lib/librte_vhost/vhost_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 <stdbool.h>
  36 #include <linux/virtio_net.h>
  37
  38 #include <rte_mbuf.h>
  39 #include <rte_memcpy.h>
  40 #include <rte_virtio_net.h>
  41
  42 #include "vhost-net.h"
  43
  44 #define MAX_PKT_BURST 32
  45
  46 static bool
  47 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t qp_nb)
  48 {
  49         return (is_tx ^ (idx & 1)) == 0 && idx < qp_nb * VIRTIO_QNUM;
  50 }
  51
  52 /**
  53  * This function adds buffers to the virtio devices RX virtqueue. Buffers can
  54  * be received from the physical port or from another virtio device. A packet
  55  * count is returned to indicate the number of packets that are succesfully
  56  * added to the RX queue. This function works when the mbuf is scattered, but
  57  * it doesn't support the mergeable feature.
  58  */
  59 static inline uint32_t __attribute__((always_inline))
  60 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
  61         struct rte_mbuf **pkts, uint32_t count)
  62 {
  63         struct vhost_virtqueue *vq;
  64         struct vring_desc *desc;
  65         struct rte_mbuf *buff;
  66         /* The virtio_hdr is initialised to 0. */
  67         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
  68         uint64_t buff_addr = 0;
  69         uint64_t buff_hdr_addr = 0;
  70         uint32_t head[MAX_PKT_BURST];
  71         uint32_t head_idx, packet_success = 0;
  72         uint16_t avail_idx, res_cur_idx;
  73         uint16_t res_base_idx, res_end_idx;
  74         uint16_t free_entries;
  75         uint8_t success = 0;
  76
  77         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_rx()\n", dev->device_fh);
  78         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
  79                 RTE_LOG(ERR, VHOST_DATA,
  80                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
  81                         __func__, dev->device_fh, queue_id);
  82                 return 0;
  83         }
  84
  85         vq = dev->virtqueue[queue_id];
  86         count = (count > MAX_PKT_BURST) ? MAX_PKT_BURST : count;
  87
  88         /*
  89          * As many data cores may want access to available buffers,
  90          * they need to be reserved.
  91          */
  92         do {
  93                 res_base_idx = vq->last_used_idx_res;
  94                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
  95
  96                 free_entries = (avail_idx - res_base_idx);
  97                 /*check that we have enough buffers*/
  98                 if (unlikely(count > free_entries))
  99                         count = free_entries;
 100
 101                 if (count == 0)
 102                         return 0;
 103
 104                 res_end_idx = res_base_idx + count;
 105                 /* vq->last_used_idx_res is atomically updated. */
 106                 /* TODO: Allow to disable cmpset if no concurrency in application. */
 107                 success = rte_atomic16_cmpset(&vq->last_used_idx_res,
 108                                 res_base_idx, res_end_idx);
 109         } while (unlikely(success == 0));
 110         res_cur_idx = res_base_idx;
 111         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Current Index %d| End Index %d\n",
 112                         dev->device_fh, res_cur_idx, res_end_idx);
 113
 114         /* Prefetch available ring to retrieve indexes. */
 115         rte_prefetch0(&vq->avail->ring[res_cur_idx & (vq->size - 1)]);
 116
 117         /* Retrieve all of the head indexes first to avoid caching issues. */
 118         for (head_idx = 0; head_idx < count; head_idx++)
 119                 head[head_idx] = vq->avail->ring[(res_cur_idx + head_idx) &
 120                                         (vq->size - 1)];
 121
 122         /*Prefetch descriptor index. */
 123         rte_prefetch0(&vq->desc[head[packet_success]]);
 124
 125         while (res_cur_idx != res_end_idx) {
 126                 uint32_t offset = 0, vb_offset = 0;
 127                 uint32_t pkt_len, len_to_cpy, data_len, total_copied = 0;
 128                 uint8_t hdr = 0, uncompleted_pkt = 0;
 129
 130                 /* Get descriptor from available ring */
 131                 desc = &vq->desc[head[packet_success]];
 132
 133                 buff = pkts[packet_success];
 134
 135                 /* Convert from gpa to vva (guest physical addr -> vhost virtual addr) */
 136                 buff_addr = gpa_to_vva(dev, desc->addr);
 137                 /* Prefetch buffer address. */
 138                 rte_prefetch0((void *)(uintptr_t)buff_addr);
 139
 140                 /* Copy virtio_hdr to packet and increment buffer address */
 141                 buff_hdr_addr = buff_addr;
 142
 143                 /*
 144                  * If the descriptors are chained the header and data are
 145                  * placed in separate buffers.
 146                  */
 147                 if ((desc->flags & VRING_DESC_F_NEXT) &&
 148                         (desc->len == vq->vhost_hlen)) {
 149                         desc = &vq->desc[desc->next];
 150                         /* Buffer address translation. */
 151                         buff_addr = gpa_to_vva(dev, desc->addr);
 152                 } else {
 153                         vb_offset += vq->vhost_hlen;
 154                         hdr = 1;
 155                 }
 156
 157                 pkt_len = rte_pktmbuf_pkt_len(buff);
 158                 data_len = rte_pktmbuf_data_len(buff);
 159                 len_to_cpy = RTE_MIN(data_len,
 160                         hdr ? desc->len - vq->vhost_hlen : desc->len);
 161                 while (total_copied < pkt_len) {
 162                         /* Copy mbuf data to buffer */
 163                         rte_memcpy((void *)(uintptr_t)(buff_addr + vb_offset),
 164                                 rte_pktmbuf_mtod_offset(buff, const void *, offset),
 165                                 len_to_cpy);
 166                         PRINT_PACKET(dev, (uintptr_t)(buff_addr + vb_offset),
 167                                 len_to_cpy, 0);
 168
 169                         offset += len_to_cpy;
 170                         vb_offset += len_to_cpy;
 171                         total_copied += len_to_cpy;
 172
 173                         /* The whole packet completes */
 174                         if (total_copied == pkt_len)
 175                                 break;
 176
 177                         /* The current segment completes */
 178                         if (offset == data_len) {
 179                                 buff = buff->next;
 180                                 offset = 0;
 181                                 data_len = rte_pktmbuf_data_len(buff);
 182                         }
 183
 184                         /* The current vring descriptor done */
 185                         if (vb_offset == desc->len) {
 186                                 if (desc->flags & VRING_DESC_F_NEXT) {
 187                                         desc = &vq->desc[desc->next];
 188                                         buff_addr = gpa_to_vva(dev, desc->addr);
 189                                         vb_offset = 0;
 190                                 } else {
 191                                         /* Room in vring buffer is not enough */
 192                                         uncompleted_pkt = 1;
 193                                         break;
 194                                 }
 195                         }
 196                         len_to_cpy = RTE_MIN(data_len - offset, desc->len - vb_offset);
 197                 }
 198
 199                 /* Update used ring with desc information */
 200                 vq->used->ring[res_cur_idx & (vq->size - 1)].id =
 201                                                         head[packet_success];
 202
 203                 /* Drop the packet if it is uncompleted */
 204                 if (unlikely(uncompleted_pkt == 1))
 205                         vq->used->ring[res_cur_idx & (vq->size - 1)].len =
 206                                                         vq->vhost_hlen;
 207                 else
 208                         vq->used->ring[res_cur_idx & (vq->size - 1)].len =
 209                                                         pkt_len + vq->vhost_hlen;
 210
 211                 res_cur_idx++;
 212                 packet_success++;
 213
 214                 if (unlikely(uncompleted_pkt == 1))
 215                         continue;
 216
 217                 rte_memcpy((void *)(uintptr_t)buff_hdr_addr,
 218                         (const void *)&virtio_hdr, vq->vhost_hlen);
 219
 220                 PRINT_PACKET(dev, (uintptr_t)buff_hdr_addr, vq->vhost_hlen, 1);
 221
 222                 if (res_cur_idx < res_end_idx) {
 223                         /* Prefetch descriptor index. */
 224                         rte_prefetch0(&vq->desc[head[packet_success]]);
 225                 }
 226         }
 227
 228         rte_compiler_barrier();
 229
 230         /* Wait until it's our turn to add our buffer to the used ring. */
 231         while (unlikely(vq->last_used_idx != res_base_idx))
 232                 rte_pause();
 233
 234         *(volatile uint16_t *)&vq->used->idx += count;
 235         vq->last_used_idx = res_end_idx;
 236
 237         /* flush used->idx update before we read avail->flags. */
 238         rte_mb();
 239
 240         /* Kick the guest if necessary. */
 241         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
 242                 eventfd_write(vq->callfd, (eventfd_t)1);
 243         return count;
 244 }
 245
 246 static inline uint32_t __attribute__((always_inline))
 247 copy_from_mbuf_to_vring(struct virtio_net *dev, uint32_t queue_id,
 248                         uint16_t res_base_idx, uint16_t res_end_idx,
 249                         struct rte_mbuf *pkt)
 250 {
 251         uint32_t vec_idx = 0;
 252         uint32_t entry_success = 0;
 253         struct vhost_virtqueue *vq;
 254         /* The virtio_hdr is initialised to 0. */
 255         struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {
 256                 {0, 0, 0, 0, 0, 0}, 0};
 257         uint16_t cur_idx = res_base_idx;
 258         uint64_t vb_addr = 0;
 259         uint64_t vb_hdr_addr = 0;
 260         uint32_t seg_offset = 0;
 261         uint32_t vb_offset = 0;
 262         uint32_t seg_avail;
 263         uint32_t vb_avail;
 264         uint32_t cpy_len, entry_len;
 265
 266         if (pkt == NULL)
 267                 return 0;
 268
 269         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Current Index %d| "
 270                 "End Index %d\n",
 271                 dev->device_fh, cur_idx, res_end_idx);
 272
 273         /*
 274          * Convert from gpa to vva
 275          * (guest physical addr -> vhost virtual addr)
 276          */
 277         vq = dev->virtqueue[queue_id];
 278         vb_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 279         vb_hdr_addr = vb_addr;
 280
 281         /* Prefetch buffer address. */
 282         rte_prefetch0((void *)(uintptr_t)vb_addr);
 283
 284         virtio_hdr.num_buffers = res_end_idx - res_base_idx;
 285
 286         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") RX: Num merge buffers %d\n",
 287                 dev->device_fh, virtio_hdr.num_buffers);
 288
 289         rte_memcpy((void *)(uintptr_t)vb_hdr_addr,
 290                 (const void *)&virtio_hdr, vq->vhost_hlen);
 291
 292         PRINT_PACKET(dev, (uintptr_t)vb_hdr_addr, vq->vhost_hlen, 1);
 293
 294         seg_avail = rte_pktmbuf_data_len(pkt);
 295         vb_offset = vq->vhost_hlen;
 296         vb_avail = vq->buf_vec[vec_idx].buf_len - vq->vhost_hlen;
 297
 298         entry_len = vq->vhost_hlen;
 299
 300         if (vb_avail == 0) {
 301                 uint32_t desc_idx =
 302                         vq->buf_vec[vec_idx].desc_idx;
 303
 304                 if ((vq->desc[desc_idx].flags
 305                         & VRING_DESC_F_NEXT) == 0) {
 306                         /* Update used ring with desc information */
 307                         vq->used->ring[cur_idx & (vq->size - 1)].id
 308                                 = vq->buf_vec[vec_idx].desc_idx;
 309                         vq->used->ring[cur_idx & (vq->size - 1)].len
 310                                 = entry_len;
 311
 312                         entry_len = 0;
 313                         cur_idx++;
 314                         entry_success++;
 315                 }
 316
 317                 vec_idx++;
 318                 vb_addr = gpa_to_vva(dev, vq->buf_vec[vec_idx].buf_addr);
 319
 320                 /* Prefetch buffer address. */
 321                 rte_prefetch0((void *)(uintptr_t)vb_addr);
 322                 vb_offset = 0;
 323                 vb_avail = vq->buf_vec[vec_idx].buf_len;
 324         }
 325
 326         cpy_len = RTE_MIN(vb_avail, seg_avail);
 327
 328         while (cpy_len > 0) {
 329                 /* Copy mbuf data to vring buffer */
 330                 rte_memcpy((void *)(uintptr_t)(vb_addr + vb_offset),
 331                         rte_pktmbuf_mtod_offset(pkt, const void *, seg_offset),
 332                         cpy_len);
 333
 334                 PRINT_PACKET(dev,
 335                         (uintptr_t)(vb_addr + vb_offset),
 336                         cpy_len, 0);
 337
 338                 seg_offset += cpy_len;
 339                 vb_offset += cpy_len;
 340                 seg_avail -= cpy_len;
 341                 vb_avail -= cpy_len;
 342                 entry_len += cpy_len;
 343
 344                 if (seg_avail != 0) {
 345                         /*
 346                          * The virtio buffer in this vring
 347                          * entry reach to its end.
 348                          * But the segment doesn't complete.
 349                          */
 350                         if ((vq->desc[vq->buf_vec[vec_idx].desc_idx].flags &
 351                                 VRING_DESC_F_NEXT) == 0) {
 352                                 /* Update used ring with desc information */
 353                                 vq->used->ring[cur_idx & (vq->size - 1)].id
 354                                         = vq->buf_vec[vec_idx].desc_idx;
 355                                 vq->used->ring[cur_idx & (vq->size - 1)].len
 356                                         = entry_len;
 357                                 entry_len = 0;
 358                                 cur_idx++;
 359                                 entry_success++;
 360                         }
 361
 362                         vec_idx++;
 363                         vb_addr = gpa_to_vva(dev,
 364                                 vq->buf_vec[vec_idx].buf_addr);
 365                         vb_offset = 0;
 366                         vb_avail = vq->buf_vec[vec_idx].buf_len;
 367                         cpy_len = RTE_MIN(vb_avail, seg_avail);
 368                 } else {
 369                         /*
 370                          * This current segment complete, need continue to
 371                          * check if the whole packet complete or not.
 372                          */
 373                         pkt = pkt->next;
 374                         if (pkt != NULL) {
 375                                 /*
 376                                  * There are more segments.
 377                                  */
 378                                 if (vb_avail == 0) {
 379                                         /*
 380                                          * This current buffer from vring is
 381                                          * used up, need fetch next buffer
 382                                          * from buf_vec.
 383                                          */
 384                                         uint32_t desc_idx =
 385                                                 vq->buf_vec[vec_idx].desc_idx;
 386
 387                                         if ((vq->desc[desc_idx].flags &
 388                                                 VRING_DESC_F_NEXT) == 0) {
 389                                                 uint16_t wrapped_idx =
 390                                                         cur_idx & (vq->size - 1);
 391                                                 /*
 392                                                  * Update used ring with the
 393                                                  * descriptor information
 394                                                  */
 395                                                 vq->used->ring[wrapped_idx].id
 396                                                         = desc_idx;
 397                                                 vq->used->ring[wrapped_idx].len
 398                                                         = entry_len;
 399                                                 entry_success++;
 400                                                 entry_len = 0;
 401                                                 cur_idx++;
 402                                         }
 403
 404                                         /* Get next buffer from buf_vec. */
 405                                         vec_idx++;
 406                                         vb_addr = gpa_to_vva(dev,
 407                                                 vq->buf_vec[vec_idx].buf_addr);
 408                                         vb_avail =
 409                                                 vq->buf_vec[vec_idx].buf_len;
 410                                         vb_offset = 0;
 411                                 }
 412
 413                                 seg_offset = 0;
 414                                 seg_avail = rte_pktmbuf_data_len(pkt);
 415                                 cpy_len = RTE_MIN(vb_avail, seg_avail);
 416                         } else {
 417                                 /*
 418                                  * This whole packet completes.
 419                                  */
 420                                 /* Update used ring with desc information */
 421                                 vq->used->ring[cur_idx & (vq->size - 1)].id
 422                                         = vq->buf_vec[vec_idx].desc_idx;
 423                                 vq->used->ring[cur_idx & (vq->size - 1)].len
 424                                         = entry_len;
 425                                 entry_success++;
 426                                 break;
 427                         }
 428                 }
 429         }
 430
 431         return entry_success;
 432 }
 433
 434 static inline void __attribute__((always_inline))
 435 update_secure_len(struct vhost_virtqueue *vq, uint32_t id,
 436         uint32_t *secure_len, uint32_t *vec_idx)
 437 {
 438         uint16_t wrapped_idx = id & (vq->size - 1);
 439         uint32_t idx = vq->avail->ring[wrapped_idx];
 440         uint8_t next_desc;
 441         uint32_t len = *secure_len;
 442         uint32_t vec_id = *vec_idx;
 443
 444         do {
 445                 next_desc = 0;
 446                 len += vq->desc[idx].len;
 447                 vq->buf_vec[vec_id].buf_addr = vq->desc[idx].addr;
 448                 vq->buf_vec[vec_id].buf_len = vq->desc[idx].len;
 449                 vq->buf_vec[vec_id].desc_idx = idx;
 450                 vec_id++;
 451
 452                 if (vq->desc[idx].flags & VRING_DESC_F_NEXT) {
 453                         idx = vq->desc[idx].next;
 454                         next_desc = 1;
 455                 }
 456         } while (next_desc);
 457
 458         *secure_len = len;
 459         *vec_idx = vec_id;
 460 }
 461
 462 /*
 463  * This function works for mergeable RX.
 464  */
 465 static inline uint32_t __attribute__((always_inline))
 466 virtio_dev_merge_rx(struct virtio_net *dev, uint16_t queue_id,
 467         struct rte_mbuf **pkts, uint32_t count)
 468 {
 469         struct vhost_virtqueue *vq;
 470         uint32_t pkt_idx = 0, entry_success = 0;
 471         uint16_t avail_idx;
 472         uint16_t res_base_idx, res_cur_idx;
 473         uint8_t success = 0;
 474
 475         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") virtio_dev_merge_rx()\n",
 476                 dev->device_fh);
 477         if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->virt_qp_nb))) {
 478                 RTE_LOG(ERR, VHOST_DATA,
 479                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 480                         __func__, dev->device_fh, queue_id);
 481                 return 0;
 482         }
 483
 484         vq = dev->virtqueue[queue_id];
 485         count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
 486
 487         if (count == 0)
 488                 return 0;
 489
 490         for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
 491                 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + vq->vhost_hlen;
 492
 493                 do {
 494                         /*
 495                          * As many data cores may want access to available
 496                          * buffers, they need to be reserved.
 497                          */
 498                         uint32_t secure_len = 0;
 499                         uint32_t vec_idx = 0;
 500
 501                         res_base_idx = vq->last_used_idx_res;
 502                         res_cur_idx = res_base_idx;
 503
 504                         do {
 505                                 avail_idx = *((volatile uint16_t *)&vq->avail->idx);
 506                                 if (unlikely(res_cur_idx == avail_idx)) {
 507                                         LOG_DEBUG(VHOST_DATA,
 508                                                 "(%"PRIu64") Failed "
 509                                                 "to get enough desc from "
 510                                                 "vring\n",
 511                                                 dev->device_fh);
 512                                         goto merge_rx_exit;
 513                                 } else {
 514                                         update_secure_len(vq, res_cur_idx, &secure_len, &vec_idx);
 515                                         res_cur_idx++;
 516                                 }
 517                         } while (pkt_len > secure_len);
 518
 519                         /* vq->last_used_idx_res is atomically updated. */
 520                         success = rte_atomic16_cmpset(&vq->last_used_idx_res,
 521                                                         res_base_idx,
 522                                                         res_cur_idx);
 523                 } while (success == 0);
 524
 525                 entry_success = copy_from_mbuf_to_vring(dev, queue_id,
 526                         res_base_idx, res_cur_idx, pkts[pkt_idx]);
 527
 528                 rte_compiler_barrier();
 529
 530                 /*
 531                  * Wait until it's our turn to add our buffer
 532                  * to the used ring.
 533                  */
 534                 while (unlikely(vq->last_used_idx != res_base_idx))
 535                         rte_pause();
 536
 537                 *(volatile uint16_t *)&vq->used->idx += entry_success;
 538                 vq->last_used_idx = res_cur_idx;
 539         }
 540
 541 merge_rx_exit:
 542         if (likely(pkt_idx)) {
 543                 /* flush used->idx update before we read avail->flags. */
 544                 rte_mb();
 545
 546                 /* Kick the guest if necessary. */
 547                 if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
 548                         eventfd_write(vq->callfd, (eventfd_t)1);
 549         }
 550
 551         return pkt_idx;
 552 }
 553
 554 uint16_t
 555 rte_vhost_enqueue_burst(struct virtio_net *dev, uint16_t queue_id,
 556         struct rte_mbuf **pkts, uint16_t count)
 557 {
 558         if (unlikely(dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF)))
 559                 return virtio_dev_merge_rx(dev, queue_id, pkts, count);
 560         else
 561                 return virtio_dev_rx(dev, queue_id, pkts, count);
 562 }
 563
 564 uint16_t
 565 rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t queue_id,
 566         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 567 {
 568         struct rte_mbuf *m, *prev;
 569         struct vhost_virtqueue *vq;
 570         struct vring_desc *desc;
 571         uint64_t vb_addr = 0;
 572         uint32_t head[MAX_PKT_BURST];
 573         uint32_t used_idx;
 574         uint32_t i;
 575         uint16_t free_entries, entry_success = 0;
 576         uint16_t avail_idx;
 577
 578         if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->virt_qp_nb))) {
 579                 RTE_LOG(ERR, VHOST_DATA,
 580                         "%s (%"PRIu64"): virtqueue idx:%d invalid.\n",
 581                         __func__, dev->device_fh, queue_id);
 582                 return 0;
 583         }
 584
 585         vq = dev->virtqueue[queue_id];
 586         avail_idx =  *((volatile uint16_t *)&vq->avail->idx);
 587
 588         /* If there are no available buffers then return. */
 589         if (vq->last_used_idx == avail_idx)
 590                 return 0;
 591
 592         LOG_DEBUG(VHOST_DATA, "%s (%"PRIu64")\n", __func__,
 593                 dev->device_fh);
 594
 595         /* Prefetch available ring to retrieve head indexes. */
 596         rte_prefetch0(&vq->avail->ring[vq->last_used_idx & (vq->size - 1)]);
 597
 598         /*get the number of free entries in the ring*/
 599         free_entries = (avail_idx - vq->last_used_idx);
 600
 601         free_entries = RTE_MIN(free_entries, count);
 602         /* Limit to MAX_PKT_BURST. */
 603         free_entries = RTE_MIN(free_entries, MAX_PKT_BURST);
 604
 605         LOG_DEBUG(VHOST_DATA, "(%"PRIu64") Buffers available %d\n",
 606                         dev->device_fh, free_entries);
 607         /* Retrieve all of the head indexes first to avoid caching issues. */
 608         for (i = 0; i < free_entries; i++)
 609                 head[i] = vq->avail->ring[(vq->last_used_idx + i) & (vq->size - 1)];
 610
 611         /* Prefetch descriptor index. */
 612         rte_prefetch0(&vq->desc[head[entry_success]]);
 613         rte_prefetch0(&vq->used->ring[vq->last_used_idx & (vq->size - 1)]);
 614
 615         while (entry_success < free_entries) {
 616                 uint32_t vb_avail, vb_offset;
 617                 uint32_t seg_avail, seg_offset;
 618                 uint32_t cpy_len;
 619                 uint32_t seg_num = 0;
 620                 struct rte_mbuf *cur;
 621                 uint8_t alloc_err = 0;
 622
 623                 desc = &vq->desc[head[entry_success]];
 624
 625                 /* Discard first buffer as it is the virtio header */
 626                 if (desc->flags & VRING_DESC_F_NEXT) {
 627                         desc = &vq->desc[desc->next];
 628                         vb_offset = 0;
 629                         vb_avail = desc->len;
 630                 } else {
 631                         vb_offset = vq->vhost_hlen;
 632                         vb_avail = desc->len - vb_offset;
 633                 }
 634
 635                 /* Buffer address translation. */
 636                 vb_addr = gpa_to_vva(dev, desc->addr);
 637                 /* Prefetch buffer address. */
 638                 rte_prefetch0((void *)(uintptr_t)vb_addr);
 639
 640                 used_idx = vq->last_used_idx & (vq->size - 1);
 641
 642                 if (entry_success < (free_entries - 1)) {
 643                         /* Prefetch descriptor index. */
 644                         rte_prefetch0(&vq->desc[head[entry_success+1]]);
 645                         rte_prefetch0(&vq->used->ring[(used_idx + 1) & (vq->size - 1)]);
 646                 }
 647
 648                 /* Update used index buffer information. */
 649                 vq->used->ring[used_idx].id = head[entry_success];
 650                 vq->used->ring[used_idx].len = 0;
 651
 652                 /* Allocate an mbuf and populate the structure. */
 653                 m = rte_pktmbuf_alloc(mbuf_pool);
 654                 if (unlikely(m == NULL)) {
 655                         RTE_LOG(ERR, VHOST_DATA,
 656                                 "Failed to allocate memory for mbuf.\n");
 657                         break;
 658                 }
 659                 seg_offset = 0;
 660                 seg_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
 661                 cpy_len = RTE_MIN(vb_avail, seg_avail);
 662
 663                 PRINT_PACKET(dev, (uintptr_t)vb_addr, desc->len, 0);
 664
 665                 seg_num++;
 666                 cur = m;
 667                 prev = m;
 668                 while (cpy_len != 0) {
 669                         rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *, seg_offset),
 670                                 (void *)((uintptr_t)(vb_addr + vb_offset)),
 671                                 cpy_len);
 672
 673                         seg_offset += cpy_len;
 674                         vb_offset += cpy_len;
 675                         vb_avail -= cpy_len;
 676                         seg_avail -= cpy_len;
 677
 678                         if (vb_avail != 0) {
 679                                 /*
 680                                  * The segment reachs to its end,
 681                                  * while the virtio buffer in TX vring has
 682                                  * more data to be copied.
 683                                  */
 684                                 cur->data_len = seg_offset;
 685                                 m->pkt_len += seg_offset;
 686                                 /* Allocate mbuf and populate the structure. */
 687                                 cur = rte_pktmbuf_alloc(mbuf_pool);
 688                                 if (unlikely(cur == NULL)) {
 689                                         RTE_LOG(ERR, VHOST_DATA, "Failed to "
 690                                                 "allocate memory for mbuf.\n");
 691                                         rte_pktmbuf_free(m);
 692                                         alloc_err = 1;
 693                                         break;
 694                                 }
 695
 696                                 seg_num++;
 697                                 prev->next = cur;
 698                                 prev = cur;
 699                                 seg_offset = 0;
 700                                 seg_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 701                         } else {
 702                                 if (desc->flags & VRING_DESC_F_NEXT) {
 703                                         /*
 704                                          * There are more virtio buffers in
 705                                          * same vring entry need to be copied.
 706                                          */
 707                                         if (seg_avail == 0) {
 708                                                 /*
 709                                                  * The current segment hasn't
 710                                                  * room to accomodate more
 711                                                  * data.
 712                                                  */
 713                                                 cur->data_len = seg_offset;
 714                                                 m->pkt_len += seg_offset;
 715                                                 /*
 716                                                  * Allocate an mbuf and
 717                                                  * populate the structure.
 718                                                  */
 719                                                 cur = rte_pktmbuf_alloc(mbuf_pool);
 720                                                 if (unlikely(cur == NULL)) {
 721                                                         RTE_LOG(ERR,
 722                                                                 VHOST_DATA,
 723                                                                 "Failed to "
 724                                                                 "allocate memory "
 725                                                                 "for mbuf\n");
 726                                                         rte_pktmbuf_free(m);
 727                                                         alloc_err = 1;
 728                                                         break;
 729                                                 }
 730                                                 seg_num++;
 731                                                 prev->next = cur;
 732                                                 prev = cur;
 733                                                 seg_offset = 0;
 734                                                 seg_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 735                                         }
 736
 737                                         desc = &vq->desc[desc->next];
 738
 739                                         /* Buffer address translation. */
 740                                         vb_addr = gpa_to_vva(dev, desc->addr);
 741                                         /* Prefetch buffer address. */
 742                                         rte_prefetch0((void *)(uintptr_t)vb_addr);
 743                                         vb_offset = 0;
 744                                         vb_avail = desc->len;
 745
 746                                         PRINT_PACKET(dev, (uintptr_t)vb_addr,
 747                                                 desc->len, 0);
 748                                 } else {
 749                                         /* The whole packet completes. */
 750                                         cur->data_len = seg_offset;
 751                                         m->pkt_len += seg_offset;
 752                                         vb_avail = 0;
 753                                 }
 754                         }
 755
 756                         cpy_len = RTE_MIN(vb_avail, seg_avail);
 757                 }
 758
 759                 if (unlikely(alloc_err == 1))
 760                         break;
 761
 762                 m->nb_segs = seg_num;
 763
 764                 pkts[entry_success] = m;
 765                 vq->last_used_idx++;
 766                 entry_success++;
 767         }
 768
 769         rte_compiler_barrier();
 770         vq->used->idx += entry_success;
 771         /* Kick guest if required. */
 772         if (!(vq->avail->flags & VRING_AVAIL_F_NO_INTERRUPT))
 773                 eventfd_write(vq->callfd, (eventfd_t)1);
 774         return entry_success;
 775 }