drivers/net/virtio/virtio_rxtx_simple.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <stdint.h>
  35 #include <stdio.h>
  36 #include <stdlib.h>
  37 #include <string.h>
  38 #include <errno.h>
  39
  40 #include <tmmintrin.h>
  41
  42 #include <rte_cycles.h>
  43 #include <rte_memory.h>
  44 #include <rte_memzone.h>
  45 #include <rte_branch_prediction.h>
  46 #include <rte_mempool.h>
  47 #include <rte_malloc.h>
  48 #include <rte_mbuf.h>
  49 #include <rte_ether.h>
  50 #include <rte_ethdev.h>
  51 #include <rte_prefetch.h>
  52 #include <rte_string_fns.h>
  53 #include <rte_errno.h>
  54 #include <rte_byteorder.h>
  55
  56 #include "virtio_logs.h"
  57 #include "virtio_ethdev.h"
  58 #include "virtqueue.h"
  59 #include "virtio_rxtx.h"
  60
  61 #define RTE_VIRTIO_VPMD_RX_BURST 32
  62 #define RTE_VIRTIO_DESC_PER_LOOP 8
  63 #define RTE_VIRTIO_VPMD_RX_REARM_THRESH RTE_VIRTIO_VPMD_RX_BURST
  64
  65 #ifndef __INTEL_COMPILER
  66 #pragma GCC diagnostic ignored "-Wcast-qual"
  67 #endif
  68
  69 int __attribute__((cold))
  70 virtqueue_enqueue_recv_refill_simple(struct virtqueue *vq,
  71         struct rte_mbuf *cookie)
  72 {
  73         struct vq_desc_extra *dxp;
  74         struct vring_desc *start_dp;
  75         uint16_t desc_idx;
  76
  77         desc_idx = vq->vq_avail_idx & (vq->vq_nentries - 1);
  78         dxp = &vq->vq_descx[desc_idx];
  79         dxp->cookie = (void *)cookie;
  80         vq->sw_ring[desc_idx] = cookie;
  81
  82         start_dp = vq->vq_ring.desc;
  83         start_dp[desc_idx].addr = (uint64_t)((uintptr_t)cookie->buf_physaddr +
  84                 RTE_PKTMBUF_HEADROOM - sizeof(struct virtio_net_hdr));
  85         start_dp[desc_idx].len = cookie->buf_len -
  86                 RTE_PKTMBUF_HEADROOM + sizeof(struct virtio_net_hdr);
  87
  88         vq->vq_free_cnt--;
  89         vq->vq_avail_idx++;
  90
  91         return 0;
  92 }
  93
  94 static inline void
  95 virtio_rxq_rearm_vec(struct virtqueue *rxvq)
  96 {
  97         int i;
  98         uint16_t desc_idx;
  99         struct rte_mbuf **sw_ring;
 100         struct vring_desc *start_dp;
 101         int ret;
 102
 103         desc_idx = rxvq->vq_avail_idx & (rxvq->vq_nentries - 1);
 104         sw_ring = &rxvq->sw_ring[desc_idx];
 105         start_dp = &rxvq->vq_ring.desc[desc_idx];
 106
 107         ret = rte_mempool_get_bulk(rxvq->mpool, (void **)sw_ring,
 108                 RTE_VIRTIO_VPMD_RX_REARM_THRESH);
 109         if (unlikely(ret)) {
 110                 rte_eth_devices[rxvq->port_id].data->rx_mbuf_alloc_failed +=
 111                         RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 112                 return;
 113         }
 114
 115         for (i = 0; i < RTE_VIRTIO_VPMD_RX_REARM_THRESH; i++) {
 116                 uintptr_t p;
 117
 118                 p = (uintptr_t)&sw_ring[i]->rearm_data;
 119                 *(uint64_t *)p = rxvq->mbuf_initializer;
 120
 121                 start_dp[i].addr =
 122                         (uint64_t)((uintptr_t)sw_ring[i]->buf_physaddr +
 123                         RTE_PKTMBUF_HEADROOM - sizeof(struct virtio_net_hdr));
 124                 start_dp[i].len = sw_ring[i]->buf_len -
 125                         RTE_PKTMBUF_HEADROOM + sizeof(struct virtio_net_hdr);
 126         }
 127
 128         rxvq->vq_avail_idx += RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 129         rxvq->vq_free_cnt -= RTE_VIRTIO_VPMD_RX_REARM_THRESH;
 130         vq_update_avail_idx(rxvq);
 131 }
 132
 133 /* virtio vPMD receive routine, only accept(nb_pkts >= RTE_VIRTIO_DESC_PER_LOOP)
 134  *
 135  * This routine is for non-mergeable RX, one desc for each guest buffer.
 136  * This routine is based on the RX ring layout optimization. Each entry in the
 137  * avail ring points to the desc with the same index in the desc ring and this
 138  * will never be changed in the driver.
 139  *
 140  * - nb_pkts < RTE_VIRTIO_DESC_PER_LOOP, just return no packet
 141  */
 142 uint16_t
 143 virtio_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
 144         uint16_t nb_pkts)
 145 {
 146         struct virtqueue *rxvq = rx_queue;
 147         uint16_t nb_used;
 148         uint16_t desc_idx;
 149         struct vring_used_elem *rused;
 150         struct rte_mbuf **sw_ring;
 151         struct rte_mbuf **sw_ring_end;
 152         uint16_t nb_pkts_received;
 153         __m128i shuf_msk1, shuf_msk2, len_adjust;
 154
 155         shuf_msk1 = _mm_set_epi8(
 156                 0xFF, 0xFF, 0xFF, 0xFF,
 157                 0xFF, 0xFF,             /* vlan tci */
 158                 5, 4,                   /* dat len */
 159                 0xFF, 0xFF, 5, 4,       /* pkt len */
 160                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
 161
 162         );
 163
 164         shuf_msk2 = _mm_set_epi8(
 165                 0xFF, 0xFF, 0xFF, 0xFF,
 166                 0xFF, 0xFF,             /* vlan tci */
 167                 13, 12,                 /* dat len */
 168                 0xFF, 0xFF, 13, 12,     /* pkt len */
 169                 0xFF, 0xFF, 0xFF, 0xFF  /* packet type */
 170         );
 171
 172         /* Subtract the header length.
 173         *  In which case do we need the header length in used->len ?
 174         */
 175         len_adjust = _mm_set_epi16(
 176                 0, 0,
 177                 0,
 178                 (uint16_t) -sizeof(struct virtio_net_hdr),
 179                 0, (uint16_t) -sizeof(struct virtio_net_hdr),
 180                 0, 0);
 181
 182         if (unlikely(nb_pkts < RTE_VIRTIO_DESC_PER_LOOP))
 183                 return 0;
 184
 185         nb_used = *(volatile uint16_t *)&rxvq->vq_ring.used->idx -
 186                 rxvq->vq_used_cons_idx;
 187
 188         rte_compiler_barrier();
 189
 190         if (unlikely(nb_used == 0))
 191                 return 0;
 192
 193         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_VIRTIO_DESC_PER_LOOP);
 194         nb_used = RTE_MIN(nb_used, nb_pkts);
 195
 196         desc_idx = (uint16_t)(rxvq->vq_used_cons_idx & (rxvq->vq_nentries - 1));
 197         rused = &rxvq->vq_ring.used->ring[desc_idx];
 198         sw_ring  = &rxvq->sw_ring[desc_idx];
 199         sw_ring_end = &rxvq->sw_ring[rxvq->vq_nentries];
 200
 201         _mm_prefetch((const void *)rused, _MM_HINT_T0);
 202
 203         if (rxvq->vq_free_cnt >= RTE_VIRTIO_VPMD_RX_REARM_THRESH) {
 204                 virtio_rxq_rearm_vec(rxvq);
 205                 if (unlikely(virtqueue_kick_prepare(rxvq)))
 206                         virtqueue_notify(rxvq);
 207         }
 208
 209         for (nb_pkts_received = 0;
 210                 nb_pkts_received < nb_used;) {
 211                 __m128i desc[RTE_VIRTIO_DESC_PER_LOOP / 2];
 212                 __m128i mbp[RTE_VIRTIO_DESC_PER_LOOP / 2];
 213                 __m128i pkt_mb[RTE_VIRTIO_DESC_PER_LOOP];
 214
 215                 mbp[0] = _mm_loadu_si128((__m128i *)(sw_ring + 0));
 216                 desc[0] = _mm_loadu_si128((__m128i *)(rused + 0));
 217                 _mm_storeu_si128((__m128i *)&rx_pkts[0], mbp[0]);
 218
 219                 mbp[1] = _mm_loadu_si128((__m128i *)(sw_ring + 2));
 220                 desc[1] = _mm_loadu_si128((__m128i *)(rused + 2));
 221                 _mm_storeu_si128((__m128i *)&rx_pkts[2], mbp[1]);
 222
 223                 mbp[2] = _mm_loadu_si128((__m128i *)(sw_ring + 4));
 224                 desc[2] = _mm_loadu_si128((__m128i *)(rused + 4));
 225                 _mm_storeu_si128((__m128i *)&rx_pkts[4], mbp[2]);
 226
 227                 mbp[3] = _mm_loadu_si128((__m128i *)(sw_ring + 6));
 228                 desc[3] = _mm_loadu_si128((__m128i *)(rused + 6));
 229                 _mm_storeu_si128((__m128i *)&rx_pkts[6], mbp[3]);
 230
 231                 pkt_mb[1] = _mm_shuffle_epi8(desc[0], shuf_msk2);
 232                 pkt_mb[0] = _mm_shuffle_epi8(desc[0], shuf_msk1);
 233                 pkt_mb[1] = _mm_add_epi16(pkt_mb[1], len_adjust);
 234                 pkt_mb[0] = _mm_add_epi16(pkt_mb[0], len_adjust);
 235                 _mm_storeu_si128((void *)&rx_pkts[1]->rx_descriptor_fields1,
 236                         pkt_mb[1]);
 237                 _mm_storeu_si128((void *)&rx_pkts[0]->rx_descriptor_fields1,
 238                         pkt_mb[0]);
 239
 240                 pkt_mb[3] = _mm_shuffle_epi8(desc[1], shuf_msk2);
 241                 pkt_mb[2] = _mm_shuffle_epi8(desc[1], shuf_msk1);
 242                 pkt_mb[3] = _mm_add_epi16(pkt_mb[3], len_adjust);
 243                 pkt_mb[2] = _mm_add_epi16(pkt_mb[2], len_adjust);
 244                 _mm_storeu_si128((void *)&rx_pkts[3]->rx_descriptor_fields1,
 245                         pkt_mb[3]);
 246                 _mm_storeu_si128((void *)&rx_pkts[2]->rx_descriptor_fields1,
 247                         pkt_mb[2]);
 248
 249                 pkt_mb[5] = _mm_shuffle_epi8(desc[2], shuf_msk2);
 250                 pkt_mb[4] = _mm_shuffle_epi8(desc[2], shuf_msk1);
 251                 pkt_mb[5] = _mm_add_epi16(pkt_mb[5], len_adjust);
 252                 pkt_mb[4] = _mm_add_epi16(pkt_mb[4], len_adjust);
 253                 _mm_storeu_si128((void *)&rx_pkts[5]->rx_descriptor_fields1,
 254                         pkt_mb[5]);
 255                 _mm_storeu_si128((void *)&rx_pkts[4]->rx_descriptor_fields1,
 256                         pkt_mb[4]);
 257
 258                 pkt_mb[7] = _mm_shuffle_epi8(desc[3], shuf_msk2);
 259                 pkt_mb[6] = _mm_shuffle_epi8(desc[3], shuf_msk1);
 260                 pkt_mb[7] = _mm_add_epi16(pkt_mb[7], len_adjust);
 261                 pkt_mb[6] = _mm_add_epi16(pkt_mb[6], len_adjust);
 262                 _mm_storeu_si128((void *)&rx_pkts[7]->rx_descriptor_fields1,
 263                         pkt_mb[7]);
 264                 _mm_storeu_si128((void *)&rx_pkts[6]->rx_descriptor_fields1,
 265                         pkt_mb[6]);
 266
 267                 if (unlikely(nb_used <= RTE_VIRTIO_DESC_PER_LOOP)) {
 268                         if (sw_ring + nb_used <= sw_ring_end)
 269                                 nb_pkts_received += nb_used;
 270                         else
 271                                 nb_pkts_received += sw_ring_end - sw_ring;
 272                         break;
 273                 } else {
 274                         if (unlikely(sw_ring + RTE_VIRTIO_DESC_PER_LOOP >=
 275                                 sw_ring_end)) {
 276                                 nb_pkts_received += sw_ring_end - sw_ring;
 277                                 break;
 278                         } else {
 279                                 nb_pkts_received += RTE_VIRTIO_DESC_PER_LOOP;
 280
 281                                 rx_pkts += RTE_VIRTIO_DESC_PER_LOOP;
 282                                 sw_ring += RTE_VIRTIO_DESC_PER_LOOP;
 283                                 rused   += RTE_VIRTIO_DESC_PER_LOOP;
 284                                 nb_used -= RTE_VIRTIO_DESC_PER_LOOP;
 285                         }
 286                 }
 287         }
 288
 289         rxvq->vq_used_cons_idx += nb_pkts_received;
 290         rxvq->vq_free_cnt += nb_pkts_received;
 291         rxvq->packets += nb_pkts_received;
 292         return nb_pkts_received;
 293 }
 294
 295 #define VIRTIO_TX_FREE_THRESH 32
 296 #define VIRTIO_TX_MAX_FREE_BUF_SZ 32
 297 #define VIRTIO_TX_FREE_NR 32
 298 /* TODO: vq->tx_free_cnt could mean num of free slots so we could avoid shift */
 299 static inline void
 300 virtio_xmit_cleanup(struct virtqueue *vq)
 301 {
 302         uint16_t i, desc_idx;
 303         int nb_free = 0;
 304         struct rte_mbuf *m, *free[VIRTIO_TX_MAX_FREE_BUF_SZ];
 305
 306         desc_idx = (uint16_t)(vq->vq_used_cons_idx &
 307                    ((vq->vq_nentries >> 1) - 1));
 308         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 309         m = __rte_pktmbuf_prefree_seg(m);
 310         if (likely(m != NULL)) {
 311                 free[0] = m;
 312                 nb_free = 1;
 313                 for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
 314                         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 315                         m = __rte_pktmbuf_prefree_seg(m);
 316                         if (likely(m != NULL)) {
 317                                 if (likely(m->pool == free[0]->pool))
 318                                         free[nb_free++] = m;
 319                                 else {
 320                                         rte_mempool_put_bulk(free[0]->pool,
 321                                                 (void **)free, nb_free);
 322                                         free[0] = m;
 323                                         nb_free = 1;
 324                                 }
 325                         }
 326                 }
 327                 rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
 328         } else {
 329                 for (i = 1; i < VIRTIO_TX_FREE_NR; i++) {
 330                         m = (struct rte_mbuf *)vq->vq_descx[desc_idx++].cookie;
 331                         m = __rte_pktmbuf_prefree_seg(m);
 332                         if (m != NULL)
 333                                 rte_mempool_put(m->pool, m);
 334                 }
 335         }
 336
 337         vq->vq_used_cons_idx += VIRTIO_TX_FREE_NR;
 338         vq->vq_free_cnt += (VIRTIO_TX_FREE_NR << 1);
 339 }
 340
 341 uint16_t
 342 virtio_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts,
 343         uint16_t nb_pkts)
 344 {
 345         struct virtqueue *txvq = tx_queue;
 346         uint16_t nb_used;
 347         uint16_t desc_idx;
 348         struct vring_desc *start_dp;
 349         uint16_t nb_tail, nb_commit;
 350         int i;
 351         uint16_t desc_idx_max = (txvq->vq_nentries >> 1) - 1;
 352
 353         nb_used = VIRTQUEUE_NUSED(txvq);
 354         rte_compiler_barrier();
 355
 356         if (nb_used >= VIRTIO_TX_FREE_THRESH)
 357                 virtio_xmit_cleanup(tx_queue);
 358
 359         nb_commit = nb_pkts = RTE_MIN((txvq->vq_free_cnt >> 1), nb_pkts);
 360         desc_idx = (uint16_t) (txvq->vq_avail_idx & desc_idx_max);
 361         start_dp = txvq->vq_ring.desc;
 362         nb_tail = (uint16_t) (desc_idx_max + 1 - desc_idx);
 363
 364         if (nb_commit >= nb_tail) {
 365                 for (i = 0; i < nb_tail; i++)
 366                         txvq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
 367                 for (i = 0; i < nb_tail; i++) {
 368                         start_dp[desc_idx].addr =
 369                                 RTE_MBUF_DATA_DMA_ADDR(*tx_pkts);
 370                         start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
 371                         tx_pkts++;
 372                         desc_idx++;
 373                 }
 374                 nb_commit -= nb_tail;
 375                 desc_idx = 0;
 376         }
 377         for (i = 0; i < nb_commit; i++)
 378                 txvq->vq_descx[desc_idx + i].cookie = tx_pkts[i];
 379         for (i = 0; i < nb_commit; i++) {
 380                 start_dp[desc_idx].addr = RTE_MBUF_DATA_DMA_ADDR(*tx_pkts);
 381                 start_dp[desc_idx].len = (*tx_pkts)->pkt_len;
 382                 tx_pkts++;
 383                 desc_idx++;
 384         }
 385
 386         rte_compiler_barrier();
 387
 388         txvq->vq_free_cnt -= (uint16_t)(nb_pkts << 1);
 389         txvq->vq_avail_idx += nb_pkts;
 390         txvq->vq_ring.avail->idx = txvq->vq_avail_idx;
 391         txvq->packets += nb_pkts;
 392
 393         if (likely(nb_pkts)) {
 394                 if (unlikely(virtqueue_kick_prepare(txvq)))
 395                         virtqueue_notify(txvq);
 396         }
 397
 398         return nb_pkts;
 399 }
 400
 401 int __attribute__((cold))
 402 virtio_rxq_vec_setup(struct virtqueue *rxq)
 403 {
 404         uintptr_t p;
 405         struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
 406
 407         mb_def.nb_segs = 1;
 408         mb_def.data_off = RTE_PKTMBUF_HEADROOM;
 409         mb_def.port = rxq->port_id;
 410         rte_mbuf_refcnt_set(&mb_def, 1);
 411
 412         /* prevent compiler reordering: rearm_data covers previous fields */
 413         rte_compiler_barrier();
 414         p = (uintptr_t)&mb_def.rearm_data;
 415         rxq->mbuf_initializer = *(uint64_t *)p;
 416
 417         return 0;
 418 }