drivers/net/virtio/virtio_rxtx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <stdlib.h>
   8 #include <string.h>
   9 #include <errno.h>
  10
  11 #include <rte_cycles.h>
  12 #include <rte_memory.h>
  13 #include <rte_branch_prediction.h>
  14 #include <rte_mempool.h>
  15 #include <rte_malloc.h>
  16 #include <rte_mbuf.h>
  17 #include <rte_ether.h>
  18 #include <rte_ethdev_driver.h>
  19 #include <rte_prefetch.h>
  20 #include <rte_string_fns.h>
  21 #include <rte_errno.h>
  22 #include <rte_byteorder.h>
  23 #include <rte_net.h>
  24 #include <rte_ip.h>
  25 #include <rte_udp.h>
  26 #include <rte_tcp.h>
  27
  28 #include "virtio_logs.h"
  29 #include "virtio_ethdev.h"
  30 #include "virtio_pci.h"
  31 #include "virtqueue.h"
  32 #include "virtio_rxtx.h"
  33 #include "virtio_rxtx_simple.h"
  34
  35 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
  36 #define VIRTIO_DUMP_PACKET(m, len) rte_pktmbuf_dump(stdout, m, len)
  37 #else
  38 #define  VIRTIO_DUMP_PACKET(m, len) do { } while (0)
  39 #endif
  40
  41 int
  42 virtio_dev_rx_queue_done(void *rxq, uint16_t offset)
  43 {
  44         struct virtnet_rx *rxvq = rxq;
  45         struct virtqueue *vq = rxvq->vq;
  46
  47         return VIRTQUEUE_NUSED(vq) >= offset;
  48 }
  49
  50 void
  51 vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx, uint16_t num)
  52 {
  53         vq->vq_free_cnt += num;
  54         vq->vq_desc_tail_idx = desc_idx & (vq->vq_nentries - 1);
  55 }
  56
  57 void
  58 vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx)
  59 {
  60         struct vring_desc *dp, *dp_tail;
  61         struct vq_desc_extra *dxp;
  62         uint16_t desc_idx_last = desc_idx;
  63
  64         dp  = &vq->vq_ring.desc[desc_idx];
  65         dxp = &vq->vq_descx[desc_idx];
  66         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt + dxp->ndescs);
  67         if ((dp->flags & VRING_DESC_F_INDIRECT) == 0) {
  68                 while (dp->flags & VRING_DESC_F_NEXT) {
  69                         desc_idx_last = dp->next;
  70                         dp = &vq->vq_ring.desc[dp->next];
  71                 }
  72         }
  73         dxp->ndescs = 0;
  74
  75         /*
  76          * We must append the existing free chain, if any, to the end of
  77          * newly freed chain. If the virtqueue was completely used, then
  78          * head would be VQ_RING_DESC_CHAIN_END (ASSERTed above).
  79          */
  80         if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END) {
  81                 vq->vq_desc_head_idx = desc_idx;
  82         } else {
  83                 dp_tail = &vq->vq_ring.desc[vq->vq_desc_tail_idx];
  84                 dp_tail->next = desc_idx;
  85         }
  86
  87         vq->vq_desc_tail_idx = desc_idx_last;
  88         dp->next = VQ_RING_DESC_CHAIN_END;
  89 }
  90
  91 static uint16_t
  92 virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
  93                            uint32_t *len, uint16_t num)
  94 {
  95         struct vring_used_elem *uep;
  96         struct rte_mbuf *cookie;
  97         uint16_t used_idx, desc_idx;
  98         uint16_t i;
  99
 100         /*  Caller does the check */
 101         for (i = 0; i < num ; i++) {
 102                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 103                 uep = &vq->vq_ring.used->ring[used_idx];
 104                 desc_idx = (uint16_t) uep->id;
 105                 len[i] = uep->len;
 106                 cookie = (struct rte_mbuf *)vq->vq_descx[desc_idx].cookie;
 107
 108                 if (unlikely(cookie == NULL)) {
 109                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 110                                 vq->vq_used_cons_idx);
 111                         break;
 112                 }
 113
 114                 rte_prefetch0(cookie);
 115                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 116                 rx_pkts[i]  = cookie;
 117                 vq->vq_used_cons_idx++;
 118                 vq_ring_free_chain(vq, desc_idx);
 119                 vq->vq_descx[desc_idx].cookie = NULL;
 120         }
 121
 122         return i;
 123 }
 124
 125 static uint16_t
 126 virtqueue_dequeue_rx_inorder(struct virtqueue *vq,
 127                         struct rte_mbuf **rx_pkts,
 128                         uint32_t *len,
 129                         uint16_t num)
 130 {
 131         struct vring_used_elem *uep;
 132         struct rte_mbuf *cookie;
 133         uint16_t used_idx = 0;
 134         uint16_t i;
 135
 136         if (unlikely(num == 0))
 137                 return 0;
 138
 139         for (i = 0; i < num; i++) {
 140                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 141                 /* Desc idx same as used idx */
 142                 uep = &vq->vq_ring.used->ring[used_idx];
 143                 len[i] = uep->len;
 144                 cookie = (struct rte_mbuf *)vq->vq_descx[used_idx].cookie;
 145
 146                 if (unlikely(cookie == NULL)) {
 147                         PMD_DRV_LOG(ERR, "vring descriptor with no mbuf cookie at %u",
 148                                 vq->vq_used_cons_idx);
 149                         break;
 150                 }
 151
 152                 rte_prefetch0(cookie);
 153                 rte_packet_prefetch(rte_pktmbuf_mtod(cookie, void *));
 154                 rx_pkts[i]  = cookie;
 155                 vq->vq_used_cons_idx++;
 156                 vq->vq_descx[used_idx].cookie = NULL;
 157         }
 158
 159         vq_ring_free_inorder(vq, used_idx, i);
 160         return i;
 161 }
 162
 163 #ifndef DEFAULT_TX_FREE_THRESH
 164 #define DEFAULT_TX_FREE_THRESH 32
 165 #endif
 166
 167 /* Cleanup from completed transmits. */
 168 static void
 169 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
 170 {
 171         uint16_t i, used_idx, desc_idx;
 172         for (i = 0; i < num; i++) {
 173                 struct vring_used_elem *uep;
 174                 struct vq_desc_extra *dxp;
 175
 176                 used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
 177                 uep = &vq->vq_ring.used->ring[used_idx];
 178
 179                 desc_idx = (uint16_t) uep->id;
 180                 dxp = &vq->vq_descx[desc_idx];
 181                 vq->vq_used_cons_idx++;
 182                 vq_ring_free_chain(vq, desc_idx);
 183
 184                 if (dxp->cookie != NULL) {
 185                         rte_pktmbuf_free(dxp->cookie);
 186                         dxp->cookie = NULL;
 187                 }
 188         }
 189 }
 190
 191 /* Cleanup from completed inorder transmits. */
 192 static void
 193 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
 194 {
 195         uint16_t i, used_idx, desc_idx = 0, last_idx;
 196         int16_t free_cnt = 0;
 197         struct vq_desc_extra *dxp = NULL;
 198
 199         if (unlikely(num == 0))
 200                 return;
 201
 202         for (i = 0; i < num; i++) {
 203                 struct vring_used_elem *uep;
 204
 205                 used_idx = vq->vq_used_cons_idx & (vq->vq_nentries - 1);
 206                 uep = &vq->vq_ring.used->ring[used_idx];
 207                 desc_idx = (uint16_t)uep->id;
 208
 209                 dxp = &vq->vq_descx[desc_idx];
 210                 vq->vq_used_cons_idx++;
 211
 212                 if (dxp->cookie != NULL) {
 213                         rte_pktmbuf_free(dxp->cookie);
 214                         dxp->cookie = NULL;
 215                 }
 216         }
 217
 218         last_idx = desc_idx + dxp->ndescs - 1;
 219         free_cnt = last_idx - vq->vq_desc_tail_idx;
 220         if (free_cnt <= 0)
 221                 free_cnt += vq->vq_nentries;
 222
 223         vq_ring_free_inorder(vq, last_idx, free_cnt);
 224 }
 225
 226 static inline int
 227 virtqueue_enqueue_refill_inorder(struct virtqueue *vq,
 228                         struct rte_mbuf **cookies,
 229                         uint16_t num)
 230 {
 231         struct vq_desc_extra *dxp;
 232         struct virtio_hw *hw = vq->hw;
 233         struct vring_desc *start_dp;
 234         uint16_t head_idx, idx, i = 0;
 235
 236         if (unlikely(vq->vq_free_cnt == 0))
 237                 return -ENOSPC;
 238         if (unlikely(vq->vq_free_cnt < num))
 239                 return -EMSGSIZE;
 240
 241         head_idx = vq->vq_desc_head_idx & (vq->vq_nentries - 1);
 242         start_dp = vq->vq_ring.desc;
 243
 244         while (i < num) {
 245                 idx = head_idx & (vq->vq_nentries - 1);
 246                 dxp = &vq->vq_descx[idx];
 247                 dxp->cookie = (void *)cookies[i];
 248                 dxp->ndescs = 1;
 249
 250                 start_dp[idx].addr =
 251                                 VIRTIO_MBUF_ADDR(cookies[i], vq) +
 252                                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 253                 start_dp[idx].len =
 254                                 cookies[i]->buf_len -
 255                                 RTE_PKTMBUF_HEADROOM +
 256                                 hw->vtnet_hdr_size;
 257                 start_dp[idx].flags =  VRING_DESC_F_WRITE;
 258
 259                 vq_update_avail_ring(vq, idx);
 260                 head_idx++;
 261                 i++;
 262         }
 263
 264         vq->vq_desc_head_idx += num;
 265         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 266         return 0;
 267 }
 268
 269 static inline int
 270 virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 271 {
 272         struct vq_desc_extra *dxp;
 273         struct virtio_hw *hw = vq->hw;
 274         struct vring_desc *start_dp;
 275         uint16_t needed = 1;
 276         uint16_t head_idx, idx;
 277
 278         if (unlikely(vq->vq_free_cnt == 0))
 279                 return -ENOSPC;
 280         if (unlikely(vq->vq_free_cnt < needed))
 281                 return -EMSGSIZE;
 282
 283         head_idx = vq->vq_desc_head_idx;
 284         if (unlikely(head_idx >= vq->vq_nentries))
 285                 return -EFAULT;
 286
 287         idx = head_idx;
 288         dxp = &vq->vq_descx[idx];
 289         dxp->cookie = (void *)cookie;
 290         dxp->ndescs = needed;
 291
 292         start_dp = vq->vq_ring.desc;
 293         start_dp[idx].addr =
 294                 VIRTIO_MBUF_ADDR(cookie, vq) +
 295                 RTE_PKTMBUF_HEADROOM - hw->vtnet_hdr_size;
 296         start_dp[idx].len =
 297                 cookie->buf_len - RTE_PKTMBUF_HEADROOM + hw->vtnet_hdr_size;
 298         start_dp[idx].flags =  VRING_DESC_F_WRITE;
 299         idx = start_dp[idx].next;
 300         vq->vq_desc_head_idx = idx;
 301         if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 302                 vq->vq_desc_tail_idx = idx;
 303         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 304         vq_update_avail_ring(vq, head_idx);
 305
 306         return 0;
 307 }
 308
 309 /* When doing TSO, the IP length is not included in the pseudo header
 310  * checksum of the packet given to the PMD, but for virtio it is
 311  * expected.
 312  */
 313 static void
 314 virtio_tso_fix_cksum(struct rte_mbuf *m)
 315 {
 316         /* common case: header is not fragmented */
 317         if (likely(rte_pktmbuf_data_len(m) >= m->l2_len + m->l3_len +
 318                         m->l4_len)) {
 319                 struct ipv4_hdr *iph;
 320                 struct ipv6_hdr *ip6h;
 321                 struct tcp_hdr *th;
 322                 uint16_t prev_cksum, new_cksum, ip_len, ip_paylen;
 323                 uint32_t tmp;
 324
 325                 iph = rte_pktmbuf_mtod_offset(m, struct ipv4_hdr *, m->l2_len);
 326                 th = RTE_PTR_ADD(iph, m->l3_len);
 327                 if ((iph->version_ihl >> 4) == 4) {
 328                         iph->hdr_checksum = 0;
 329                         iph->hdr_checksum = rte_ipv4_cksum(iph);
 330                         ip_len = iph->total_length;
 331                         ip_paylen = rte_cpu_to_be_16(rte_be_to_cpu_16(ip_len) -
 332                                 m->l3_len);
 333                 } else {
 334                         ip6h = (struct ipv6_hdr *)iph;
 335                         ip_paylen = ip6h->payload_len;
 336                 }
 337
 338                 /* calculate the new phdr checksum not including ip_paylen */
 339                 prev_cksum = th->cksum;
 340                 tmp = prev_cksum;
 341                 tmp += ip_paylen;
 342                 tmp = (tmp & 0xffff) + (tmp >> 16);
 343                 new_cksum = tmp;
 344
 345                 /* replace it in the packet */
 346                 th->cksum = new_cksum;
 347         }
 348 }
 349
 350 static inline int
 351 tx_offload_enabled(struct virtio_hw *hw)
 352 {
 353         return vtpci_with_feature(hw, VIRTIO_NET_F_CSUM) ||
 354                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO4) ||
 355                 vtpci_with_feature(hw, VIRTIO_NET_F_HOST_TSO6);
 356 }
 357
 358 /* avoid write operation when necessary, to lessen cache issues */
 359 #define ASSIGN_UNLESS_EQUAL(var, val) do {      \
 360         if ((var) != (val))                     \
 361                 (var) = (val);                  \
 362 } while (0)
 363
 364 static inline void
 365 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
 366                         struct rte_mbuf *cookie,
 367                         int offload)
 368 {
 369         if (offload) {
 370                 if (cookie->ol_flags & PKT_TX_TCP_SEG)
 371                         cookie->ol_flags |= PKT_TX_TCP_CKSUM;
 372
 373                 switch (cookie->ol_flags & PKT_TX_L4_MASK) {
 374                 case PKT_TX_UDP_CKSUM:
 375                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 376                         hdr->csum_offset = offsetof(struct udp_hdr,
 377                                 dgram_cksum);
 378                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 379                         break;
 380
 381                 case PKT_TX_TCP_CKSUM:
 382                         hdr->csum_start = cookie->l2_len + cookie->l3_len;
 383                         hdr->csum_offset = offsetof(struct tcp_hdr, cksum);
 384                         hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
 385                         break;
 386
 387                 default:
 388                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 389                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 390                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 391                         break;
 392                 }
 393
 394                 /* TCP Segmentation Offload */
 395                 if (cookie->ol_flags & PKT_TX_TCP_SEG) {
 396                         virtio_tso_fix_cksum(cookie);
 397                         hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
 398                                 VIRTIO_NET_HDR_GSO_TCPV6 :
 399                                 VIRTIO_NET_HDR_GSO_TCPV4;
 400                         hdr->gso_size = cookie->tso_segsz;
 401                         hdr->hdr_len =
 402                                 cookie->l2_len +
 403                                 cookie->l3_len +
 404                                 cookie->l4_len;
 405                 } else {
 406                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 407                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 408                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 409                 }
 410         }
 411 }
 412
 413 static inline void
 414 virtqueue_enqueue_xmit_inorder(struct virtnet_tx *txvq,
 415                         struct rte_mbuf **cookies,
 416                         uint16_t num)
 417 {
 418         struct vq_desc_extra *dxp;
 419         struct virtqueue *vq = txvq->vq;
 420         struct vring_desc *start_dp;
 421         struct virtio_net_hdr *hdr;
 422         uint16_t idx;
 423         uint16_t head_size = vq->hw->vtnet_hdr_size;
 424         int offload;
 425         uint16_t i = 0;
 426
 427         idx = vq->vq_desc_head_idx;
 428         start_dp = vq->vq_ring.desc;
 429
 430         offload = tx_offload_enabled(vq->hw);
 431
 432         while (i < num) {
 433                 idx = idx & (vq->vq_nentries - 1);
 434                 dxp = &vq->vq_descx[idx];
 435                 dxp->cookie = (void *)cookies[i];
 436                 dxp->ndescs = 1;
 437
 438                 hdr = (struct virtio_net_hdr *)
 439                         rte_pktmbuf_prepend(cookies[i], head_size);
 440                 cookies[i]->pkt_len -= head_size;
 441
 442                 /* if offload disabled, it is not zeroed below, do it now */
 443                 if (offload == 0) {
 444                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 445                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 446                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 447                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 448                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 449                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 450                 }
 451
 452                 virtqueue_xmit_offload(hdr, cookies[i], offload);
 453
 454                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookies[i], vq);
 455                 start_dp[idx].len   = cookies[i]->data_len;
 456                 start_dp[idx].flags = 0;
 457
 458                 vq_update_avail_ring(vq, idx);
 459
 460                 idx++;
 461                 i++;
 462         };
 463
 464         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - num);
 465         vq->vq_desc_head_idx = idx & (vq->vq_nentries - 1);
 466 }
 467
 468 static inline void
 469 virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
 470                         uint16_t needed, int use_indirect, int can_push,
 471                         int in_order)
 472 {
 473         struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
 474         struct vq_desc_extra *dxp;
 475         struct virtqueue *vq = txvq->vq;
 476         struct vring_desc *start_dp;
 477         uint16_t seg_num = cookie->nb_segs;
 478         uint16_t head_idx, idx;
 479         uint16_t head_size = vq->hw->vtnet_hdr_size;
 480         struct virtio_net_hdr *hdr;
 481         int offload;
 482
 483         offload = tx_offload_enabled(vq->hw);
 484
 485         head_idx = vq->vq_desc_head_idx;
 486         idx = head_idx;
 487         dxp = &vq->vq_descx[idx];
 488         dxp->cookie = (void *)cookie;
 489         dxp->ndescs = needed;
 490
 491         start_dp = vq->vq_ring.desc;
 492
 493         if (can_push) {
 494                 /* prepend cannot fail, checked by caller */
 495                 hdr = (struct virtio_net_hdr *)
 496                         rte_pktmbuf_prepend(cookie, head_size);
 497                 /* rte_pktmbuf_prepend() counts the hdr size to the pkt length,
 498                  * which is wrong. Below subtract restores correct pkt size.
 499                  */
 500                 cookie->pkt_len -= head_size;
 501
 502                 /* if offload disabled, it is not zeroed below, do it now */
 503                 if (offload == 0) {
 504                         ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
 505                         ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
 506                         ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
 507                         ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
 508                         ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
 509                         ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
 510                 }
 511         } else if (use_indirect) {
 512                 /* setup tx ring slot to point to indirect
 513                  * descriptor list stored in reserved region.
 514                  *
 515                  * the first slot in indirect ring is already preset
 516                  * to point to the header in reserved region
 517                  */
 518                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 519                         RTE_PTR_DIFF(&txr[idx].tx_indir, txr);
 520                 start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
 521                 start_dp[idx].flags = VRING_DESC_F_INDIRECT;
 522                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 523
 524                 /* loop below will fill in rest of the indirect elements */
 525                 start_dp = txr[idx].tx_indir;
 526                 idx = 1;
 527         } else {
 528                 /* setup first tx ring slot to point to header
 529                  * stored in reserved region.
 530                  */
 531                 start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
 532                         RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
 533                 start_dp[idx].len   = vq->hw->vtnet_hdr_size;
 534                 start_dp[idx].flags = VRING_DESC_F_NEXT;
 535                 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
 536
 537                 idx = start_dp[idx].next;
 538         }
 539
 540         virtqueue_xmit_offload(hdr, cookie, offload);
 541
 542         do {
 543                 start_dp[idx].addr  = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
 544                 start_dp[idx].len   = cookie->data_len;
 545                 start_dp[idx].flags = cookie->next ? VRING_DESC_F_NEXT : 0;
 546                 idx = start_dp[idx].next;
 547         } while ((cookie = cookie->next) != NULL);
 548
 549         if (use_indirect)
 550                 idx = vq->vq_ring.desc[head_idx].next;
 551
 552         vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
 553
 554         vq->vq_desc_head_idx = idx;
 555         vq_update_avail_ring(vq, head_idx);
 556
 557         if (!in_order) {
 558                 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 559                         vq->vq_desc_tail_idx = idx;
 560         }
 561 }
 562
 563 void
 564 virtio_dev_cq_start(struct rte_eth_dev *dev)
 565 {
 566         struct virtio_hw *hw = dev->data->dev_private;
 567
 568         if (hw->cvq && hw->cvq->vq) {
 569                 rte_spinlock_init(&hw->cvq->lock);
 570                 VIRTQUEUE_DUMP((struct virtqueue *)hw->cvq->vq);
 571         }
 572 }
 573
 574 int
 575 virtio_dev_rx_queue_setup(struct rte_eth_dev *dev,
 576                         uint16_t queue_idx,
 577                         uint16_t nb_desc,
 578                         unsigned int socket_id __rte_unused,
 579                         const struct rte_eth_rxconf *rx_conf __rte_unused,
 580                         struct rte_mempool *mp)
 581 {
 582         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 583         struct virtio_hw *hw = dev->data->dev_private;
 584         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 585         struct virtnet_rx *rxvq;
 586
 587         PMD_INIT_FUNC_TRACE();
 588
 589         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 590                 nb_desc = vq->vq_nentries;
 591         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 592
 593         rxvq = &vq->rxq;
 594         rxvq->queue_id = queue_idx;
 595         rxvq->mpool = mp;
 596         if (rxvq->mpool == NULL) {
 597                 rte_exit(EXIT_FAILURE,
 598                         "Cannot allocate mbufs for rx virtqueue");
 599         }
 600
 601         dev->data->rx_queues[queue_idx] = rxvq;
 602
 603         return 0;
 604 }
 605
 606 int
 607 virtio_dev_rx_queue_setup_finish(struct rte_eth_dev *dev, uint16_t queue_idx)
 608 {
 609         uint16_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_RQ_QUEUE_IDX;
 610         struct virtio_hw *hw = dev->data->dev_private;
 611         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 612         struct virtnet_rx *rxvq = &vq->rxq;
 613         struct rte_mbuf *m;
 614         uint16_t desc_idx;
 615         int error, nbufs, i;
 616
 617         PMD_INIT_FUNC_TRACE();
 618
 619         /* Allocate blank mbufs for the each rx descriptor */
 620         nbufs = 0;
 621
 622         if (hw->use_simple_rx) {
 623                 for (desc_idx = 0; desc_idx < vq->vq_nentries;
 624                      desc_idx++) {
 625                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 626                         vq->vq_ring.desc[desc_idx].flags =
 627                                 VRING_DESC_F_WRITE;
 628                 }
 629
 630                 virtio_rxq_vec_setup(rxvq);
 631         }
 632
 633         memset(&rxvq->fake_mbuf, 0, sizeof(rxvq->fake_mbuf));
 634         for (desc_idx = 0; desc_idx < RTE_PMD_VIRTIO_RX_MAX_BURST;
 635              desc_idx++) {
 636                 vq->sw_ring[vq->vq_nentries + desc_idx] =
 637                         &rxvq->fake_mbuf;
 638         }
 639
 640         if (hw->use_simple_rx) {
 641                 while (vq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 642                         virtio_rxq_rearm_vec(rxvq);
 643                         nbufs += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 644                 }
 645         } else if (hw->use_inorder_rx) {
 646                 if ((!virtqueue_full(vq))) {
 647                         uint16_t free_cnt = vq->vq_free_cnt;
 648                         struct rte_mbuf *pkts[free_cnt];
 649
 650                         if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, pkts,
 651                                 free_cnt)) {
 652                                 error = virtqueue_enqueue_refill_inorder(vq,
 653                                                 pkts,
 654                                                 free_cnt);
 655                                 if (unlikely(error)) {
 656                                         for (i = 0; i < free_cnt; i++)
 657                                                 rte_pktmbuf_free(pkts[i]);
 658                                 }
 659                         }
 660
 661                         nbufs += free_cnt;
 662                         vq_update_avail_idx(vq);
 663                 }
 664         } else {
 665                 while (!virtqueue_full(vq)) {
 666                         m = rte_mbuf_raw_alloc(rxvq->mpool);
 667                         if (m == NULL)
 668                                 break;
 669
 670                         /* Enqueue allocated buffers */
 671                         error = virtqueue_enqueue_recv_refill(vq, m);
 672                         if (error) {
 673                                 rte_pktmbuf_free(m);
 674                                 break;
 675                         }
 676                         nbufs++;
 677                 }
 678
 679                 vq_update_avail_idx(vq);
 680         }
 681
 682         PMD_INIT_LOG(DEBUG, "Allocated %d bufs", nbufs);
 683
 684         VIRTQUEUE_DUMP(vq);
 685
 686         return 0;
 687 }
 688
 689 /*
 690  * struct rte_eth_dev *dev: Used to update dev
 691  * uint16_t nb_desc: Defaults to values read from config space
 692  * unsigned int socket_id: Used to allocate memzone
 693  * const struct rte_eth_txconf *tx_conf: Used to setup tx engine
 694  * uint16_t queue_idx: Just used as an index in dev txq list
 695  */
 696 int
 697 virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
 698                         uint16_t queue_idx,
 699                         uint16_t nb_desc,
 700                         unsigned int socket_id __rte_unused,
 701                         const struct rte_eth_txconf *tx_conf)
 702 {
 703         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 704         struct virtio_hw *hw = dev->data->dev_private;
 705         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 706         struct virtnet_tx *txvq;
 707         uint16_t tx_free_thresh;
 708
 709         PMD_INIT_FUNC_TRACE();
 710
 711         /* cannot use simple rxtx funcs with multisegs or offloads */
 712         if (dev->data->dev_conf.txmode.offloads)
 713                 hw->use_simple_tx = 0;
 714
 715         if (nb_desc == 0 || nb_desc > vq->vq_nentries)
 716                 nb_desc = vq->vq_nentries;
 717         vq->vq_free_cnt = RTE_MIN(vq->vq_free_cnt, nb_desc);
 718
 719         txvq = &vq->txq;
 720         txvq->queue_id = queue_idx;
 721
 722         tx_free_thresh = tx_conf->tx_free_thresh;
 723         if (tx_free_thresh == 0)
 724                 tx_free_thresh =
 725                         RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
 726
 727         if (tx_free_thresh >= (vq->vq_nentries - 3)) {
 728                 RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
 729                         "number of TX entries minus 3 (%u)."
 730                         " (tx_free_thresh=%u port=%u queue=%u)\n",
 731                         vq->vq_nentries - 3,
 732                         tx_free_thresh, dev->data->port_id, queue_idx);
 733                 return -EINVAL;
 734         }
 735
 736         vq->vq_free_thresh = tx_free_thresh;
 737
 738         dev->data->tx_queues[queue_idx] = txvq;
 739         return 0;
 740 }
 741
 742 int
 743 virtio_dev_tx_queue_setup_finish(struct rte_eth_dev *dev,
 744                                 uint16_t queue_idx)
 745 {
 746         uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
 747         struct virtio_hw *hw = dev->data->dev_private;
 748         struct virtqueue *vq = hw->vqs[vtpci_queue_idx];
 749         uint16_t mid_idx = vq->vq_nentries >> 1;
 750         struct virtnet_tx *txvq = &vq->txq;
 751         uint16_t desc_idx;
 752
 753         PMD_INIT_FUNC_TRACE();
 754
 755         if (hw->use_simple_tx) {
 756                 for (desc_idx = 0; desc_idx < mid_idx; desc_idx++) {
 757                         vq->vq_ring.avail->ring[desc_idx] =
 758                                 desc_idx + mid_idx;
 759                         vq->vq_ring.desc[desc_idx + mid_idx].next =
 760                                 desc_idx;
 761                         vq->vq_ring.desc[desc_idx + mid_idx].addr =
 762                                 txvq->virtio_net_hdr_mem +
 763                                 offsetof(struct virtio_tx_region, tx_hdr);
 764                         vq->vq_ring.desc[desc_idx + mid_idx].len =
 765                                 vq->hw->vtnet_hdr_size;
 766                         vq->vq_ring.desc[desc_idx + mid_idx].flags =
 767                                 VRING_DESC_F_NEXT;
 768                         vq->vq_ring.desc[desc_idx].flags = 0;
 769                 }
 770                 for (desc_idx = mid_idx; desc_idx < vq->vq_nentries;
 771                      desc_idx++)
 772                         vq->vq_ring.avail->ring[desc_idx] = desc_idx;
 773         } else if (hw->use_inorder_tx) {
 774                 vq->vq_ring.desc[vq->vq_nentries - 1].next = 0;
 775         }
 776
 777         VIRTQUEUE_DUMP(vq);
 778
 779         return 0;
 780 }
 781
 782 static void
 783 virtio_discard_rxbuf(struct virtqueue *vq, struct rte_mbuf *m)
 784 {
 785         int error;
 786         /*
 787          * Requeue the discarded mbuf. This should always be
 788          * successful since it was just dequeued.
 789          */
 790         error = virtqueue_enqueue_recv_refill(vq, m);
 791
 792         if (unlikely(error)) {
 793                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 794                 rte_pktmbuf_free(m);
 795         }
 796 }
 797
 798 static void
 799 virtio_discard_rxbuf_inorder(struct virtqueue *vq, struct rte_mbuf *m)
 800 {
 801         int error;
 802
 803         error = virtqueue_enqueue_refill_inorder(vq, &m, 1);
 804         if (unlikely(error)) {
 805                 RTE_LOG(ERR, PMD, "cannot requeue discarded mbuf");
 806                 rte_pktmbuf_free(m);
 807         }
 808 }
 809
 810 static void
 811 virtio_update_packet_stats(struct virtnet_stats *stats, struct rte_mbuf *mbuf)
 812 {
 813         uint32_t s = mbuf->pkt_len;
 814         struct ether_addr *ea;
 815
 816         if (s == 64) {
 817                 stats->size_bins[1]++;
 818         } else if (s > 64 && s < 1024) {
 819                 uint32_t bin;
 820
 821                 /* count zeros, and offset into correct bin */
 822                 bin = (sizeof(s) * 8) - __builtin_clz(s) - 5;
 823                 stats->size_bins[bin]++;
 824         } else {
 825                 if (s < 64)
 826                         stats->size_bins[0]++;
 827                 else if (s < 1519)
 828                         stats->size_bins[6]++;
 829                 else if (s >= 1519)
 830                         stats->size_bins[7]++;
 831         }
 832
 833         ea = rte_pktmbuf_mtod(mbuf, struct ether_addr *);
 834         if (is_multicast_ether_addr(ea)) {
 835                 if (is_broadcast_ether_addr(ea))
 836                         stats->broadcast++;
 837                 else
 838                         stats->multicast++;
 839         }
 840 }
 841
 842 static inline void
 843 virtio_rx_stats_updated(struct virtnet_rx *rxvq, struct rte_mbuf *m)
 844 {
 845         VIRTIO_DUMP_PACKET(m, m->data_len);
 846
 847         rxvq->stats.bytes += m->pkt_len;
 848         virtio_update_packet_stats(&rxvq->stats, m);
 849 }
 850
 851 /* Optionally fill offload information in structure */
 852 static int
 853 virtio_rx_offload(struct rte_mbuf *m, struct virtio_net_hdr *hdr)
 854 {
 855         struct rte_net_hdr_lens hdr_lens;
 856         uint32_t hdrlen, ptype;
 857         int l4_supported = 0;
 858
 859         /* nothing to do */
 860         if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
 861                 return 0;
 862
 863         m->ol_flags |= PKT_RX_IP_CKSUM_UNKNOWN;
 864
 865         ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
 866         m->packet_type = ptype;
 867         if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
 868             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
 869             (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
 870                 l4_supported = 1;
 871
 872         if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
 873                 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
 874                 if (hdr->csum_start <= hdrlen && l4_supported) {
 875                         m->ol_flags |= PKT_RX_L4_CKSUM_NONE;
 876                 } else {
 877                         /* Unknown proto or tunnel, do sw cksum. We can assume
 878                          * the cksum field is in the first segment since the
 879                          * buffers we provided to the host are large enough.
 880                          * In case of SCTP, this will be wrong since it's a CRC
 881                          * but there's nothing we can do.
 882                          */
 883                         uint16_t csum = 0, off;
 884
 885                         rte_raw_cksum_mbuf(m, hdr->csum_start,
 886                                 rte_pktmbuf_pkt_len(m) - hdr->csum_start,
 887                                 &csum);
 888                         if (likely(csum != 0xffff))
 889                                 csum = ~csum;
 890                         off = hdr->csum_offset + hdr->csum_start;
 891                         if (rte_pktmbuf_data_len(m) >= off + 1)
 892                                 *rte_pktmbuf_mtod_offset(m, uint16_t *,
 893                                         off) = csum;
 894                 }
 895         } else if (hdr->flags & VIRTIO_NET_HDR_F_DATA_VALID && l4_supported) {
 896                 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
 897         }
 898
 899         /* GSO request, save required information in mbuf */
 900         if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
 901                 /* Check unsupported modes */
 902                 if ((hdr->gso_type & VIRTIO_NET_HDR_GSO_ECN) ||
 903                     (hdr->gso_size == 0)) {
 904                         return -EINVAL;
 905                 }
 906
 907                 /* Update mss lengthes in mbuf */
 908                 m->tso_segsz = hdr->gso_size;
 909                 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
 910                         case VIRTIO_NET_HDR_GSO_TCPV4:
 911                         case VIRTIO_NET_HDR_GSO_TCPV6:
 912                                 m->ol_flags |= PKT_RX_LRO | \
 913                                         PKT_RX_L4_CKSUM_NONE;
 914                                 break;
 915                         default:
 916                                 return -EINVAL;
 917                 }
 918         }
 919
 920         return 0;
 921 }
 922
 923 static inline int
 924 rx_offload_enabled(struct virtio_hw *hw)
 925 {
 926         return vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_CSUM) ||
 927                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO4) ||
 928                 vtpci_with_feature(hw, VIRTIO_NET_F_GUEST_TSO6);
 929 }
 930
 931 #define VIRTIO_MBUF_BURST_SZ 64
 932 #define DESC_PER_CACHELINE (RTE_CACHE_LINE_SIZE / sizeof(struct vring_desc))
 933 uint16_t
 934 virtio_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 935 {
 936         struct virtnet_rx *rxvq = rx_queue;
 937         struct virtqueue *vq = rxvq->vq;
 938         struct virtio_hw *hw = vq->hw;
 939         struct rte_mbuf *rxm, *new_mbuf;
 940         uint16_t nb_used, num, nb_rx;
 941         uint32_t len[VIRTIO_MBUF_BURST_SZ];
 942         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
 943         int error;
 944         uint32_t i, nb_enqueued;
 945         uint32_t hdr_size;
 946         int offload;
 947         struct virtio_net_hdr *hdr;
 948
 949         nb_rx = 0;
 950         if (unlikely(hw->started == 0))
 951                 return nb_rx;
 952
 953         nb_used = VIRTQUEUE_NUSED(vq);
 954
 955         virtio_rmb();
 956
 957         num = likely(nb_used <= nb_pkts) ? nb_used : nb_pkts;
 958         if (unlikely(num > VIRTIO_MBUF_BURST_SZ))
 959                 num = VIRTIO_MBUF_BURST_SZ;
 960         if (likely(num > DESC_PER_CACHELINE))
 961                 num = num - ((vq->vq_used_cons_idx + num) % DESC_PER_CACHELINE);
 962
 963         num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, num);
 964         PMD_RX_LOG(DEBUG, "used:%d dequeue:%d", nb_used, num);
 965
 966         nb_enqueued = 0;
 967         hdr_size = hw->vtnet_hdr_size;
 968         offload = rx_offload_enabled(hw);
 969
 970         for (i = 0; i < num ; i++) {
 971                 rxm = rcv_pkts[i];
 972
 973                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
 974
 975                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
 976                         PMD_RX_LOG(ERR, "Packet drop");
 977                         nb_enqueued++;
 978                         virtio_discard_rxbuf(vq, rxm);
 979                         rxvq->stats.errors++;
 980                         continue;
 981                 }
 982
 983                 rxm->port = rxvq->port_id;
 984                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
 985                 rxm->ol_flags = 0;
 986                 rxm->vlan_tci = 0;
 987
 988                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
 989                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
 990
 991                 hdr = (struct virtio_net_hdr *)((char *)rxm->buf_addr +
 992                         RTE_PKTMBUF_HEADROOM - hdr_size);
 993
 994                 if (hw->vlan_strip)
 995                         rte_vlan_strip(rxm);
 996
 997                 if (offload && virtio_rx_offload(rxm, hdr) < 0) {
 998                         virtio_discard_rxbuf(vq, rxm);
 999                         rxvq->stats.errors++;
1000                         continue;
1001                 }
1002
1003                 virtio_rx_stats_updated(rxvq, rxm);
1004
1005                 rx_pkts[nb_rx++] = rxm;
1006         }
1007
1008         rxvq->stats.packets += nb_rx;
1009
1010         /* Allocate new mbuf for the used descriptor */
1011         error = ENOSPC;
1012         while (likely(!virtqueue_full(vq))) {
1013                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1014                 if (unlikely(new_mbuf == NULL)) {
1015                         struct rte_eth_dev *dev
1016                                 = &rte_eth_devices[rxvq->port_id];
1017                         dev->data->rx_mbuf_alloc_failed++;
1018                         break;
1019                 }
1020                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
1021                 if (unlikely(error)) {
1022                         rte_pktmbuf_free(new_mbuf);
1023                         break;
1024                 }
1025                 nb_enqueued++;
1026         }
1027
1028         if (likely(nb_enqueued)) {
1029                 vq_update_avail_idx(vq);
1030
1031                 if (unlikely(virtqueue_kick_prepare(vq))) {
1032                         virtqueue_notify(vq);
1033                         PMD_RX_LOG(DEBUG, "Notified");
1034                 }
1035         }
1036
1037         return nb_rx;
1038 }
1039
1040 uint16_t
1041 virtio_recv_mergeable_pkts_inorder(void *rx_queue,
1042                         struct rte_mbuf **rx_pkts,
1043                         uint16_t nb_pkts)
1044 {
1045         struct virtnet_rx *rxvq = rx_queue;
1046         struct virtqueue *vq = rxvq->vq;
1047         struct virtio_hw *hw = vq->hw;
1048         struct rte_mbuf *rxm;
1049         struct rte_mbuf *prev;
1050         uint16_t nb_used, num, nb_rx;
1051         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1052         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1053         int error;
1054         uint32_t nb_enqueued;
1055         uint32_t seg_num;
1056         uint32_t seg_res;
1057         uint32_t hdr_size;
1058         int32_t i;
1059         int offload;
1060
1061         nb_rx = 0;
1062         if (unlikely(hw->started == 0))
1063                 return nb_rx;
1064
1065         nb_used = VIRTQUEUE_NUSED(vq);
1066         nb_used = RTE_MIN(nb_used, nb_pkts);
1067         nb_used = RTE_MIN(nb_used, VIRTIO_MBUF_BURST_SZ);
1068
1069         virtio_rmb();
1070
1071         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1072
1073         nb_enqueued = 0;
1074         seg_num = 1;
1075         seg_res = 0;
1076         hdr_size = hw->vtnet_hdr_size;
1077         offload = rx_offload_enabled(hw);
1078
1079         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len, nb_used);
1080
1081         for (i = 0; i < num; i++) {
1082                 struct virtio_net_hdr_mrg_rxbuf *header;
1083
1084                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1085                 PMD_RX_LOG(DEBUG, "packet len:%d", len[i]);
1086
1087                 rxm = rcv_pkts[i];
1088
1089                 if (unlikely(len[i] < hdr_size + ETHER_HDR_LEN)) {
1090                         PMD_RX_LOG(ERR, "Packet drop");
1091                         nb_enqueued++;
1092                         virtio_discard_rxbuf_inorder(vq, rxm);
1093                         rxvq->stats.errors++;
1094                         continue;
1095                 }
1096
1097                 header = (struct virtio_net_hdr_mrg_rxbuf *)
1098                          ((char *)rxm->buf_addr + RTE_PKTMBUF_HEADROOM
1099                          - hdr_size);
1100                 seg_num = header->num_buffers;
1101
1102                 if (seg_num == 0)
1103                         seg_num = 1;
1104
1105                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1106                 rxm->nb_segs = seg_num;
1107                 rxm->ol_flags = 0;
1108                 rxm->vlan_tci = 0;
1109                 rxm->pkt_len = (uint32_t)(len[i] - hdr_size);
1110                 rxm->data_len = (uint16_t)(len[i] - hdr_size);
1111
1112                 rxm->port = rxvq->port_id;
1113
1114                 rx_pkts[nb_rx] = rxm;
1115                 prev = rxm;
1116
1117                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
1118                         virtio_discard_rxbuf_inorder(vq, rxm);
1119                         rxvq->stats.errors++;
1120                         continue;
1121                 }
1122
1123                 if (hw->vlan_strip)
1124                         rte_vlan_strip(rx_pkts[nb_rx]);
1125
1126                 seg_res = seg_num - 1;
1127
1128                 /* Merge remaining segments */
1129                 while (seg_res != 0 && i < (num - 1)) {
1130                         i++;
1131
1132                         rxm = rcv_pkts[i];
1133                         rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1134                         rxm->pkt_len = (uint32_t)(len[i]);
1135                         rxm->data_len = (uint16_t)(len[i]);
1136
1137                         rx_pkts[nb_rx]->pkt_len += (uint32_t)(len[i]);
1138                         rx_pkts[nb_rx]->data_len += (uint16_t)(len[i]);
1139
1140                         if (prev)
1141                                 prev->next = rxm;
1142
1143                         prev = rxm;
1144                         seg_res -= 1;
1145                 }
1146
1147                 if (!seg_res) {
1148                         virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1149                         nb_rx++;
1150                 }
1151         }
1152
1153         /* Last packet still need merge segments */
1154         while (seg_res != 0) {
1155                 uint16_t rcv_cnt = RTE_MIN((uint16_t)seg_res,
1156                                         VIRTIO_MBUF_BURST_SZ);
1157
1158                 prev = rcv_pkts[nb_rx];
1159                 if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1160                         num = virtqueue_dequeue_rx_inorder(vq, rcv_pkts, len,
1161                                                            rcv_cnt);
1162                         uint16_t extra_idx = 0;
1163
1164                         rcv_cnt = num;
1165                         while (extra_idx < rcv_cnt) {
1166                                 rxm = rcv_pkts[extra_idx];
1167                                 rxm->data_off =
1168                                         RTE_PKTMBUF_HEADROOM - hdr_size;
1169                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1170                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1171                                 prev->next = rxm;
1172                                 prev = rxm;
1173                                 rx_pkts[nb_rx]->pkt_len += len[extra_idx];
1174                                 rx_pkts[nb_rx]->data_len += len[extra_idx];
1175                                 extra_idx += 1;
1176                         };
1177                         seg_res -= rcv_cnt;
1178
1179                         if (!seg_res) {
1180                                 virtio_rx_stats_updated(rxvq, rx_pkts[nb_rx]);
1181                                 nb_rx++;
1182                         }
1183                 } else {
1184                         PMD_RX_LOG(ERR,
1185                                         "No enough segments for packet.");
1186                         virtio_discard_rxbuf_inorder(vq, prev);
1187                         rxvq->stats.errors++;
1188                         break;
1189                 }
1190         }
1191
1192         rxvq->stats.packets += nb_rx;
1193
1194         /* Allocate new mbuf for the used descriptor */
1195
1196         if (likely(!virtqueue_full(vq))) {
1197                 /* free_cnt may include mrg descs */
1198                 uint16_t free_cnt = vq->vq_free_cnt;
1199                 struct rte_mbuf *new_pkts[free_cnt];
1200
1201                 if (!rte_pktmbuf_alloc_bulk(rxvq->mpool, new_pkts, free_cnt)) {
1202                         error = virtqueue_enqueue_refill_inorder(vq, new_pkts,
1203                                         free_cnt);
1204                         if (unlikely(error)) {
1205                                 for (i = 0; i < free_cnt; i++)
1206                                         rte_pktmbuf_free(new_pkts[i]);
1207                         }
1208                         nb_enqueued += free_cnt;
1209                 } else {
1210                         struct rte_eth_dev *dev =
1211                                 &rte_eth_devices[rxvq->port_id];
1212                         dev->data->rx_mbuf_alloc_failed += free_cnt;
1213                 }
1214         }
1215
1216         if (likely(nb_enqueued)) {
1217                 vq_update_avail_idx(vq);
1218
1219                 if (unlikely(virtqueue_kick_prepare(vq))) {
1220                         virtqueue_notify(vq);
1221                         PMD_RX_LOG(DEBUG, "Notified");
1222                 }
1223         }
1224
1225         return nb_rx;
1226 }
1227
1228 uint16_t
1229 virtio_recv_mergeable_pkts(void *rx_queue,
1230                         struct rte_mbuf **rx_pkts,
1231                         uint16_t nb_pkts)
1232 {
1233         struct virtnet_rx *rxvq = rx_queue;
1234         struct virtqueue *vq = rxvq->vq;
1235         struct virtio_hw *hw = vq->hw;
1236         struct rte_mbuf *rxm, *new_mbuf;
1237         uint16_t nb_used, num, nb_rx;
1238         uint32_t len[VIRTIO_MBUF_BURST_SZ];
1239         struct rte_mbuf *rcv_pkts[VIRTIO_MBUF_BURST_SZ];
1240         struct rte_mbuf *prev;
1241         int error;
1242         uint32_t i, nb_enqueued;
1243         uint32_t seg_num;
1244         uint16_t extra_idx;
1245         uint32_t seg_res;
1246         uint32_t hdr_size;
1247         int offload;
1248
1249         nb_rx = 0;
1250         if (unlikely(hw->started == 0))
1251                 return nb_rx;
1252
1253         nb_used = VIRTQUEUE_NUSED(vq);
1254
1255         virtio_rmb();
1256
1257         PMD_RX_LOG(DEBUG, "used:%d", nb_used);
1258
1259         i = 0;
1260         nb_enqueued = 0;
1261         seg_num = 0;
1262         extra_idx = 0;
1263         seg_res = 0;
1264         hdr_size = hw->vtnet_hdr_size;
1265         offload = rx_offload_enabled(hw);
1266
1267         while (i < nb_used) {
1268                 struct virtio_net_hdr_mrg_rxbuf *header;
1269
1270                 if (nb_rx == nb_pkts)
1271                         break;
1272
1273                 num = virtqueue_dequeue_burst_rx(vq, rcv_pkts, len, 1);
1274                 if (num != 1)
1275                         continue;
1276
1277                 i++;
1278
1279                 PMD_RX_LOG(DEBUG, "dequeue:%d", num);
1280                 PMD_RX_LOG(DEBUG, "packet len:%d", len[0]);
1281
1282                 rxm = rcv_pkts[0];
1283
1284                 if (unlikely(len[0] < hdr_size + ETHER_HDR_LEN)) {
1285                         PMD_RX_LOG(ERR, "Packet drop");
1286                         nb_enqueued++;
1287                         virtio_discard_rxbuf(vq, rxm);
1288                         rxvq->stats.errors++;
1289                         continue;
1290                 }
1291
1292                 header = (struct virtio_net_hdr_mrg_rxbuf *)((char *)rxm->buf_addr +
1293                         RTE_PKTMBUF_HEADROOM - hdr_size);
1294                 seg_num = header->num_buffers;
1295
1296                 if (seg_num == 0)
1297                         seg_num = 1;
1298
1299                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1300                 rxm->nb_segs = seg_num;
1301                 rxm->ol_flags = 0;
1302                 rxm->vlan_tci = 0;
1303                 rxm->pkt_len = (uint32_t)(len[0] - hdr_size);
1304                 rxm->data_len = (uint16_t)(len[0] - hdr_size);
1305
1306                 rxm->port = rxvq->port_id;
1307                 rx_pkts[nb_rx] = rxm;
1308                 prev = rxm;
1309
1310                 if (offload && virtio_rx_offload(rxm, &header->hdr) < 0) {
1311                         virtio_discard_rxbuf(vq, rxm);
1312                         rxvq->stats.errors++;
1313                         continue;
1314                 }
1315
1316                 seg_res = seg_num - 1;
1317
1318                 while (seg_res != 0) {
1319                         /*
1320                          * Get extra segments for current uncompleted packet.
1321                          */
1322                         uint16_t  rcv_cnt =
1323                                 RTE_MIN(seg_res, RTE_DIM(rcv_pkts));
1324                         if (likely(VIRTQUEUE_NUSED(vq) >= rcv_cnt)) {
1325                                 uint32_t rx_num =
1326                                         virtqueue_dequeue_burst_rx(vq,
1327                                         rcv_pkts, len, rcv_cnt);
1328                                 i += rx_num;
1329                                 rcv_cnt = rx_num;
1330                         } else {
1331                                 PMD_RX_LOG(ERR,
1332                                            "No enough segments for packet.");
1333                                 nb_enqueued++;
1334                                 virtio_discard_rxbuf(vq, rxm);
1335                                 rxvq->stats.errors++;
1336                                 break;
1337                         }
1338
1339                         extra_idx = 0;
1340
1341                         while (extra_idx < rcv_cnt) {
1342                                 rxm = rcv_pkts[extra_idx];
1343
1344                                 rxm->data_off = RTE_PKTMBUF_HEADROOM - hdr_size;
1345                                 rxm->pkt_len = (uint32_t)(len[extra_idx]);
1346                                 rxm->data_len = (uint16_t)(len[extra_idx]);
1347
1348                                 if (prev)
1349                                         prev->next = rxm;
1350
1351                                 prev = rxm;
1352                                 rx_pkts[nb_rx]->pkt_len += rxm->pkt_len;
1353                                 extra_idx++;
1354                         };
1355                         seg_res -= rcv_cnt;
1356                 }
1357
1358                 if (hw->vlan_strip)
1359                         rte_vlan_strip(rx_pkts[nb_rx]);
1360
1361                 VIRTIO_DUMP_PACKET(rx_pkts[nb_rx],
1362                         rx_pkts[nb_rx]->data_len);
1363
1364                 rxvq->stats.bytes += rx_pkts[nb_rx]->pkt_len;
1365                 virtio_update_packet_stats(&rxvq->stats, rx_pkts[nb_rx]);
1366                 nb_rx++;
1367         }
1368
1369         rxvq->stats.packets += nb_rx;
1370
1371         /* Allocate new mbuf for the used descriptor */
1372         error = ENOSPC;
1373         while (likely(!virtqueue_full(vq))) {
1374                 new_mbuf = rte_mbuf_raw_alloc(rxvq->mpool);
1375                 if (unlikely(new_mbuf == NULL)) {
1376                         struct rte_eth_dev *dev
1377                                 = &rte_eth_devices[rxvq->port_id];
1378                         dev->data->rx_mbuf_alloc_failed++;
1379                         break;
1380                 }
1381                 error = virtqueue_enqueue_recv_refill(vq, new_mbuf);
1382                 if (unlikely(error)) {
1383                         rte_pktmbuf_free(new_mbuf);
1384                         break;
1385                 }
1386                 nb_enqueued++;
1387         }
1388
1389         if (likely(nb_enqueued)) {
1390                 vq_update_avail_idx(vq);
1391
1392                 if (unlikely(virtqueue_kick_prepare(vq))) {
1393                         virtqueue_notify(vq);
1394                         PMD_RX_LOG(DEBUG, "Notified");
1395                 }
1396         }
1397
1398         return nb_rx;
1399 }
1400
1401 uint16_t
1402 virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1403 {
1404         struct virtnet_tx *txvq = tx_queue;
1405         struct virtqueue *vq = txvq->vq;
1406         struct virtio_hw *hw = vq->hw;
1407         uint16_t hdr_size = hw->vtnet_hdr_size;
1408         uint16_t nb_used, nb_tx = 0;
1409         int error;
1410
1411         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1412                 return nb_tx;
1413
1414         if (unlikely(nb_pkts < 1))
1415                 return nb_pkts;
1416
1417         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1418         nb_used = VIRTQUEUE_NUSED(vq);
1419
1420         virtio_rmb();
1421         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1422                 virtio_xmit_cleanup(vq, nb_used);
1423
1424         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1425                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1426                 int can_push = 0, use_indirect = 0, slots, need;
1427
1428                 /* Do VLAN tag insertion */
1429                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1430                         error = rte_vlan_insert(&txm);
1431                         if (unlikely(error)) {
1432                                 rte_pktmbuf_free(txm);
1433                                 continue;
1434                         }
1435                 }
1436
1437                 /* optimize ring usage */
1438                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1439                       vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1440                     rte_mbuf_refcnt_read(txm) == 1 &&
1441                     RTE_MBUF_DIRECT(txm) &&
1442                     txm->nb_segs == 1 &&
1443                     rte_pktmbuf_headroom(txm) >= hdr_size &&
1444                     rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1445                                    __alignof__(struct virtio_net_hdr_mrg_rxbuf)))
1446                         can_push = 1;
1447                 else if (vtpci_with_feature(hw, VIRTIO_RING_F_INDIRECT_DESC) &&
1448                          txm->nb_segs < VIRTIO_MAX_TX_INDIRECT)
1449                         use_indirect = 1;
1450
1451                 /* How many main ring entries are needed to this Tx?
1452                  * any_layout => number of segments
1453                  * indirect   => 1
1454                  * default    => number of segments + 1
1455                  */
1456                 slots = use_indirect ? 1 : (txm->nb_segs + !can_push);
1457                 need = slots - vq->vq_free_cnt;
1458
1459                 /* Positive value indicates it need free vring descriptors */
1460                 if (unlikely(need > 0)) {
1461                         nb_used = VIRTQUEUE_NUSED(vq);
1462                         virtio_rmb();
1463                         need = RTE_MIN(need, (int)nb_used);
1464
1465                         virtio_xmit_cleanup(vq, need);
1466                         need = slots - vq->vq_free_cnt;
1467                         if (unlikely(need > 0)) {
1468                                 PMD_TX_LOG(ERR,
1469                                            "No free tx descriptors to transmit");
1470                                 break;
1471                         }
1472                 }
1473
1474                 /* Enqueue Packet buffers */
1475                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
1476                         can_push, 0);
1477
1478                 txvq->stats.bytes += txm->pkt_len;
1479                 virtio_update_packet_stats(&txvq->stats, txm);
1480         }
1481
1482         txvq->stats.packets += nb_tx;
1483
1484         if (likely(nb_tx)) {
1485                 vq_update_avail_idx(vq);
1486
1487                 if (unlikely(virtqueue_kick_prepare(vq))) {
1488                         virtqueue_notify(vq);
1489                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1490                 }
1491         }
1492
1493         return nb_tx;
1494 }
1495
1496 uint16_t
1497 virtio_xmit_pkts_inorder(void *tx_queue,
1498                         struct rte_mbuf **tx_pkts,
1499                         uint16_t nb_pkts)
1500 {
1501         struct virtnet_tx *txvq = tx_queue;
1502         struct virtqueue *vq = txvq->vq;
1503         struct virtio_hw *hw = vq->hw;
1504         uint16_t hdr_size = hw->vtnet_hdr_size;
1505         uint16_t nb_used, nb_avail, nb_tx = 0, nb_inorder_pkts = 0;
1506         struct rte_mbuf *inorder_pkts[nb_pkts];
1507         int error;
1508
1509         if (unlikely(hw->started == 0 && tx_pkts != hw->inject_pkts))
1510                 return nb_tx;
1511
1512         if (unlikely(nb_pkts < 1))
1513                 return nb_pkts;
1514
1515         VIRTQUEUE_DUMP(vq);
1516         PMD_TX_LOG(DEBUG, "%d packets to xmit", nb_pkts);
1517         nb_used = VIRTQUEUE_NUSED(vq);
1518
1519         virtio_rmb();
1520         if (likely(nb_used > vq->vq_nentries - vq->vq_free_thresh))
1521                 virtio_xmit_cleanup_inorder(vq, nb_used);
1522
1523         if (unlikely(!vq->vq_free_cnt))
1524                 virtio_xmit_cleanup_inorder(vq, nb_used);
1525
1526         nb_avail = RTE_MIN(vq->vq_free_cnt, nb_pkts);
1527
1528         for (nb_tx = 0; nb_tx < nb_avail; nb_tx++) {
1529                 struct rte_mbuf *txm = tx_pkts[nb_tx];
1530                 int slots, need;
1531
1532                 /* Do VLAN tag insertion */
1533                 if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
1534                         error = rte_vlan_insert(&txm);
1535                         if (unlikely(error)) {
1536                                 rte_pktmbuf_free(txm);
1537                                 continue;
1538                         }
1539                 }
1540
1541                 /* optimize ring usage */
1542                 if ((vtpci_with_feature(hw, VIRTIO_F_ANY_LAYOUT) ||
1543                      vtpci_with_feature(hw, VIRTIO_F_VERSION_1)) &&
1544                      rte_mbuf_refcnt_read(txm) == 1 &&
1545                      RTE_MBUF_DIRECT(txm) &&
1546                      txm->nb_segs == 1 &&
1547                      rte_pktmbuf_headroom(txm) >= hdr_size &&
1548                      rte_is_aligned(rte_pktmbuf_mtod(txm, char *),
1549                                 __alignof__(struct virtio_net_hdr_mrg_rxbuf))) {
1550                         inorder_pkts[nb_inorder_pkts] = txm;
1551                         nb_inorder_pkts++;
1552
1553                         txvq->stats.bytes += txm->pkt_len;
1554                         virtio_update_packet_stats(&txvq->stats, txm);
1555                         continue;
1556                 }
1557
1558                 if (nb_inorder_pkts) {
1559                         virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1560                                                         nb_inorder_pkts);
1561                         nb_inorder_pkts = 0;
1562                 }
1563
1564                 slots = txm->nb_segs + 1;
1565                 need = slots - vq->vq_free_cnt;
1566                 if (unlikely(need > 0)) {
1567                         nb_used = VIRTQUEUE_NUSED(vq);
1568                         virtio_rmb();
1569                         need = RTE_MIN(need, (int)nb_used);
1570
1571                         virtio_xmit_cleanup_inorder(vq, need);
1572
1573                         need = slots - vq->vq_free_cnt;
1574
1575                         if (unlikely(need > 0)) {
1576                                 PMD_TX_LOG(ERR,
1577                                         "No free tx descriptors to transmit");
1578                                 break;
1579                         }
1580                 }
1581                 /* Enqueue Packet buffers */
1582                 virtqueue_enqueue_xmit(txvq, txm, slots, 0, 0, 1);
1583
1584                 txvq->stats.bytes += txm->pkt_len;
1585                 virtio_update_packet_stats(&txvq->stats, txm);
1586         }
1587
1588         /* Transmit all inorder packets */
1589         if (nb_inorder_pkts)
1590                 virtqueue_enqueue_xmit_inorder(txvq, inorder_pkts,
1591                                                 nb_inorder_pkts);
1592
1593         txvq->stats.packets += nb_tx;
1594
1595         if (likely(nb_tx)) {
1596                 vq_update_avail_idx(vq);
1597
1598                 if (unlikely(virtqueue_kick_prepare(vq))) {
1599                         virtqueue_notify(vq);
1600                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
1601                 }
1602         }
1603
1604         VIRTQUEUE_DUMP(vq);
1605
1606         return nb_tx;
1607 }