drivers/net/fm10k/fm10k_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2013-2015 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 <inttypes.h>
  35
  36 #include <rte_ethdev.h>
  37 #include <rte_common.h>
  38 #include "fm10k.h"
  39 #include "base/fm10k_type.h"
  40
  41 #ifdef RTE_PMD_PACKET_PREFETCH
  42 #define rte_packet_prefetch(p)  rte_prefetch1(p)
  43 #else
  44 #define rte_packet_prefetch(p)  do {} while (0)
  45 #endif
  46
  47 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
  48 static inline void dump_rxd(union fm10k_rx_desc *rxd)
  49 {
  50         PMD_RX_LOG(DEBUG, "+----------------|----------------+");
  51         PMD_RX_LOG(DEBUG, "|     GLORT      | PKT HDR & TYPE |");
  52         PMD_RX_LOG(DEBUG, "|   0x%08x   |   0x%08x   |", rxd->d.glort,
  53                         rxd->d.data);
  54         PMD_RX_LOG(DEBUG, "+----------------|----------------+");
  55         PMD_RX_LOG(DEBUG, "|   VLAN & LEN   |     STATUS     |");
  56         PMD_RX_LOG(DEBUG, "|   0x%08x   |   0x%08x   |", rxd->d.vlan_len,
  57                         rxd->d.staterr);
  58         PMD_RX_LOG(DEBUG, "+----------------|----------------+");
  59         PMD_RX_LOG(DEBUG, "|    RESERVED    |    RSS_HASH    |");
  60         PMD_RX_LOG(DEBUG, "|   0x%08x   |   0x%08x   |", 0, rxd->d.rss);
  61         PMD_RX_LOG(DEBUG, "+----------------|----------------+");
  62         PMD_RX_LOG(DEBUG, "|            TIME TAG             |");
  63         PMD_RX_LOG(DEBUG, "|       0x%016"PRIx64"        |", rxd->q.timestamp);
  64         PMD_RX_LOG(DEBUG, "+----------------|----------------+");
  65 }
  66 #endif
  67
  68 static inline void
  69 rx_desc_to_ol_flags(struct rte_mbuf *m, const union fm10k_rx_desc *d)
  70 {
  71 #ifdef RTE_NEXT_ABI
  72         static const uint32_t
  73                 ptype_table[FM10K_RXD_PKTTYPE_MASK >> FM10K_RXD_PKTTYPE_SHIFT]
  74                         __rte_cache_aligned = {
  75                 [FM10K_PKTTYPE_OTHER] = RTE_PTYPE_L2_ETHER,
  76                 [FM10K_PKTTYPE_IPV4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4,
  77                 [FM10K_PKTTYPE_IPV4_EX] = RTE_PTYPE_L2_ETHER |
  78                         RTE_PTYPE_L3_IPV4_EXT,
  79                 [FM10K_PKTTYPE_IPV6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6,
  80                 [FM10K_PKTTYPE_IPV6_EX] = RTE_PTYPE_L2_ETHER |
  81                         RTE_PTYPE_L3_IPV6_EXT,
  82                 [FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
  83                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP,
  84                 [FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
  85                         RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_TCP,
  86                 [FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
  87                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP,
  88                 [FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
  89                         RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_UDP,
  90         };
  91
  92         m->packet_type = ptype_table[(d->w.pkt_info & FM10K_RXD_PKTTYPE_MASK)
  93                                                 >> FM10K_RXD_PKTTYPE_SHIFT];
  94 #else /* RTE_NEXT_ABI */
  95         uint16_t ptype;
  96         static const uint16_t pt_lut[] = { 0,
  97                 PKT_RX_IPV4_HDR, PKT_RX_IPV4_HDR_EXT,
  98                 PKT_RX_IPV6_HDR, PKT_RX_IPV6_HDR_EXT,
  99                 0, 0, 0
 100         };
 101 #endif /* RTE_NEXT_ABI */
 102
 103         if (d->w.pkt_info & FM10K_RXD_RSSTYPE_MASK)
 104                 m->ol_flags |= PKT_RX_RSS_HASH;
 105
 106         if (unlikely((d->d.staterr &
 107                 (FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)) ==
 108                 (FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)))
 109                 m->ol_flags |= PKT_RX_IP_CKSUM_BAD;
 110
 111         if (unlikely((d->d.staterr &
 112                 (FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)) ==
 113                 (FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)))
 114                 m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
 115
 116         if (d->d.staterr & FM10K_RXD_STATUS_VEXT)
 117                 m->ol_flags |= PKT_RX_VLAN_PKT;
 118
 119         if (unlikely(d->d.staterr & FM10K_RXD_STATUS_HBO))
 120                 m->ol_flags |= PKT_RX_HBUF_OVERFLOW;
 121
 122         if (unlikely(d->d.staterr & FM10K_RXD_STATUS_RXE))
 123                 m->ol_flags |= PKT_RX_RECIP_ERR;
 124
 125 #ifndef RTE_NEXT_ABI
 126         ptype = (d->d.data & FM10K_RXD_PKTTYPE_MASK_L3) >>
 127                                                 FM10K_RXD_PKTTYPE_SHIFT;
 128         m->ol_flags |= pt_lut[(uint8_t)ptype];
 129 #endif
 130 }
 131
 132 uint16_t
 133 fm10k_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 134         uint16_t nb_pkts)
 135 {
 136         struct rte_mbuf *mbuf;
 137         union fm10k_rx_desc desc;
 138         struct fm10k_rx_queue *q = rx_queue;
 139         uint16_t count = 0;
 140         int alloc = 0;
 141         uint16_t next_dd;
 142         int ret;
 143
 144         next_dd = q->next_dd;
 145
 146         nb_pkts = RTE_MIN(nb_pkts, q->alloc_thresh);
 147         for (count = 0; count < nb_pkts; ++count) {
 148                 mbuf = q->sw_ring[next_dd];
 149                 desc = q->hw_ring[next_dd];
 150                 if (!(desc.d.staterr & FM10K_RXD_STATUS_DD))
 151                         break;
 152 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
 153                 dump_rxd(&desc);
 154 #endif
 155                 rte_pktmbuf_pkt_len(mbuf) = desc.w.length;
 156                 rte_pktmbuf_data_len(mbuf) = desc.w.length;
 157
 158                 mbuf->ol_flags = 0;
 159 #ifdef RTE_LIBRTE_FM10K_RX_OLFLAGS_ENABLE
 160                 rx_desc_to_ol_flags(mbuf, &desc);
 161 #endif
 162
 163                 mbuf->hash.rss = desc.d.rss;
 164
 165                 rx_pkts[count] = mbuf;
 166                 if (++next_dd == q->nb_desc) {
 167                         next_dd = 0;
 168                         alloc = 1;
 169                 }
 170
 171                 /* Prefetch next mbuf while processing current one. */
 172                 rte_prefetch0(q->sw_ring[next_dd]);
 173
 174                 /*
 175                  * When next RX descriptor is on a cache-line boundary,
 176                  * prefetch the next 4 RX descriptors and the next 8 pointers
 177                  * to mbufs.
 178                  */
 179                 if ((next_dd & 0x3) == 0) {
 180                         rte_prefetch0(&q->hw_ring[next_dd]);
 181                         rte_prefetch0(&q->sw_ring[next_dd]);
 182                 }
 183         }
 184
 185         q->next_dd = next_dd;
 186
 187         if ((q->next_dd > q->next_trigger) || (alloc == 1)) {
 188                 ret = rte_mempool_get_bulk(q->mp,
 189                                         (void **)&q->sw_ring[q->next_alloc],
 190                                         q->alloc_thresh);
 191
 192                 if (unlikely(ret != 0)) {
 193                         uint8_t port = q->port_id;
 194                         PMD_RX_LOG(ERR, "Failed to alloc mbuf");
 195                         /*
 196                          * Need to restore next_dd if we cannot allocate new
 197                          * buffers to replenish the old ones.
 198                          */
 199                         q->next_dd = (q->next_dd + q->nb_desc - count) %
 200                                                                 q->nb_desc;
 201                         rte_eth_devices[port].data->rx_mbuf_alloc_failed++;
 202                         return 0;
 203                 }
 204
 205                 for (; q->next_alloc <= q->next_trigger; ++q->next_alloc) {
 206                         mbuf = q->sw_ring[q->next_alloc];
 207
 208                         /* setup static mbuf fields */
 209                         fm10k_pktmbuf_reset(mbuf, q->port_id);
 210
 211                         /* write descriptor */
 212                         desc.q.pkt_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
 213                         desc.q.hdr_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
 214                         q->hw_ring[q->next_alloc] = desc;
 215                 }
 216                 FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_trigger);
 217                 q->next_trigger += q->alloc_thresh;
 218                 if (q->next_trigger >= q->nb_desc) {
 219                         q->next_trigger = q->alloc_thresh - 1;
 220                         q->next_alloc = 0;
 221                 }
 222         }
 223
 224         return count;
 225 }
 226
 227 uint16_t
 228 fm10k_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 229                                 uint16_t nb_pkts)
 230 {
 231         struct rte_mbuf *mbuf;
 232         union fm10k_rx_desc desc;
 233         struct fm10k_rx_queue *q = rx_queue;
 234         uint16_t count = 0;
 235         uint16_t nb_rcv, nb_seg;
 236         int alloc = 0;
 237         uint16_t next_dd;
 238         struct rte_mbuf *first_seg = q->pkt_first_seg;
 239         struct rte_mbuf *last_seg = q->pkt_last_seg;
 240         int ret;
 241
 242         next_dd = q->next_dd;
 243         nb_rcv = 0;
 244
 245         nb_seg = RTE_MIN(nb_pkts, q->alloc_thresh);
 246         for (count = 0; count < nb_seg; count++) {
 247                 mbuf = q->sw_ring[next_dd];
 248                 desc = q->hw_ring[next_dd];
 249                 if (!(desc.d.staterr & FM10K_RXD_STATUS_DD))
 250                         break;
 251 #ifdef RTE_LIBRTE_FM10K_DEBUG_RX
 252                 dump_rxd(&desc);
 253 #endif
 254
 255                 if (++next_dd == q->nb_desc) {
 256                         next_dd = 0;
 257                         alloc = 1;
 258                 }
 259
 260                 /* Prefetch next mbuf while processing current one. */
 261                 rte_prefetch0(q->sw_ring[next_dd]);
 262
 263                 /*
 264                  * When next RX descriptor is on a cache-line boundary,
 265                  * prefetch the next 4 RX descriptors and the next 8 pointers
 266                  * to mbufs.
 267                  */
 268                 if ((next_dd & 0x3) == 0) {
 269                         rte_prefetch0(&q->hw_ring[next_dd]);
 270                         rte_prefetch0(&q->sw_ring[next_dd]);
 271                 }
 272
 273                 /* Fill data length */
 274                 rte_pktmbuf_data_len(mbuf) = desc.w.length;
 275
 276                 /*
 277                  * If this is the first buffer of the received packet,
 278                  * set the pointer to the first mbuf of the packet and
 279                  * initialize its context.
 280                  * Otherwise, update the total length and the number of segments
 281                  * of the current scattered packet, and update the pointer to
 282                  * the last mbuf of the current packet.
 283                  */
 284                 if (!first_seg) {
 285                         first_seg = mbuf;
 286                         first_seg->pkt_len = desc.w.length;
 287                 } else {
 288                         first_seg->pkt_len =
 289                                         (uint16_t)(first_seg->pkt_len +
 290                                         rte_pktmbuf_data_len(mbuf));
 291                         first_seg->nb_segs++;
 292                         last_seg->next = mbuf;
 293                 }
 294
 295                 /*
 296                  * If this is not the last buffer of the received packet,
 297                  * update the pointer to the last mbuf of the current scattered
 298                  * packet and continue to parse the RX ring.
 299                  */
 300                 if (!(desc.d.staterr & FM10K_RXD_STATUS_EOP)) {
 301                         last_seg = mbuf;
 302                         continue;
 303                 }
 304
 305                 first_seg->ol_flags = 0;
 306 #ifdef RTE_LIBRTE_FM10K_RX_OLFLAGS_ENABLE
 307                 rx_desc_to_ol_flags(first_seg, &desc);
 308 #endif
 309                 first_seg->hash.rss = desc.d.rss;
 310
 311                 /* Prefetch data of first segment, if configured to do so. */
 312                 rte_packet_prefetch((char *)first_seg->buf_addr +
 313                         first_seg->data_off);
 314
 315                 /*
 316                  * Store the mbuf address into the next entry of the array
 317                  * of returned packets.
 318                  */
 319                 rx_pkts[nb_rcv++] = first_seg;
 320
 321                 /*
 322                  * Setup receipt context for a new packet.
 323                  */
 324                 first_seg = NULL;
 325         }
 326
 327         q->next_dd = next_dd;
 328
 329         if ((q->next_dd > q->next_trigger) || (alloc == 1)) {
 330                 ret = rte_mempool_get_bulk(q->mp,
 331                                         (void **)&q->sw_ring[q->next_alloc],
 332                                         q->alloc_thresh);
 333
 334                 if (unlikely(ret != 0)) {
 335                         uint8_t port = q->port_id;
 336                         PMD_RX_LOG(ERR, "Failed to alloc mbuf");
 337                         /*
 338                          * Need to restore next_dd if we cannot allocate new
 339                          * buffers to replenish the old ones.
 340                          */
 341                         q->next_dd = (q->next_dd + q->nb_desc - count) %
 342                                                                 q->nb_desc;
 343                         rte_eth_devices[port].data->rx_mbuf_alloc_failed++;
 344                         return 0;
 345                 }
 346
 347                 for (; q->next_alloc <= q->next_trigger; ++q->next_alloc) {
 348                         mbuf = q->sw_ring[q->next_alloc];
 349
 350                         /* setup static mbuf fields */
 351                         fm10k_pktmbuf_reset(mbuf, q->port_id);
 352
 353                         /* write descriptor */
 354                         desc.q.pkt_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
 355                         desc.q.hdr_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
 356                         q->hw_ring[q->next_alloc] = desc;
 357                 }
 358                 FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_trigger);
 359                 q->next_trigger += q->alloc_thresh;
 360                 if (q->next_trigger >= q->nb_desc) {
 361                         q->next_trigger = q->alloc_thresh - 1;
 362                         q->next_alloc = 0;
 363                 }
 364         }
 365
 366         q->pkt_first_seg = first_seg;
 367         q->pkt_last_seg = last_seg;
 368
 369         return nb_rcv;
 370 }
 371
 372 static inline void tx_free_descriptors(struct fm10k_tx_queue *q)
 373 {
 374         uint16_t next_rs, count = 0;
 375
 376         next_rs = fifo_peek(&q->rs_tracker);
 377         if (!(q->hw_ring[next_rs].flags & FM10K_TXD_FLAG_DONE))
 378                 return;
 379
 380         /* the DONE flag is set on this descriptor so remove the ID
 381          * from the RS bit tracker and free the buffers */
 382         fifo_remove(&q->rs_tracker);
 383
 384         /* wrap around? if so, free buffers from last_free up to but NOT
 385          * including nb_desc */
 386         if (q->last_free > next_rs) {
 387                 count = q->nb_desc - q->last_free;
 388                 while (q->last_free < q->nb_desc) {
 389                         rte_pktmbuf_free_seg(q->sw_ring[q->last_free]);
 390                         q->sw_ring[q->last_free] = NULL;
 391                         ++q->last_free;
 392                 }
 393                 q->last_free = 0;
 394         }
 395
 396         /* adjust free descriptor count before the next loop */
 397         q->nb_free += count + (next_rs + 1 - q->last_free);
 398
 399         /* free buffers from last_free, up to and including next_rs */
 400         while (q->last_free <= next_rs) {
 401                 rte_pktmbuf_free_seg(q->sw_ring[q->last_free]);
 402                 q->sw_ring[q->last_free] = NULL;
 403                 ++q->last_free;
 404         }
 405
 406         if (q->last_free == q->nb_desc)
 407                 q->last_free = 0;
 408 }
 409
 410 static inline void tx_xmit_pkt(struct fm10k_tx_queue *q, struct rte_mbuf *mb)
 411 {
 412         uint16_t last_id;
 413         uint8_t flags;
 414
 415         /* always set the LAST flag on the last descriptor used to
 416          * transmit the packet */
 417         flags = FM10K_TXD_FLAG_LAST;
 418         last_id = q->next_free + mb->nb_segs - 1;
 419         if (last_id >= q->nb_desc)
 420                 last_id = last_id - q->nb_desc;
 421
 422         /* but only set the RS flag on the last descriptor if rs_thresh
 423          * descriptors will be used since the RS flag was last set */
 424         if ((q->nb_used + mb->nb_segs) >= q->rs_thresh) {
 425                 flags |= FM10K_TXD_FLAG_RS;
 426                 fifo_insert(&q->rs_tracker, last_id);
 427                 q->nb_used = 0;
 428         } else {
 429                 q->nb_used = q->nb_used + mb->nb_segs;
 430         }
 431
 432         q->nb_free -= mb->nb_segs;
 433
 434         q->hw_ring[q->next_free].flags = 0;
 435         /* set checksum flags on first descriptor of packet. SCTP checksum
 436          * offload is not supported, but we do not explicitly check for this
 437          * case in favor of greatly simplified processing. */
 438         if (mb->ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK))
 439                 q->hw_ring[q->next_free].flags |= FM10K_TXD_FLAG_CSUM;
 440
 441         /* set vlan if requested */
 442         if (mb->ol_flags & PKT_TX_VLAN_PKT)
 443                 q->hw_ring[q->next_free].vlan = mb->vlan_tci;
 444
 445         q->sw_ring[q->next_free] = mb;
 446         q->hw_ring[q->next_free].buffer_addr =
 447                         rte_cpu_to_le_64(MBUF_DMA_ADDR(mb));
 448         q->hw_ring[q->next_free].buflen =
 449                         rte_cpu_to_le_16(rte_pktmbuf_data_len(mb));
 450         if (++q->next_free == q->nb_desc)
 451                 q->next_free = 0;
 452
 453         /* fill up the rings */
 454         for (mb = mb->next; mb != NULL; mb = mb->next) {
 455                 q->sw_ring[q->next_free] = mb;
 456                 q->hw_ring[q->next_free].buffer_addr =
 457                                 rte_cpu_to_le_64(MBUF_DMA_ADDR(mb));
 458                 q->hw_ring[q->next_free].buflen =
 459                                 rte_cpu_to_le_16(rte_pktmbuf_data_len(mb));
 460                 q->hw_ring[q->next_free].flags = 0;
 461                 if (++q->next_free == q->nb_desc)
 462                         q->next_free = 0;
 463         }
 464
 465         q->hw_ring[last_id].flags = flags;
 466 }
 467
 468 uint16_t
 469 fm10k_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 470         uint16_t nb_pkts)
 471 {
 472         struct fm10k_tx_queue *q = tx_queue;
 473         struct rte_mbuf *mb;
 474         uint16_t count;
 475
 476         for (count = 0; count < nb_pkts; ++count) {
 477                 mb = tx_pkts[count];
 478
 479                 /* running low on descriptors? try to free some... */
 480                 if (q->nb_free < q->free_thresh)
 481                         tx_free_descriptors(q);
 482
 483                 /* make sure there are enough free descriptors to transmit the
 484                  * entire packet before doing anything */
 485                 if (q->nb_free < mb->nb_segs)
 486                         break;
 487
 488                 /* sanity check to make sure the mbuf is valid */
 489                 if ((mb->nb_segs == 0) ||
 490                     ((mb->nb_segs > 1) && (mb->next == NULL)))
 491                         break;
 492
 493                 /* process the packet */
 494                 tx_xmit_pkt(q, mb);
 495         }
 496
 497         /* update the tail pointer if any packets were processed */
 498         if (likely(count > 0))
 499                 FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_free);
 500
 501         return count;
 502 }