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