drivers/net/liquidio/lio_rxtx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Cavium, Inc
   3  */
   4
   5 #include <ethdev_driver.h>
   6 #include <rte_cycles.h>
   7 #include <rte_malloc.h>
   8
   9 #include "lio_logs.h"
  10 #include "lio_struct.h"
  11 #include "lio_ethdev.h"
  12 #include "lio_rxtx.h"
  13
  14 #define LIO_MAX_SG 12
  15 /* Flush iq if available tx_desc fall below LIO_FLUSH_WM */
  16 #define LIO_FLUSH_WM(_iq) ((_iq)->nb_desc / 2)
  17 #define LIO_PKT_IN_DONE_CNT_MASK 0x00000000FFFFFFFFULL
  18
  19 static void
  20 lio_droq_compute_max_packet_bufs(struct lio_droq *droq)
  21 {
  22         uint32_t count = 0;
  23
  24         do {
  25                 count += droq->buffer_size;
  26         } while (count < LIO_MAX_RX_PKTLEN);
  27 }
  28
  29 static void
  30 lio_droq_reset_indices(struct lio_droq *droq)
  31 {
  32         droq->read_idx  = 0;
  33         droq->write_idx = 0;
  34         droq->refill_idx = 0;
  35         droq->refill_count = 0;
  36         rte_atomic64_set(&droq->pkts_pending, 0);
  37 }
  38
  39 static void
  40 lio_droq_destroy_ring_buffers(struct lio_droq *droq)
  41 {
  42         uint32_t i;
  43
  44         for (i = 0; i < droq->nb_desc; i++) {
  45                 if (droq->recv_buf_list[i].buffer) {
  46                         rte_pktmbuf_free((struct rte_mbuf *)
  47                                          droq->recv_buf_list[i].buffer);
  48                         droq->recv_buf_list[i].buffer = NULL;
  49                 }
  50         }
  51
  52         lio_droq_reset_indices(droq);
  53 }
  54
  55 static int
  56 lio_droq_setup_ring_buffers(struct lio_device *lio_dev,
  57                             struct lio_droq *droq)
  58 {
  59         struct lio_droq_desc *desc_ring = droq->desc_ring;
  60         uint32_t i;
  61         void *buf;
  62
  63         for (i = 0; i < droq->nb_desc; i++) {
  64                 buf = rte_pktmbuf_alloc(droq->mpool);
  65                 if (buf == NULL) {
  66                         lio_dev_err(lio_dev, "buffer alloc failed\n");
  67                         droq->stats.rx_alloc_failure++;
  68                         lio_droq_destroy_ring_buffers(droq);
  69                         return -ENOMEM;
  70                 }
  71
  72                 droq->recv_buf_list[i].buffer = buf;
  73                 droq->info_list[i].length = 0;
  74
  75                 /* map ring buffers into memory */
  76                 desc_ring[i].info_ptr = lio_map_ring_info(droq, i);
  77                 desc_ring[i].buffer_ptr =
  78                         lio_map_ring(droq->recv_buf_list[i].buffer);
  79         }
  80
  81         lio_droq_reset_indices(droq);
  82
  83         lio_droq_compute_max_packet_bufs(droq);
  84
  85         return 0;
  86 }
  87
  88 static void
  89 lio_dma_zone_free(struct lio_device *lio_dev, const struct rte_memzone *mz)
  90 {
  91         const struct rte_memzone *mz_tmp;
  92         int ret = 0;
  93
  94         if (mz == NULL) {
  95                 lio_dev_err(lio_dev, "Memzone NULL\n");
  96                 return;
  97         }
  98
  99         mz_tmp = rte_memzone_lookup(mz->name);
 100         if (mz_tmp == NULL) {
 101                 lio_dev_err(lio_dev, "Memzone %s Not Found\n", mz->name);
 102                 return;
 103         }
 104
 105         ret = rte_memzone_free(mz);
 106         if (ret)
 107                 lio_dev_err(lio_dev, "Memzone free Failed ret %d\n", ret);
 108 }
 109
 110 /**
 111  *  Frees the space for descriptor ring for the droq.
 112  *
 113  *  @param lio_dev      - pointer to the lio device structure
 114  *  @param q_no         - droq no.
 115  */
 116 static void
 117 lio_delete_droq(struct lio_device *lio_dev, uint32_t q_no)
 118 {
 119         struct lio_droq *droq = lio_dev->droq[q_no];
 120
 121         lio_dev_dbg(lio_dev, "OQ[%d]\n", q_no);
 122
 123         lio_droq_destroy_ring_buffers(droq);
 124         rte_free(droq->recv_buf_list);
 125         droq->recv_buf_list = NULL;
 126         lio_dma_zone_free(lio_dev, droq->info_mz);
 127         lio_dma_zone_free(lio_dev, droq->desc_ring_mz);
 128
 129         memset(droq, 0, LIO_DROQ_SIZE);
 130 }
 131
 132 static void *
 133 lio_alloc_info_buffer(struct lio_device *lio_dev,
 134                       struct lio_droq *droq, unsigned int socket_id)
 135 {
 136         droq->info_mz = rte_eth_dma_zone_reserve(lio_dev->eth_dev,
 137                                                  "info_list", droq->q_no,
 138                                                  (droq->nb_desc *
 139                                                         LIO_DROQ_INFO_SIZE),
 140                                                  RTE_CACHE_LINE_SIZE,
 141                                                  socket_id);
 142
 143         if (droq->info_mz == NULL)
 144                 return NULL;
 145
 146         droq->info_list_dma = droq->info_mz->iova;
 147         droq->info_alloc_size = droq->info_mz->len;
 148         droq->info_base_addr = (size_t)droq->info_mz->addr;
 149
 150         return droq->info_mz->addr;
 151 }
 152
 153 /**
 154  *  Allocates space for the descriptor ring for the droq and
 155  *  sets the base addr, num desc etc in Octeon registers.
 156  *
 157  * @param lio_dev       - pointer to the lio device structure
 158  * @param q_no          - droq no.
 159  * @param app_ctx       - pointer to application context
 160  * @return Success: 0   Failure: -1
 161  */
 162 static int
 163 lio_init_droq(struct lio_device *lio_dev, uint32_t q_no,
 164               uint32_t num_descs, uint32_t desc_size,
 165               struct rte_mempool *mpool, unsigned int socket_id)
 166 {
 167         uint32_t c_refill_threshold;
 168         uint32_t desc_ring_size;
 169         struct lio_droq *droq;
 170
 171         lio_dev_dbg(lio_dev, "OQ[%d]\n", q_no);
 172
 173         droq = lio_dev->droq[q_no];
 174         droq->lio_dev = lio_dev;
 175         droq->q_no = q_no;
 176         droq->mpool = mpool;
 177
 178         c_refill_threshold = LIO_OQ_REFILL_THRESHOLD_CFG(lio_dev);
 179
 180         droq->nb_desc = num_descs;
 181         droq->buffer_size = desc_size;
 182
 183         desc_ring_size = droq->nb_desc * LIO_DROQ_DESC_SIZE;
 184         droq->desc_ring_mz = rte_eth_dma_zone_reserve(lio_dev->eth_dev,
 185                                                       "droq", q_no,
 186                                                       desc_ring_size,
 187                                                       RTE_CACHE_LINE_SIZE,
 188                                                       socket_id);
 189
 190         if (droq->desc_ring_mz == NULL) {
 191                 lio_dev_err(lio_dev,
 192                             "Output queue %d ring alloc failed\n", q_no);
 193                 return -1;
 194         }
 195
 196         droq->desc_ring_dma = droq->desc_ring_mz->iova;
 197         droq->desc_ring = (struct lio_droq_desc *)droq->desc_ring_mz->addr;
 198
 199         lio_dev_dbg(lio_dev, "droq[%d]: desc_ring: virt: 0x%p, dma: %lx\n",
 200                     q_no, droq->desc_ring, (unsigned long)droq->desc_ring_dma);
 201         lio_dev_dbg(lio_dev, "droq[%d]: num_desc: %d\n", q_no,
 202                     droq->nb_desc);
 203
 204         droq->info_list = lio_alloc_info_buffer(lio_dev, droq, socket_id);
 205         if (droq->info_list == NULL) {
 206                 lio_dev_err(lio_dev, "Cannot allocate memory for info list.\n");
 207                 goto init_droq_fail;
 208         }
 209
 210         droq->recv_buf_list = rte_zmalloc_socket("recv_buf_list",
 211                                                  (droq->nb_desc *
 212                                                         LIO_DROQ_RECVBUF_SIZE),
 213                                                  RTE_CACHE_LINE_SIZE,
 214                                                  socket_id);
 215         if (droq->recv_buf_list == NULL) {
 216                 lio_dev_err(lio_dev,
 217                             "Output queue recv buf list alloc failed\n");
 218                 goto init_droq_fail;
 219         }
 220
 221         if (lio_droq_setup_ring_buffers(lio_dev, droq))
 222                 goto init_droq_fail;
 223
 224         droq->refill_threshold = c_refill_threshold;
 225
 226         rte_spinlock_init(&droq->lock);
 227
 228         lio_dev->fn_list.setup_oq_regs(lio_dev, q_no);
 229
 230         lio_dev->io_qmask.oq |= (1ULL << q_no);
 231
 232         return 0;
 233
 234 init_droq_fail:
 235         lio_delete_droq(lio_dev, q_no);
 236
 237         return -1;
 238 }
 239
 240 int
 241 lio_setup_droq(struct lio_device *lio_dev, int oq_no, int num_descs,
 242                int desc_size, struct rte_mempool *mpool, unsigned int socket_id)
 243 {
 244         struct lio_droq *droq;
 245
 246         PMD_INIT_FUNC_TRACE();
 247
 248         /* Allocate the DS for the new droq. */
 249         droq = rte_zmalloc_socket("ethdev RX queue", sizeof(*droq),
 250                                   RTE_CACHE_LINE_SIZE, socket_id);
 251         if (droq == NULL)
 252                 return -ENOMEM;
 253
 254         lio_dev->droq[oq_no] = droq;
 255
 256         /* Initialize the Droq */
 257         if (lio_init_droq(lio_dev, oq_no, num_descs, desc_size, mpool,
 258                           socket_id)) {
 259                 lio_dev_err(lio_dev, "Droq[%u] Initialization Failed\n", oq_no);
 260                 rte_free(lio_dev->droq[oq_no]);
 261                 lio_dev->droq[oq_no] = NULL;
 262                 return -ENOMEM;
 263         }
 264
 265         lio_dev->num_oqs++;
 266
 267         lio_dev_dbg(lio_dev, "Total number of OQ: %d\n", lio_dev->num_oqs);
 268
 269         /* Send credit for octeon output queues. credits are always
 270          * sent after the output queue is enabled.
 271          */
 272         rte_write32(lio_dev->droq[oq_no]->nb_desc,
 273                     lio_dev->droq[oq_no]->pkts_credit_reg);
 274         rte_wmb();
 275
 276         return 0;
 277 }
 278
 279 static inline uint32_t
 280 lio_droq_get_bufcount(uint32_t buf_size, uint32_t total_len)
 281 {
 282         uint32_t buf_cnt = 0;
 283
 284         while (total_len > (buf_size * buf_cnt))
 285                 buf_cnt++;
 286
 287         return buf_cnt;
 288 }
 289
 290 /* If we were not able to refill all buffers, try to move around
 291  * the buffers that were not dispatched.
 292  */
 293 static inline uint32_t
 294 lio_droq_refill_pullup_descs(struct lio_droq *droq,
 295                              struct lio_droq_desc *desc_ring)
 296 {
 297         uint32_t refill_index = droq->refill_idx;
 298         uint32_t desc_refilled = 0;
 299
 300         while (refill_index != droq->read_idx) {
 301                 if (droq->recv_buf_list[refill_index].buffer) {
 302                         droq->recv_buf_list[droq->refill_idx].buffer =
 303                                 droq->recv_buf_list[refill_index].buffer;
 304                         desc_ring[droq->refill_idx].buffer_ptr =
 305                                 desc_ring[refill_index].buffer_ptr;
 306                         droq->recv_buf_list[refill_index].buffer = NULL;
 307                         desc_ring[refill_index].buffer_ptr = 0;
 308                         do {
 309                                 droq->refill_idx = lio_incr_index(
 310                                                         droq->refill_idx, 1,
 311                                                         droq->nb_desc);
 312                                 desc_refilled++;
 313                                 droq->refill_count--;
 314                         } while (droq->recv_buf_list[droq->refill_idx].buffer);
 315                 }
 316                 refill_index = lio_incr_index(refill_index, 1,
 317                                               droq->nb_desc);
 318         }       /* while */
 319
 320         return desc_refilled;
 321 }
 322
 323 /* lio_droq_refill
 324  *
 325  * @param droq          - droq in which descriptors require new buffers.
 326  *
 327  * Description:
 328  *  Called during normal DROQ processing in interrupt mode or by the poll
 329  *  thread to refill the descriptors from which buffers were dispatched
 330  *  to upper layers. Attempts to allocate new buffers. If that fails, moves
 331  *  up buffers (that were not dispatched) to form a contiguous ring.
 332  *
 333  * Returns:
 334  *  No of descriptors refilled.
 335  *
 336  * Locks:
 337  * This routine is called with droq->lock held.
 338  */
 339 static uint32_t
 340 lio_droq_refill(struct lio_droq *droq)
 341 {
 342         struct lio_droq_desc *desc_ring;
 343         uint32_t desc_refilled = 0;
 344         void *buf = NULL;
 345
 346         desc_ring = droq->desc_ring;
 347
 348         while (droq->refill_count && (desc_refilled < droq->nb_desc)) {
 349                 /* If a valid buffer exists (happens if there is no dispatch),
 350                  * reuse the buffer, else allocate.
 351                  */
 352                 if (droq->recv_buf_list[droq->refill_idx].buffer == NULL) {
 353                         buf = rte_pktmbuf_alloc(droq->mpool);
 354                         /* If a buffer could not be allocated, no point in
 355                          * continuing
 356                          */
 357                         if (buf == NULL) {
 358                                 droq->stats.rx_alloc_failure++;
 359                                 break;
 360                         }
 361
 362                         droq->recv_buf_list[droq->refill_idx].buffer = buf;
 363                 }
 364
 365                 desc_ring[droq->refill_idx].buffer_ptr =
 366                     lio_map_ring(droq->recv_buf_list[droq->refill_idx].buffer);
 367                 /* Reset any previous values in the length field. */
 368                 droq->info_list[droq->refill_idx].length = 0;
 369
 370                 droq->refill_idx = lio_incr_index(droq->refill_idx, 1,
 371                                                   droq->nb_desc);
 372                 desc_refilled++;
 373                 droq->refill_count--;
 374         }
 375
 376         if (droq->refill_count)
 377                 desc_refilled += lio_droq_refill_pullup_descs(droq, desc_ring);
 378
 379         /* if droq->refill_count
 380          * The refill count would not change in pass two. We only moved buffers
 381          * to close the gap in the ring, but we would still have the same no. of
 382          * buffers to refill.
 383          */
 384         return desc_refilled;
 385 }
 386
 387 static int
 388 lio_droq_fast_process_packet(struct lio_device *lio_dev,
 389                              struct lio_droq *droq,
 390                              struct rte_mbuf **rx_pkts)
 391 {
 392         struct rte_mbuf *nicbuf = NULL;
 393         struct lio_droq_info *info;
 394         uint32_t total_len = 0;
 395         int data_total_len = 0;
 396         uint32_t pkt_len = 0;
 397         union octeon_rh *rh;
 398         int data_pkts = 0;
 399
 400         info = &droq->info_list[droq->read_idx];
 401         lio_swap_8B_data((uint64_t *)info, 2);
 402
 403         if (!info->length)
 404                 return -1;
 405
 406         /* Len of resp hdr in included in the received data len. */
 407         info->length -= OCTEON_RH_SIZE;
 408         rh = &info->rh;
 409
 410         total_len += (uint32_t)info->length;
 411
 412         if (lio_opcode_slow_path(rh)) {
 413                 uint32_t buf_cnt;
 414
 415                 buf_cnt = lio_droq_get_bufcount(droq->buffer_size,
 416                                                 (uint32_t)info->length);
 417                 droq->read_idx = lio_incr_index(droq->read_idx, buf_cnt,
 418                                                 droq->nb_desc);
 419                 droq->refill_count += buf_cnt;
 420         } else {
 421                 if (info->length <= droq->buffer_size) {
 422                         if (rh->r_dh.has_hash)
 423                                 pkt_len = (uint32_t)(info->length - 8);
 424                         else
 425                                 pkt_len = (uint32_t)info->length;
 426
 427                         nicbuf = droq->recv_buf_list[droq->read_idx].buffer;
 428                         droq->recv_buf_list[droq->read_idx].buffer = NULL;
 429                         droq->read_idx = lio_incr_index(
 430                                                 droq->read_idx, 1,
 431                                                 droq->nb_desc);
 432                         droq->refill_count++;
 433
 434                         if (likely(nicbuf != NULL)) {
 435                                 /* We don't have a way to pass flags yet */
 436                                 nicbuf->ol_flags = 0;
 437                                 if (rh->r_dh.has_hash) {
 438                                         uint64_t *hash_ptr;
 439
 440                                         nicbuf->ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
 441                                         hash_ptr = rte_pktmbuf_mtod(nicbuf,
 442                                                                     uint64_t *);
 443                                         lio_swap_8B_data(hash_ptr, 1);
 444                                         nicbuf->hash.rss = (uint32_t)*hash_ptr;
 445                                         nicbuf->data_off += 8;
 446                                 }
 447
 448                                 nicbuf->pkt_len = pkt_len;
 449                                 nicbuf->data_len = pkt_len;
 450                                 nicbuf->port = lio_dev->port_id;
 451                                 /* Store the mbuf */
 452                                 rx_pkts[data_pkts++] = nicbuf;
 453                                 data_total_len += pkt_len;
 454                         }
 455
 456                         /* Prefetch buffer pointers when on a cache line
 457                          * boundary
 458                          */
 459                         if ((droq->read_idx & 3) == 0) {
 460                                 rte_prefetch0(
 461                                     &droq->recv_buf_list[droq->read_idx]);
 462                                 rte_prefetch0(
 463                                     &droq->info_list[droq->read_idx]);
 464                         }
 465                 } else {
 466                         struct rte_mbuf *first_buf = NULL;
 467                         struct rte_mbuf *last_buf = NULL;
 468
 469                         while (pkt_len < info->length) {
 470                                 int cpy_len = 0;
 471
 472                                 cpy_len = ((pkt_len + droq->buffer_size) >
 473                                                 info->length)
 474                                                 ? ((uint32_t)info->length -
 475                                                         pkt_len)
 476                                                 : droq->buffer_size;
 477
 478                                 nicbuf =
 479                                     droq->recv_buf_list[droq->read_idx].buffer;
 480                                 droq->recv_buf_list[droq->read_idx].buffer =
 481                                     NULL;
 482
 483                                 if (likely(nicbuf != NULL)) {
 484                                         /* Note the first seg */
 485                                         if (!pkt_len)
 486                                                 first_buf = nicbuf;
 487
 488                                         nicbuf->port = lio_dev->port_id;
 489                                         /* We don't have a way to pass
 490                                          * flags yet
 491                                          */
 492                                         nicbuf->ol_flags = 0;
 493                                         if ((!pkt_len) && (rh->r_dh.has_hash)) {
 494                                                 uint64_t *hash_ptr;
 495
 496                                                 nicbuf->ol_flags |=
 497                                                     RTE_MBUF_F_RX_RSS_HASH;
 498                                                 hash_ptr = rte_pktmbuf_mtod(
 499                                                     nicbuf, uint64_t *);
 500                                                 lio_swap_8B_data(hash_ptr, 1);
 501                                                 nicbuf->hash.rss =
 502                                                     (uint32_t)*hash_ptr;
 503                                                 nicbuf->data_off += 8;
 504                                                 nicbuf->pkt_len = cpy_len - 8;
 505                                                 nicbuf->data_len = cpy_len - 8;
 506                                         } else {
 507                                                 nicbuf->pkt_len = cpy_len;
 508                                                 nicbuf->data_len = cpy_len;
 509                                         }
 510
 511                                         if (pkt_len)
 512                                                 first_buf->nb_segs++;
 513
 514                                         if (last_buf)
 515                                                 last_buf->next = nicbuf;
 516
 517                                         last_buf = nicbuf;
 518                                 } else {
 519                                         PMD_RX_LOG(lio_dev, ERR, "no buf\n");
 520                                 }
 521
 522                                 pkt_len += cpy_len;
 523                                 droq->read_idx = lio_incr_index(
 524                                                         droq->read_idx,
 525                                                         1, droq->nb_desc);
 526                                 droq->refill_count++;
 527
 528                                 /* Prefetch buffer pointers when on a
 529                                  * cache line boundary
 530                                  */
 531                                 if ((droq->read_idx & 3) == 0) {
 532                                         rte_prefetch0(&droq->recv_buf_list
 533                                                               [droq->read_idx]);
 534
 535                                         rte_prefetch0(
 536                                             &droq->info_list[droq->read_idx]);
 537                                 }
 538                         }
 539                         rx_pkts[data_pkts++] = first_buf;
 540                         if (rh->r_dh.has_hash)
 541                                 data_total_len += (pkt_len - 8);
 542                         else
 543                                 data_total_len += pkt_len;
 544                 }
 545
 546                 /* Inform upper layer about packet checksum verification */
 547                 struct rte_mbuf *m = rx_pkts[data_pkts - 1];
 548
 549                 if (rh->r_dh.csum_verified & LIO_IP_CSUM_VERIFIED)
 550                         m->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_GOOD;
 551
 552                 if (rh->r_dh.csum_verified & LIO_L4_CSUM_VERIFIED)
 553                         m->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_GOOD;
 554         }
 555
 556         if (droq->refill_count >= droq->refill_threshold) {
 557                 int desc_refilled = lio_droq_refill(droq);
 558
 559                 /* Flush the droq descriptor data to memory to be sure
 560                  * that when we update the credits the data in memory is
 561                  * accurate.
 562                  */
 563                 rte_wmb();
 564                 rte_write32(desc_refilled, droq->pkts_credit_reg);
 565                 /* make sure mmio write completes */
 566                 rte_wmb();
 567         }
 568
 569         info->length = 0;
 570         info->rh.rh64 = 0;
 571
 572         droq->stats.pkts_received++;
 573         droq->stats.rx_pkts_received += data_pkts;
 574         droq->stats.rx_bytes_received += data_total_len;
 575         droq->stats.bytes_received += total_len;
 576
 577         return data_pkts;
 578 }
 579
 580 static uint32_t
 581 lio_droq_fast_process_packets(struct lio_device *lio_dev,
 582                               struct lio_droq *droq,
 583                               struct rte_mbuf **rx_pkts,
 584                               uint32_t pkts_to_process)
 585 {
 586         int ret, data_pkts = 0;
 587         uint32_t pkt;
 588
 589         for (pkt = 0; pkt < pkts_to_process; pkt++) {
 590                 ret = lio_droq_fast_process_packet(lio_dev, droq,
 591                                                    &rx_pkts[data_pkts]);
 592                 if (ret < 0) {
 593                         lio_dev_err(lio_dev, "Port[%d] DROQ[%d] idx: %d len:0, pkt_cnt: %d\n",
 594                                     lio_dev->port_id, droq->q_no,
 595                                     droq->read_idx, pkts_to_process);
 596                         break;
 597                 }
 598                 data_pkts += ret;
 599         }
 600
 601         rte_atomic64_sub(&droq->pkts_pending, pkt);
 602
 603         return data_pkts;
 604 }
 605
 606 static inline uint32_t
 607 lio_droq_check_hw_for_pkts(struct lio_droq *droq)
 608 {
 609         uint32_t last_count;
 610         uint32_t pkt_count;
 611
 612         pkt_count = rte_read32(droq->pkts_sent_reg);
 613
 614         last_count = pkt_count - droq->pkt_count;
 615         droq->pkt_count = pkt_count;
 616
 617         if (last_count)
 618                 rte_atomic64_add(&droq->pkts_pending, last_count);
 619
 620         return last_count;
 621 }
 622
 623 uint16_t
 624 lio_dev_recv_pkts(void *rx_queue,
 625                   struct rte_mbuf **rx_pkts,
 626                   uint16_t budget)
 627 {
 628         struct lio_droq *droq = rx_queue;
 629         struct lio_device *lio_dev = droq->lio_dev;
 630         uint32_t pkts_processed = 0;
 631         uint32_t pkt_count = 0;
 632
 633         lio_droq_check_hw_for_pkts(droq);
 634
 635         pkt_count = rte_atomic64_read(&droq->pkts_pending);
 636         if (!pkt_count)
 637                 return 0;
 638
 639         if (pkt_count > budget)
 640                 pkt_count = budget;
 641
 642         /* Grab the lock */
 643         rte_spinlock_lock(&droq->lock);
 644         pkts_processed = lio_droq_fast_process_packets(lio_dev,
 645                                                        droq, rx_pkts,
 646                                                        pkt_count);
 647
 648         if (droq->pkt_count) {
 649                 rte_write32(droq->pkt_count, droq->pkts_sent_reg);
 650                 droq->pkt_count = 0;
 651         }
 652
 653         /* Release the spin lock */
 654         rte_spinlock_unlock(&droq->lock);
 655
 656         return pkts_processed;
 657 }
 658
 659 void
 660 lio_delete_droq_queue(struct lio_device *lio_dev,
 661                       int oq_no)
 662 {
 663         lio_delete_droq(lio_dev, oq_no);
 664         lio_dev->num_oqs--;
 665         rte_free(lio_dev->droq[oq_no]);
 666         lio_dev->droq[oq_no] = NULL;
 667 }
 668
 669 /**
 670  *  lio_init_instr_queue()
 671  *  @param lio_dev      - pointer to the lio device structure.
 672  *  @param txpciq       - queue to be initialized.
 673  *
 674  *  Called at driver init time for each input queue. iq_conf has the
 675  *  configuration parameters for the queue.
 676  *
 677  *  @return  Success: 0 Failure: -1
 678  */
 679 static int
 680 lio_init_instr_queue(struct lio_device *lio_dev,
 681                      union octeon_txpciq txpciq,
 682                      uint32_t num_descs, unsigned int socket_id)
 683 {
 684         uint32_t iq_no = (uint32_t)txpciq.s.q_no;
 685         struct lio_instr_queue *iq;
 686         uint32_t instr_type;
 687         uint32_t q_size;
 688
 689         instr_type = LIO_IQ_INSTR_TYPE(lio_dev);
 690
 691         q_size = instr_type * num_descs;
 692         iq = lio_dev->instr_queue[iq_no];
 693         iq->iq_mz = rte_eth_dma_zone_reserve(lio_dev->eth_dev,
 694                                              "instr_queue", iq_no, q_size,
 695                                              RTE_CACHE_LINE_SIZE,
 696                                              socket_id);
 697         if (iq->iq_mz == NULL) {
 698                 lio_dev_err(lio_dev, "Cannot allocate memory for instr queue %d\n",
 699                             iq_no);
 700                 return -1;
 701         }
 702
 703         iq->base_addr_dma = iq->iq_mz->iova;
 704         iq->base_addr = (uint8_t *)iq->iq_mz->addr;
 705
 706         iq->nb_desc = num_descs;
 707
 708         /* Initialize a list to holds requests that have been posted to Octeon
 709          * but has yet to be fetched by octeon
 710          */
 711         iq->request_list = rte_zmalloc_socket("request_list",
 712                                               sizeof(*iq->request_list) *
 713                                                         num_descs,
 714                                               RTE_CACHE_LINE_SIZE,
 715                                               socket_id);
 716         if (iq->request_list == NULL) {
 717                 lio_dev_err(lio_dev, "Alloc failed for IQ[%d] nr free list\n",
 718                             iq_no);
 719                 lio_dma_zone_free(lio_dev, iq->iq_mz);
 720                 return -1;
 721         }
 722
 723         lio_dev_dbg(lio_dev, "IQ[%d]: base: %p basedma: %lx count: %d\n",
 724                     iq_no, iq->base_addr, (unsigned long)iq->base_addr_dma,
 725                     iq->nb_desc);
 726
 727         iq->lio_dev = lio_dev;
 728         iq->txpciq.txpciq64 = txpciq.txpciq64;
 729         iq->fill_cnt = 0;
 730         iq->host_write_index = 0;
 731         iq->lio_read_index = 0;
 732         iq->flush_index = 0;
 733
 734         rte_atomic64_set(&iq->instr_pending, 0);
 735
 736         /* Initialize the spinlock for this instruction queue */
 737         rte_spinlock_init(&iq->lock);
 738         rte_spinlock_init(&iq->post_lock);
 739
 740         rte_atomic64_clear(&iq->iq_flush_running);
 741
 742         lio_dev->io_qmask.iq |= (1ULL << iq_no);
 743
 744         /* Set the 32B/64B mode for each input queue */
 745         lio_dev->io_qmask.iq64B |= ((instr_type == 64) << iq_no);
 746         iq->iqcmd_64B = (instr_type == 64);
 747
 748         lio_dev->fn_list.setup_iq_regs(lio_dev, iq_no);
 749
 750         return 0;
 751 }
 752
 753 int
 754 lio_setup_instr_queue0(struct lio_device *lio_dev)
 755 {
 756         union octeon_txpciq txpciq;
 757         uint32_t num_descs = 0;
 758         uint32_t iq_no = 0;
 759
 760         num_descs = LIO_NUM_DEF_TX_DESCS_CFG(lio_dev);
 761
 762         lio_dev->num_iqs = 0;
 763
 764         lio_dev->instr_queue[0] = rte_zmalloc(NULL,
 765                                         sizeof(struct lio_instr_queue), 0);
 766         if (lio_dev->instr_queue[0] == NULL)
 767                 return -ENOMEM;
 768
 769         lio_dev->instr_queue[0]->q_index = 0;
 770         lio_dev->instr_queue[0]->app_ctx = (void *)(size_t)0;
 771         txpciq.txpciq64 = 0;
 772         txpciq.s.q_no = iq_no;
 773         txpciq.s.pkind = lio_dev->pfvf_hsword.pkind;
 774         txpciq.s.use_qpg = 0;
 775         txpciq.s.qpg = 0;
 776         if (lio_init_instr_queue(lio_dev, txpciq, num_descs, SOCKET_ID_ANY)) {
 777                 rte_free(lio_dev->instr_queue[0]);
 778                 lio_dev->instr_queue[0] = NULL;
 779                 return -1;
 780         }
 781
 782         lio_dev->num_iqs++;
 783
 784         return 0;
 785 }
 786
 787 /**
 788  *  lio_delete_instr_queue()
 789  *  @param lio_dev      - pointer to the lio device structure.
 790  *  @param iq_no        - queue to be deleted.
 791  *
 792  *  Called at driver unload time for each input queue. Deletes all
 793  *  allocated resources for the input queue.
 794  */
 795 static void
 796 lio_delete_instr_queue(struct lio_device *lio_dev, uint32_t iq_no)
 797 {
 798         struct lio_instr_queue *iq = lio_dev->instr_queue[iq_no];
 799
 800         rte_free(iq->request_list);
 801         iq->request_list = NULL;
 802         lio_dma_zone_free(lio_dev, iq->iq_mz);
 803 }
 804
 805 void
 806 lio_free_instr_queue0(struct lio_device *lio_dev)
 807 {
 808         lio_delete_instr_queue(lio_dev, 0);
 809         rte_free(lio_dev->instr_queue[0]);
 810         lio_dev->instr_queue[0] = NULL;
 811         lio_dev->num_iqs--;
 812 }
 813
 814 /* Return 0 on success, -1 on failure */
 815 int
 816 lio_setup_iq(struct lio_device *lio_dev, int q_index,
 817              union octeon_txpciq txpciq, uint32_t num_descs, void *app_ctx,
 818              unsigned int socket_id)
 819 {
 820         uint32_t iq_no = (uint32_t)txpciq.s.q_no;
 821
 822         lio_dev->instr_queue[iq_no] = rte_zmalloc_socket("ethdev TX queue",
 823                                                 sizeof(struct lio_instr_queue),
 824                                                 RTE_CACHE_LINE_SIZE, socket_id);
 825         if (lio_dev->instr_queue[iq_no] == NULL)
 826                 return -1;
 827
 828         lio_dev->instr_queue[iq_no]->q_index = q_index;
 829         lio_dev->instr_queue[iq_no]->app_ctx = app_ctx;
 830
 831         if (lio_init_instr_queue(lio_dev, txpciq, num_descs, socket_id)) {
 832                 rte_free(lio_dev->instr_queue[iq_no]);
 833                 lio_dev->instr_queue[iq_no] = NULL;
 834                 return -1;
 835         }
 836
 837         lio_dev->num_iqs++;
 838
 839         return 0;
 840 }
 841
 842 int
 843 lio_wait_for_instr_fetch(struct lio_device *lio_dev)
 844 {
 845         int pending, instr_cnt;
 846         int i, retry = 1000;
 847
 848         do {
 849                 instr_cnt = 0;
 850
 851                 for (i = 0; i < LIO_MAX_INSTR_QUEUES(lio_dev); i++) {
 852                         if (!(lio_dev->io_qmask.iq & (1ULL << i)))
 853                                 continue;
 854
 855                         if (lio_dev->instr_queue[i] == NULL)
 856                                 break;
 857
 858                         pending = rte_atomic64_read(
 859                             &lio_dev->instr_queue[i]->instr_pending);
 860                         if (pending)
 861                                 lio_flush_iq(lio_dev, lio_dev->instr_queue[i]);
 862
 863                         instr_cnt += pending;
 864                 }
 865
 866                 if (instr_cnt == 0)
 867                         break;
 868
 869                 rte_delay_ms(1);
 870
 871         } while (retry-- && instr_cnt);
 872
 873         return instr_cnt;
 874 }
 875
 876 static inline void
 877 lio_ring_doorbell(struct lio_device *lio_dev,
 878                   struct lio_instr_queue *iq)
 879 {
 880         if (rte_atomic64_read(&lio_dev->status) == LIO_DEV_RUNNING) {
 881                 rte_write32(iq->fill_cnt, iq->doorbell_reg);
 882                 /* make sure doorbell write goes through */
 883                 rte_wmb();
 884                 iq->fill_cnt = 0;
 885         }
 886 }
 887
 888 static inline void
 889 copy_cmd_into_iq(struct lio_instr_queue *iq, uint8_t *cmd)
 890 {
 891         uint8_t *iqptr, cmdsize;
 892
 893         cmdsize = ((iq->iqcmd_64B) ? 64 : 32);
 894         iqptr = iq->base_addr + (cmdsize * iq->host_write_index);
 895
 896         rte_memcpy(iqptr, cmd, cmdsize);
 897 }
 898
 899 static inline struct lio_iq_post_status
 900 post_command2(struct lio_instr_queue *iq, uint8_t *cmd)
 901 {
 902         struct lio_iq_post_status st;
 903
 904         st.status = LIO_IQ_SEND_OK;
 905
 906         /* This ensures that the read index does not wrap around to the same
 907          * position if queue gets full before Octeon could fetch any instr.
 908          */
 909         if (rte_atomic64_read(&iq->instr_pending) >=
 910                         (int32_t)(iq->nb_desc - 1)) {
 911                 st.status = LIO_IQ_SEND_FAILED;
 912                 st.index = -1;
 913                 return st;
 914         }
 915
 916         if (rte_atomic64_read(&iq->instr_pending) >=
 917                         (int32_t)(iq->nb_desc - 2))
 918                 st.status = LIO_IQ_SEND_STOP;
 919
 920         copy_cmd_into_iq(iq, cmd);
 921
 922         /* "index" is returned, host_write_index is modified. */
 923         st.index = iq->host_write_index;
 924         iq->host_write_index = lio_incr_index(iq->host_write_index, 1,
 925                                               iq->nb_desc);
 926         iq->fill_cnt++;
 927
 928         /* Flush the command into memory. We need to be sure the data is in
 929          * memory before indicating that the instruction is pending.
 930          */
 931         rte_wmb();
 932
 933         rte_atomic64_inc(&iq->instr_pending);
 934
 935         return st;
 936 }
 937
 938 static inline void
 939 lio_add_to_request_list(struct lio_instr_queue *iq,
 940                         int idx, void *buf, int reqtype)
 941 {
 942         iq->request_list[idx].buf = buf;
 943         iq->request_list[idx].reqtype = reqtype;
 944 }
 945
 946 static inline void
 947 lio_free_netsgbuf(void *buf)
 948 {
 949         struct lio_buf_free_info *finfo = buf;
 950         struct lio_device *lio_dev = finfo->lio_dev;
 951         struct rte_mbuf *m = finfo->mbuf;
 952         struct lio_gather *g = finfo->g;
 953         uint8_t iq = finfo->iq_no;
 954
 955         /* This will take care of multiple segments also */
 956         rte_pktmbuf_free(m);
 957
 958         rte_spinlock_lock(&lio_dev->glist_lock[iq]);
 959         STAILQ_INSERT_TAIL(&lio_dev->glist_head[iq], &g->list, entries);
 960         rte_spinlock_unlock(&lio_dev->glist_lock[iq]);
 961         rte_free(finfo);
 962 }
 963
 964 /* Can only run in process context */
 965 static int
 966 lio_process_iq_request_list(struct lio_device *lio_dev,
 967                             struct lio_instr_queue *iq)
 968 {
 969         struct octeon_instr_irh *irh = NULL;
 970         uint32_t old = iq->flush_index;
 971         struct lio_soft_command *sc;
 972         uint32_t inst_count = 0;
 973         int reqtype;
 974         void *buf;
 975
 976         while (old != iq->lio_read_index) {
 977                 reqtype = iq->request_list[old].reqtype;
 978                 buf     = iq->request_list[old].buf;
 979
 980                 if (reqtype == LIO_REQTYPE_NONE)
 981                         goto skip_this;
 982
 983                 switch (reqtype) {
 984                 case LIO_REQTYPE_NORESP_NET:
 985                         rte_pktmbuf_free((struct rte_mbuf *)buf);
 986                         break;
 987                 case LIO_REQTYPE_NORESP_NET_SG:
 988                         lio_free_netsgbuf(buf);
 989                         break;
 990                 case LIO_REQTYPE_SOFT_COMMAND:
 991                         sc = buf;
 992                         irh = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh;
 993                         if (irh->rflag) {
 994                                 /* We're expecting a response from Octeon.
 995                                  * It's up to lio_process_ordered_list() to
 996                                  * process sc. Add sc to the ordered soft
 997                                  * command response list because we expect
 998                                  * a response from Octeon.
 999                                  */
1000                                 rte_spinlock_lock(&lio_dev->response_list.lock);
1001                                 rte_atomic64_inc(
1002                                     &lio_dev->response_list.pending_req_count);
1003                                 STAILQ_INSERT_TAIL(
1004                                         &lio_dev->response_list.head,
1005                                         &sc->node, entries);
1006                                 rte_spinlock_unlock(
1007                                                 &lio_dev->response_list.lock);
1008                         } else {
1009                                 if (sc->callback) {
1010                                         /* This callback must not sleep */
1011                                         sc->callback(LIO_REQUEST_DONE,
1012                                                      sc->callback_arg);
1013                                 }
1014                         }
1015                         break;
1016                 default:
1017                         lio_dev_err(lio_dev,
1018                                     "Unknown reqtype: %d buf: %p at idx %d\n",
1019                                     reqtype, buf, old);
1020                 }
1021
1022                 iq->request_list[old].buf = NULL;
1023                 iq->request_list[old].reqtype = 0;
1024
1025 skip_this:
1026                 inst_count++;
1027                 old = lio_incr_index(old, 1, iq->nb_desc);
1028         }
1029
1030         iq->flush_index = old;
1031
1032         return inst_count;
1033 }
1034
1035 static void
1036 lio_update_read_index(struct lio_instr_queue *iq)
1037 {
1038         uint32_t pkt_in_done = rte_read32(iq->inst_cnt_reg);
1039         uint32_t last_done;
1040
1041         last_done = pkt_in_done - iq->pkt_in_done;
1042         iq->pkt_in_done = pkt_in_done;
1043
1044         /* Add last_done and modulo with the IQ size to get new index */
1045         iq->lio_read_index = (iq->lio_read_index +
1046                         (uint32_t)(last_done & LIO_PKT_IN_DONE_CNT_MASK)) %
1047                         iq->nb_desc;
1048 }
1049
1050 int
1051 lio_flush_iq(struct lio_device *lio_dev, struct lio_instr_queue *iq)
1052 {
1053         uint32_t inst_processed = 0;
1054         int tx_done = 1;
1055
1056         if (rte_atomic64_test_and_set(&iq->iq_flush_running) == 0)
1057                 return tx_done;
1058
1059         rte_spinlock_lock(&iq->lock);
1060
1061         lio_update_read_index(iq);
1062
1063         do {
1064                 /* Process any outstanding IQ packets. */
1065                 if (iq->flush_index == iq->lio_read_index)
1066                         break;
1067
1068                 inst_processed = lio_process_iq_request_list(lio_dev, iq);
1069
1070                 if (inst_processed) {
1071                         rte_atomic64_sub(&iq->instr_pending, inst_processed);
1072                         iq->stats.instr_processed += inst_processed;
1073                 }
1074
1075                 inst_processed = 0;
1076
1077         } while (1);
1078
1079         rte_spinlock_unlock(&iq->lock);
1080
1081         rte_atomic64_clear(&iq->iq_flush_running);
1082
1083         return tx_done;
1084 }
1085
1086 static int
1087 lio_send_command(struct lio_device *lio_dev, uint32_t iq_no, void *cmd,
1088                  void *buf, uint32_t datasize, uint32_t reqtype)
1089 {
1090         struct lio_instr_queue *iq = lio_dev->instr_queue[iq_no];
1091         struct lio_iq_post_status st;
1092
1093         rte_spinlock_lock(&iq->post_lock);
1094
1095         st = post_command2(iq, cmd);
1096
1097         if (st.status != LIO_IQ_SEND_FAILED) {
1098                 lio_add_to_request_list(iq, st.index, buf, reqtype);
1099                 LIO_INCR_INSTRQUEUE_PKT_COUNT(lio_dev, iq_no, bytes_sent,
1100                                               datasize);
1101                 LIO_INCR_INSTRQUEUE_PKT_COUNT(lio_dev, iq_no, instr_posted, 1);
1102
1103                 lio_ring_doorbell(lio_dev, iq);
1104         } else {
1105                 LIO_INCR_INSTRQUEUE_PKT_COUNT(lio_dev, iq_no, instr_dropped, 1);
1106         }
1107
1108         rte_spinlock_unlock(&iq->post_lock);
1109
1110         return st.status;
1111 }
1112
1113 void
1114 lio_prepare_soft_command(struct lio_device *lio_dev,
1115                          struct lio_soft_command *sc, uint8_t opcode,
1116                          uint8_t subcode, uint32_t irh_ossp, uint64_t ossp0,
1117                          uint64_t ossp1)
1118 {
1119         struct octeon_instr_pki_ih3 *pki_ih3;
1120         struct octeon_instr_ih3 *ih3;
1121         struct octeon_instr_irh *irh;
1122         struct octeon_instr_rdp *rdp;
1123
1124         RTE_ASSERT(opcode <= 15);
1125         RTE_ASSERT(subcode <= 127);
1126
1127         ih3       = (struct octeon_instr_ih3 *)&sc->cmd.cmd3.ih3;
1128
1129         ih3->pkind = lio_dev->instr_queue[sc->iq_no]->txpciq.s.pkind;
1130
1131         pki_ih3 = (struct octeon_instr_pki_ih3 *)&sc->cmd.cmd3.pki_ih3;
1132
1133         pki_ih3->w      = 1;
1134         pki_ih3->raw    = 1;
1135         pki_ih3->utag   = 1;
1136         pki_ih3->uqpg   = lio_dev->instr_queue[sc->iq_no]->txpciq.s.use_qpg;
1137         pki_ih3->utt    = 1;
1138
1139         pki_ih3->tag    = LIO_CONTROL;
1140         pki_ih3->tagtype = OCTEON_ATOMIC_TAG;
1141         pki_ih3->qpg    = lio_dev->instr_queue[sc->iq_no]->txpciq.s.qpg;
1142         pki_ih3->pm     = 0x7;
1143         pki_ih3->sl     = 8;
1144
1145         if (sc->datasize)
1146                 ih3->dlengsz = sc->datasize;
1147
1148         irh             = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh;
1149         irh->opcode     = opcode;
1150         irh->subcode    = subcode;
1151
1152         /* opcode/subcode specific parameters (ossp) */
1153         irh->ossp = irh_ossp;
1154         sc->cmd.cmd3.ossp[0] = ossp0;
1155         sc->cmd.cmd3.ossp[1] = ossp1;
1156
1157         if (sc->rdatasize) {
1158                 rdp = (struct octeon_instr_rdp *)&sc->cmd.cmd3.rdp;
1159                 rdp->pcie_port = lio_dev->pcie_port;
1160                 rdp->rlen      = sc->rdatasize;
1161                 irh->rflag = 1;
1162                 /* PKI IH3 */
1163                 ih3->fsz    = OCTEON_SOFT_CMD_RESP_IH3;
1164         } else {
1165                 irh->rflag = 0;
1166                 /* PKI IH3 */
1167                 ih3->fsz    = OCTEON_PCI_CMD_O3;
1168         }
1169 }
1170
1171 int
1172 lio_send_soft_command(struct lio_device *lio_dev,
1173                       struct lio_soft_command *sc)
1174 {
1175         struct octeon_instr_ih3 *ih3;
1176         struct octeon_instr_irh *irh;
1177         uint32_t len = 0;
1178
1179         ih3 = (struct octeon_instr_ih3 *)&sc->cmd.cmd3.ih3;
1180         if (ih3->dlengsz) {
1181                 RTE_ASSERT(sc->dmadptr);
1182                 sc->cmd.cmd3.dptr = sc->dmadptr;
1183         }
1184
1185         irh = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh;
1186         if (irh->rflag) {
1187                 RTE_ASSERT(sc->dmarptr);
1188                 RTE_ASSERT(sc->status_word != NULL);
1189                 *sc->status_word = LIO_COMPLETION_WORD_INIT;
1190                 sc->cmd.cmd3.rptr = sc->dmarptr;
1191         }
1192
1193         len = (uint32_t)ih3->dlengsz;
1194
1195         if (sc->wait_time)
1196                 sc->timeout = lio_uptime + sc->wait_time;
1197
1198         return lio_send_command(lio_dev, sc->iq_no, &sc->cmd, sc, len,
1199                                 LIO_REQTYPE_SOFT_COMMAND);
1200 }
1201
1202 int
1203 lio_setup_sc_buffer_pool(struct lio_device *lio_dev)
1204 {
1205         char sc_pool_name[RTE_MEMPOOL_NAMESIZE];
1206         uint16_t buf_size;
1207
1208         buf_size = LIO_SOFT_COMMAND_BUFFER_SIZE + RTE_PKTMBUF_HEADROOM;
1209         snprintf(sc_pool_name, sizeof(sc_pool_name),
1210                  "lio_sc_pool_%u", lio_dev->port_id);
1211         lio_dev->sc_buf_pool = rte_pktmbuf_pool_create(sc_pool_name,
1212                                                 LIO_MAX_SOFT_COMMAND_BUFFERS,
1213                                                 0, 0, buf_size, SOCKET_ID_ANY);
1214         return 0;
1215 }
1216
1217 void
1218 lio_free_sc_buffer_pool(struct lio_device *lio_dev)
1219 {
1220         rte_mempool_free(lio_dev->sc_buf_pool);
1221 }
1222
1223 struct lio_soft_command *
1224 lio_alloc_soft_command(struct lio_device *lio_dev, uint32_t datasize,
1225                        uint32_t rdatasize, uint32_t ctxsize)
1226 {
1227         uint32_t offset = sizeof(struct lio_soft_command);
1228         struct lio_soft_command *sc;
1229         struct rte_mbuf *m;
1230         uint64_t dma_addr;
1231
1232         RTE_ASSERT((offset + datasize + rdatasize + ctxsize) <=
1233                    LIO_SOFT_COMMAND_BUFFER_SIZE);
1234
1235         m = rte_pktmbuf_alloc(lio_dev->sc_buf_pool);
1236         if (m == NULL) {
1237                 lio_dev_err(lio_dev, "Cannot allocate mbuf for sc\n");
1238                 return NULL;
1239         }
1240
1241         /* set rte_mbuf data size and there is only 1 segment */
1242         m->pkt_len = LIO_SOFT_COMMAND_BUFFER_SIZE;
1243         m->data_len = LIO_SOFT_COMMAND_BUFFER_SIZE;
1244
1245         /* use rte_mbuf buffer for soft command */
1246         sc = rte_pktmbuf_mtod(m, struct lio_soft_command *);
1247         memset(sc, 0, LIO_SOFT_COMMAND_BUFFER_SIZE);
1248         sc->size = LIO_SOFT_COMMAND_BUFFER_SIZE;
1249         sc->dma_addr = rte_mbuf_data_iova(m);
1250         sc->mbuf = m;
1251
1252         dma_addr = sc->dma_addr;
1253
1254         if (ctxsize) {
1255                 sc->ctxptr = (uint8_t *)sc + offset;
1256                 sc->ctxsize = ctxsize;
1257         }
1258
1259         /* Start data at 128 byte boundary */
1260         offset = (offset + ctxsize + 127) & 0xffffff80;
1261
1262         if (datasize) {
1263                 sc->virtdptr = (uint8_t *)sc + offset;
1264                 sc->dmadptr = dma_addr + offset;
1265                 sc->datasize = datasize;
1266         }
1267
1268         /* Start rdata at 128 byte boundary */
1269         offset = (offset + datasize + 127) & 0xffffff80;
1270
1271         if (rdatasize) {
1272                 RTE_ASSERT(rdatasize >= 16);
1273                 sc->virtrptr = (uint8_t *)sc + offset;
1274                 sc->dmarptr = dma_addr + offset;
1275                 sc->rdatasize = rdatasize;
1276                 sc->status_word = (uint64_t *)((uint8_t *)(sc->virtrptr) +
1277                                                rdatasize - 8);
1278         }
1279
1280         return sc;
1281 }
1282
1283 void
1284 lio_free_soft_command(struct lio_soft_command *sc)
1285 {
1286         rte_pktmbuf_free(sc->mbuf);
1287 }
1288
1289 void
1290 lio_setup_response_list(struct lio_device *lio_dev)
1291 {
1292         STAILQ_INIT(&lio_dev->response_list.head);
1293         rte_spinlock_init(&lio_dev->response_list.lock);
1294         rte_atomic64_set(&lio_dev->response_list.pending_req_count, 0);
1295 }
1296
1297 int
1298 lio_process_ordered_list(struct lio_device *lio_dev)
1299 {
1300         int resp_to_process = LIO_MAX_ORD_REQS_TO_PROCESS;
1301         struct lio_response_list *ordered_sc_list;
1302         struct lio_soft_command *sc;
1303         int request_complete = 0;
1304         uint64_t status64;
1305         uint32_t status;
1306
1307         ordered_sc_list = &lio_dev->response_list;
1308
1309         do {
1310                 rte_spinlock_lock(&ordered_sc_list->lock);
1311
1312                 if (STAILQ_EMPTY(&ordered_sc_list->head)) {
1313                         /* ordered_sc_list is empty; there is
1314                          * nothing to process
1315                          */
1316                         rte_spinlock_unlock(&ordered_sc_list->lock);
1317                         return -1;
1318                 }
1319
1320                 sc = LIO_STQUEUE_FIRST_ENTRY(&ordered_sc_list->head,
1321                                              struct lio_soft_command, node);
1322
1323                 status = LIO_REQUEST_PENDING;
1324
1325                 /* check if octeon has finished DMA'ing a response
1326                  * to where rptr is pointing to
1327                  */
1328                 status64 = *sc->status_word;
1329
1330                 if (status64 != LIO_COMPLETION_WORD_INIT) {
1331                         /* This logic ensures that all 64b have been written.
1332                          * 1. check byte 0 for non-FF
1333                          * 2. if non-FF, then swap result from BE to host order
1334                          * 3. check byte 7 (swapped to 0) for non-FF
1335                          * 4. if non-FF, use the low 32-bit status code
1336                          * 5. if either byte 0 or byte 7 is FF, don't use status
1337                          */
1338                         if ((status64 & 0xff) != 0xff) {
1339                                 lio_swap_8B_data(&status64, 1);
1340                                 if (((status64 & 0xff) != 0xff)) {
1341                                         /* retrieve 16-bit firmware status */
1342                                         status = (uint32_t)(status64 &
1343                                                             0xffffULL);
1344                                         if (status) {
1345                                                 status =
1346                                                 LIO_FIRMWARE_STATUS_CODE(
1347                                                                         status);
1348                                         } else {
1349                                                 /* i.e. no error */
1350                                                 status = LIO_REQUEST_DONE;
1351                                         }
1352                                 }
1353                         }
1354                 } else if ((sc->timeout && lio_check_timeout(lio_uptime,
1355                                                              sc->timeout))) {
1356                         lio_dev_err(lio_dev,
1357                                     "cmd failed, timeout (%ld, %ld)\n",
1358                                     (long)lio_uptime, (long)sc->timeout);
1359                         status = LIO_REQUEST_TIMEOUT;
1360                 }
1361
1362                 if (status != LIO_REQUEST_PENDING) {
1363                         /* we have received a response or we have timed out.
1364                          * remove node from linked list
1365                          */
1366                         STAILQ_REMOVE(&ordered_sc_list->head,
1367                                       &sc->node, lio_stailq_node, entries);
1368                         rte_atomic64_dec(
1369                             &lio_dev->response_list.pending_req_count);
1370                         rte_spinlock_unlock(&ordered_sc_list->lock);
1371
1372                         if (sc->callback)
1373                                 sc->callback(status, sc->callback_arg);
1374
1375                         request_complete++;
1376                 } else {
1377                         /* no response yet */
1378                         request_complete = 0;
1379                         rte_spinlock_unlock(&ordered_sc_list->lock);
1380                 }
1381
1382                 /* If we hit the Max Ordered requests to process every loop,
1383                  * we quit and let this function be invoked the next time
1384                  * the poll thread runs to process the remaining requests.
1385                  * This function can take up the entire CPU if there is
1386                  * no upper limit to the requests processed.
1387                  */
1388                 if (request_complete >= resp_to_process)
1389                         break;
1390         } while (request_complete);
1391
1392         return 0;
1393 }
1394
1395 static inline struct lio_stailq_node *
1396 list_delete_first_node(struct lio_stailq_head *head)
1397 {
1398         struct lio_stailq_node *node;
1399
1400         if (STAILQ_EMPTY(head))
1401                 node = NULL;
1402         else
1403                 node = STAILQ_FIRST(head);
1404
1405         if (node)
1406                 STAILQ_REMOVE(head, node, lio_stailq_node, entries);
1407
1408         return node;
1409 }
1410
1411 void
1412 lio_delete_sglist(struct lio_instr_queue *txq)
1413 {
1414         struct lio_device *lio_dev = txq->lio_dev;
1415         int iq_no = txq->q_index;
1416         struct lio_gather *g;
1417
1418         if (lio_dev->glist_head == NULL)
1419                 return;
1420
1421         do {
1422                 g = (struct lio_gather *)list_delete_first_node(
1423                                                 &lio_dev->glist_head[iq_no]);
1424                 if (g) {
1425                         if (g->sg)
1426                                 rte_free(
1427                                     (void *)((unsigned long)g->sg - g->adjust));
1428                         rte_free(g);
1429                 }
1430         } while (g);
1431 }
1432
1433 /**
1434  * \brief Setup gather lists
1435  * @param lio per-network private data
1436  */
1437 int
1438 lio_setup_sglists(struct lio_device *lio_dev, int iq_no,
1439                   int fw_mapped_iq, int num_descs, unsigned int socket_id)
1440 {
1441         struct lio_gather *g;
1442         int i;
1443
1444         rte_spinlock_init(&lio_dev->glist_lock[iq_no]);
1445
1446         STAILQ_INIT(&lio_dev->glist_head[iq_no]);
1447
1448         for (i = 0; i < num_descs; i++) {
1449                 g = rte_zmalloc_socket(NULL, sizeof(*g), RTE_CACHE_LINE_SIZE,
1450                                        socket_id);
1451                 if (g == NULL) {
1452                         lio_dev_err(lio_dev,
1453                                     "lio_gather memory allocation failed for qno %d\n",
1454                                     iq_no);
1455                         break;
1456                 }
1457
1458                 g->sg_size =
1459                     ((ROUNDUP4(LIO_MAX_SG) >> 2) * LIO_SG_ENTRY_SIZE);
1460
1461                 g->sg = rte_zmalloc_socket(NULL, g->sg_size + 8,
1462                                            RTE_CACHE_LINE_SIZE, socket_id);
1463                 if (g->sg == NULL) {
1464                         lio_dev_err(lio_dev,
1465                                     "sg list memory allocation failed for qno %d\n",
1466                                     iq_no);
1467                         rte_free(g);
1468                         break;
1469                 }
1470
1471                 /* The gather component should be aligned on 64-bit boundary */
1472                 if (((unsigned long)g->sg) & 7) {
1473                         g->adjust = 8 - (((unsigned long)g->sg) & 7);
1474                         g->sg =
1475                             (struct lio_sg_entry *)((unsigned long)g->sg +
1476                                                        g->adjust);
1477                 }
1478
1479                 STAILQ_INSERT_TAIL(&lio_dev->glist_head[iq_no], &g->list,
1480                                    entries);
1481         }
1482
1483         if (i != num_descs) {
1484                 lio_delete_sglist(lio_dev->instr_queue[fw_mapped_iq]);
1485                 return -ENOMEM;
1486         }
1487
1488         return 0;
1489 }
1490
1491 void
1492 lio_delete_instruction_queue(struct lio_device *lio_dev, int iq_no)
1493 {
1494         lio_delete_instr_queue(lio_dev, iq_no);
1495         rte_free(lio_dev->instr_queue[iq_no]);
1496         lio_dev->instr_queue[iq_no] = NULL;
1497         lio_dev->num_iqs--;
1498 }
1499
1500 static inline uint32_t
1501 lio_iq_get_available(struct lio_device *lio_dev, uint32_t q_no)
1502 {
1503         return ((lio_dev->instr_queue[q_no]->nb_desc - 1) -
1504                 (uint32_t)rte_atomic64_read(
1505                                 &lio_dev->instr_queue[q_no]->instr_pending));
1506 }
1507
1508 static inline int
1509 lio_iq_is_full(struct lio_device *lio_dev, uint32_t q_no)
1510 {
1511         return ((uint32_t)rte_atomic64_read(
1512                                 &lio_dev->instr_queue[q_no]->instr_pending) >=
1513                                 (lio_dev->instr_queue[q_no]->nb_desc - 2));
1514 }
1515
1516 static int
1517 lio_dev_cleanup_iq(struct lio_device *lio_dev, int iq_no)
1518 {
1519         struct lio_instr_queue *iq = lio_dev->instr_queue[iq_no];
1520         uint32_t count = 10000;
1521
1522         while ((lio_iq_get_available(lio_dev, iq_no) < LIO_FLUSH_WM(iq)) &&
1523                         --count)
1524                 lio_flush_iq(lio_dev, iq);
1525
1526         return count ? 0 : 1;
1527 }
1528
1529 static void
1530 lio_ctrl_cmd_callback(uint32_t status __rte_unused, void *sc_ptr)
1531 {
1532         struct lio_soft_command *sc = sc_ptr;
1533         struct lio_dev_ctrl_cmd *ctrl_cmd;
1534         struct lio_ctrl_pkt *ctrl_pkt;
1535
1536         ctrl_pkt = (struct lio_ctrl_pkt *)sc->ctxptr;
1537         ctrl_cmd = ctrl_pkt->ctrl_cmd;
1538         ctrl_cmd->cond = 1;
1539
1540         lio_free_soft_command(sc);
1541 }
1542
1543 static inline struct lio_soft_command *
1544 lio_alloc_ctrl_pkt_sc(struct lio_device *lio_dev,
1545                       struct lio_ctrl_pkt *ctrl_pkt)
1546 {
1547         struct lio_soft_command *sc = NULL;
1548         uint32_t uddsize, datasize;
1549         uint32_t rdatasize;
1550         uint8_t *data;
1551
1552         uddsize = (uint32_t)(ctrl_pkt->ncmd.s.more * 8);
1553
1554         datasize = OCTEON_CMD_SIZE + uddsize;
1555         rdatasize = (ctrl_pkt->wait_time) ? 16 : 0;
1556
1557         sc = lio_alloc_soft_command(lio_dev, datasize,
1558                                     rdatasize, sizeof(struct lio_ctrl_pkt));
1559         if (sc == NULL)
1560                 return NULL;
1561
1562         rte_memcpy(sc->ctxptr, ctrl_pkt, sizeof(struct lio_ctrl_pkt));
1563
1564         data = (uint8_t *)sc->virtdptr;
1565
1566         rte_memcpy(data, &ctrl_pkt->ncmd, OCTEON_CMD_SIZE);
1567
1568         lio_swap_8B_data((uint64_t *)data, OCTEON_CMD_SIZE >> 3);
1569
1570         if (uddsize) {
1571                 /* Endian-Swap for UDD should have been done by caller. */
1572                 rte_memcpy(data + OCTEON_CMD_SIZE, ctrl_pkt->udd, uddsize);
1573         }
1574
1575         sc->iq_no = (uint32_t)ctrl_pkt->iq_no;
1576
1577         lio_prepare_soft_command(lio_dev, sc,
1578                                  LIO_OPCODE, LIO_OPCODE_CMD,
1579                                  0, 0, 0);
1580
1581         sc->callback = lio_ctrl_cmd_callback;
1582         sc->callback_arg = sc;
1583         sc->wait_time = ctrl_pkt->wait_time;
1584
1585         return sc;
1586 }
1587
1588 int
1589 lio_send_ctrl_pkt(struct lio_device *lio_dev, struct lio_ctrl_pkt *ctrl_pkt)
1590 {
1591         struct lio_soft_command *sc = NULL;
1592         int retval;
1593
1594         sc = lio_alloc_ctrl_pkt_sc(lio_dev, ctrl_pkt);
1595         if (sc == NULL) {
1596                 lio_dev_err(lio_dev, "soft command allocation failed\n");
1597                 return -1;
1598         }
1599
1600         retval = lio_send_soft_command(lio_dev, sc);
1601         if (retval == LIO_IQ_SEND_FAILED) {
1602                 lio_free_soft_command(sc);
1603                 lio_dev_err(lio_dev, "Port: %d soft command: %d send failed status: %x\n",
1604                             lio_dev->port_id, ctrl_pkt->ncmd.s.cmd, retval);
1605                 return -1;
1606         }
1607
1608         return retval;
1609 }
1610
1611 /** Send data packet to the device
1612  *  @param lio_dev - lio device pointer
1613  *  @param ndata   - control structure with queueing, and buffer information
1614  *
1615  *  @returns IQ_FAILED if it failed to add to the input queue. IQ_STOP if it the
1616  *  queue should be stopped, and LIO_IQ_SEND_OK if it sent okay.
1617  */
1618 static inline int
1619 lio_send_data_pkt(struct lio_device *lio_dev, struct lio_data_pkt *ndata)
1620 {
1621         return lio_send_command(lio_dev, ndata->q_no, &ndata->cmd,
1622                                 ndata->buf, ndata->datasize, ndata->reqtype);
1623 }
1624
1625 uint16_t
1626 lio_dev_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts)
1627 {
1628         struct lio_instr_queue *txq = tx_queue;
1629         union lio_cmd_setup cmdsetup;
1630         struct lio_device *lio_dev;
1631         struct lio_iq_stats *stats;
1632         struct lio_data_pkt ndata;
1633         int i, processed = 0;
1634         struct rte_mbuf *m;
1635         uint32_t tag = 0;
1636         int status = 0;
1637         int iq_no;
1638
1639         lio_dev = txq->lio_dev;
1640         iq_no = txq->txpciq.s.q_no;
1641         stats = &lio_dev->instr_queue[iq_no]->stats;
1642
1643         if (!lio_dev->intf_open || !lio_dev->linfo.link.s.link_up) {
1644                 PMD_TX_LOG(lio_dev, ERR, "Transmit failed link_status : %d\n",
1645                            lio_dev->linfo.link.s.link_up);
1646                 goto xmit_failed;
1647         }
1648
1649         lio_dev_cleanup_iq(lio_dev, iq_no);
1650
1651         for (i = 0; i < nb_pkts; i++) {
1652                 uint32_t pkt_len = 0;
1653
1654                 m = pkts[i];
1655
1656                 /* Prepare the attributes for the data to be passed to BASE. */
1657                 memset(&ndata, 0, sizeof(struct lio_data_pkt));
1658
1659                 ndata.buf = m;
1660
1661                 ndata.q_no = iq_no;
1662                 if (lio_iq_is_full(lio_dev, ndata.q_no)) {
1663                         stats->tx_iq_busy++;
1664                         if (lio_dev_cleanup_iq(lio_dev, iq_no)) {
1665                                 PMD_TX_LOG(lio_dev, ERR,
1666                                            "Transmit failed iq:%d full\n",
1667                                            ndata.q_no);
1668                                 break;
1669                         }
1670                 }
1671
1672                 cmdsetup.cmd_setup64 = 0;
1673                 cmdsetup.s.iq_no = iq_no;
1674
1675                 /* check checksum offload flags to form cmd */
1676                 if (m->ol_flags & RTE_MBUF_F_TX_IP_CKSUM)
1677                         cmdsetup.s.ip_csum = 1;
1678
1679                 if (m->ol_flags & RTE_MBUF_F_TX_OUTER_IP_CKSUM)
1680                         cmdsetup.s.tnl_csum = 1;
1681                 else if ((m->ol_flags & RTE_MBUF_F_TX_TCP_CKSUM) ||
1682                                 (m->ol_flags & RTE_MBUF_F_TX_UDP_CKSUM))
1683                         cmdsetup.s.transport_csum = 1;
1684
1685                 if (m->nb_segs == 1) {
1686                         pkt_len = rte_pktmbuf_data_len(m);
1687                         cmdsetup.s.u.datasize = pkt_len;
1688                         lio_prepare_pci_cmd(lio_dev, &ndata.cmd,
1689                                             &cmdsetup, tag);
1690                         ndata.cmd.cmd3.dptr = rte_mbuf_data_iova(m);
1691                         ndata.reqtype = LIO_REQTYPE_NORESP_NET;
1692                 } else {
1693                         struct lio_buf_free_info *finfo;
1694                         struct lio_gather *g;
1695                         rte_iova_t phyaddr;
1696                         int i, frags;
1697
1698                         finfo = (struct lio_buf_free_info *)rte_malloc(NULL,
1699                                                         sizeof(*finfo), 0);
1700                         if (finfo == NULL) {
1701                                 PMD_TX_LOG(lio_dev, ERR,
1702                                            "free buffer alloc failed\n");
1703                                 goto xmit_failed;
1704                         }
1705
1706                         rte_spinlock_lock(&lio_dev->glist_lock[iq_no]);
1707                         g = (struct lio_gather *)list_delete_first_node(
1708                                                 &lio_dev->glist_head[iq_no]);
1709                         rte_spinlock_unlock(&lio_dev->glist_lock[iq_no]);
1710                         if (g == NULL) {
1711                                 PMD_TX_LOG(lio_dev, ERR,
1712                                            "Transmit scatter gather: glist null!\n");
1713                                 goto xmit_failed;
1714                         }
1715
1716                         cmdsetup.s.gather = 1;
1717                         cmdsetup.s.u.gatherptrs = m->nb_segs;
1718                         lio_prepare_pci_cmd(lio_dev, &ndata.cmd,
1719                                             &cmdsetup, tag);
1720
1721                         memset(g->sg, 0, g->sg_size);
1722                         g->sg[0].ptr[0] = rte_mbuf_data_iova(m);
1723                         lio_add_sg_size(&g->sg[0], m->data_len, 0);
1724                         pkt_len = m->data_len;
1725                         finfo->mbuf = m;
1726
1727                         /* First seg taken care above */
1728                         frags = m->nb_segs - 1;
1729                         i = 1;
1730                         m = m->next;
1731                         while (frags--) {
1732                                 g->sg[(i >> 2)].ptr[(i & 3)] =
1733                                                 rte_mbuf_data_iova(m);
1734                                 lio_add_sg_size(&g->sg[(i >> 2)],
1735                                                 m->data_len, (i & 3));
1736                                 pkt_len += m->data_len;
1737                                 i++;
1738                                 m = m->next;
1739                         }
1740
1741                         phyaddr = rte_mem_virt2iova(g->sg);
1742                         if (phyaddr == RTE_BAD_IOVA) {
1743                                 PMD_TX_LOG(lio_dev, ERR, "bad phys addr\n");
1744                                 goto xmit_failed;
1745                         }
1746
1747                         ndata.cmd.cmd3.dptr = phyaddr;
1748                         ndata.reqtype = LIO_REQTYPE_NORESP_NET_SG;
1749
1750                         finfo->g = g;
1751                         finfo->lio_dev = lio_dev;
1752                         finfo->iq_no = (uint64_t)iq_no;
1753                         ndata.buf = finfo;
1754                 }
1755
1756                 ndata.datasize = pkt_len;
1757
1758                 status = lio_send_data_pkt(lio_dev, &ndata);
1759
1760                 if (unlikely(status == LIO_IQ_SEND_FAILED)) {
1761                         PMD_TX_LOG(lio_dev, ERR, "send failed\n");
1762                         break;
1763                 }
1764
1765                 if (unlikely(status == LIO_IQ_SEND_STOP)) {
1766                         PMD_TX_LOG(lio_dev, DEBUG, "iq full\n");
1767                         /* create space as iq is full */
1768                         lio_dev_cleanup_iq(lio_dev, iq_no);
1769                 }
1770
1771                 stats->tx_done++;
1772                 stats->tx_tot_bytes += pkt_len;
1773                 processed++;
1774         }
1775
1776 xmit_failed:
1777         stats->tx_dropped += (nb_pkts - processed);
1778
1779         return processed;
1780 }
1781
1782 void
1783 lio_dev_clear_queues(struct rte_eth_dev *eth_dev)
1784 {
1785         struct lio_instr_queue *txq;
1786         struct lio_droq *rxq;
1787         uint16_t i;
1788
1789         for (i = 0; i < eth_dev->data->nb_tx_queues; i++) {
1790                 txq = eth_dev->data->tx_queues[i];
1791                 if (txq != NULL) {
1792                         lio_dev_tx_queue_release(eth_dev, i);
1793                         eth_dev->data->tx_queues[i] = NULL;
1794                 }
1795         }
1796
1797         for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
1798                 rxq = eth_dev->data->rx_queues[i];
1799                 if (rxq != NULL) {
1800                         lio_dev_rx_queue_release(eth_dev, i);
1801                         eth_dev->data->rx_queues[i] = NULL;
1802                 }
1803         }
1804 }