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 static void
  71 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  72 {
  73         struct vring_desc *dp, *dp_tail;
  74         struct vq_desc_extra *dxp;
  75         uint16_t desc_idx_last = desc_idx;
  76
  77         dp  = &vq->vq_ring.desc[desc_idx];
  78         dxp = &vq->vq_descx[desc_idx];
  79         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  80         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  81                 while (dp->flags & VRING_DESC_F_NEXT) {
  82                         desc_idx_last = dp->next;
  83                         dp = &vq->vq_ring.desc[dp->next];
  84                 }
  85         }
  86         dxp->ndescs = 0;
  87
  88         /*
  89          * We must append the existing free chain, if any, to the end of
  90          * newly freed chain. If the virtqueue was completely used, then
  91          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  92          */
  93         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  94                 vq->vq_desc_head_idx = desc_idx;
  95         } else {
  96                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
  97                 dp_tail->next = desc_idx;
  98         }
  99
 100         vq->vq_desc_tail_idx = desc_idx_last;
 101         dp->next = VQ_RING_DESC_CHAIN_END;
 102 }
 103
 104 static uint16_t
 105 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 106                            uint32_t *len, uint16_t num)
 107 {
 108         struct vring_used_elem *uep;
 109         struct rte_mbuf *cookie;
 110         uint16_t used_idx, desc_idx;
 111         uint16_t i;
 112
 113         /*  Caller does the check */
 114         for (i = 0; i < num ; i++) {
 115                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 116                 uep = &vq->vq_ring.used->ring[used_idx];
 117                 desc_idx = (uint16_t) uep->id;
 118                 len[i] = uep->len;
 119                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 120
 121                 if (unlikely(cookie == NULL)) {
 122                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 123                                 vq->vq_used_cons_idx);
 124                         break;
 125                 }
 126
 127                 rte_prefetch0(cookie);
 128                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 129                 rx_pkts[i]  = cookie;
 130                 vq->vq_used_cons_idx++;
 131                 vq_ring_free_chain(vq, desc_idx);
 132                 vq->vq_descx[desc_idx].cookie = NULL;
 133         }
 134
 135         return i;
 136 }
 137
 138 #ifndef DEFAULT_TX_FREE_THRESH
 139 #define DEFAULT_TX_FREE_THRESH 32
 140 #endif
 141
 142 /* Cleanup from completed transmits. */
 143 static void
 144 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 145 {
 146         uint16_t i, used_idx, desc_idx;
 147         for (i = 0; i < num; i++) {
 148                 struct vring_used_elem *uep;
 149                 struct vq_desc_extra *dxp;
 150
 151                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 152                 uep = &vq->vq_ring.used->ring[used_idx];
 153
 154                 desc_idx = (uint16_t) uep->id;
 155                 dxp = &vq->vq_descx[desc_idx];
 156                 vq->vq_used_cons_idx++;
 157                 vq_ring_free_chain(vq, desc_idx);
 158
 159                 if (dxp->cookie != NULL) {
 160                         rte_pktmbuf_free(dxp->cookie);
 161                         dxp->cookie = NULL;
 162                 }
 163         }
 164 }
 165
 166
 167 static inline int
 168 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 169 {
 170         struct vq_desc_extra *dxp;
 171         struct virtio_hw *hw = vq->hw;
 172         struct vring_desc *start_dp;
 173         uint16_t needed = 1;
 174         uint16_t head_idx, idx;
 175
 176         if (unlikely(vq->vq_free_cnt == 0))
 177                 return -ENOSPC;
 178         if (unlikely(vq->vq_free_cnt < needed))
 179                 return -EMSGSIZE;
 180
 181         head_idx = vq->vq_desc_head_idx;
 182         if (unlikely(head_idx >= vq->vq_nentries))
 183                 return -EFAULT;
 184
 185         idx = head_idx;
 186         dxp = &vq->vq_descx[idx];
 187         dxp->cookie = (void *)cookie;
 188         dxp->ndescs = needed;
 189
 190         start_dp = vq->vq_ring.desc;
 191         start_dp[idx].addr =
 192                 VIRTIO_MBUF_ADDR(cookie, vq) +
 193                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 194         start_dp[idx].len =
 195                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 196         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 197         idx = start_dp[idx].next;
 198         vq->vq_desc_head_idx = idx;
 199         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 200                 vq->vq_desc_tail_idx = idx;
 201         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 202         vq_update_avail_ring(vq, head_idx);
 203
 204         return 0;
 205 }
 206
 207 static inline void
 208 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 209                        uint16_t needed, int use_indirect, int can_push)
 210 {
 211         struct vq_desc_extra *dxp;
 212         struct virtqueue *vq = txvq->vq;
 213         struct vring_desc *start_dp;
 214         uint16_t seg_num = cookie->nb_segs;
 215         uint16_t head_idx, idx;
 216         uint16_t head_size = vq->hw->vtnet_hdr_size;
 217         unsigned long offs;
 218
 219         head_idx = vq->vq_desc_head_idx;
 220         idx = head_idx;
 221         dxp = &vq->vq_descx[idx];
 222         dxp->cookie = (void *)cookie;
 223         dxp->ndescs = needed;
 224
 225         start_dp = vq->vq_ring.desc;
 226
 227         if (can_push) {
 228                 /* put on zero'd transmit header (no offloads) */
 229                 void *hdr = rte_pktmbuf_prepend(cookie, head_size);
 230
 231                 memset(hdr, 0, head_size);
 232         } else if (use_indirect) {
 233                 /* setup tx ring slot to point to indirect
 234                  * descriptor list stored in reserved region.
 235                  *
 236                  * the first slot in indirect ring is already preset
 237                  * to point to the header in reserved region
 238                  */
 239                 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 240
 241                 offs = idx * sizeof(struct virtio_tx_region)
 242                         + offsetof(struct virtio_tx_region, tx_indir);
 243
 244                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 245                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 246                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 247
 248                 /* loop below will fill in rest of the indirect elements */
 249                 start_dp = txr[idx].tx_indir;
 250                 idx = 1;
 251         } else {
 252                 /* setup first tx ring slot to point to header
 253                  * stored in reserved region.
 254                  */
 255                 offs = idx * sizeof(struct virtio_tx_region)
 256                         + offsetof(struct virtio_tx_region, tx_hdr);
 257
 258                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 259                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 260                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 261                 idx = start_dp[idx].next;
 262         }
 263
 264         do {
 265                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 266                 start_dp[idx].len   = cookie->data_len;
 267                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 268                 idx = start_dp[idx].next;
 269         } while ((cookie = cookie->next) != NULL);
 270
 271         if (use_indirect)
 272                 idx = vq->vq_ring.desc[head_idx].next;
 273
 274         vq->vq_desc_head_idx = idx;
 275         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 276                 vq->vq_desc_tail_idx = idx;
 277         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 278         vq_update_avail_ring(vq, head_idx);
 279 }
 280
 281 static void
 282 virtio_dev_vring_start(struct virtqueue *vq)
 283 {
 284         int size = vq->vq_nentries;
 285         struct vring *vr = &vq->vq_ring;
 286         uint8_t *ring_mem = vq->vq_ring_virt_mem;
 287
 288         PMD_INIT_FUNC_TRACE();
 289
 290         /*
 291          * Reinitialise since virtio port might have been stopped and restarted
 292          */
 293         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
 294         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
 295         vq->vq_used_cons_idx = 0;
 296         vq->vq_desc_head_idx = 0;
 297         vq->vq_avail_idx = 0;
 298         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
 299         vq->vq_free_cnt = vq->vq_nentries;
 300         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 301
 302         vring_desc_init(vr->desc, size);
 303
 304         /*
 305          * Disable device(host) interrupting guest
 306          */
 307         virtqueue_disable_intr(vq);
 308 }
 309
 310 void
 311 virtio_dev_cq_start(struct rte_eth_dev *dev)
 312 {
 313         struct virtio_hw *hw = dev->data->dev_private;
 314
 315         if (hw->cvq && hw->cvq->vq) {
 316                 virtio_dev_vring_start(hw->cvq->vq);
 317                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 318         }
 319 }
 320
 321 void
 322 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 323 {
 324         /*
 325          * Start receive and transmit vrings
 326          * -    Setup vring structure for all queues
 327          * -    Initialize descriptor for the rx vring
 328          * -    Allocate blank mbufs for the each rx descriptor
 329          *
 330          */
 331         uint16_t i;
 332         uint16_t desc_idx;
 333         struct virtio_hw *hw = dev->data->dev_private;
 334
 335         PMD_INIT_FUNC_TRACE();
 336
 337         /* Start rx vring. */
 338         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 339                 struct virtnet_rx *rxvq = dev->data->rx_queues[i];
 340                 struct virtqueue *vq = rxvq->vq;
 341                 int error, nbufs;
 342                 struct rte_mbuf *m;
 343
 344                 virtio_dev_vring_start(vq);
 345                 if (rxvq->mpool == NULL) {
 346                         rte_exit(EXIT_FAILURE,
 347                                 "Cannot allocate mbufs for rx virtqueue");
 348                 }
 349
 350                 /* Allocate blank mbufs for the each rx descriptor */
 351                 nbufs = 0;
 352                 error = ENOSPC;
 353
 354                 if (hw->use_simple_rxtx) {
 355                         for (desc_idx = 0; desc_idx < vq->vq_nentries;
 356                              desc_idx++) {
 357                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 358                                 vq->vq_ring.desc[desc_idx].flags =
 359                                         VRING_DESC_F_WRITE;
 360                         }
 361                 }
 362
 363                 memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 364                 for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 365                      desc_idx++) {
 366                         vq->sw_ring[vq->vq_nentries + desc_idx] =
 367                                 &rxvq->fake_mbuf;
 368                 }
 369
 370                 while (!virtqueue_full(vq)) {
 371                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 372                         if (m == NULL)
 373                                 break;
 374
 375                         /******************************************
 376                         *         Enqueue allocated buffers        *
 377                         *******************************************/
 378                         if (hw->use_simple_rxtx)
 379                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 380                         else
 381                                 error = virtqueue_enqueue_recv_refill(vq, m);
 382
 383                         if (error) {
 384                                 rte_pktmbuf_free(m);
 385                                 break;
 386                         }
 387                         nbufs++;
 388                 }
 389
 390                 vq_update_avail_idx(vq);
 391
 392                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 393
 394                 VIRTQUEUE_DUMP(vq);
 395         }
 396
 397         /* Start tx vring. */
 398         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 399                 struct virtnet_tx *txvq = dev->data->tx_queues[i];
 400                 struct virtqueue *vq = txvq->vq;
 401
 402                 virtio_dev_vring_start(vq);
 403                 if (hw->use_simple_rxtx) {
 404                         uint16_t mid_idx  = vq->vq_nentries >> 1;
 405
 406                         for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 407                                 vq->vq_ring.avail->ring[desc_idx] =
 408                                         desc_idx + mid_idx;
 409                                 vq->vq_ring.desc[desc_idx + mid_idx].next =
 410                                         desc_idx;
 411                                 vq->vq_ring.desc[desc_idx + mid_idx].addr =
 412                                         txvq->virtio_net_hdr_mem +
 413                                         offsetof(struct virtio_tx_region, tx_hdr);
 414                                 vq->vq_ring.desc[desc_idx + mid_idx].len =
 415                                         vq->hw->vtnet_hdr_size;
 416                                 vq->vq_ring.desc[desc_idx + mid_idx].flags =
 417                                         VRING_DESC_F_NEXT;
 418                                 vq->vq_ring.desc[desc_idx].flags = 0;
 419                         }
 420                         for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 421                              desc_idx++)
 422                                 vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 423                 }
 424
 425                 VIRTQUEUE_DUMP(vq);
 426         }
 427 }
 428
 429 int
 430 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 431                         uint16_t queue_idx,
 432                         uint16_t nb_desc,
 433                         unsigned int socket_id,
 434                         __rte_unused const struct rte_eth_rxconf *rx_conf,
 435                         struct rte_mempool *mp)
 436 {
 437         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 438         struct virtnet_rx *rxvq;
 439         int ret;
 440
 441         PMD_INIT_FUNC_TRACE();
 442         ret = virtio_dev_queue_setup(dev, VTNET_RQ, queue_idx, vtpci_queue_idx,
 443                         nb_desc, socket_id, (void **)&rxvq);
 444         if (ret < 0) {
 445                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 446                 return ret;
 447         }
 448
 449         /* Create mempool for rx mbuf allocation */
 450         rxvq->mpool = mp;
 451
 452         dev->data->rx_queues[queue_idx] = rxvq;
 453
 454         virtio_rxq_vec_setup(rxvq);
 455
 456         return 0;
 457 }
 458
 459 void
 460 virtio_dev_rx_queue_release(void *rxq)
 461 {
 462         struct virtnet_rx *rxvq = rxq;
 463         struct virtqueue *vq;
 464         const struct rte_memzone *mz;
 465
 466         if (rxvq == NULL)
 467                 return;
 468
 469         /*
 470          * rxvq is freed when vq is freed, and as mz should be freed after the
 471          * del_queue, so we reserve the mz pointer first.
 472          */
 473         vq = rxvq->vq;
 474         mz = rxvq->mz;
 475
 476         virtio_dev_queue_release(vq);
 477         rte_memzone_free(mz);
 478 }
 479
 480 /*
 481  * struct rte_eth_dev *dev: Used to update dev
 482  * uint16_t nb_desc: Defaults to values read from config space
 483  * unsigned int socket_id: Used to allocate memzone
 484  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 485  * uint16_t queue_idx: Just used as an index in dev txq list
 486  */
 487 int
 488 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 489                         uint16_t queue_idx,
 490                         uint16_t nb_desc,
 491                         unsigned int socket_id,
 492                         const struct rte_eth_txconf *tx_conf)
 493 {
 494         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 495
 496 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 497         struct virtio_hw *hw = dev->data->dev_private;
 498 #endif
 499         struct virtnet_tx *txvq;
 500         struct virtqueue *vq;
 501         uint16_t tx_free_thresh;
 502         int ret;
 503
 504         PMD_INIT_FUNC_TRACE();
 505
 506         if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
 507             != ETH_TXQ_FLAGS_NOXSUMS) {
 508                 PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
 509                 return -EINVAL;
 510         }
 511
 512 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 513         /* Use simple rx/tx func if single segment and no offloads */
 514         if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 515              !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 516                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 517                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 518                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 519                 hw->use_simple_rxtx = 1;
 520         }
 521 #endif
 522
 523         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
 524                         nb_desc, socket_id, (void **)&txvq);
 525         if (ret < 0) {
 526                 PMD_INIT_LOG(ERR, "tvq initialization failed");
 527                 return ret;
 528         }
 529         vq = txvq->vq;
 530
 531         tx_free_thresh = tx_conf->tx_free_thresh;
 532         if (tx_free_thresh == 0)
 533                 tx_free_thresh =
 534                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 535
 536         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 537                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 538                         "number of TX entries minus 3 (%u)."
 539                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 540                         vq->vq_nentries - 3,
 541                         tx_free_thresh, dev->data->port_id, queue_idx);
 542                 return -EINVAL;
 543         }
 544
 545         vq->vq_free_thresh = tx_free_thresh;
 546
 547         dev->data->tx_queues[queue_idx] = txvq;
 548         return 0;
 549 }
 550
 551 void
 552 virtio_dev_tx_queue_release(void *txq)
 553 {
 554         struct virtnet_tx *txvq = txq;
 555         struct virtqueue *vq;
 556         const struct rte_memzone *mz;
 557         const struct rte_memzone *hdr_mz;
 558
 559         if (txvq == NULL)
 560                 return;
 561
 562         /*
 563          * txvq is freed when vq is freed, and as mz should be freed after the
 564          * del_queue, so we reserve the mz pointer first.
 565          */
 566         vq = txvq->vq;
 567         mz = txvq->mz;
 568         hdr_mz = txvq->virtio_net_hdr_mz;
 569
 570         virtio_dev_queue_release(vq);
 571         rte_memzone_free(mz);
 572         rte_memzone_free(hdr_mz);
 573 }
 574
 575 static void
 576 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 577 {
 578         int error;
 579         /*
 580          * Requeue the discarded mbuf. This should always be
 581          * successful since it was just dequeued.
 582          */
 583         error = virtqueue_enqueue_recv_refill(vq, m);
 584         if (unlikely(error)) {
 585                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 586                 rte_pktmbuf_free(m);
 587         }
 588 }
 589
 590 static void
 591 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 592 {
 593         uint32_t s = mbuf->pkt_len;
 594         struct ether_addr *ea;
 595
 596         if (s == 64) {
 597                 stats->size_bins[1]++;
 598         } else if (s > 64 && s < 1024) {
 599                 uint32_t bin;
 600
 601                 /* count zeros, and offset into correct bin */
 602                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 603                 stats->size_bins[bin]++;
 604         } else {
 605                 if (s < 64)
 606                         stats->size_bins[0]++;
 607                 else if (s < 1519)
 608                         stats->size_bins[6]++;
 609                 else if (s >= 1519)
 610                         stats->size_bins[7]++;
 611         }
 612
 613         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 614         if (is_multicast_ether_addr(ea)) {
 615                 if (is_broadcast_ether_addr(ea))
 616                         stats->broadcast++;
 617                 else
 618                         stats->multicast++;
 619         }
 620 }
 621
 622 #define VIRTIO_MBUF_BURST_SZ 64
 623 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 624 uint16_t
 625 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 626 {
 627         struct virtnet_rx *rxvq = rx_queue;
 628         struct virtqueue *vq = rxvq->vq;
 629         struct virtio_hw *hw;
 630         struct rte_mbuf *rxm, *new_mbuf;
 631         uint16_t nb_used, num, nb_rx;
 632         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 633         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 634         int error;
 635         uint32_t i, nb_enqueued;
 636         uint32_t hdr_size;
 637
 638         nb_used = VIRTQUEUE_NUSED(vq);
 639
 640         virtio_rmb();
 641
 642         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 643         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 644         if (likely(num > DESC_PER_CACHELINE))
 645                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 646
 647         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 648         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 649
 650         hw = vq->hw;
 651         nb_rx = 0;
 652         nb_enqueued = 0;
 653         hdr_size = hw->vtnet_hdr_size;
 654
 655         for (i = 0; i < num ; i++) {
 656                 rxm = rcv_pkts[i];
 657
 658                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 659
 660                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 661                         PMD_RX_LOG(ERR, "Packet drop");
 662                         nb_enqueued++;
 663                         virtio_discard_rxbuf(vq, rxm);
 664                         rxvq->stats.errors++;
 665                         continue;
 666                 }
 667
 668                 rxm->port = rxvq->port_id;
 669                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 670                 rxm->ol_flags = 0;
 671                 rxm->vlan_tci = 0;
 672
 673                 rxm->nb_segs = 1;
 674                 rxm->next = NULL;
 675                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 676                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 677
 678                 if (hw->vlan_strip)
 679                         rte_vlan_strip(rxm);
 680
 681                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 682
 683                 rx_pkts[nb_rx++] = rxm;
 684
 685                 rxvq->stats.bytes += rx_pkts[nb_rx - 1]->pkt_len;
 686                 virtio_update_packet_stats(&rxvq->stats, rxm);
 687         }
 688
 689         rxvq->stats.packets += nb_rx;
 690
 691         /* Allocate new mbuf for the used descriptor */
 692         error = ENOSPC;
 693         while (likely(!virtqueue_full(vq))) {
 694                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 695                 if (unlikely(new_mbuf == NULL)) {
 696                         struct rte_eth_dev *dev
 697                                 = &rte_eth_devices[rxvq->port_id];
 698                         dev->data->rx_mbuf_alloc_failed++;
 699                         break;
 700                 }
 701                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 702                 if (unlikely(error)) {
 703                         rte_pktmbuf_free(new_mbuf);
 704                         break;
 705                 }
 706                 nb_enqueued++;
 707         }
 708
 709         if (likely(nb_enqueued)) {
 710                 vq_update_avail_idx(vq);
 711
 712                 if (unlikely(virtqueue_kick_prepare(vq))) {
 713                         virtqueue_notify(vq);
 714                         PMD_RX_LOG(DEBUG, "Notified");
 715                 }
 716         }
 717
 718         return nb_rx;
 719 }
 720
 721 uint16_t
 722 virtio_recv_mergeable_pkts(void *rx_queue,
 723                         struct rte_mbuf **rx_pkts,
 724                         uint16_t nb_pkts)
 725 {
 726         struct virtnet_rx *rxvq = rx_queue;
 727         struct virtqueue *vq = rxvq->vq;
 728         struct virtio_hw *hw;
 729         struct rte_mbuf *rxm, *new_mbuf;
 730         uint16_t nb_used, num, nb_rx;
 731         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 732         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 733         struct rte_mbuf *prev;
 734         int error;
 735         uint32_t i, nb_enqueued;
 736         uint32_t seg_num;
 737         uint16_t extra_idx;
 738         uint32_t seg_res;
 739         uint32_t hdr_size;
 740
 741         nb_used = VIRTQUEUE_NUSED(vq);
 742
 743         virtio_rmb();
 744
 745         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 746
 747         hw = vq->hw;
 748         nb_rx = 0;
 749         i = 0;
 750         nb_enqueued = 0;
 751         seg_num = 0;
 752         extra_idx = 0;
 753         seg_res = 0;
 754         hdr_size = hw->vtnet_hdr_size;
 755
 756         while (i < nb_used) {
 757                 struct virtio_net_hdr_mrg_rxbuf *header;
 758
 759                 if (nb_rx == nb_pkts)
 760                         break;
 761
 762                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
 763                 if (num != 1)
 764                         continue;
 765
 766                 i++;
 767
 768                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 769                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 770
 771                 rxm = rcv_pkts[0];
 772
 773                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 774                         PMD_RX_LOG(ERR, "Packet drop");
 775                         nb_enqueued++;
 776                         virtio_discard_rxbuf(vq, rxm);
 777                         rxvq->stats.errors++;
 778                         continue;
 779                 }
 780
 781                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 782                         RTE_PKTMBUF_HEADROOM - hdr_size);
 783                 seg_num = header->num_buffers;
 784
 785                 if (seg_num == 0)
 786                         seg_num = 1;
 787
 788                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 789                 rxm->nb_segs = seg_num;
 790                 rxm->next = NULL;
 791                 rxm->ol_flags = 0;
 792                 rxm->vlan_tci = 0;
 793                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 794                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 795
 796                 rxm->port = rxvq->port_id;
 797                 rx_pkts[nb_rx] = rxm;
 798                 prev = rxm;
 799
 800                 seg_res = seg_num - 1;
 801
 802                 while (seg_res != 0) {
 803                         /*
 804                          * Get extra segments for current uncompleted packet.
 805                          */
 806                         uint16_t  rcv_cnt =
 807                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 808                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
 809                                 uint32_t rx_num =
 810                                         virtqueue_dequeue_burst_rx(vq,
 811                                         rcv_pkts, len, rcv_cnt);
 812                                 i += rx_num;
 813                                 rcv_cnt = rx_num;
 814                         } else {
 815                                 PMD_RX_LOG(ERR,
 816                                            "No enough segments for packet.");
 817                                 nb_enqueued++;
 818                                 virtio_discard_rxbuf(vq, rxm);
 819                                 rxvq->stats.errors++;
 820                                 break;
 821                         }
 822
 823                         extra_idx = 0;
 824
 825                         while (extra_idx < rcv_cnt) {
 826                                 rxm = rcv_pkts[extra_idx];
 827
 828                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 829                                 rxm->next = NULL;
 830                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 831                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 832
 833                                 if (prev)
 834                                         prev->next = rxm;
 835
 836                                 prev = rxm;
 837                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 838                                 extra_idx++;
 839                         };
 840                         seg_res -= rcv_cnt;
 841                 }
 842
 843                 if (hw->vlan_strip)
 844                         rte_vlan_strip(rx_pkts[nb_rx]);
 845
 846                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 847                         rx_pkts[nb_rx]->data_len);
 848
 849                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
 850                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
 851                 nb_rx++;
 852         }
 853
 854         rxvq->stats.packets += nb_rx;
 855
 856         /* Allocate new mbuf for the used descriptor */
 857         error = ENOSPC;
 858         while (likely(!virtqueue_full(vq))) {
 859                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
 860                 if (unlikely(new_mbuf == NULL)) {
 861                         struct rte_eth_dev *dev
 862                                 = &rte_eth_devices[rxvq->port_id];
 863                         dev->data->rx_mbuf_alloc_failed++;
 864                         break;
 865                 }
 866                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
 867                 if (unlikely(error)) {
 868                         rte_pktmbuf_free(new_mbuf);
 869                         break;
 870                 }
 871                 nb_enqueued++;
 872         }
 873
 874         if (likely(nb_enqueued)) {
 875                 vq_update_avail_idx(vq);
 876
 877                 if (unlikely(virtqueue_kick_prepare(vq))) {
 878                         virtqueue_notify(vq);
 879                         PMD_RX_LOG(DEBUG, "Notified");
 880                 }
 881         }
 882
 883         return nb_rx;
 884 }
 885
 886 uint16_t
 887 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 888 {
 889         struct virtnet_tx *txvq = tx_queue;
 890         struct virtqueue *vq = txvq->vq;
 891         struct virtio_hw *hw = vq->hw;
 892         uint16_t hdr_size = hw->vtnet_hdr_size;
 893         uint16_t nb_used, nb_tx;
 894         int error;
 895
 896         if (unlikely(nb_pkts < 1))
 897                 return nb_pkts;
 898
 899         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
 900         nb_used = VIRTQUEUE_NUSED(vq);
 901
 902         virtio_rmb();
 903         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
 904                 virtio_xmit_cleanup(vq, nb_used);
 905
 906         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 907                 struct rte_mbuf *txm = tx_pkts[nb_tx];
 908                 int can_push = 0, use_indirect = 0, slots, need;
 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                 /* optimize ring usage */
 920                 if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
 921                     rte_mbuf_refcnt_read(txm) == 1 &&
 922                     RTE_MBUF_DIRECT(txm) &&
 923                     txm->nb_segs == 1 &&
 924                     rte_pktmbuf_headroom(txm) >= hdr_size &&
 925                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
 926                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
 927                         can_push = 1;
 928                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
 929                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
 930                         use_indirect = 1;
 931
 932                 /* How many main ring entries are needed to this Tx?
 933                  * any_layout => number of segments
 934                  * indirect   => 1
 935                  * default    => number of segments + 1
 936                  */
 937                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
 938                 need = slots - vq->vq_free_cnt;
 939
 940                 /* Positive value indicates it need free vring descriptors */
 941                 if (unlikely(need > 0)) {
 942                         nb_used = VIRTQUEUE_NUSED(vq);
 943                         virtio_rmb();
 944                         need = RTE_MIN(need, (int)nb_used);
 945
 946                         virtio_xmit_cleanup(vq, need);
 947                         need = slots - vq->vq_free_cnt;
 948                         if (unlikely(need > 0)) {
 949                                 PMD_TX_LOG(ERR,
 950                                            "No free tx descriptors to transmit");
 951                                 break;
 952                         }
 953                 }
 954
 955                 /* Enqueue Packet buffers */
 956                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
 957
 958                 txvq->stats.bytes += txm->pkt_len;
 959                 virtio_update_packet_stats(&txvq->stats, txm);
 960         }
 961
 962         txvq->stats.packets += nb_tx;
 963
 964         if (likely(nb_tx)) {
 965                 vq_update_avail_idx(vq);
 966
 967                 if (unlikely(virtqueue_kick_prepare(vq))) {
 968                         virtqueue_notify(vq);
 969                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
 970                 }
 971         }
 972
 973         return nb_tx;
 974 }