drivers/net/virtio/virtio_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38 #include <errno.h>
  39
  40 #include <rte_cycles.h>
  41 #include <rte_memory.h>
  42 #include <rte_memzone.h>
  43 #include <rte_branch_prediction.h>
  44 #include <rte_mempool.h>
  45 #include <rte_malloc.h>
  46 #include <rte_mbuf.h>
  47 #include <rte_ether.h>
  48 #include <rte_ethdev.h>
  49 #include <rte_prefetch.h>
  50 #include <rte_string_fns.h>
  51 #include <rte_errno.h>
  52 #include <rte_byteorder.h>
  53
  54 #include "virtio_logs.h"
  55 #include "virtio_ethdev.h"
  56 #include "virtio_pci.h"
  57 #include "virtqueue.h"
  58 #include "virtio_rxtx.h"
  59
  60 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  61 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  62 #else
  63 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  64 #endif
  65
  66
  67 #define VIRTIO_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
  68         ETH_TXQ_FLAGS_NOOFFLOADS)
  69
  70 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
  71 static int use_simple_rxtx;
  72 #endif
  73
  74 static void
  75 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  76 {
  77         struct vring_desc *dp, *dp_tail;
  78         struct vq_desc_extra *dxp;
  79         uint16_t desc_idx_last = desc_idx;
  80
  81         dp  = &vq->vq_ring.desc[desc_idx];
  82         dxp = &vq->vq_descx[desc_idx];
  83         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  84         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  85                 while (dp->flags & VRING_DESC_F_NEXT) {
  86                         desc_idx_last = dp->next;
  87                         dp = &vq->vq_ring.desc[dp->next];
  88                 }
  89         }
  90         dxp->ndescs = 0;
  91
  92         /*
  93          * We must append the existing free chain, if any, to the end of
  94          * newly freed chain. If the virtqueue was completely used, then
  95          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  96          */
  97         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  98                 vq->vq_desc_head_idx = desc_idx;
  99         } else {
 100                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
 101                 dp_tail->next = desc_idx;
 102         }
 103
 104         vq->vq_desc_tail_idx = desc_idx_last;
 105         dp->next = VQ_RING_DESC_CHAIN_END;
 106 }
 107
 108 static uint16_t
 109 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
 110                            uint32_t *len, uint16_t num)
 111 {
 112         struct vring_used_elem *uep;
 113         struct rte_mbuf *cookie;
 114         uint16_t used_idx, desc_idx;
 115         uint16_t i;
 116
 117         /*  Caller does the check */
 118         for (i = 0; i < num ; i++) {
 119                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 120                 uep = &vq->vq_ring.used->ring[used_idx];
 121                 desc_idx = (uint16_t) uep->id;
 122                 len[i] = uep->len;
 123                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 124
 125                 if (unlikely(cookie == NULL)) {
 126                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u\n",
 127                                 vq->vq_used_cons_idx);
 128                         break;
 129                 }
 130
 131                 rte_prefetch0(cookie);
 132                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 133                 rx_pkts[i]  = cookie;
 134                 vq->vq_used_cons_idx++;
 135                 vq_ring_free_chain(vq, desc_idx);
 136                 vq->vq_descx[desc_idx].cookie = NULL;
 137         }
 138
 139         return i;
 140 }
 141
 142 #ifndef DEFAULT_TX_FREE_THRESH
 143 #define DEFAULT_TX_FREE_THRESH 32
 144 #endif
 145
 146 /* Cleanup from completed transmits. */
 147 static void
 148 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 149 {
 150         uint16_t i, used_idx, desc_idx;
 151         for (i = 0; i < num; i++) {
 152                 struct vring_used_elem *uep;
 153                 struct vq_desc_extra *dxp;
 154
 155                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 156                 uep = &vq->vq_ring.used->ring[used_idx];
 157
 158                 desc_idx = (uint16_t) uep->id;
 159                 dxp = &vq->vq_descx[desc_idx];
 160                 vq->vq_used_cons_idx++;
 161                 vq_ring_free_chain(vq, desc_idx);
 162
 163                 if (dxp->cookie != NULL) {
 164                         rte_pktmbuf_free(dxp->cookie);
 165                         dxp->cookie = NULL;
 166                 }
 167         }
 168 }
 169
 170
 171 static inline int
 172 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 173 {
 174         struct vq_desc_extra *dxp;
 175         struct virtio_hw *hw = vq->hw;
 176         struct vring_desc *start_dp;
 177         uint16_t needed = 1;
 178         uint16_t head_idx, idx;
 179
 180         if (unlikely(vq->vq_free_cnt == 0))
 181                 return -ENOSPC;
 182         if (unlikely(vq->vq_free_cnt < needed))
 183                 return -EMSGSIZE;
 184
 185         head_idx = vq->vq_desc_head_idx;
 186         if (unlikely(head_idx >= vq->vq_nentries))
 187                 return -EFAULT;
 188
 189         idx = head_idx;
 190         dxp = &vq->vq_descx[idx];
 191         dxp->cookie = (void *)cookie;
 192         dxp->ndescs = needed;
 193
 194         start_dp = vq->vq_ring.desc;
 195         start_dp[idx].addr =
 196                 (uint64_t)(cookie->buf_physaddr + RTE_PKTMBUF_HEADROOM
 197                 - hw->vtnet_hdr_size);
 198         start_dp[idx].len =
 199                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 200         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 201         idx = start_dp[idx].next;
 202         vq->vq_desc_head_idx = idx;
 203         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 204                 vq->vq_desc_tail_idx = idx;
 205         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 206         vq_update_avail_ring(vq, head_idx);
 207
 208         return 0;
 209 }
 210
 211 static inline void
 212 virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie,
 213                        uint16_t needed, int use_indirect, int can_push)
 214 {
 215         struct vq_desc_extra *dxp;
 216         struct vring_desc *start_dp;
 217         uint16_t seg_num = cookie->nb_segs;
 218         uint16_t head_idx, idx;
 219         uint16_t head_size = txvq->hw->vtnet_hdr_size;
 220         unsigned long offs;
 221
 222         head_idx = txvq->vq_desc_head_idx;
 223         idx = head_idx;
 224         dxp = &txvq->vq_descx[idx];
 225         dxp->cookie = (void *)cookie;
 226         dxp->ndescs = needed;
 227
 228         start_dp = txvq->vq_ring.desc;
 229
 230         if (can_push) {
 231                 /* put on zero'd transmit header (no offloads) */
 232                 void *hdr = rte_pktmbuf_prepend(cookie, head_size);
 233
 234                 memset(hdr, 0, head_size);
 235         } else if (use_indirect) {
 236                 /* setup tx ring slot to point to indirect
 237                  * descriptor list stored in reserved region.
 238                  *
 239                  * the first slot in indirect ring is already preset
 240                  * to point to the header in reserved region
 241                  */
 242                 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 243
 244                 offs = idx * sizeof(struct virtio_tx_region)
 245                         + offsetof(struct virtio_tx_region, tx_indir);
 246
 247                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 248                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 249                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 250
 251                 /* loop below will fill in rest of the indirect elements */
 252                 start_dp = txr[idx].tx_indir;
 253                 idx = 1;
 254         } else {
 255                 /* setup first tx ring slot to point to header
 256                  * stored in reserved region.
 257                  */
 258                 offs = idx * sizeof(struct virtio_tx_region)
 259                         + offsetof(struct virtio_tx_region, tx_hdr);
 260
 261                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
 262                 start_dp[idx].len   = txvq->hw->vtnet_hdr_size;
 263                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 264                 idx = start_dp[idx].next;
 265         }
 266
 267         do {
 268                 start_dp[idx].addr  = rte_mbuf_data_dma_addr(cookie);
 269                 start_dp[idx].len   = cookie->data_len;
 270                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 271                 idx = start_dp[idx].next;
 272         } while ((cookie = cookie->next) != NULL);
 273
 274         if (use_indirect)
 275                 idx = txvq->vq_ring.desc[head_idx].next;
 276
 277         txvq->vq_desc_head_idx = idx;
 278         if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 279                 txvq->vq_desc_tail_idx = idx;
 280         txvq->vq_free_cnt = (uint16_t)(txvq->vq_free_cnt - needed);
 281         vq_update_avail_ring(txvq, head_idx);
 282 }
 283
 284 static inline struct rte_mbuf *
 285 rte_rxmbuf_alloc(struct rte_mempool *mp)
 286 {
 287         struct rte_mbuf *m;
 288
 289         m = __rte_mbuf_raw_alloc(mp);
 290         __rte_mbuf_sanity_check_raw(m, 0);
 291
 292         return m;
 293 }
 294
 295 static void
 296 virtio_dev_vring_start(struct virtqueue *vq, int queue_type)
 297 {
 298         struct rte_mbuf *m;
 299         int i, nbufs, error, size = vq->vq_nentries;
 300         struct vring *vr = &vq->vq_ring;
 301         uint8_t *ring_mem = vq->vq_ring_virt_mem;
 302
 303         PMD_INIT_FUNC_TRACE();
 304
 305         /*
 306          * Reinitialise since virtio port might have been stopped and restarted
 307          */
 308         memset(vq->vq_ring_virt_mem, 0, vq->vq_ring_size);
 309         vring_init(vr, size, ring_mem, VIRTIO_PCI_VRING_ALIGN);
 310         vq->vq_used_cons_idx = 0;
 311         vq->vq_desc_head_idx = 0;
 312         vq->vq_avail_idx = 0;
 313         vq->vq_desc_tail_idx = (uint16_t)(vq->vq_nentries - 1);
 314         vq->vq_free_cnt = vq->vq_nentries;
 315         memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 316
 317         vring_desc_init(vr->desc, size);
 318
 319         /*
 320          * Disable device(host) interrupting guest
 321          */
 322         virtqueue_disable_intr(vq);
 323
 324         /* Only rx virtqueue needs mbufs to be allocated at initialization */
 325         if (queue_type == VTNET_RQ) {
 326                 if (vq->mpool == NULL)
 327                         rte_exit(EXIT_FAILURE,
 328                         "Cannot allocate initial mbufs for rx virtqueue");
 329
 330                 /* Allocate blank mbufs for the each rx descriptor */
 331                 nbufs = 0;
 332                 error = ENOSPC;
 333
 334 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 335                 if (use_simple_rxtx)
 336                         for (i = 0; i < vq->vq_nentries; i++) {
 337                                 vq->vq_ring.avail->ring[i] = i;
 338                                 vq->vq_ring.desc[i].flags = VRING_DESC_F_WRITE;
 339                         }
 340 #endif
 341                 memset(&vq->fake_mbuf, 0, sizeof(vq->fake_mbuf));
 342                 for (i = 0; i < RTE_PMD_VIRTIO_RX_MAX_BURST; i++)
 343                         vq->sw_ring[vq->vq_nentries + i] = &vq->fake_mbuf;
 344
 345                 while (!virtqueue_full(vq)) {
 346                         m = rte_rxmbuf_alloc(vq->mpool);
 347                         if (m == NULL)
 348                                 break;
 349
 350                         /******************************************
 351                         *         Enqueue allocated buffers        *
 352                         *******************************************/
 353 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 354                         if (use_simple_rxtx)
 355                                 error = virtqueue_enqueue_recv_refill_simple(vq, m);
 356                         else
 357 #endif
 358                                 error = virtqueue_enqueue_recv_refill(vq, m);
 359                         if (error) {
 360                                 rte_pktmbuf_free(m);
 361                                 break;
 362                         }
 363                         nbufs++;
 364                 }
 365
 366                 vq_update_avail_idx(vq);
 367
 368                 PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 369         } else if (queue_type == VTNET_TQ) {
 370 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 371                 if (use_simple_rxtx) {
 372                         int mid_idx  = vq->vq_nentries >> 1;
 373                         for (i = 0; i < mid_idx; i++) {
 374                                 vq->vq_ring.avail->ring[i] = i + mid_idx;
 375                                 vq->vq_ring.desc[i + mid_idx].next = i;
 376                                 vq->vq_ring.desc[i + mid_idx].addr =
 377                                         vq->virtio_net_hdr_mem +
 378                                         offsetof(struct virtio_tx_region, tx_hdr);
 379                                 vq->vq_ring.desc[i + mid_idx].len =
 380                                         vq->hw->vtnet_hdr_size;
 381                                 vq->vq_ring.desc[i + mid_idx].flags =
 382                                         VRING_DESC_F_NEXT;
 383                                 vq->vq_ring.desc[i].flags = 0;
 384                         }
 385                         for (i = mid_idx; i < vq->vq_nentries; i++)
 386                                 vq->vq_ring.avail->ring[i] = i;
 387                 }
 388 #endif
 389         }
 390 }
 391
 392 void
 393 virtio_dev_cq_start(struct rte_eth_dev *dev)
 394 {
 395         struct virtio_hw *hw = dev->data->dev_private;
 396
 397         if (hw->cvq) {
 398                 virtio_dev_vring_start(hw->cvq, VTNET_CQ);
 399                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq);
 400         }
 401 }
 402
 403 void
 404 virtio_dev_rxtx_start(struct rte_eth_dev *dev)
 405 {
 406         /*
 407          * Start receive and transmit vrings
 408          * -    Setup vring structure for all queues
 409          * -    Initialize descriptor for the rx vring
 410          * -    Allocate blank mbufs for the each rx descriptor
 411          *
 412          */
 413         int i;
 414
 415         PMD_INIT_FUNC_TRACE();
 416
 417         /* Start rx vring. */
 418         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 419                 virtio_dev_vring_start(dev->data->rx_queues[i], VTNET_RQ);
 420                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->rx_queues[i]);
 421         }
 422
 423         /* Start tx vring. */
 424         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 425                 virtio_dev_vring_start(dev->data->tx_queues[i], VTNET_TQ);
 426                 VIRTQUEUE_DUMP((struct virtqueue *)dev->data->tx_queues[i]);
 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 virtqueue *vq;
 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, &vq);
 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         vq->mpool = mp;
 452
 453         dev->data->rx_queues[queue_idx] = vq;
 454
 455 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 456         virtio_rxq_vec_setup(vq);
 457 #endif
 458
 459         return 0;
 460 }
 461
 462 void
 463 virtio_dev_rx_queue_release(void *rxq)
 464 {
 465         virtio_dev_queue_release(rxq);
 466 }
 467
 468 /*
 469  * struct rte_eth_dev *dev: Used to update dev
 470  * uint16_t nb_desc: Defaults to values read from config space
 471  * unsigned int socket_id: Used to allocate memzone
 472  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 473  * uint16_t queue_idx: Just used as an index in dev txq list
 474  */
 475 int
 476 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 477                         uint16_t queue_idx,
 478                         uint16_t nb_desc,
 479                         unsigned int socket_id,
 480                         const struct rte_eth_txconf *tx_conf)
 481 {
 482         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 483
 484 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 485         struct virtio_hw *hw = dev->data->dev_private;
 486 #endif
 487         struct virtqueue *vq;
 488         uint16_t tx_free_thresh;
 489         int ret;
 490
 491         PMD_INIT_FUNC_TRACE();
 492
 493         if ((tx_conf->txq_flags & ETH_TXQ_FLAGS_NOXSUMS)
 494             != ETH_TXQ_FLAGS_NOXSUMS) {
 495                 PMD_INIT_LOG(ERR, "TX checksum offload not supported\n");
 496                 return -EINVAL;
 497         }
 498
 499 #ifdef RTE_MACHINE_CPUFLAG_SSSE3
 500         /* Use simple rx/tx func if single segment and no offloads */
 501         if ((tx_conf->txq_flags & VIRTIO_SIMPLE_FLAGS) == VIRTIO_SIMPLE_FLAGS &&
 502              !vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF)) {
 503                 PMD_INIT_LOG(INFO, "Using simple rx/tx path");
 504                 dev->tx_pkt_burst = virtio_xmit_pkts_simple;
 505                 dev->rx_pkt_burst = virtio_recv_pkts_vec;
 506                 use_simple_rxtx = 1;
 507         }
 508 #endif
 509
 510         ret = virtio_dev_queue_setup(dev, VTNET_TQ, queue_idx, vtpci_queue_idx,
 511                         nb_desc, socket_id, &vq);
 512         if (ret < 0) {
 513                 PMD_INIT_LOG(ERR, "rvq initialization failed");
 514                 return ret;
 515         }
 516
 517         tx_free_thresh = tx_conf->tx_free_thresh;
 518         if (tx_free_thresh == 0)
 519                 tx_free_thresh =
 520                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 521
 522         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 523                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 524                         "number of TX entries minus 3 (%u)."
 525                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 526                         vq->vq_nentries - 3,
 527                         tx_free_thresh, dev->data->port_id, queue_idx);
 528                 return -EINVAL;
 529         }
 530
 531         vq->vq_free_thresh = tx_free_thresh;
 532
 533         dev->data->tx_queues[queue_idx] = vq;
 534         return 0;
 535 }
 536
 537 void
 538 virtio_dev_tx_queue_release(void *txq)
 539 {
 540         virtio_dev_queue_release(txq);
 541 }
 542
 543 static void
 544 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 545 {
 546         int error;
 547         /*
 548          * Requeue the discarded mbuf. This should always be
 549          * successful since it was just dequeued.
 550          */
 551         error = virtqueue_enqueue_recv_refill(vq, m);
 552         if (unlikely(error)) {
 553                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 554                 rte_pktmbuf_free(m);
 555         }
 556 }
 557
 558 static void
 559 virtio_update_packet_stats(struct virtqueue *vq, struct rte_mbuf *mbuf)
 560 {
 561         uint32_t s = mbuf->pkt_len;
 562         struct ether_addr *ea;
 563
 564         if (s == 64) {
 565                 vq->size_bins[1]++;
 566         } else if (s > 64 && s < 1024) {
 567                 uint32_t bin;
 568
 569                 /* count zeros, and offset into correct bin */
 570                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 571                 vq->size_bins[bin]++;
 572         } else {
 573                 if (s < 64)
 574                         vq->size_bins[0]++;
 575                 else if (s < 1519)
 576                         vq->size_bins[6]++;
 577                 else if (s >= 1519)
 578                         vq->size_bins[7]++;
 579         }
 580
 581         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 582         if (is_multicast_ether_addr(ea)) {
 583                 if (is_broadcast_ether_addr(ea))
 584                         vq->broadcast++;
 585                 else
 586                         vq->multicast++;
 587         }
 588 }
 589
 590 #define VIRTIO_MBUF_BURST_SZ 64
 591 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 592 uint16_t
 593 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 594 {
 595         struct virtqueue *rxvq = rx_queue;
 596         struct virtio_hw *hw;
 597         struct rte_mbuf *rxm, *new_mbuf;
 598         uint16_t nb_used, num, nb_rx;
 599         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 600         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 601         int error;
 602         uint32_t i, nb_enqueued;
 603         uint32_t hdr_size;
 604
 605         nb_used = VIRTQUEUE_NUSED(rxvq);
 606
 607         virtio_rmb();
 608
 609         num = (uint16_t)(likely(nb_used <= nb_pkts) ? nb_used : nb_pkts);
 610         num = (uint16_t)(likely(num <= VIRTIO_MBUF_BURST_SZ) ? num : VIRTIO_MBUF_BURST_SZ);
 611         if (likely(num > DESC_PER_CACHELINE))
 612                 num = num - ((rxvq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 613
 614         num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, num);
 615         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 616
 617         hw = rxvq->hw;
 618         nb_rx = 0;
 619         nb_enqueued = 0;
 620         hdr_size = hw->vtnet_hdr_size;
 621
 622         for (i = 0; i < num ; i++) {
 623                 rxm = rcv_pkts[i];
 624
 625                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 626
 627                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 628                         PMD_RX_LOG(ERR, "Packet drop");
 629                         nb_enqueued++;
 630                         virtio_discard_rxbuf(rxvq, rxm);
 631                         rxvq->errors++;
 632                         continue;
 633                 }
 634
 635                 rxm->port = rxvq->port_id;
 636                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 637                 rxm->ol_flags = 0;
 638                 rxm->vlan_tci = 0;
 639
 640                 rxm->nb_segs = 1;
 641                 rxm->next = NULL;
 642                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 643                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 644
 645                 if (hw->vlan_strip)
 646                         rte_vlan_strip(rxm);
 647
 648                 VIRTIO_DUMP_PACKET(rxm, rxm->data_len);
 649
 650                 rx_pkts[nb_rx++] = rxm;
 651
 652                 rxvq->bytes += rx_pkts[nb_rx - 1]->pkt_len;
 653                 virtio_update_packet_stats(rxvq, rxm);
 654         }
 655
 656         rxvq->packets += nb_rx;
 657
 658         /* Allocate new mbuf for the used descriptor */
 659         error = ENOSPC;
 660         while (likely(!virtqueue_full(rxvq))) {
 661                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
 662                 if (unlikely(new_mbuf == NULL)) {
 663                         struct rte_eth_dev *dev
 664                                 = &rte_eth_devices[rxvq->port_id];
 665                         dev->data->rx_mbuf_alloc_failed++;
 666                         break;
 667                 }
 668                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 669                 if (unlikely(error)) {
 670                         rte_pktmbuf_free(new_mbuf);
 671                         break;
 672                 }
 673                 nb_enqueued++;
 674         }
 675
 676         if (likely(nb_enqueued)) {
 677                 vq_update_avail_idx(rxvq);
 678
 679                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 680                         virtqueue_notify(rxvq);
 681                         PMD_RX_LOG(DEBUG, "Notified");
 682                 }
 683         }
 684
 685         return nb_rx;
 686 }
 687
 688 uint16_t
 689 virtio_recv_mergeable_pkts(void *rx_queue,
 690                         struct rte_mbuf **rx_pkts,
 691                         uint16_t nb_pkts)
 692 {
 693         struct virtqueue *rxvq = rx_queue;
 694         struct virtio_hw *hw;
 695         struct rte_mbuf *rxm, *new_mbuf;
 696         uint16_t nb_used, num, nb_rx;
 697         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 698         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 699         struct rte_mbuf *prev;
 700         int error;
 701         uint32_t i, nb_enqueued;
 702         uint32_t seg_num;
 703         uint16_t extra_idx;
 704         uint32_t seg_res;
 705         uint32_t hdr_size;
 706
 707         nb_used = VIRTQUEUE_NUSED(rxvq);
 708
 709         virtio_rmb();
 710
 711         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
 712
 713         hw = rxvq->hw;
 714         nb_rx = 0;
 715         i = 0;
 716         nb_enqueued = 0;
 717         seg_num = 0;
 718         extra_idx = 0;
 719         seg_res = 0;
 720         hdr_size = hw->vtnet_hdr_size;
 721
 722         while (i < nb_used) {
 723                 struct virtio_net_hdr_mrg_rxbuf *header;
 724
 725                 if (nb_rx == nb_pkts)
 726                         break;
 727
 728                 num = virtqueue_dequeue_burst_rx(rxvq, rcv_pkts, len, 1);
 729                 if (num != 1)
 730                         continue;
 731
 732                 i++;
 733
 734                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
 735                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
 736
 737                 rxm = rcv_pkts[0];
 738
 739                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
 740                         PMD_RX_LOG(ERR, "Packet drop");
 741                         nb_enqueued++;
 742                         virtio_discard_rxbuf(rxvq, rxm);
 743                         rxvq->errors++;
 744                         continue;
 745                 }
 746
 747                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
 748                         RTE_PKTMBUF_HEADROOM - hdr_size);
 749                 seg_num = header->num_buffers;
 750
 751                 if (seg_num == 0)
 752                         seg_num = 1;
 753
 754                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 755                 rxm->nb_segs = seg_num;
 756                 rxm->next = NULL;
 757                 rxm->ol_flags = 0;
 758                 rxm->vlan_tci = 0;
 759                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
 760                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
 761
 762                 rxm->port = rxvq->port_id;
 763                 rx_pkts[nb_rx] = rxm;
 764                 prev = rxm;
 765
 766                 seg_res = seg_num - 1;
 767
 768                 while (seg_res != 0) {
 769                         /*
 770                          * Get extra segments for current uncompleted packet.
 771                          */
 772                         uint16_t  rcv_cnt =
 773                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
 774                         if (likely(VIRTQUEUE_NUSED(rxvq) >= rcv_cnt)) {
 775                                 uint32_t rx_num =
 776                                         virtqueue_dequeue_burst_rx(rxvq,
 777                                         rcv_pkts, len, rcv_cnt);
 778                                 i += rx_num;
 779                                 rcv_cnt = rx_num;
 780                         } else {
 781                                 PMD_RX_LOG(ERR,
 782                                            "No enough segments for packet.");
 783                                 nb_enqueued++;
 784                                 virtio_discard_rxbuf(rxvq, rxm);
 785                                 rxvq->errors++;
 786                                 break;
 787                         }
 788
 789                         extra_idx = 0;
 790
 791                         while (extra_idx < rcv_cnt) {
 792                                 rxm = rcv_pkts[extra_idx];
 793
 794                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
 795                                 rxm->next = NULL;
 796                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
 797                                 rxm->data_len = (uint16_t)(len[extra_idx]);
 798
 799                                 if (prev)
 800                                         prev->next = rxm;
 801
 802                                 prev = rxm;
 803                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
 804                                 extra_idx++;
 805                         };
 806                         seg_res -= rcv_cnt;
 807                 }
 808
 809                 if (hw->vlan_strip)
 810                         rte_vlan_strip(rx_pkts[nb_rx]);
 811
 812                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
 813                         rx_pkts[nb_rx]->data_len);
 814
 815                 rxvq->bytes += rx_pkts[nb_rx]->pkt_len;
 816                 virtio_update_packet_stats(rxvq, rx_pkts[nb_rx]);
 817                 nb_rx++;
 818         }
 819
 820         rxvq->packets += nb_rx;
 821
 822         /* Allocate new mbuf for the used descriptor */
 823         error = ENOSPC;
 824         while (likely(!virtqueue_full(rxvq))) {
 825                 new_mbuf = rte_rxmbuf_alloc(rxvq->mpool);
 826                 if (unlikely(new_mbuf == NULL)) {
 827                         struct rte_eth_dev *dev
 828                                 = &rte_eth_devices[rxvq->port_id];
 829                         dev->data->rx_mbuf_alloc_failed++;
 830                         break;
 831                 }
 832                 error = virtqueue_enqueue_recv_refill(rxvq, new_mbuf);
 833                 if (unlikely(error)) {
 834                         rte_pktmbuf_free(new_mbuf);
 835                         break;
 836                 }
 837                 nb_enqueued++;
 838         }
 839
 840         if (likely(nb_enqueued)) {
 841                 vq_update_avail_idx(rxvq);
 842
 843                 if (unlikely(virtqueue_kick_prepare(rxvq))) {
 844                         virtqueue_notify(rxvq);
 845                         PMD_RX_LOG(DEBUG, "Notified");
 846                 }
 847         }
 848
 849         return nb_rx;
 850 }
 851
 852 uint16_t
 853 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 854 {
 855         struct virtqueue *txvq = tx_queue;
 856         struct virtio_hw *hw = txvq->hw;
 857         uint16_t hdr_size = hw->vtnet_hdr_size;
 858         uint16_t nb_used, nb_tx;
 859         int error;
 860
 861         if (unlikely(nb_pkts < 1))
 862                 return nb_pkts;
 863
 864         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
 865         nb_used = VIRTQUEUE_NUSED(txvq);
 866
 867         virtio_rmb();
 868         if (likely(nb_used > txvq->vq_nentries - txvq->vq_free_thresh))
 869                 virtio_xmit_cleanup(txvq, nb_used);
 870
 871         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 872                 struct rte_mbuf *txm = tx_pkts[nb_tx];
 873                 int can_push = 0, use_indirect = 0, slots, need;
 874
 875                 /* Do VLAN tag insertion */
 876                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
 877                         error = rte_vlan_insert(&txm);
 878                         if (unlikely(error)) {
 879                                 rte_pktmbuf_free(txm);
 880                                 continue;
 881                         }
 882                 }
 883
 884                 /* optimize ring usage */
 885                 if (vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) &&
 886                     rte_mbuf_refcnt_read(txm) == 1 &&
 887                     txm->nb_segs == 1 &&
 888                     rte_pktmbuf_headroom(txm) >= hdr_size &&
 889                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
 890                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
 891                         can_push = 1;
 892                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
 893                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
 894                         use_indirect = 1;
 895
 896                 /* How many main ring entries are needed to this Tx?
 897                  * any_layout => number of segments
 898                  * indirect   => 1
 899                  * default    => number of segments + 1
 900                  */
 901                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
 902                 need = slots - txvq->vq_free_cnt;
 903
 904                 /* Positive value indicates it need free vring descriptors */
 905                 if (unlikely(need > 0)) {
 906                         nb_used = VIRTQUEUE_NUSED(txvq);
 907                         virtio_rmb();
 908                         need = RTE_MIN(need, (int)nb_used);
 909
 910                         virtio_xmit_cleanup(txvq, need);
 911                         need = slots - txvq->vq_free_cnt;
 912                         if (unlikely(need > 0)) {
 913                                 PMD_TX_LOG(ERR,
 914                                            "No free tx descriptors to transmit");
 915                                 break;
 916                         }
 917                 }
 918
 919                 /* Enqueue Packet buffers */
 920                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect, can_push);
 921
 922                 txvq->bytes += txm->pkt_len;
 923                 virtio_update_packet_stats(txvq, txm);
 924         }
 925
 926         txvq->packets += nb_tx;
 927
 928         if (likely(nb_tx)) {
 929                 vq_update_avail_idx(txvq);
 930
 931                 if (unlikely(virtqueue_kick_prepare(txvq))) {
 932                         virtqueue_notify(txvq);
 933                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
 934                 }
 935         }
 936
 937         return nb_tx;
 938 }