examples/vhost/virtio_net.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2017 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <linux/virtio_net.h>
   8
   9 #include <rte_mbuf.h>
  10 #include <rte_memcpy.h>
  11 #include <rte_vhost.h>
  12
  13 #include "main.h"
  14
  15 /*
  16  * A very simple vhost-user net driver implementation, without
  17  * any extra features being enabled, such as TSO and mrg-Rx.
  18  */
  19
  20 void
  21 vs_vhost_net_setup(struct vhost_dev *dev)
  22 {
  23         uint16_t i;
  24         int vid = dev->vid;
  25         struct vhost_queue *queue;
  26         int ret;
  27
  28         RTE_LOG(INFO, VHOST_CONFIG,
  29                 "setting builtin vhost-user net driver\n");
  30
  31         rte_vhost_get_negotiated_features(vid, &dev->features);
  32         if (dev->features & (1 << VIRTIO_NET_F_MRG_RXBUF))
  33                 dev->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
  34         else
  35                 dev->hdr_len = sizeof(struct virtio_net_hdr);
  36
  37         ret = rte_vhost_get_mem_table(vid, &dev->mem);
  38         if (ret < 0) {
  39                 RTE_LOG(ERR, VHOST_CONFIG, "Failed to get "
  40                         "VM memory layout for device(%d)\n", vid);
  41                 return;
  42         }
  43
  44         dev->nr_vrings = rte_vhost_get_vring_num(vid);
  45         for (i = 0; i < dev->nr_vrings; i++) {
  46                 queue = &dev->queues[i];
  47
  48                 queue->last_used_idx  = 0;
  49                 queue->last_avail_idx = 0;
  50                 rte_vhost_get_vhost_vring(vid, i, &queue->vr);
  51         }
  52 }
  53
  54 void
  55 vs_vhost_net_remove(struct vhost_dev *dev)
  56 {
  57         free(dev->mem);
  58 }
  59
  60 static __rte_always_inline int
  61 enqueue_pkt(struct vhost_dev *dev, struct rte_vhost_vring *vr,
  62             struct rte_mbuf *m, uint16_t desc_idx)
  63 {
  64         uint32_t desc_avail, desc_offset;
  65         uint64_t desc_chunck_len;
  66         uint32_t mbuf_avail, mbuf_offset;
  67         uint32_t cpy_len;
  68         struct vring_desc *desc;
  69         uint64_t desc_addr, desc_gaddr;
  70         struct virtio_net_hdr virtio_hdr = {0, 0, 0, 0, 0, 0};
  71         /* A counter to avoid desc dead loop chain */
  72         uint16_t nr_desc = 1;
  73
  74         desc = &vr->desc[desc_idx];
  75         desc_chunck_len = desc->len;
  76         desc_gaddr = desc->addr;
  77         desc_addr = rte_vhost_va_from_guest_pa(
  78                         dev->mem, desc_gaddr, &desc_chunck_len);
  79         /*
  80          * Checking of 'desc_addr' placed outside of 'unlikely' macro to avoid
  81          * performance issue with some versions of gcc (4.8.4 and 5.3.0) which
  82          * otherwise stores offset on the stack instead of in a register.
  83          */
  84         if (unlikely(desc->len < dev->hdr_len) || !desc_addr)
  85                 return -1;
  86
  87         rte_prefetch0((void *)(uintptr_t)desc_addr);
  88
  89         /* write virtio-net header */
  90         if (likely(desc_chunck_len >= dev->hdr_len)) {
  91                 *(struct virtio_net_hdr *)(uintptr_t)desc_addr = virtio_hdr;
  92                 desc_offset = dev->hdr_len;
  93         } else {
  94                 uint64_t len;
  95                 uint64_t remain = dev->hdr_len;
  96                 uint64_t src = (uint64_t)(uintptr_t)&virtio_hdr, dst;
  97                 uint64_t guest_addr = desc_gaddr;
  98
  99                 while (remain) {
 100                         len = remain;
 101                         dst = rte_vhost_va_from_guest_pa(dev->mem,
 102                                         guest_addr, &len);
 103                         if (unlikely(!dst || !len))
 104                                 return -1;
 105
 106                         rte_memcpy((void *)(uintptr_t)dst,
 107                                         (void *)(uintptr_t)src,
 108                                         len);
 109
 110                         remain -= len;
 111                         guest_addr += len;
 112                         src += len;
 113                 }
 114
 115                 desc_chunck_len = desc->len - dev->hdr_len;
 116                 desc_gaddr += dev->hdr_len;
 117                 desc_addr = rte_vhost_va_from_guest_pa(
 118                                 dev->mem, desc_gaddr,
 119                                 &desc_chunck_len);
 120                 if (unlikely(!desc_addr))
 121                         return -1;
 122
 123                 desc_offset = 0;
 124         }
 125
 126         desc_avail  = desc->len - dev->hdr_len;
 127
 128         mbuf_avail  = rte_pktmbuf_data_len(m);
 129         mbuf_offset = 0;
 130         while (mbuf_avail != 0 || m->next != NULL) {
 131                 /* done with current mbuf, fetch next */
 132                 if (mbuf_avail == 0) {
 133                         m = m->next;
 134
 135                         mbuf_offset = 0;
 136                         mbuf_avail  = rte_pktmbuf_data_len(m);
 137                 }
 138
 139                 /* done with current desc buf, fetch next */
 140                 if (desc_avail == 0) {
 141                         if ((desc->flags & VRING_DESC_F_NEXT) == 0) {
 142                                 /* Room in vring buffer is not enough */
 143                                 return -1;
 144                         }
 145                         if (unlikely(desc->next >= vr->size ||
 146                                      ++nr_desc > vr->size))
 147                                 return -1;
 148
 149                         desc = &vr->desc[desc->next];
 150                         desc_chunck_len = desc->len;
 151                         desc_gaddr = desc->addr;
 152                         desc_addr = rte_vhost_va_from_guest_pa(
 153                                         dev->mem, desc_gaddr, &desc_chunck_len);
 154                         if (unlikely(!desc_addr))
 155                                 return -1;
 156
 157                         desc_offset = 0;
 158                         desc_avail  = desc->len;
 159                 } else if (unlikely(desc_chunck_len == 0)) {
 160                         desc_chunck_len = desc_avail;
 161                         desc_gaddr += desc_offset;
 162                         desc_addr = rte_vhost_va_from_guest_pa(dev->mem,
 163                                         desc_gaddr,
 164                                         &desc_chunck_len);
 165                         if (unlikely(!desc_addr))
 166                                 return -1;
 167
 168                         desc_offset = 0;
 169                 }
 170
 171                 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
 172                 rte_memcpy((void *)((uintptr_t)(desc_addr + desc_offset)),
 173                         rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 174                         cpy_len);
 175
 176                 mbuf_avail  -= cpy_len;
 177                 mbuf_offset += cpy_len;
 178                 desc_avail  -= cpy_len;
 179                 desc_offset += cpy_len;
 180                 desc_chunck_len -= cpy_len;
 181         }
 182
 183         return 0;
 184 }
 185
 186 uint16_t
 187 vs_enqueue_pkts(struct vhost_dev *dev, uint16_t queue_id,
 188                 struct rte_mbuf **pkts, uint32_t count)
 189 {
 190         struct vhost_queue *queue;
 191         struct rte_vhost_vring *vr;
 192         uint16_t avail_idx, free_entries, start_idx;
 193         uint16_t desc_indexes[MAX_PKT_BURST];
 194         uint16_t used_idx;
 195         uint32_t i;
 196
 197         queue = &dev->queues[queue_id];
 198         vr    = &queue->vr;
 199
 200         avail_idx = __atomic_load_n(&vr->avail->idx, __ATOMIC_ACQUIRE);
 201         start_idx = queue->last_used_idx;
 202         free_entries = avail_idx - start_idx;
 203         count = RTE_MIN(count, free_entries);
 204         count = RTE_MIN(count, (uint32_t)MAX_PKT_BURST);
 205         if (count == 0)
 206                 return 0;
 207
 208         /* Retrieve all of the desc indexes first to avoid caching issues. */
 209         rte_prefetch0(&vr->avail->ring[start_idx & (vr->size - 1)]);
 210         for (i = 0; i < count; i++) {
 211                 used_idx = (start_idx + i) & (vr->size - 1);
 212                 desc_indexes[i] = vr->avail->ring[used_idx];
 213                 vr->used->ring[used_idx].id = desc_indexes[i];
 214                 vr->used->ring[used_idx].len = pkts[i]->pkt_len +
 215                                                dev->hdr_len;
 216         }
 217
 218         rte_prefetch0(&vr->desc[desc_indexes[0]]);
 219         for (i = 0; i < count; i++) {
 220                 uint16_t desc_idx = desc_indexes[i];
 221                 int err;
 222
 223                 err = enqueue_pkt(dev, vr, pkts[i], desc_idx);
 224                 if (unlikely(err)) {
 225                         used_idx = (start_idx + i) & (vr->size - 1);
 226                         vr->used->ring[used_idx].len = dev->hdr_len;
 227                 }
 228
 229                 if (i + 1 < count)
 230                         rte_prefetch0(&vr->desc[desc_indexes[i+1]]);
 231         }
 232
 233         __atomic_add_fetch(&vr->used->idx, count, __ATOMIC_RELEASE);
 234         queue->last_used_idx += count;
 235
 236         rte_vhost_vring_call(dev->vid, queue_id);
 237
 238         return count;
 239 }
 240
 241 static __rte_always_inline int
 242 dequeue_pkt(struct vhost_dev *dev, struct rte_vhost_vring *vr,
 243             struct rte_mbuf *m, uint16_t desc_idx,
 244             struct rte_mempool *mbuf_pool)
 245 {
 246         struct vring_desc *desc;
 247         uint64_t desc_addr, desc_gaddr;
 248         uint32_t desc_avail, desc_offset;
 249         uint64_t desc_chunck_len;
 250         uint32_t mbuf_avail, mbuf_offset;
 251         uint32_t cpy_len;
 252         struct rte_mbuf *cur = m, *prev = m;
 253         /* A counter to avoid desc dead loop chain */
 254         uint32_t nr_desc = 1;
 255
 256         desc = &vr->desc[desc_idx];
 257         if (unlikely((desc->len < dev->hdr_len)) ||
 258                         (desc->flags & VRING_DESC_F_INDIRECT))
 259                 return -1;
 260
 261         desc_chunck_len = desc->len;
 262         desc_gaddr = desc->addr;
 263         desc_addr = rte_vhost_va_from_guest_pa(
 264                         dev->mem, desc_gaddr, &desc_chunck_len);
 265         if (unlikely(!desc_addr))
 266                 return -1;
 267
 268         /*
 269          * We don't support ANY_LAYOUT, neither VERSION_1, meaning
 270          * a Tx packet from guest must have 2 desc buffers at least:
 271          * the first for storing the header and the others for
 272          * storing the data.
 273          *
 274          * And since we don't support TSO, we could simply skip the
 275          * header.
 276          */
 277         desc = &vr->desc[desc->next];
 278         desc_chunck_len = desc->len;
 279         desc_gaddr = desc->addr;
 280         desc_addr = rte_vhost_va_from_guest_pa(
 281                         dev->mem, desc_gaddr, &desc_chunck_len);
 282         if (unlikely(!desc_addr))
 283                 return -1;
 284         rte_prefetch0((void *)(uintptr_t)desc_addr);
 285
 286         desc_offset = 0;
 287         desc_avail  = desc->len;
 288         nr_desc    += 1;
 289
 290         mbuf_offset = 0;
 291         mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
 292         while (1) {
 293                 cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
 294                 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 295                                                    mbuf_offset),
 296                         (void *)((uintptr_t)(desc_addr + desc_offset)),
 297                         cpy_len);
 298
 299                 mbuf_avail  -= cpy_len;
 300                 mbuf_offset += cpy_len;
 301                 desc_avail  -= cpy_len;
 302                 desc_offset += cpy_len;
 303                 desc_chunck_len -= cpy_len;
 304
 305                 /* This desc reaches to its end, get the next one */
 306                 if (desc_avail == 0) {
 307                         if ((desc->flags & VRING_DESC_F_NEXT) == 0)
 308                                 break;
 309
 310                         if (unlikely(desc->next >= vr->size ||
 311                                      ++nr_desc > vr->size))
 312                                 return -1;
 313                         desc = &vr->desc[desc->next];
 314
 315                         desc_chunck_len = desc->len;
 316                         desc_gaddr = desc->addr;
 317                         desc_addr = rte_vhost_va_from_guest_pa(
 318                                         dev->mem, desc_gaddr, &desc_chunck_len);
 319                         if (unlikely(!desc_addr))
 320                                 return -1;
 321                         rte_prefetch0((void *)(uintptr_t)desc_addr);
 322
 323                         desc_offset = 0;
 324                         desc_avail  = desc->len;
 325                 } else if (unlikely(desc_chunck_len == 0)) {
 326                         desc_chunck_len = desc_avail;
 327                         desc_gaddr += desc_offset;
 328                         desc_addr = rte_vhost_va_from_guest_pa(dev->mem,
 329                                         desc_gaddr,
 330                                         &desc_chunck_len);
 331                         if (unlikely(!desc_addr))
 332                                 return -1;
 333
 334                         desc_offset = 0;
 335                 }
 336
 337                 /*
 338                  * This mbuf reaches to its end, get a new one
 339                  * to hold more data.
 340                  */
 341                 if (mbuf_avail == 0) {
 342                         cur = rte_pktmbuf_alloc(mbuf_pool);
 343                         if (unlikely(cur == NULL)) {
 344                                 RTE_LOG(ERR, VHOST_DATA, "Failed to "
 345                                         "allocate memory for mbuf.\n");
 346                                 return -1;
 347                         }
 348
 349                         prev->next = cur;
 350                         prev->data_len = mbuf_offset;
 351                         m->nb_segs += 1;
 352                         m->pkt_len += mbuf_offset;
 353                         prev = cur;
 354
 355                         mbuf_offset = 0;
 356                         mbuf_avail  = cur->buf_len - RTE_PKTMBUF_HEADROOM;
 357                 }
 358         }
 359
 360         prev->data_len = mbuf_offset;
 361         m->pkt_len    += mbuf_offset;
 362
 363         return 0;
 364 }
 365
 366 uint16_t
 367 vs_dequeue_pkts(struct vhost_dev *dev, uint16_t queue_id,
 368         struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
 369 {
 370         struct vhost_queue *queue;
 371         struct rte_vhost_vring *vr;
 372         uint32_t desc_indexes[MAX_PKT_BURST];
 373         uint32_t used_idx;
 374         uint32_t i = 0;
 375         uint16_t free_entries;
 376         uint16_t avail_idx;
 377
 378         queue = &dev->queues[queue_id];
 379         vr    = &queue->vr;
 380
 381         free_entries = __atomic_load_n(&vr->avail->idx, __ATOMIC_ACQUIRE) -
 382                         queue->last_avail_idx;
 383         if (free_entries == 0)
 384                 return 0;
 385
 386         /* Prefetch available and used ring */
 387         avail_idx = queue->last_avail_idx & (vr->size - 1);
 388         used_idx  = queue->last_used_idx  & (vr->size - 1);
 389         rte_prefetch0(&vr->avail->ring[avail_idx]);
 390         rte_prefetch0(&vr->used->ring[used_idx]);
 391
 392         count = RTE_MIN(count, MAX_PKT_BURST);
 393         count = RTE_MIN(count, free_entries);
 394
 395         if (unlikely(count == 0))
 396                 return 0;
 397
 398         /*
 399          * Retrieve all of the head indexes first and pre-update used entries
 400          * to avoid caching issues.
 401          */
 402         for (i = 0; i < count; i++) {
 403                 avail_idx = (queue->last_avail_idx + i) & (vr->size - 1);
 404                 used_idx  = (queue->last_used_idx  + i) & (vr->size - 1);
 405                 desc_indexes[i] = vr->avail->ring[avail_idx];
 406
 407                 vr->used->ring[used_idx].id  = desc_indexes[i];
 408                 vr->used->ring[used_idx].len = 0;
 409         }
 410
 411         /* Prefetch descriptor index. */
 412         rte_prefetch0(&vr->desc[desc_indexes[0]]);
 413         for (i = 0; i < count; i++) {
 414                 int err;
 415
 416                 if (likely(i + 1 < count))
 417                         rte_prefetch0(&vr->desc[desc_indexes[i + 1]]);
 418
 419                 pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 420                 if (unlikely(pkts[i] == NULL)) {
 421                         RTE_LOG(ERR, VHOST_DATA,
 422                                 "Failed to allocate memory for mbuf.\n");
 423                         break;
 424                 }
 425
 426                 err = dequeue_pkt(dev, vr, pkts[i], desc_indexes[i], mbuf_pool);
 427                 if (unlikely(err)) {
 428                         rte_pktmbuf_free(pkts[i]);
 429                         break;
 430                 }
 431
 432         }
 433
 434         queue->last_avail_idx += i;
 435         queue->last_used_idx += i;
 436
 437         __atomic_add_fetch(&vr->used->idx, i, __ATOMIC_ACQ_REL);
 438
 439         rte_vhost_vring_call(dev->vid, queue_id);
 440
 441         return i;
 442 }