virtio_dev_vring_start(vq);
 444                 if (hw->use_simple_rxtx) {
 445                         uint16_t mid_idx  = vq->vq_nentries >> 1;
 446
 447                         for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 448                                 vq->vq_ring.avail->ring[desc_idx] =
 449                                         desc_idx + mid_idx;
 450                                 vq->vq_ring.desc[desc_idx + mid_idx].next =
 451                                         desc_idx;
 452                                 vq->vq_ring.desc[desc_idx + mid_idx].addr =
 453                                         txvq->virtio_net_hdr_mem +
 454                                         offsetof(struct virtio_tx_region, tx_hdr);
 455                                 vq->vq_ring.desc[desc_idx + mid_idx].len =
 456                                         vq->hw->vtnet_hdr_size;
 457                                 vq->vq_ring.desc[desc_idx + mid_idx].flags =
 458                                         VRING_DESC_F_NEXT;
 459                                 vq->vq_ring.desc[desc_idx].flags = 0;
 460                         }
 461                         for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 462                              desc_idx++)
 463                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 464                 }
 465
 466                 VIRTQUEUE_DUMP(vq);
 467         }
 468 }
 469
 470 int
 471 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 472                         uint16_t queue_idx,
 473                         uint16_t nb_desc,
 474                         unsigned int socket_id,
 475                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 476                         struct rte_mempool *mp)
 477 {
 478         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 479         struct virtnet_rx *rxvq;
 480         int ret;
 481
 482         PMD_INIT_FUNC_TRACE();
 483         ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
 484                         nb_desc, socket_id, (void **)&rxvq);
 485         if (ret < 0) {
 486                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 487                 return ret;
 488         }
 489
 490         /* Create mempool for rx mbuf allocation */
 491         rxvq->mpool = mp;
 492
 493         dev->data->rx_queues[queue_idx] = rxvq;
 494
 495         virtio_rxq_vec_setup(rxvq);
 496
 497         return 0;
 498 }
 499
 500 void
 501 virtio_dev_rx_queue_release(void *rxq)
 502 {
 503         struct virtnet_rx *rxvq = rxq;
 504         struct virtqueue *vq;
 505         const struct rte_memzone *mz;
 506
 507         if (rxvq == NULL)
 508                 return;
 509
 510         /*
 511          * rxvq is freed when vq is freed, and as mz should be freed after the
 512          * del_queue, so we reserve the mz pointer first.
 513          */
 514         vq = rxvq->vq;
 515         mz = rxvq->mz;
 516
 517         virtio_dev_queue_release(vq);
 518         rte_memzone_free(mz);
 519 }
 520
 521 static void
 522 virtio_update_rxtx_handler(struct rte_eth_dev *dev,
 523                            const struct rte_eth_txconf *tx_conf)
 524 {
 525         uint8_t use_simple_rxtx = 0;
 526         struct virtio_hw *hw = dev->data->dev_private;
 527
 528 #if defined RTE_ARCH_X86
 529         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE3))
 530                 use_simple_rxtx = 1;
 531 #elif defined RTE_ARCH_ARM64 || defined CONFIG_RTE_ARCH_ARM
 532         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
 533                 use_simple_rxtx = 1;
 534 #endif
 535         /* Use simple rx/tx func if single segment and no offloads */
 536         if (use_simple_rxtx &&
 537             (tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 538             !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 539                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 540                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 541                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 542                 hw->use_simple_rxtx = use_simple_rxtx;
 543         }
 544 }
 545
 546 /*
 547  * struct rte_eth_dev *dev: Used to update dev
 548  * uint16_t nb_desc: Defaults to values read from config space
 549  * unsigned int socket_id: Used to allocate memzone
 550  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 551  * uint16_t queue_idx: Just used as an index in dev txq list
 552  */
 553 int
 554 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 555                         uint16_t queue_idx,
 556                         uint16_t nb_desc,
 557                         unsigned int socket_id,
 558                         const struct rte_eth_txconf *tx_conf)
 559 {
 560         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 561         struct virtnet_tx *txvq;
 562         struct virtqueue *vq;
 563         uint16_t tx_free_thresh;
 564         int ret;
 565
 566         PMD_INIT_FUNC_TRACE();
 567
 568
 569         virtio_update_rxtx_handler(dev, tx_conf);
 570
 571         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
 572                         nb_desc, socket_id, (void **)&txvq);
 573         if (ret < 0) {
 574                 PMD_INIT_LOG(ERR, "tvq initialization failed");
 575                 return ret;
 576         }
 577         vq = txvq->vq;
 578
 579         tx_free_thresh = tx_conf->tx_free_thresh;
 580         if (tx_free_thresh == 0)
 581                 tx_free_thresh =
 582                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 583
 584         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 585                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 586                         "number of TX entries minus 3 (%u)."
 587                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 588                         vq->vq_nentries - 3,
 589                         tx_free_thresh, dev->data->port_id, queue_idx);
 590                 return -EINVAL;
 591         }
 592
 593         vq->vq_free_thresh = tx_free_thresh;
 594
 595         dev->data->tx_queues[queue_idx] = txvq;
 596         return 0;
 597 }
 598
 599 void
 600 virtio_dev_tx_queue_release(void *txq)
 601 {
 602         struct virtnet_tx *txvq = txq;
 603         struct virtqueue *vq;
 604         const struct rte_memzone *mz;
 605         const struct rte_memzone *hdr_mz;
 606
 607         if (txvq == NULL)
 608                 return;
 609
 610         /*
 611          * txvq is freed when vq is freed, and as mz should be freed after the
 612          * del_queue, so we reserve the mz pointer first.
 613          */
 614         vq = txvq->vq;
 615         mz = txvq->mz;
 616         hdr_mz = txvq->virtio_net_hdr_mz;
 617
 618         virtio_dev_queue_release(vq);
 619         rte_memzone_free(mz);
 620         rte_memzone_free(hdr_mz);
 621 }
 622
 623 static void
 624 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 625 {
 626         int error;
 627         /*
 628          * Requeue the discarded mbuf. This should always be
 629          * successful since it was just dequeued.
 630          */
 631         error = virtqueue_enqueue_recv_refill(vq, m);
 632         if (unlikely(error)) {
 633                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 634                 rte_pktmbuf_free(m);
 635         }
 636 }
 637
 638 static void
 639 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 640 {
 641         uint32_t s = mbuf->pkt_len;
 642         struct ether_addr *ea;
 643
 644         if (s == 64) {
 645                 stats->size_bins[1]++;
 646         } else if (s > 64 && s < 1024) {
 647                 uint32_t bin;
 648
 649                 /* count zeros, and offset into correct bin */
 650                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 651                 stats->size_bins[bin]++;
 652         } else {
 653                 if (s < 64)
 654                         stats->size_bins[0]++;
 655                 else if (s < 1519)
 656                         stats->size_bins[6]++;
 657                 else if (s >= 1519)
 658                         stats->size_bins[7]++;
 659         }
 660
 661         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 662         if (is_multicast_ether_addr(ea)) {
 663                 if (is_broadcast_ether_addr(ea))
 664                         stats->broadcast++;
 665                 else
 666                         stats->multicast++;
 667         }
 668 }
 669
 670 /* Optionally fill offload information in structure */
 671 static int
 672 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 673 {
 674         struct rte_net_hdr_lens hdr_lens;
 675         uint32_t hdrlen, ptype;
 676         int l4_supported = 0;
 677
 678         /* nothing to do */
 679         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 680                 return 0;
 681
 682         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 683
 684         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 685         m->packet_type = ptype;
 686         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 687             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 688             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 689                 l4_supported = 1;
 690
 691         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 692                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 693                 if (hdr->csum_start <= hdrlen && l4_supported) {
 694                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 695                 } else {
 696                         /* Unknown proto or tunnel, do sw cksum. We can assume
 697                          * the cksum field is in the first segment since the
 698                          * buffers we provided to the host are large enough.
 699                          * In case of SCTP, this will be wrong since it's a CRC
 700                          * but there's nothing we can do.
 701                          */
 702                         uint16_t csum, off;
 703
 704                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 705                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 706                                 &csum);
 707                         if (likely(csum != 0xffff))
 708                                 csum = ~csum;
 709                         off = hdr->csum_offset + hdr->csum_start;
 710                         if (rte_pktmbuf_data_len(m) >= off + 1)
 711                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 712                                         off) = csum;
 713                 }
 714         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 715                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 716         }
 717
 718         return 0;
 719 }
 720
 721 static inline int
 722 rx_offload_enabled(struct virtio_hw *hw)
 723 {
 724         return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM);
 725 }
 726
 727 #define VIRTIO_MBUF_BURST_SZ 64
 728 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 729 uint16_t
 730 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 731 {
 732         struct virtnet_rx *rxvq = rx_queue;
 733         struct virtqueue *vq = rxvq->vq;
 734         struct virtio_hw *hw;
 735         struct rte_mbuf *rxm, *new_mbuf;
 736         uint16_t nb_used, num, nb_rx;
 737         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 738         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 739         int error;
 740         uint32_t i, nb_enqueued;
 741         uint32_t hdr_size;
 742         int offload;
 743         struct virtio_net_hdr *hdr;
 744
 745         nb_used = VIRTQUEUE_NUSED(vq);
 746
 747         virtio_rmb();
 748
 749         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 750         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 751         if (likely(num > DESC_PER_CACHELINE))
 752                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 753
 754         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 755         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 756
 757         hw = vq->hw;
 758         nb_rx = 0;
 759         nb_enqueued = 0;
 760         hdr_size = hw->vtnet_hdr_size;
 761         offload = rx_offload_enabled(hw);
 762
 763         for (i = 0; i < num ; i++) {
 764                 rxm = rcv_pkts[i];
 765
 766                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 767
 768                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 769                         PMD_RX_LOG(ERR, "Packet drop");
 770                         nb_enqueued++;
 771                         virtio_discard_rxbuf(vq, rxm);
 772                         rxvq->stats.errors++;
 773                         continue;
 774                 }
 775
 776                 rxm->port = rxvq->port_id;
 777                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 778                 rxm->ol_flags = 0;
 779                 rxm->vlan_tci = 0;
 780
 781                 rxm->nb_segs = 1;
 782                 rxm->next = NULL;
 783                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 784                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 785
 786                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 787                         RTE_PKTMBUF_HEADROOM - hdr_size);
 788
 789                 if (hw->vlan_strip)
 790                         rte_vlan_strip(rxm);
 791
 792                 if (offload && virtio_rx_offload(rxm, hdr) < 0) {
 793                         virtio_discard_rxbuf(vq, rxm);
 794                         rxvq->stats.errors++;
 795                         continue;
 796                 }
 797
 798                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 799
 800                 rx_pkts[nb_rx++] = rxm;
 801
 802                 rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
 803                 virtio_update_packet_stats(&rxvq->stats, rxm);
 804         }
 805
 806         rxvq->stats.packets += nb_rx;
 807
 808         /* Allocate new mbuf for the used descriptor */
 809         error = ENOSPC;
 810         while (likely(!virtqueue_full(vq))) {
 811                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 812                 if (unlikely(new_mbuf == NULL)) {
 813                         struct rte_eth_dev *dev
 814                                 = &rte_eth_devices[rxvq->port_id];
 815                         dev->data->rx_mbuf_alloc_failed++;
 816                         break;
 817                 }
 818                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 819                 if (unlikely(error)) {
 820                         rte_pktmbuf_free(new_mbuf);
 821                         break;
 822                 }
 823                 nb_enqueued++;
 824         }
 825
 826         if (likely(nb_enqueued)) {
 827                 vq_update_avail_idx(vq);
 828
 829                 if (unlikely(virtqueue_kick_prepare(vq))) {
 830                         virtqueue_notify(vq);
 831                         PMD_RX_LOG(DEBUG, "Notified");
 832                 }
 833         }
 834
 835         return nb_rx;
 836 }
 837
 838 uint16_t
 839 virtio_recv_mergeable_pkts(void *rx_queue,
 840                         struct rte_mbuf **rx_pkts,
 841                         uint16_t nb_pkts)
 842 {
 843         struct virtnet_rx *rxvq = rx_queue;
 844         struct virtqueue *vq = rxvq->vq;
 845         struct virtio_hw *hw;
 846         struct rte_mbuf *rxm, *new_mbuf;
 847         uint16_t nb_used, num, nb_rx;
 848         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 849         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 850         struct rte_mbuf *prev;
 851         int error;
 852         uint32_t i, nb_enqueued;
 853         uint32_t seg_num;
 854         uint16_t extra_idx;
 855         uint32_t seg_res;
 856         uint32_t hdr_size;
 857         int offload;
 858
 859         nb_used = VIRTQUEUE_NUSED(vq);
 860
 861         virtio_rmb();
 862
 863         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 864
 865         hw = vq->hw;
 866         nb_rx = 0;
 867         i = 0;
 868         nb_enqueued = 0;
 869         seg_num = 0;
 870         extra_idx = 0;
 871         seg_res = 0;
 872         hdr_size = hw->vtnet_hdr_size;
 873         offload = rx_offload_enabled(hw);
 874
 875         while (i < nb_used) {
 876                 struct virtio_net_hdr_mrg_rxbuf *header;
 877
 878                 if (nb_rx == nb_pkts)
 879                         break;
 880
 881                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
 882                 if (num != 1)
 883                         continue;
 884
 885                 i++;
 886
 887                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 888                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 889
 890                 rxm = rcv_pkts[0];
 891
 892                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 893                         PMD_RX_LOG(ERR, "Packet drop");
 894                         nb_enqueued++;
 895                         virtio_discard_rxbuf(vq, rxm);
 896                         rxvq->stats.errors++;
 897                         continue;
 898                 }
 899
 900                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 901                         RTE_PKTMBUF_HEADROOM - hdr_size);
 902                 seg_num = header->num_buffers;
 903
 904                 if (seg_num == 0)
 905                         seg_num = 1;
 906
 907                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 908                 rxm->nb_segs = seg_num;
 909                 rxm->next = NULL;
 910                 rxm->ol_flags = 0;
 911                 rxm->vlan_tci = 0;
 912                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 913                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 914
 915                 rxm->port = rxvq->port_id;
 916                 rx_pkts[nb_rx] = rxm;
 917                 prev = rxm;
 918
 919                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
 920                         virtio_discard_rxbuf(vq, rxm);
 921                         rxvq->stats.errors++;
 922                         continue;
 923                 }
 924
 925                 seg_res = seg_num - 1;
 926
 927                 while (seg_res != 0) {
 928                         /*
 929                          * Get extra segments for current uncompleted packet.
 930                          */
 931                         uint16_t  rcv_cnt =
 932                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 933                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
 934                                 uint32_t rx_num =
 935                                         virtqueue_dequeue_burst_rx(vq,
 936                                         rcv_pkts, len, rcv_cnt);
 937                                 i += rx_num;
 938                                 rcv_cnt = rx_num;
 939                         } else {
 940                                 PMD_RX_LOG(ERR,
 941                                            "No enough segments for packet.");
 942                                 nb_enqueued++;
 943                                 virtio_discard_rxbuf(vq, rxm);
 944                                 rxvq->stats.errors++;
 945                                 break;
 946                         }
 947
 948                         extra_idx = 0;
 949
 950                         while (extra_idx < rcv_cnt) {
 951                                 rxm = rcv_pkts[extra_idx];
 952
 953                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 954                                 rxm->next = NULL;
 955                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 956                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 957
 958                                 if (prev)
 959                                         prev->next = rxm;
 960
 961                                 prev = rxm;
 962                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 963                                 extra_idx++;
 964                         };
 965                         seg_res -= rcv_cnt;
 966                 }
 967
 968                 if (hw->vlan_strip)
 969                         rte_vlan_strip(rx_pkts[nb_rx]);
 970
 971                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 972                         rx_pkts[nb_rx]->data_len);
 973
 974                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
 975                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
 976                 nb_rx++;
 977         }
 978
 979         rxvq->stats.packets += nb_rx;
 980
 981         /* Allocate new mbuf for the used descriptor */
 982         error = ENOSPC;
 983         while (likely(!virtqueue_full(vq))) {
 984                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 985                 if (unlikely(new_mbuf == NULL)) {
 986                         struct rte_eth_dev *dev
 987                                 = &rte_eth_devices[rxvq->port_id];
 988                         dev->data->rx_mbuf_alloc_failed++;
 989                         break;
 990                 }
 991                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 992                 if (unlikely(error)) {
 993                         rte_pktmbuf_free(new_mbuf);
 994                         break;
 995                 }
 996                 nb_enqueued++;
 997         }
 998
 999         if (likely(nb_enqueued)) {
1000                 vq_update_avail_idx(vq);
1001
1002                 if (unlikely(virtqueue_kick_prepare(vq))) {
1003                         virtqueue_notify(vq);
1004                         PMD_RX_LOG(DEBUG, "Notified");
1005                 }
1006         }
1007
1008         return nb_rx;
1009 }
1010
1011 uint16_t
1012 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1013 {
1014         struct virtnet_tx *txvq = tx_queue;
1015         struct virtqueue *vq = txvq->vq;
1016         struct virtio_hw *hw = vq->hw;
1017         uint16_t hdr_size = hw->vtnet_hdr_size;
1018         uint16_t nb_used, nb_tx;
1019         int error;
1020
1021         if (unlikely(nb_pkts < 1))
1022                 return nb_pkts;
1023
1024         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1025         nb_used = VIRTQUEUE_NUSED(vq);
1026
1027         virtio_rmb();
1028         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1029                 virtio_xmit_cleanup(vq, nb_used);
1030
1031         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1032                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1033                 int can_push = 0, use_indirect = 0, slots, need;
1034
1035                 /* Do VLAN tag insertion */
1036                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1037                         error = rte_vlan_insert(&txm);
1038                         if (unlikely(error)) {
1039                                 rte_pktmbuf_free(txm);
1040                                 continue;
1041                         }
1042                 }
1043
1044                 /* optimize ring usage */
1045                 if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
1046                     rte_mbuf_refcnt_read(txm) == 1 &&
1047                     RTE_MBUF_DIRECT(txm) &&
1048                     txm->nb_segs == 1 &&
1049                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1050                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1051                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1052                         can_push = 1;
1053                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1054                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1055                         use_indirect = 1;
1056
1057                 /* How many main ring entries are needed to this Tx?
1058                  * any_layout => number of segments
1059                  * indirect   => 1
1060                  * default    => number of segments + 1
1061                  */
1062                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1063                 need = slots - vq->vq_free_cnt;
1064
1065                 /* Positive value indicates it need free vring descriptors */
1066                 if (unlikely(need > 0)) {
1067                         nb_used = VIRTQUEUE_NUSED(vq);
1068                         virtio_rmb();
1069                         need = RTE_MIN(need, (int)nb_used);
1070
1071                         virtio_xmit_cleanup(vq, need);
1072                         need = slots - vq->vq_free_cnt;
1073                         if (unlikely(need > 0)) {
1074                                 PMD_TX_LOG(ERR,
1075                                            "No free tx descriptors to transmit");
1076                                 break;
1077                         }
1078                 }
1079
1080                 /* Enqueue Packet buffers */
1081                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
1082
1083                 txvq->stats.bytes += txm->pkt_len;
1084                 virtio_update_packet_stats(&txvq->stats, txm);
1085         }
1086
1087         txvq->stats.packets += nb_tx;
1088
1089         if (likely(nb_tx)) {
1090                 vq_update_avail_idx(vq);
1091
1092                 if (unlikely(virtqueue_kick_prepare(vq))) {
1093                         virtqueue_notify(vq);
1094                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1095                 }
1096         }
1097
1098         return nb_tx;
1099 }