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
  55 #include "virtio_logs.h"
  56 #include "virtio_ethdev.h"
  57 #include "virtio_pci.h"
  58 #include "virtqueue.h"
  59 #include "virtio_rxtx.h"
  60
  61 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  62 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  63 #else
  64 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  65 #endif
  66
  67
  68 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
  69         ETH_TXQ_FLAGS_NOOFFLOADS)
  70
  71 static void
  72 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  73 {
  74         struct vring_desc *dp, *dp_tail;
  75         struct vq_desc_extra *dxp;
  76         uint16_t desc_idx_last = desc_idx;
  77
  78         dp  = &vq->vq_ring.desc[desc_idx];
  79         dxp = &vq->vq_descx[desc_idx];
  80         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  81         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  82                 while (dp->flags & VRING_DESC_F_NEXT) {
  83                         desc_idx_last = dp->next;
  84                         dp = &vq->vq_ring.desc[dp->next];
  85                 }
  86         }
  87         dxp->ndescs = 0;
  88
  89         /*
  90          * We must append the existing free chain, if any, to the end of
  91          * newly freed chain. If the virtqueue was completely used, then
  92          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  93          */
  94         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  95                 vq->vq_desc_head_idx = desc_idx;
  96         } else {
  97                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
  98                 dp_tail->next = desc_idx;
  99         }
 100
 101         vq->vq_desc_tail_idx = desc_idx_last;
 102         dp->next = VQ_RING_DESC_CHAIN_END;
 103 }
 104
 105 static uint16_t
 106 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 107                            uint32_t *len, uint16_t num)
 108 {
 109         struct vring_used_elem *uep;
 110         struct rte_mbuf *cookie;
 111         uint16_t used_idx, desc_idx;
 112         uint16_t i;
 113
 114         /*  Caller does the check */
 115         for (i = 0; i < num ; i++) {
 116                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 117                 uep = &vq->vq_ring.used->ring[used_idx];
 118                 desc_idx = (uint16_t) uep->id;
 119                 len[i] = uep->len;
 120                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 121
 122                 if (unlikely(cookie == NULL)) {
 123                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 124                                 vq->vq_used_cons_idx);
 125                         break;
 126                 }
 127
 128                 rte_prefetch0(cookie);
 129                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 130                 rx_pkts[i]  = cookie;
 131                 vq->vq_used_cons_idx++;
 132                 vq_ring_free_chain(vq, desc_idx);
 133                 vq->vq_descx[desc_idx].cookie = NULL;
 134         }
 135
 136         return i;
 137 }
 138
 139 #ifndef DEFAULT_TX_FREE_THRESH
 140 #define DEFAULT_TX_FREE_THRESH 32
 141 #endif
 142
 143 /* Cleanup from completed transmits. */
 144 static void
 145 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 146 {
 147         uint16_t i, used_idx, desc_idx;
 148         for (i = 0; i < num; i++) {
 149                 struct vring_used_elem *uep;
 150                 struct vq_desc_extra *dxp;
 151
 152                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 153                 uep = &vq->vq_ring.used->ring[used_idx];
 154
 155                 desc_idx = (uint16_t) uep->id;
 156                 dxp = &vq->vq_descx[desc_idx];
 157                 vq->vq_used_cons_idx++;
 158                 vq_ring_free_chain(vq, desc_idx);
 159
 160                 if (dxp->cookie != NULL) {
 161                         rte_pktmbuf_free(dxp->cookie);
 162                         dxp->cookie = NULL;
 163                 }
 164         }
 165 }
 166
 167
 168 static inline int
 169 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 170 {
 171         struct vq_desc_extra *dxp;
 172         struct virtio_hw *hw = vq->hw;
 173         struct vring_desc *start_dp;
 174         uint16_t needed = 1;
 175         uint16_t head_idx, idx;
 176
 177         if (unlikely(vq->vq_free_cnt == 0))
 178                 return -ENOSPC;
 179         if (unlikely(vq->vq_free_cnt < needed))
 180                 return -EMSGSIZE;
 181
 182         head_idx = vq->vq_desc_head_idx;
 183         if (unlikely(head_idx >= vq->vq_nentries))
 184                 return -EFAULT;
 185
 186         idx = head_idx;
 187         dxp = &vq->vq_descx[idx];
 188         dxp->cookie = (void *)cookie;
 189         dxp->ndescs = needed;
 190
 191         start_dp = vq->vq_ring.desc;
 192         start_dp[idx].addr =
 193                 VIRTIO_MBUF_ADDR(cookie, vq) +
 194                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 195         start_dp[idx].len =
 196                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 197         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 198         idx = start_dp[idx].next;
 199         vq->vq_desc_head_idx = idx;
 200         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 201                 vq->vq_desc_tail_idx = idx;
 202         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 203         vq_update_avail_ring(vq, head_idx);
 204
 205         return 0;
 206 }
 207
 208 static inline void
 209 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 210                        uint16_t needed, int use_indirect, int can_push)
 211 {
 212         struct vq_desc_extra *dxp;
 213         struct virtqueue *vq = txvq->vq;
 214         struct vring_desc *start_dp;
 215         uint16_t seg_num = cookie->nb_segs;
 216         uint16_t head_idx, idx;
 217         uint16_t head_size = vq->hw->vtnet_hdr_size;
 218         unsigned long offs;
 219
 220         head_idx = vq->vq_desc_head_idx;
 221         idx = head_idx;
 222         dxp = &vq->vq_descx[idx];
 223         dxp->cookie = (void *)cookie;
 224         dxp->ndescs = needed;
 225
 226         start_dp = vq->vq_ring.desc;
 227
 228         if (can_push) {
 229                 /* put on zero'd transmit header (no offloads) */
 230                 void *hdr = rte_pktmbuf_prepend(cookie, head_size);
 231
 232                 memset(hdr, 0, head_size);
 233         } else if (use_indirect) {
 234                 /* setup tx ring slot to point to indirect
 235                  * descriptor list stored in reserved region.
 236                  *
 237                  * the first slot in indirect ring is already preset
 238                  * to point to the header in reserved region
 239                  */
 240                 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 241
 242                 offs = idx * sizeof(struct virtio_tx_region)
 243                         + offsetof(struct virtio_tx_region, tx_indir);
 244
 245                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 246                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 247                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 248
 249                 /* loop below will fill in rest of the indirect elements */
 250                 start_dp = txr[idx].tx_indir;
 251                 idx = 1;
 252         } else {
 253                 /* setup first tx ring slot to point to header
 254                  * stored in reserved region.
 255                  */
 256                 offs = idx * sizeof(struct virtio_tx_region)
 257                         + offsetof(struct virtio_tx_region, tx_hdr);
 258
 259                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 260                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 261                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 262                 idx = start_dp[idx].next;
 263         }
 264
 265         do {
 266                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 267                 start_dp[idx].len   = cookie->data_len;
 268                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 269                 idx = start_dp[idx].next;
 270         } while ((cookie = cookie->next) != NULL);
 271
 272         if (use_indirect)
 273                 idx = vq->vq_ring.desc[head_idx].next;
 274
 275         vq->vq_desc_head_idx = idx;
 276         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 277                 vq->vq_desc_tail_idx = idx;
 278         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 279         vq_update_avail_ring(vq, head_idx);
 280 }
 281
 282 static void
 283 virtio_dev_vring_start(struct virtqueue *vq)
 284 {
 285         int size = vq->vq_nentries;
 286         struct vring *vr = &vq->vq_ring;
 287         uint8_t *ring_mem = vq->vq_ring_virt_mem;
 288
 289         PMD_INIT_FUNC_TRACE();
 290
 291         /*
 292          * Reinitialise since virtio port might have been stopped and restarted
 293          */
 294         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
 295         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
 296         vq->vq_used_cons_idx = 0;
 297         vq->vq_desc_head_idx = 0;
 298         vq->vq_avail_idx = 0;
 299         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
 300         vq->vq_free_cnt = vq->vq_nentries;
 301         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 302
 303         vring_desc_init(vr->desc, size);
 304
 305         /*
 306          * Disable device(host) interrupting guest
 307          */
 308         virtqueue_disable_intr(vq);
 309 }
 310
 311 void
 312 virtio_dev_cq_start(struct rte_eth_dev *dev)
 313 {
 314         struct virtio_hw *hw = dev->data->dev_private;
 315
 316         if (hw->cvq && hw->cvq->vq) {
 317                 virtio_dev_vring_start(hw->cvq->vq);
 318                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 319         }
 320 }
 321
 322 void
 323 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 324 {
 325         /*
 326          * Start receive and transmit vrings
 327          * -    Setup vring structure for all queues
 328          * -    Initialize descriptor for the rx vring
 329          * -    Allocate blank mbufs for the each rx descriptor
 330          *
 331          */
 332         uint16_t i;
 333         uint16_t desc_idx;
 334         struct virtio_hw *hw = dev->data->dev_private;
 335
 336         PMD_INIT_FUNC_TRACE();
 337
 338         /* Start rx vring. */
 339         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 340                 struct virtnet_rx *rxvq = dev->data->rx_queues[i];
 341                 struct virtqueue *vq = rxvq->vq;
 342                 int error, nbufs;
 343                 struct rte_mbuf *m;
 344
 345                 virtio_dev_vring_start(vq);
 346                 if (rxvq->mpool == NULL) {
 347                         rte_exit(EXIT_FAILURE,
 348                                 "Cannot allocate mbufs for rx virtqueue");
 349                 }
 350
 351                 /* Allocate blank mbufs for the each rx descriptor */
 352                 nbufs = 0;
 353                 error = ENOSPC;
 354
 355                 if (hw->use_simple_rxtx) {
 356                         for (desc_idx = 0; desc_idx < vq->vq_nentries;
 357                              desc_idx++) {
 358                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 359                                 vq->vq_ring.desc[desc_idx].flags =
 360                                         VRING_DESC_F_WRITE;
 361                         }
 362                 }
 363
 364                 memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 365                 for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 366                      desc_idx++) {
 367                         vq->sw_ring[vq->vq_nentries + desc_idx] =
 368                                 &rxvq->fake_mbuf;
 369                 }
 370
 371                 while (!virtqueue_full(vq)) {
 372                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 373                         if (m == NULL)
 374                                 break;
 375
 376                         /******************************************
 377                         *         Enqueue allocated buffers        *
 378                         *******************************************/
 379                         if (hw->use_simple_rxtx)
 380                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 381                         else
 382                                 error = virtqueue_enqueue_recv_refill(vq, m);
 383
 384                         if (error) {
 385                                 rte_pktmbuf_free(m);
 386                                 break;
 387                         }
 388                         nbufs++;
 389                 }
 390
 391                 vq_update_avail_idx(vq);
 392
 393                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 394
 395                 VIRTQUEUE_DUMP(vq);
 396         }
 397
 398         /* Start tx vring. */
 399         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 400                 struct virtnet_tx *txvq = dev->data->tx_queues[i];
 401                 struct virtqueue *vq = txvq->vq;
 402
 403                 virtio_dev_vring_start(vq);
 404                 if (hw->use_simple_rxtx) {
 405                         uint16_t mid_idx  = vq->vq_nentries >> 1;
 406
 407                         for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 408                                 vq->vq_ring.avail->ring[desc_idx] =
 409                                         desc_idx + mid_idx;
 410                                 vq->vq_ring.desc[desc_idx + mid_idx].next =
 411                                         desc_idx;
 412                                 vq->vq_ring.desc[desc_idx + mid_idx].addr =
 413                                         txvq->virtio_net_hdr_mem +
 414                                         offsetof(struct virtio_tx_region, tx_hdr);
 415                                 vq->vq_ring.desc[desc_idx + mid_idx].len =
 416                                         vq->hw->vtnet_hdr_size;
 417                                 vq->vq_ring.desc[desc_idx + mid_idx].flags =
 418                                         VRING_DESC_F_NEXT;
 419                                 vq->vq_ring.desc[desc_idx].flags = 0;
 420                         }
 421                         for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 422                              desc_idx++)
 423                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 424                 }
 425
 426                 VIRTQUEUE_DUMP(vq);
 427         }
 428 }
 429
 430 int
 431 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 432                         uint16_t queue_idx,
 433                         uint16_t nb_desc,
 434                         unsigned int socket_id,
 435                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 436                         struct rte_mempool *mp)
 437 {
 438         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 439         struct virtnet_rx *rxvq;
 440         int ret;
 441
 442         PMD_INIT_FUNC_TRACE();
 443         ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
 444                         nb_desc, socket_id, (void **)&rxvq);
 445         if (ret < 0) {
 446                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 447                 return ret;
 448         }
 449
 450         /* Create mempool for rx mbuf allocation */
 451         rxvq->mpool = mp;
 452
 453         dev->data->rx_queues[queue_idx] = rxvq;
 454
 455         virtio_rxq_vec_setup(rxvq);
 456
 457         return 0;
 458 }
 459
 460 void
 461 virtio_dev_rx_queue_release(void *rxq)
 462 {
 463         struct virtnet_rx *rxvq = rxq;
 464         struct virtqueue *vq;
 465         const struct rte_memzone *mz;
 466
 467         if (rxvq == NULL)
 468                 return;
 469
 470         /*
 471          * rxvq is freed when vq is freed, and as mz should be freed after the
 472          * del_queue, so we reserve the mz pointer first.
 473          */
 474         vq = rxvq->vq;
 475         mz = rxvq->mz;
 476
 477         virtio_dev_queue_release(vq);
 478         rte_memzone_free(mz);
 479 }
 480
 481 static void
 482 virtio_update_rxtx_handler(struct rte_eth_dev *dev,
 483                            const struct rte_eth_txconf *tx_conf)
 484 {
 485         uint8_t use_simple_rxtx = 0;
 486         struct virtio_hw *hw = dev->data->dev_private;
 487
 488 #if defined RTE_ARCH_X86
 489         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE3))
 490                 use_simple_rxtx = 1;
 491 #elif defined RTE_ARCH_ARM64 || defined CONFIG_RTE_ARCH_ARM
 492         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
 493                 use_simple_rxtx = 1;
 494 #endif
 495         /* Use simple rx/tx func if single segment and no offloads */
 496         if (use_simple_rxtx &&
 497             (tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 498             !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 499                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 500                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 501                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 502                 hw->use_simple_rxtx = use_simple_rxtx;
 503         }
 504 }
 505
 506 /*
 507  * struct rte_eth_dev *dev: Used to update dev
 508  * uint16_t nb_desc: Defaults to values read from config space
 509  * unsigned int socket_id: Used to allocate memzone
 510  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 511  * uint16_t queue_idx: Just used as an index in dev txq list
 512  */
 513 int
 514 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 515                         uint16_t queue_idx,
 516                         uint16_t nb_desc,
 517                         unsigned int socket_id,
 518                         const struct rte_eth_txconf *tx_conf)
 519 {
 520         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 521         struct virtnet_tx *txvq;
 522         struct virtqueue *vq;
 523         uint16_t tx_free_thresh;
 524         int ret;
 525
 526         PMD_INIT_FUNC_TRACE();
 527
 528         if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
 529             != ETH_TXQ_FLAGS_NOXSUMS) {
 530                 PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
 531                 return -EINVAL;
 532         }
 533
 534         virtio_update_rxtx_handler(dev, tx_conf);
 535
 536         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
 537                         nb_desc, socket_id, (void **)&txvq);
 538         if (ret < 0) {
 539                 PMD_INIT_LOG(ERR, "tvq initialization failed");
 540                 return ret;
 541         }
 542         vq = txvq->vq;
 543
 544         tx_free_thresh = tx_conf->tx_free_thresh;
 545         if (tx_free_thresh == 0)
 546                 tx_free_thresh =
 547                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 548
 549         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 550                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 551                         "number of TX entries minus 3 (%u)."
 552                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 553                         vq->vq_nentries - 3,
 554                         tx_free_thresh, dev->data->port_id, queue_idx);
 555                 return -EINVAL;
 556         }
 557
 558         vq->vq_free_thresh = tx_free_thresh;
 559
 560         dev->data->tx_queues[queue_idx] = txvq;
 561         return 0;
 562 }
 563
 564 void
 565 virtio_dev_tx_queue_release(void *txq)
 566 {
 567         struct virtnet_tx *txvq = txq;
 568         struct virtqueue *vq;
 569         const struct rte_memzone *mz;
 570         const struct rte_memzone *hdr_mz;
 571
 572         if (txvq == NULL)
 573                 return;
 574
 575         /*
 576          * txvq is freed when vq is freed, and as mz should be freed after the
 577          * del_queue, so we reserve the mz pointer first.
 578          */
 579         vq = txvq->vq;
 580         mz = txvq->mz;
 581         hdr_mz = txvq->virtio_net_hdr_mz;
 582
 583         virtio_dev_queue_release(vq);
 584         rte_memzone_free(mz);
 585         rte_memzone_free(hdr_mz);
 586 }
 587
 588 static void
 589 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 590 {
 591         int error;
 592         /*
 593          * Requeue the discarded mbuf. This should always be
 594          * successful since it was just dequeued.
 595          */
 596         error = virtqueue_enqueue_recv_refill(vq, m);
 597         if (unlikely(error)) {
 598                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 599                 rte_pktmbuf_free(m);
 600         }
 601 }
 602
 603 static void
 604 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 605 {
 606         uint32_t s = mbuf->pkt_len;
 607         struct ether_addr *ea;
 608
 609         if (s == 64) {
 610                 stats->size_bins[1]++;
 611         } else if (s > 64 && s < 1024) {
 612                 uint32_t bin;
 613
 614                 /* count zeros, and offset into correct bin */
 615                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 616                 stats->size_bins[bin]++;
 617         } else {
 618                 if (s < 64)
 619                         stats->size_bins[0]++;
 620                 else if (s < 1519)
 621                         stats->size_bins[6]++;
 622                 else if (s >= 1519)
 623                         stats->size_bins[7]++;
 624         }
 625
 626         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 627         if (is_multicast_ether_addr(ea)) {
 628                 if (is_broadcast_ether_addr(ea))
 629                         stats->broadcast++;
 630                 else
 631                         stats->multicast++;
 632         }
 633 }
 634
 635 #define VIRTIO_MBUF_BURST_SZ 64
 636 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 637 uint16_t
 638 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 639 {
 640         struct virtnet_rx *rxvq = rx_queue;
 641         struct virtqueue *vq = rxvq->vq;
 642         struct virtio_hw *hw;
 643         struct rte_mbuf *rxm, *new_mbuf;
 644         uint16_t nb_used, num, nb_rx;
 645         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 646         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 647         int error;
 648         uint32_t i, nb_enqueued;
 649         uint32_t hdr_size;
 650
 651         nb_used = VIRTQUEUE_NUSED(vq);
 652
 653         virtio_rmb();
 654
 655         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 656         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 657         if (likely(num > DESC_PER_CACHELINE))
 658                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 659
 660         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 661         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 662
 663         hw = vq->hw;
 664         nb_rx = 0;
 665         nb_enqueued = 0;
 666         hdr_size = hw->vtnet_hdr_size;
 667
 668         for (i = 0; i < num ; i++) {
 669                 rxm = rcv_pkts[i];
 670
 671                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 672
 673                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 674                         PMD_RX_LOG(ERR, "Packet drop");
 675                         nb_enqueued++;
 676                         virtio_discard_rxbuf(vq, rxm);
 677                         rxvq->stats.errors++;
 678                         continue;
 679                 }
 680
 681                 rxm->port = rxvq->port_id;
 682                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 683                 rxm->ol_flags = 0;
 684                 rxm->vlan_tci = 0;
 685
 686                 rxm->nb_segs = 1;
 687                 rxm->next = NULL;
 688                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 689                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 690
 691                 if (hw->vlan_strip)
 692                         rte_vlan_strip(rxm);
 693
 694                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 695
 696                 rx_pkts[nb_rx++] = rxm;
 697
 698                 rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
 699                 virtio_update_packet_stats(&rxvq->stats, rxm);
 700         }
 701
 702         rxvq->stats.packets += nb_rx;
 703
 704         /* Allocate new mbuf for the used descriptor */
 705         error = ENOSPC;
 706         while (likely(!virtqueue_full(vq))) {
 707                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 708                 if (unlikely(new_mbuf == NULL)) {
 709                         struct rte_eth_dev *dev
 710                                 = &rte_eth_devices[rxvq->port_id];
 711                         dev->data->rx_mbuf_alloc_failed++;
 712                         break;
 713                 }
 714                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 715                 if (unlikely(error)) {
 716                         rte_pktmbuf_free(new_mbuf);
 717                         break;
 718                 }
 719                 nb_enqueued++;
 720         }
 721
 722         if (likely(nb_enqueued)) {
 723                 vq_update_avail_idx(vq);
 724
 725                 if (unlikely(virtqueue_kick_prepare(vq))) {
 726                         virtqueue_notify(vq);
 727                         PMD_RX_LOG(DEBUG, "Notified");
 728                 }
 729         }
 730
 731         return nb_rx;
 732 }
 733
 734 uint16_t
 735 virtio_recv_mergeable_pkts(void *rx_queue,
 736                         struct rte_mbuf **rx_pkts,
 737                         uint16_t nb_pkts)
 738 {
 739         struct virtnet_rx *rxvq = rx_queue;
 740         struct virtqueue *vq = rxvq->vq;
 741         struct virtio_hw *hw;
 742         struct rte_mbuf *rxm, *new_mbuf;
 743         uint16_t nb_used, num, nb_rx;
 744         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 745         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 746         struct rte_mbuf *prev;
 747         int error;
 748         uint32_t i, nb_enqueued;
 749         uint32_t seg_num;
 750         uint16_t extra_idx;
 751         uint32_t seg_res;
 752         uint32_t hdr_size;
 753
 754         nb_used = VIRTQUEUE_NUSED(vq);
 755
 756         virtio_rmb();
 757
 758         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 759
 760         hw = vq->hw;
 761         nb_rx = 0;
 762         i = 0;
 763         nb_enqueued = 0;
 764         seg_num = 0;
 765         extra_idx = 0;
 766         seg_res = 0;
 767         hdr_size = hw->vtnet_hdr_size;
 768
 769         while (i < nb_used) {
 770                 struct virtio_net_hdr_mrg_rxbuf *header;
 771
 772                 if (nb_rx == nb_pkts)
 773                         break;
 774
 775                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
 776                 if (num != 1)
 777                         continue;
 778
 779                 i++;
 780
 781                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 782                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 783
 784                 rxm = rcv_pkts[0];
 785
 786                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 787                         PMD_RX_LOG(ERR, "Packet drop");
 788                         nb_enqueued++;
 789                         virtio_discard_rxbuf(vq, rxm);
 790                         rxvq->stats.errors++;
 791                         continue;
 792                 }
 793
 794                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 795                         RTE_PKTMBUF_HEADROOM - hdr_size);
 796                 seg_num = header->num_buffers;
 797
 798                 if (seg_num == 0)
 799                         seg_num = 1;
 800
 801                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 802                 rxm->nb_segs = seg_num;
 803                 rxm->next = NULL;
 804                 rxm->ol_flags = 0;
 805                 rxm->vlan_tci = 0;
 806                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 807                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 808
 809                 rxm->port = rxvq->port_id;
 810                 rx_pkts[nb_rx] = rxm;
 811                 prev = rxm;
 812
 813                 seg_res = seg_num - 1;
 814
 815                 while (seg_res != 0) {
 816                         /*
 817                          * Get extra segments for current uncompleted packet.
 818                          */
 819                         uint16_t  rcv_cnt =
 820                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 821                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
 822                                 uint32_t rx_num =
 823                                         virtqueue_dequeue_burst_rx(vq,
 824                                         rcv_pkts, len, rcv_cnt);
 825                                 i += rx_num;
 826                                 rcv_cnt = rx_num;
 827                         } else {
 828                                 PMD_RX_LOG(ERR,
 829                                            "No enough segments for packet.");
 830                                 nb_enqueued++;
 831                                 virtio_discard_rxbuf(vq, rxm);
 832                                 rxvq->stats.errors++;
 833                                 break;
 834                         }
 835
 836                         extra_idx = 0;
 837
 838                         while (extra_idx < rcv_cnt) {
 839                                 rxm = rcv_pkts[extra_idx];
 840
 841                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 842                                 rxm->next = NULL;
 843                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 844                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 845
 846                                 if (prev)
 847                                         prev->next = rxm;
 848
 849                                 prev = rxm;
 850                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 851                                 extra_idx++;
 852                         };
 853                         seg_res -= rcv_cnt;
 854                 }
 855
 856                 if (hw->vlan_strip)
 857                         rte_vlan_strip(rx_pkts[nb_rx]);
 858
 859                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 860                         rx_pkts[nb_rx]->data_len);
 861
 862                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
 863                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
 864                 nb_rx++;
 865         }
 866
 867         rxvq->stats.packets += nb_rx;
 868
 869         /* Allocate new mbuf for the used descriptor */
 870         error = ENOSPC;
 871         while (likely(!virtqueue_full(vq))) {
 872                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 873                 if (unlikely(new_mbuf == NULL)) {
 874                         struct rte_eth_dev *dev
 875                                 = &rte_eth_devices[rxvq->port_id];
 876                         dev->data->rx_mbuf_alloc_failed++;
 877                         break;
 878                 }
 879                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 880                 if (unlikely(error)) {
 881                         rte_pktmbuf_free(new_mbuf);
 882                         break;
 883                 }
 884                 nb_enqueued++;
 885         }
 886
 887         if (likely(nb_enqueued)) {
 888                 vq_update_avail_idx(vq);
 889
 890                 if (unlikely(virtqueue_kick_prepare(vq))) {
 891                         virtqueue_notify(vq);
 892                         PMD_RX_LOG(DEBUG, "Notified");
 893                 }
 894         }
 895
 896         return nb_rx;
 897 }
 898
 899 uint16_t
 900 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 901 {
 902         struct virtnet_tx *txvq = tx_queue;
 903         struct virtqueue *vq = txvq->vq;
 904         struct virtio_hw *hw = vq->hw;
 905         uint16_t hdr_size = hw->vtnet_hdr_size;
 906         uint16_t nb_used, nb_tx;
 907         int error;
 908
 909         if (unlikely(nb_pkts < 1))
 910                 return nb_pkts;
 911
 912         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
 913         nb_used = VIRTQUEUE_NUSED(vq);
 914
 915         virtio_rmb();
 916         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
 917                 virtio_xmit_cleanup(vq, nb_used);
 918
 919         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 920                 struct rte_mbuf *txm = tx_pkts[nb_tx];
 921                 int can_push = 0, use_indirect = 0, slots, need;
 922
 923                 /* Do VLAN tag insertion */
 924                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
 925                         error = rte_vlan_insert(&txm);
 926                         if (unlikely(error)) {
 927                                 rte_pktmbuf_free(txm);
 928                                 continue;
 929                         }
 930                 }
 931
 932                 /* optimize ring usage */
 933                 if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
 934                     rte_mbuf_refcnt_read(txm) == 1 &&
 935                     RTE_MBUF_DIRECT(txm) &&
 936                     txm->nb_segs == 1 &&
 937                     rte_pktmbuf_headroom(txm) >= hdr_size &&
 938                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
 939                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
 940                         can_push = 1;
 941                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
 942                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
 943                         use_indirect = 1;
 944
 945                 /* How many main ring entries are needed to this Tx?
 946                  * any_layout => number of segments
 947                  * indirect   => 1
 948                  * default    => number of segments + 1
 949                  */
 950                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
 951                 need = slots - vq->vq_free_cnt;
 952
 953                 /* Positive value indicates it need free vring descriptors */
 954                 if (unlikely(need > 0)) {
 955                         nb_used = VIRTQUEUE_NUSED(vq);
 956                         virtio_rmb();
 957                         need = RTE_MIN(need, (int)nb_used);
 958
 959                         virtio_xmit_cleanup(vq, need);
 960                         need = slots - vq->vq_free_cnt;
 961                         if (unlikely(need > 0)) {
 962                                 PMD_TX_LOG(ERR,
 963                                            "No free tx descriptors to transmit");
 964                                 break;
 965                         }
 966                 }
 967
 968                 /* Enqueue Packet buffers */
 969                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
 970
 971                 txvq->stats.bytes += txm->pkt_len;
 972                 virtio_update_packet_stats(&txvq->stats, txm);
 973         }
 974
 975         txvq->stats.packets += nb_tx;
 976
 977         if (likely(nb_tx)) {
 978                 vq_update_avail_idx(vq);
 979
 980                 if (unlikely(virtqueue_kick_prepare(vq))) {
 981                         virtqueue_notify(vq);
 982                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
 983                 }
 984         }
 985
 986         return nb_tx;
 987 }