drivers/net/hinic/hinic_pmd_tx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Huawei Technologies Co., Ltd
   3  */
   4
   5 #include <rte_mbuf.h>
   6 #include <rte_tcp.h>
   7 #include <rte_sctp.h>
   8 #include <rte_udp.h>
   9 #include <rte_ip.h>
  10 #ifdef __ARM64_NEON__
  11 #include <arm_neon.h>
  12 #endif
  13
  14 #include "base/hinic_compat.h"
  15 #include "base/hinic_pmd_hwdev.h"
  16 #include "base/hinic_pmd_hwif.h"
  17 #include "base/hinic_pmd_wq.h"
  18 #include "base/hinic_pmd_nicio.h"
  19 #include "base/hinic_pmd_niccfg.h"
  20 #include "hinic_pmd_ethdev.h"
  21 #include "hinic_pmd_tx.h"
  22
  23 /* packet header and tx offload info */
  24 #define ETHER_LEN_NO_VLAN               14
  25 #define ETHER_LEN_WITH_VLAN             18
  26 #define VXLANLEN                        8
  27 #define MAX_PLD_OFFSET                  221
  28 #define MAX_SINGLE_SGE_SIZE             65536
  29 #define TSO_ENABLE                      1
  30 #define TX_MSS_DEFAULT                  0x3E00
  31 #define TX_MSS_MIN                      0x50
  32
  33 #define HINIC_NONTSO_PKT_MAX_SGE                17      /* non-tso max sge 17 */
  34 #define HINIC_NONTSO_SEG_NUM_INVALID(num)       \
  35                         ((num) > HINIC_NONTSO_PKT_MAX_SGE)
  36
  37 #define HINIC_TSO_PKT_MAX_SGE                   127     /* tso max sge 127 */
  38 #define HINIC_TSO_SEG_NUM_INVALID(num)          ((num) > HINIC_TSO_PKT_MAX_SGE)
  39
  40 /* sizeof(struct hinic_sq_bufdesc) == 16, shift 4 */
  41 #define HINIC_BUF_DESC_SIZE(nr_descs)   (SIZE_8BYTES(((u32)nr_descs) << 4))
  42
  43 #define MASKED_SQ_IDX(sq, idx)          ((idx) & (sq)->wq->mask)
  44
  45 /* SQ_CTRL */
  46 #define SQ_CTRL_BUFDESC_SECT_LEN_SHIFT          0
  47 #define SQ_CTRL_TASKSECT_LEN_SHIFT              16
  48 #define SQ_CTRL_DATA_FORMAT_SHIFT               22
  49 #define SQ_CTRL_LEN_SHIFT                       29
  50 #define SQ_CTRL_OWNER_SHIFT                     31
  51
  52 #define SQ_CTRL_BUFDESC_SECT_LEN_MASK           0xFFU
  53 #define SQ_CTRL_TASKSECT_LEN_MASK               0x1FU
  54 #define SQ_CTRL_DATA_FORMAT_MASK                0x1U
  55 #define SQ_CTRL_LEN_MASK                        0x3U
  56 #define SQ_CTRL_OWNER_MASK                      0x1U
  57
  58 #define SQ_CTRL_SET(val, member)        \
  59         (((val) & SQ_CTRL_##member##_MASK) << SQ_CTRL_##member##_SHIFT)
  60
  61 #define SQ_CTRL_QUEUE_INFO_PLDOFF_SHIFT         2
  62 #define SQ_CTRL_QUEUE_INFO_UFO_SHIFT            10
  63 #define SQ_CTRL_QUEUE_INFO_TSO_SHIFT            11
  64 #define SQ_CTRL_QUEUE_INFO_TCPUDP_CS_SHIFT      12
  65 #define SQ_CTRL_QUEUE_INFO_MSS_SHIFT            13
  66 #define SQ_CTRL_QUEUE_INFO_SCTP_SHIFT           27
  67 #define SQ_CTRL_QUEUE_INFO_UC_SHIFT             28
  68 #define SQ_CTRL_QUEUE_INFO_PRI_SHIFT            29
  69
  70 #define SQ_CTRL_QUEUE_INFO_PLDOFF_MASK          0xFFU
  71 #define SQ_CTRL_QUEUE_INFO_UFO_MASK             0x1U
  72 #define SQ_CTRL_QUEUE_INFO_TSO_MASK             0x1U
  73 #define SQ_CTRL_QUEUE_INFO_TCPUDP_CS_MASK       0x1U
  74 #define SQ_CTRL_QUEUE_INFO_MSS_MASK             0x3FFFU
  75 #define SQ_CTRL_QUEUE_INFO_SCTP_MASK            0x1U
  76 #define SQ_CTRL_QUEUE_INFO_UC_MASK              0x1U
  77 #define SQ_CTRL_QUEUE_INFO_PRI_MASK             0x7U
  78
  79 #define SQ_CTRL_QUEUE_INFO_SET(val, member)     \
  80         (((u32)(val) & SQ_CTRL_QUEUE_INFO_##member##_MASK) <<   \
  81                         SQ_CTRL_QUEUE_INFO_##member##_SHIFT)
  82
  83 #define SQ_CTRL_QUEUE_INFO_GET(val, member)     \
  84         (((val) >> SQ_CTRL_QUEUE_INFO_##member##_SHIFT) &       \
  85                         SQ_CTRL_QUEUE_INFO_##member##_MASK)
  86
  87 #define SQ_CTRL_QUEUE_INFO_CLEAR(val, member)   \
  88         ((val) & (~(SQ_CTRL_QUEUE_INFO_##member##_MASK << \
  89                         SQ_CTRL_QUEUE_INFO_##member##_SHIFT)))
  90
  91 #define SQ_TASK_INFO0_L2HDR_LEN_SHIFT           0
  92 #define SQ_TASK_INFO0_L4OFFLOAD_SHIFT           8
  93 #define SQ_TASK_INFO0_INNER_L3TYPE_SHIFT        10
  94 #define SQ_TASK_INFO0_VLAN_OFFLOAD_SHIFT        12
  95 #define SQ_TASK_INFO0_PARSE_FLAG_SHIFT          13
  96 #define SQ_TASK_INFO0_UFO_AVD_SHIFT             14
  97 #define SQ_TASK_INFO0_TSO_UFO_SHIFT             15
  98 #define SQ_TASK_INFO0_VLAN_TAG_SHIFT            16
  99
 100 #define SQ_TASK_INFO0_L2HDR_LEN_MASK            0xFFU
 101 #define SQ_TASK_INFO0_L4OFFLOAD_MASK            0x3U
 102 #define SQ_TASK_INFO0_INNER_L3TYPE_MASK         0x3U
 103 #define SQ_TASK_INFO0_VLAN_OFFLOAD_MASK         0x1U
 104 #define SQ_TASK_INFO0_PARSE_FLAG_MASK           0x1U
 105 #define SQ_TASK_INFO0_UFO_AVD_MASK              0x1U
 106 #define SQ_TASK_INFO0_TSO_UFO_MASK              0x1U
 107 #define SQ_TASK_INFO0_VLAN_TAG_MASK             0xFFFFU
 108
 109 #define SQ_TASK_INFO0_SET(val, member)                  \
 110         (((u32)(val) & SQ_TASK_INFO0_##member##_MASK) <<        \
 111                         SQ_TASK_INFO0_##member##_SHIFT)
 112
 113 #define SQ_TASK_INFO1_MD_TYPE_SHIFT             8
 114 #define SQ_TASK_INFO1_INNER_L4LEN_SHIFT         16
 115 #define SQ_TASK_INFO1_INNER_L3LEN_SHIFT         24
 116
 117 #define SQ_TASK_INFO1_MD_TYPE_MASK              0xFFU
 118 #define SQ_TASK_INFO1_INNER_L4LEN_MASK          0xFFU
 119 #define SQ_TASK_INFO1_INNER_L3LEN_MASK          0xFFU
 120
 121 #define SQ_TASK_INFO1_SET(val, member)                  \
 122         (((val) & SQ_TASK_INFO1_##member##_MASK) <<     \
 123                         SQ_TASK_INFO1_##member##_SHIFT)
 124
 125 #define SQ_TASK_INFO2_TUNNEL_L4LEN_SHIFT        0
 126 #define SQ_TASK_INFO2_OUTER_L3LEN_SHIFT         8
 127 #define SQ_TASK_INFO2_TUNNEL_L4TYPE_SHIFT       16
 128 #define SQ_TASK_INFO2_OUTER_L3TYPE_SHIFT        24
 129
 130 #define SQ_TASK_INFO2_TUNNEL_L4LEN_MASK         0xFFU
 131 #define SQ_TASK_INFO2_OUTER_L3LEN_MASK          0xFFU
 132 #define SQ_TASK_INFO2_TUNNEL_L4TYPE_MASK        0x7U
 133 #define SQ_TASK_INFO2_OUTER_L3TYPE_MASK         0x3U
 134
 135 #define SQ_TASK_INFO2_SET(val, member)                  \
 136         (((val) & SQ_TASK_INFO2_##member##_MASK) <<     \
 137                         SQ_TASK_INFO2_##member##_SHIFT)
 138
 139 #define SQ_TASK_INFO4_L2TYPE_SHIFT              31
 140
 141 #define SQ_TASK_INFO4_L2TYPE_MASK               0x1U
 142
 143 #define SQ_TASK_INFO4_SET(val, member)          \
 144         (((u32)(val) & SQ_TASK_INFO4_##member##_MASK) << \
 145                         SQ_TASK_INFO4_##member##_SHIFT)
 146
 147 /* SQ_DB */
 148 #define SQ_DB_OFF                               0x00000800
 149 #define SQ_DB_INFO_HI_PI_SHIFT                  0
 150 #define SQ_DB_INFO_QID_SHIFT                    8
 151 #define SQ_DB_INFO_CFLAG_SHIFT                  23
 152 #define SQ_DB_INFO_COS_SHIFT                    24
 153 #define SQ_DB_INFO_TYPE_SHIFT                   27
 154
 155 #define SQ_DB_INFO_HI_PI_MASK                   0xFFU
 156 #define SQ_DB_INFO_QID_MASK                     0x3FFU
 157 #define SQ_DB_INFO_CFLAG_MASK                   0x1U
 158 #define SQ_DB_INFO_COS_MASK                     0x7U
 159 #define SQ_DB_INFO_TYPE_MASK                    0x1FU
 160 #define SQ_DB_INFO_SET(val, member)             \
 161         (((u32)(val) & SQ_DB_INFO_##member##_MASK) <<   \
 162                         SQ_DB_INFO_##member##_SHIFT)
 163
 164 #define SQ_DB                                   1
 165 #define SQ_CFLAG_DP                             0       /* CFLAG_DATA_PATH */
 166
 167 #define SQ_DB_PI_LOW_MASK                       0xFF
 168 #define SQ_DB_PI_LOW(pi)                        ((pi) & SQ_DB_PI_LOW_MASK)
 169 #define SQ_DB_PI_HI_SHIFT                       8
 170 #define SQ_DB_PI_HIGH(pi)                       ((pi) >> SQ_DB_PI_HI_SHIFT)
 171 #define SQ_DB_ADDR(sq, pi)              \
 172         ((u64 *)((u8 __iomem *)((sq)->db_addr) + SQ_DB_OFF) + SQ_DB_PI_LOW(pi))
 173
 174 /* txq wq operations */
 175 #define HINIC_GET_SQ_WQE_MASK(txq)              ((txq)->wq->mask)
 176
 177 #define HINIC_GET_SQ_HW_CI(txq) \
 178         ((be16_to_cpu(*(txq)->cons_idx_addr)) & HINIC_GET_SQ_WQE_MASK(txq))
 179
 180 #define HINIC_GET_SQ_LOCAL_CI(txq)      \
 181         (((txq)->wq->cons_idx) & HINIC_GET_SQ_WQE_MASK(txq))
 182
 183 #define HINIC_UPDATE_SQ_LOCAL_CI(txq, wqebb_cnt)        \
 184         do {                                            \
 185                 (txq)->wq->cons_idx += wqebb_cnt;       \
 186                 (txq)->wq->delta += wqebb_cnt;          \
 187         } while (0)
 188
 189 #define HINIC_GET_SQ_FREE_WQEBBS(txq)   ((txq)->wq->delta - 1)
 190
 191 #define HINIC_IS_SQ_EMPTY(txq)  (((txq)->wq->delta) == ((txq)->q_depth))
 192
 193 #define BUF_DESC_SIZE_SHIFT             4
 194
 195 #define HINIC_SQ_WQE_SIZE(num_sge)              \
 196         (sizeof(struct hinic_sq_ctrl) + sizeof(struct hinic_sq_task) +  \
 197                         (unsigned int)((num_sge) << BUF_DESC_SIZE_SHIFT))
 198
 199 #define HINIC_SQ_WQEBB_CNT(num_sge)     \
 200         (int)(ALIGN(HINIC_SQ_WQE_SIZE((u32)num_sge), \
 201                         HINIC_SQ_WQEBB_SIZE) >> HINIC_SQ_WQEBB_SHIFT)
 202
 203
 204 static inline void hinic_sq_wqe_cpu_to_be32(void *data, int nr_wqebb)
 205 {
 206 #if defined(__X86_64_SSE__)
 207         int i;
 208         __m128i *wqe_line = (__m128i *)data;
 209         __m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
 210                                         11, 4, 5, 6, 7, 0, 1, 2, 3);
 211
 212         for (i = 0; i < nr_wqebb; i++) {
 213                 /* convert 64B wqebb using 4 SSE instructions */
 214                 wqe_line[0] = _mm_shuffle_epi8(wqe_line[0], shuf_mask);
 215                 wqe_line[1] = _mm_shuffle_epi8(wqe_line[1], shuf_mask);
 216                 wqe_line[2] = _mm_shuffle_epi8(wqe_line[2], shuf_mask);
 217                 wqe_line[3] = _mm_shuffle_epi8(wqe_line[3], shuf_mask);
 218                 wqe_line += 4;
 219         }
 220 #elif defined(__ARM64_NEON__)
 221         int i;
 222         uint8x16_t *wqe_line = (uint8x16_t *)data;
 223         const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
 224                                         9, 8, 15, 14, 13, 12};
 225
 226         for (i = 0; i < nr_wqebb; i++) {
 227                 wqe_line[0] = vqtbl1q_u8(wqe_line[0], shuf_mask);
 228                 wqe_line[1] = vqtbl1q_u8(wqe_line[1], shuf_mask);
 229                 wqe_line[2] = vqtbl1q_u8(wqe_line[2], shuf_mask);
 230                 wqe_line[3] = vqtbl1q_u8(wqe_line[3], shuf_mask);
 231                 wqe_line += 4;
 232         }
 233 #else
 234         hinic_cpu_to_be32(data, nr_wqebb * HINIC_SQ_WQEBB_SIZE);
 235 #endif
 236 }
 237
 238 static inline void hinic_sge_cpu_to_be32(void *data, int nr_sge)
 239 {
 240 #if defined(__X86_64_SSE__)
 241         int i;
 242         __m128i *sge_line = (__m128i *)data;
 243         __m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
 244                                         11, 4, 5, 6, 7, 0, 1, 2, 3);
 245
 246         for (i = 0; i < nr_sge; i++) {
 247                 /* convert 16B sge using 1 SSE instructions */
 248                 *sge_line = _mm_shuffle_epi8(*sge_line, shuf_mask);
 249                 sge_line++;
 250         }
 251 #elif defined(__ARM64_NEON__)
 252         int i;
 253         uint8x16_t *sge_line = (uint8x16_t *)data;
 254         const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
 255                                         9, 8, 15, 14, 13, 12};
 256
 257         for (i = 0; i < nr_sge; i++) {
 258                 *sge_line = vqtbl1q_u8(*sge_line, shuf_mask);
 259                 sge_line++;
 260         }
 261 #else
 262         hinic_cpu_to_be32(data, nr_sge * sizeof(struct hinic_sq_bufdesc));
 263 #endif
 264 }
 265
 266 void hinic_txq_get_stats(struct hinic_txq *txq, struct hinic_txq_stats *stats)
 267 {
 268         if (!txq || !stats) {
 269                 PMD_DRV_LOG(ERR, "Txq or stats is NULL");
 270                 return;
 271         }
 272
 273         memcpy(stats, &txq->txq_stats, sizeof(txq->txq_stats));
 274 }
 275
 276 void hinic_txq_stats_reset(struct hinic_txq *txq)
 277 {
 278         struct hinic_txq_stats *txq_stats;
 279
 280         if (txq == NULL)
 281                 return;
 282
 283         txq_stats = &txq->txq_stats;
 284         memset(txq_stats, 0, sizeof(*txq_stats));
 285 }
 286
 287 static inline struct rte_mbuf *hinic_copy_tx_mbuf(struct hinic_nic_dev *nic_dev,
 288                                                   struct rte_mbuf *mbuf,
 289                                                   u16 sge_cnt)
 290 {
 291         struct rte_mbuf *dst_mbuf;
 292         u32 offset = 0;
 293         u16 i;
 294
 295         if (unlikely(!nic_dev->cpy_mpool))
 296                 return NULL;
 297
 298         dst_mbuf = rte_pktmbuf_alloc(nic_dev->cpy_mpool);
 299         if (unlikely(!dst_mbuf))
 300                 return NULL;
 301
 302         dst_mbuf->data_off = 0;
 303         for (i = 0; i < sge_cnt; i++) {
 304                 rte_memcpy((char *)dst_mbuf->buf_addr + offset,
 305                            (char *)mbuf->buf_addr + mbuf->data_off,
 306                            mbuf->data_len);
 307                 dst_mbuf->data_len += mbuf->data_len;
 308                 offset += mbuf->data_len;
 309                 mbuf = mbuf->next;
 310         }
 311
 312         dst_mbuf->pkt_len = dst_mbuf->data_len;
 313
 314         return dst_mbuf;
 315 }
 316
 317 static inline bool hinic_mbuf_dma_map_sge(struct hinic_txq *txq,
 318                                           struct rte_mbuf *mbuf,
 319                                           struct hinic_sq_bufdesc *sges,
 320                                           struct hinic_wqe_info *sqe_info)
 321 {
 322         dma_addr_t dma_addr;
 323         u16 i, around_sges;
 324         u16 nb_segs = sqe_info->sge_cnt - sqe_info->cpy_mbuf_cnt;
 325         u16 real_nb_segs = mbuf->nb_segs;
 326         struct hinic_sq_bufdesc *sge_idx = sges;
 327
 328         if (unlikely(sqe_info->around)) {
 329                 /* parts of wqe is in sq bottom while parts
 330                  * of wqe is in sq head
 331                  */
 332                 i = 0;
 333                 for (sge_idx = sges; (u64)sge_idx <= txq->sq_bot_sge_addr;
 334                      sge_idx++) {
 335                         if (unlikely(mbuf == NULL)) {
 336                                 txq->txq_stats.mbuf_null++;
 337                                 return false;
 338                         }
 339
 340                         dma_addr = rte_mbuf_data_iova(mbuf);
 341                         if (unlikely(mbuf->data_len == 0)) {
 342                                 txq->txq_stats.sge_len0++;
 343                                 return false;
 344                         }
 345                         hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
 346                                       mbuf->data_len);
 347                         mbuf = mbuf->next;
 348                         i++;
 349                 }
 350
 351                 around_sges = nb_segs - i;
 352                 sge_idx = (struct hinic_sq_bufdesc *)
 353                                 ((void *)txq->sq_head_addr);
 354                 for (; i < nb_segs; i++) {
 355                         if (unlikely(mbuf == NULL)) {
 356                                 txq->txq_stats.mbuf_null++;
 357                                 return false;
 358                         }
 359
 360                         dma_addr = rte_mbuf_data_iova(mbuf);
 361                         if (unlikely(mbuf->data_len == 0)) {
 362                                 txq->txq_stats.sge_len0++;
 363                                 return false;
 364                         }
 365                         hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
 366                                       mbuf->data_len);
 367                         mbuf = mbuf->next;
 368                         sge_idx++;
 369                 }
 370
 371                 /* covert sges at head to big endian */
 372                 hinic_sge_cpu_to_be32((void *)txq->sq_head_addr, around_sges);
 373         } else {
 374                 /* wqe is in continuous space */
 375                 for (i = 0; i < nb_segs; i++) {
 376                         if (unlikely(mbuf == NULL)) {
 377                                 txq->txq_stats.mbuf_null++;
 378                                 return false;
 379                         }
 380
 381                         dma_addr = rte_mbuf_data_iova(mbuf);
 382                         if (unlikely(mbuf->data_len == 0)) {
 383                                 txq->txq_stats.sge_len0++;
 384                                 return false;
 385                         }
 386                         hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
 387                                       mbuf->data_len);
 388                         mbuf = mbuf->next;
 389                         sge_idx++;
 390                 }
 391         }
 392
 393         /* for now: support non-tso over 17 sge, copy the last 2 mbuf */
 394         if (unlikely(sqe_info->cpy_mbuf_cnt != 0)) {
 395                 /* copy invalid mbuf segs to a valid buffer, lost performance */
 396                 txq->txq_stats.cpy_pkts += 1;
 397                 mbuf = hinic_copy_tx_mbuf(txq->nic_dev, mbuf,
 398                                           real_nb_segs - nb_segs);
 399                 if (unlikely(!mbuf))
 400                         return false;
 401
 402                 txq->tx_info[sqe_info->pi].cpy_mbuf = mbuf;
 403
 404                 /* deal with the last mbuf */
 405                 dma_addr = rte_mbuf_data_iova(mbuf);
 406                 if (unlikely(mbuf->data_len == 0)) {
 407                         txq->txq_stats.sge_len0++;
 408                         return false;
 409                 }
 410                 hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
 411                               mbuf->data_len);
 412                 if (unlikely(sqe_info->around))
 413                         hinic_sge_cpu_to_be32((void *)sge_idx, 1);
 414         }
 415
 416         return true;
 417 }
 418
 419 static inline void hinic_fill_sq_wqe_header(struct hinic_sq_ctrl *ctrl,
 420                                             u32 queue_info, int nr_descs,
 421                                             u8 owner)
 422 {
 423         u32 ctrl_size, task_size, bufdesc_size;
 424
 425         ctrl_size = SIZE_8BYTES(sizeof(struct hinic_sq_ctrl));
 426         task_size = SIZE_8BYTES(sizeof(struct hinic_sq_task));
 427         bufdesc_size = HINIC_BUF_DESC_SIZE(nr_descs);
 428
 429         ctrl->ctrl_fmt = SQ_CTRL_SET(bufdesc_size, BUFDESC_SECT_LEN) |
 430                         SQ_CTRL_SET(task_size, TASKSECT_LEN)    |
 431                         SQ_CTRL_SET(SQ_NORMAL_WQE, DATA_FORMAT) |
 432                         SQ_CTRL_SET(ctrl_size, LEN)             |
 433                         SQ_CTRL_SET(owner, OWNER);
 434
 435         ctrl->queue_info = queue_info;
 436         ctrl->queue_info |= SQ_CTRL_QUEUE_INFO_SET(1U, UC);
 437
 438         if (!SQ_CTRL_QUEUE_INFO_GET(ctrl->queue_info, MSS)) {
 439                 ctrl->queue_info |=
 440                         SQ_CTRL_QUEUE_INFO_SET(TX_MSS_DEFAULT, MSS);
 441         } else if (SQ_CTRL_QUEUE_INFO_GET(ctrl->queue_info, MSS) < TX_MSS_MIN) {
 442                 /* mss should not be less than 80 */
 443                 ctrl->queue_info =
 444                                 SQ_CTRL_QUEUE_INFO_CLEAR(ctrl->queue_info, MSS);
 445                 ctrl->queue_info |= SQ_CTRL_QUEUE_INFO_SET(TX_MSS_MIN, MSS);
 446         }
 447 }
 448
 449 static inline bool hinic_is_tso_sge_valid(struct rte_mbuf *mbuf,
 450                                           struct hinic_tx_offload_info
 451                                           *poff_info,
 452                                           struct hinic_wqe_info *sqe_info)
 453 {
 454         u32 total_len, limit_len, checked_len, left_len, adjust_mss;
 455         u32 i, first_mss_sges, left_sges;
 456         struct rte_mbuf *mbuf_head, *mbuf_pre;
 457
 458         left_sges = mbuf->nb_segs;
 459         mbuf_head = mbuf;
 460
 461         /* tso sge number validation */
 462         if (unlikely(left_sges >= HINIC_NONTSO_PKT_MAX_SGE)) {
 463                 checked_len = 0;
 464                 adjust_mss = mbuf->tso_segsz >= TX_MSS_MIN ?
 465                                 mbuf->tso_segsz : TX_MSS_MIN;
 466                 limit_len = adjust_mss + poff_info->payload_offset;
 467                 first_mss_sges = HINIC_NONTSO_PKT_MAX_SGE;
 468
 469                 /* each continues 17 mbufs segmust do one check */
 470                 while (left_sges >= HINIC_NONTSO_PKT_MAX_SGE) {
 471                         /* total len of first 16 mbufs must equal
 472                          * or more than limit_len
 473                          */
 474                         total_len = 0;
 475                         for (i = 0; i < first_mss_sges; i++) {
 476                                 total_len += mbuf->data_len;
 477                                 mbuf_pre = mbuf;
 478                                 mbuf = mbuf->next;
 479                                 if (total_len >= limit_len) {
 480                                         limit_len = adjust_mss;
 481                                         break;
 482                                 }
 483                         }
 484
 485                         checked_len += total_len;
 486
 487                         /* try to copy if not valid */
 488                         if (unlikely(first_mss_sges == i)) {
 489                                 left_sges -= first_mss_sges;
 490                                 checked_len -= mbuf_pre->data_len;
 491
 492                                 left_len = mbuf_head->pkt_len - checked_len;
 493                                 if (left_len > HINIC_COPY_MBUF_SIZE)
 494                                         return false;
 495
 496                                 sqe_info->sge_cnt = mbuf_head->nb_segs -
 497                                                         left_sges;
 498                                 sqe_info->cpy_mbuf_cnt = 1;
 499
 500                                 return true;
 501                         }
 502                         first_mss_sges = (HINIC_NONTSO_PKT_MAX_SGE - 1);
 503
 504                         /* continue next 16 mbufs */
 505                         left_sges -= (i + 1);
 506                 } /* end of while */
 507         }
 508
 509         sqe_info->sge_cnt = mbuf_head->nb_segs;
 510         return true;
 511 }
 512
 513 static inline void
 514 hinic_set_l4_csum_info(struct hinic_sq_task *task,
 515                 u32 *queue_info, struct hinic_tx_offload_info *poff_info)
 516 {
 517         u32 tcp_udp_cs, sctp = 0;
 518         u16 l2hdr_len;
 519
 520         if (unlikely(poff_info->inner_l4_type == SCTP_OFFLOAD_ENABLE))
 521                 sctp = 1;
 522
 523         tcp_udp_cs = poff_info->inner_l4_tcp_udp;
 524
 525         if (poff_info->tunnel_type == TUNNEL_UDP_CSUM ||
 526             poff_info->tunnel_type == TUNNEL_UDP_NO_CSUM) {
 527                 l2hdr_len =  poff_info->outer_l2_len;
 528
 529                 task->pkt_info2 |=
 530                 SQ_TASK_INFO2_SET(poff_info->outer_l3_type, OUTER_L3TYPE) |
 531                 SQ_TASK_INFO2_SET(poff_info->outer_l3_len, OUTER_L3LEN);
 532                 task->pkt_info2 |=
 533                 SQ_TASK_INFO2_SET(poff_info->tunnel_type, TUNNEL_L4TYPE) |
 534                 SQ_TASK_INFO2_SET(poff_info->tunnel_length, TUNNEL_L4LEN);
 535         } else {
 536                 l2hdr_len = poff_info->inner_l2_len;
 537         }
 538
 539         task->pkt_info0 |= SQ_TASK_INFO0_SET(l2hdr_len, L2HDR_LEN);
 540         task->pkt_info1 |=
 541                 SQ_TASK_INFO1_SET(poff_info->inner_l3_len, INNER_L3LEN);
 542         task->pkt_info0 |=
 543                 SQ_TASK_INFO0_SET(poff_info->inner_l3_type, INNER_L3TYPE);
 544         task->pkt_info1 |=
 545                 SQ_TASK_INFO1_SET(poff_info->inner_l4_len, INNER_L4LEN);
 546         task->pkt_info0 |=
 547                 SQ_TASK_INFO0_SET(poff_info->inner_l4_type, L4OFFLOAD);
 548         *queue_info |=
 549                 SQ_CTRL_QUEUE_INFO_SET(poff_info->payload_offset, PLDOFF) |
 550                 SQ_CTRL_QUEUE_INFO_SET(tcp_udp_cs, TCPUDP_CS) |
 551                 SQ_CTRL_QUEUE_INFO_SET(sctp, SCTP);
 552 }
 553
 554 static inline void
 555 hinic_set_tso_info(struct hinic_sq_task *task,
 556                 u32 *queue_info, struct rte_mbuf *mbuf,
 557                 struct hinic_tx_offload_info *poff_info)
 558 {
 559         hinic_set_l4_csum_info(task, queue_info, poff_info);
 560
 561         /* wqe for tso */
 562         task->pkt_info0 |=
 563                 SQ_TASK_INFO0_SET(poff_info->inner_l3_type, INNER_L3TYPE);
 564         task->pkt_info0 |= SQ_TASK_INFO0_SET(TSO_ENABLE, TSO_UFO);
 565         *queue_info |= SQ_CTRL_QUEUE_INFO_SET(TSO_ENABLE, TSO);
 566         /* qsf was initialized in prepare_sq_wqe */
 567         *queue_info = SQ_CTRL_QUEUE_INFO_CLEAR(*queue_info, MSS);
 568         *queue_info |= SQ_CTRL_QUEUE_INFO_SET(mbuf->tso_segsz, MSS);
 569 }
 570
 571 static inline void
 572 hinic_set_vlan_tx_offload(struct hinic_sq_task *task,
 573                         u32 *queue_info, u16 vlan_tag, u16 vlan_pri)
 574 {
 575         task->pkt_info0 |= SQ_TASK_INFO0_SET(vlan_tag, VLAN_TAG) |
 576                                 SQ_TASK_INFO0_SET(1U, VLAN_OFFLOAD);
 577
 578         *queue_info |= SQ_CTRL_QUEUE_INFO_SET(vlan_pri, PRI);
 579 }
 580
 581 static inline void
 582 hinic_fill_tx_offload_info(struct rte_mbuf *mbuf,
 583                 struct hinic_sq_task *task, u32 *queue_info,
 584                 struct hinic_tx_offload_info *tx_off_info)
 585 {
 586         u16 vlan_tag;
 587         uint64_t ol_flags = mbuf->ol_flags;
 588
 589         /* clear DW0~2 of task section for offload */
 590         task->pkt_info0 = 0;
 591         task->pkt_info1 = 0;
 592         task->pkt_info2 = 0;
 593
 594         /* Base VLAN */
 595         if (unlikely(ol_flags & PKT_TX_VLAN_PKT)) {
 596                 vlan_tag = mbuf->vlan_tci;
 597                 hinic_set_vlan_tx_offload(task, queue_info, vlan_tag,
 598                                           vlan_tag >> VLAN_PRIO_SHIFT);
 599         }
 600
 601         /* non checksum or tso */
 602         if (unlikely(!(ol_flags & HINIC_TX_CKSUM_OFFLOAD_MASK)))
 603                 return;
 604
 605         if ((ol_flags & PKT_TX_TCP_SEG))
 606                 /* set tso info for task and qsf */
 607                 hinic_set_tso_info(task, queue_info, mbuf, tx_off_info);
 608         else /* just support l4 checksum offload */
 609                 hinic_set_l4_csum_info(task, queue_info, tx_off_info);
 610 }
 611
 612 static inline void hinic_xmit_mbuf_cleanup(struct hinic_txq *txq)
 613 {
 614         struct hinic_tx_info *tx_info;
 615         struct rte_mbuf *mbuf, *m, *mbuf_free[HINIC_MAX_TX_FREE_BULK];
 616         int i, nb_free = 0;
 617         u16 hw_ci, sw_ci, sq_mask;
 618         int wqebb_cnt = 0;
 619
 620         hw_ci = HINIC_GET_SQ_HW_CI(txq);
 621         sw_ci = HINIC_GET_SQ_LOCAL_CI(txq);
 622         sq_mask = HINIC_GET_SQ_WQE_MASK(txq);
 623
 624         for (i = 0; i < txq->tx_free_thresh; ++i) {
 625                 tx_info = &txq->tx_info[sw_ci];
 626                 if (hw_ci == sw_ci ||
 627                         (((hw_ci - sw_ci) & sq_mask) < tx_info->wqebb_cnt))
 628                         break;
 629
 630                 sw_ci = (sw_ci + tx_info->wqebb_cnt) & sq_mask;
 631
 632                 if (unlikely(tx_info->cpy_mbuf != NULL)) {
 633                         rte_pktmbuf_free(tx_info->cpy_mbuf);
 634                         tx_info->cpy_mbuf = NULL;
 635                 }
 636
 637                 wqebb_cnt += tx_info->wqebb_cnt;
 638                 mbuf = tx_info->mbuf;
 639
 640                 if (likely(mbuf->nb_segs == 1)) {
 641                         m = rte_pktmbuf_prefree_seg(mbuf);
 642                         tx_info->mbuf = NULL;
 643
 644                         if (unlikely(m == NULL))
 645                                 continue;
 646
 647                         mbuf_free[nb_free++] = m;
 648                         if (unlikely(m->pool != mbuf_free[0]->pool ||
 649                                 nb_free >= HINIC_MAX_TX_FREE_BULK)) {
 650                                 rte_mempool_put_bulk(mbuf_free[0]->pool,
 651                                         (void **)mbuf_free, (nb_free - 1));
 652                                 nb_free = 0;
 653                                 mbuf_free[nb_free++] = m;
 654                         }
 655                 } else {
 656                         rte_pktmbuf_free(mbuf);
 657                         tx_info->mbuf = NULL;
 658                 }
 659         }
 660
 661         if (nb_free > 0)
 662                 rte_mempool_put_bulk(mbuf_free[0]->pool, (void **)mbuf_free,
 663                                      nb_free);
 664
 665         HINIC_UPDATE_SQ_LOCAL_CI(txq, wqebb_cnt);
 666 }
 667
 668 static inline struct hinic_sq_wqe *
 669 hinic_get_sq_wqe(struct hinic_txq *txq, int wqebb_cnt,
 670                  struct hinic_wqe_info *wqe_info)
 671 {
 672         u32 cur_pi, end_pi;
 673         u16 remain_wqebbs;
 674         struct hinic_sq *sq = txq->sq;
 675         struct hinic_wq *wq = txq->wq;
 676
 677         /* record current pi */
 678         cur_pi = MASKED_WQE_IDX(wq, wq->prod_idx);
 679         end_pi = cur_pi + wqebb_cnt;
 680
 681         /* update next pi and delta */
 682         wq->prod_idx += wqebb_cnt;
 683         wq->delta -= wqebb_cnt;
 684
 685         /* return current pi and owner */
 686         wqe_info->pi = cur_pi;
 687         wqe_info->owner = sq->owner;
 688         wqe_info->around = 0;
 689         wqe_info->seq_wqebbs = wqebb_cnt;
 690
 691         if (unlikely(end_pi >= txq->q_depth)) {
 692                 /* update owner of next prod_idx */
 693                 sq->owner = !sq->owner;
 694
 695                 /* turn around to head */
 696                 if (unlikely(end_pi > txq->q_depth)) {
 697                         wqe_info->around = 1;
 698                         remain_wqebbs = txq->q_depth - cur_pi;
 699                         wqe_info->seq_wqebbs = remain_wqebbs;
 700                 }
 701         }
 702
 703         return (struct hinic_sq_wqe *)WQ_WQE_ADDR(wq, cur_pi);
 704 }
 705
 706 static inline uint16_t
 707 hinic_ipv4_phdr_cksum(const struct rte_ipv4_hdr *ipv4_hdr, uint64_t ol_flags)
 708 {
 709         struct ipv4_psd_header {
 710                 uint32_t src_addr; /* IP address of source host. */
 711                 uint32_t dst_addr; /* IP address of destination host. */
 712                 uint8_t  zero;     /* zero. */
 713                 uint8_t  proto;    /* L4 protocol type. */
 714                 uint16_t len;      /* L4 length. */
 715         } psd_hdr;
 716
 717         psd_hdr.src_addr = ipv4_hdr->src_addr;
 718         psd_hdr.dst_addr = ipv4_hdr->dst_addr;
 719         psd_hdr.zero = 0;
 720         psd_hdr.proto = ipv4_hdr->next_proto_id;
 721         if (ol_flags & PKT_TX_TCP_SEG) {
 722                 psd_hdr.len = 0;
 723         } else {
 724                 psd_hdr.len =
 725                 rte_cpu_to_be_16(rte_be_to_cpu_16(ipv4_hdr->total_length) -
 726                                  rte_ipv4_hdr_len(ipv4_hdr));
 727         }
 728         return rte_raw_cksum(&psd_hdr, sizeof(psd_hdr));
 729 }
 730
 731 static inline uint16_t
 732 hinic_ipv6_phdr_cksum(const struct rte_ipv6_hdr *ipv6_hdr, uint64_t ol_flags)
 733 {
 734         uint32_t sum;
 735         struct {
 736                 uint32_t len;   /* L4 length. */
 737                 uint32_t proto; /* L4 protocol - top 3 bytes must be zero */
 738         } psd_hdr;
 739
 740         psd_hdr.proto = (ipv6_hdr->proto << 24);
 741         if (ol_flags & PKT_TX_TCP_SEG)
 742                 psd_hdr.len = 0;
 743         else
 744                 psd_hdr.len = ipv6_hdr->payload_len;
 745
 746         sum = __rte_raw_cksum(ipv6_hdr->src_addr,
 747                 sizeof(ipv6_hdr->src_addr) + sizeof(ipv6_hdr->dst_addr), 0);
 748         sum = __rte_raw_cksum(&psd_hdr, sizeof(psd_hdr), sum);
 749         return __rte_raw_cksum_reduce(sum);
 750 }
 751
 752 static inline void hinic_get_outer_cs_pld_offset(struct rte_mbuf *m,
 753                                         struct hinic_tx_offload_info *off_info)
 754 {
 755         uint64_t ol_flags = m->ol_flags;
 756
 757         if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_UDP_CKSUM)
 758                 off_info->payload_offset = m->outer_l2_len + m->outer_l3_len +
 759                                            m->l2_len + m->l3_len;
 760         else if ((ol_flags & PKT_TX_TCP_CKSUM) || (ol_flags & PKT_TX_TCP_SEG))
 761                 off_info->payload_offset = m->outer_l2_len + m->outer_l3_len +
 762                                            m->l2_len + m->l3_len + m->l4_len;
 763 }
 764
 765 static inline void hinic_get_pld_offset(struct rte_mbuf *m,
 766                                         struct hinic_tx_offload_info *off_info)
 767 {
 768         uint64_t ol_flags = m->ol_flags;
 769
 770         if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_UDP_CKSUM)
 771                 off_info->payload_offset = m->l2_len + m->l3_len;
 772         else if ((ol_flags & PKT_TX_TCP_CKSUM) || (ol_flags & PKT_TX_TCP_SEG))
 773                 off_info->payload_offset = m->l2_len + m->l3_len +
 774                                            m->l4_len;
 775 }
 776
 777 static inline void hinic_analyze_tx_info(struct rte_mbuf *mbuf,
 778                                          struct hinic_tx_offload_info *off_info)
 779 {
 780         struct rte_ether_hdr *eth_hdr;
 781         struct rte_vlan_hdr *vlan_hdr;
 782         struct rte_ipv4_hdr *ipv4_hdr;
 783         u16 eth_type;
 784
 785         eth_hdr = rte_pktmbuf_mtod(mbuf, struct rte_ether_hdr *);
 786         eth_type = rte_be_to_cpu_16(eth_hdr->ether_type);
 787
 788         if (eth_type == RTE_ETHER_TYPE_VLAN) {
 789                 off_info->outer_l2_len = ETHER_LEN_WITH_VLAN;
 790                 vlan_hdr = (struct rte_vlan_hdr *)(eth_hdr + 1);
 791                 eth_type = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 792         } else {
 793                 off_info->outer_l2_len = ETHER_LEN_NO_VLAN;
 794         }
 795
 796         if (eth_type == RTE_ETHER_TYPE_IPV4) {
 797                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_ipv4_hdr *,
 798                                                    off_info->outer_l2_len);
 799                 off_info->outer_l3_len = rte_ipv4_hdr_len(ipv4_hdr);
 800         } else if (eth_type == RTE_ETHER_TYPE_IPV6) {
 801                 /* not support ipv6 extension header */
 802                 off_info->outer_l3_len = sizeof(struct rte_ipv6_hdr);
 803         }
 804 }
 805
 806 static inline void hinic_analyze_outer_ip_vxlan(struct rte_mbuf *mbuf,
 807                                         struct hinic_tx_offload_info *off_info)
 808 {
 809         struct rte_ether_hdr *eth_hdr;
 810         struct rte_vlan_hdr *vlan_hdr;
 811         struct rte_ipv4_hdr *ipv4_hdr;
 812         struct rte_udp_hdr *udp_hdr;
 813         u16 eth_type = 0;
 814
 815         eth_hdr = rte_pktmbuf_mtod(mbuf, struct rte_ether_hdr *);
 816         eth_type = rte_be_to_cpu_16(eth_hdr->ether_type);
 817
 818         if (eth_type == RTE_ETHER_TYPE_VLAN) {
 819                 vlan_hdr = (struct rte_vlan_hdr *)(eth_hdr + 1);
 820                 eth_type = rte_be_to_cpu_16(vlan_hdr->eth_proto);
 821         }
 822
 823         if (eth_type == RTE_ETHER_TYPE_IPV4) {
 824                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_ipv4_hdr *,
 825                                                    mbuf->outer_l2_len);
 826                 off_info->outer_l3_type = IPV4_PKT_WITH_CHKSUM_OFFLOAD;
 827                 ipv4_hdr->hdr_checksum = 0;
 828
 829                 udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr +
 830                                                  mbuf->outer_l3_len);
 831                 udp_hdr->dgram_cksum = 0;
 832         } else if (eth_type == RTE_ETHER_TYPE_IPV6) {
 833                 off_info->outer_l3_type = IPV6_PKT;
 834
 835                 udp_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_udp_hdr *,
 836                                                   (mbuf->outer_l2_len +
 837                                                    mbuf->outer_l3_len));
 838                 udp_hdr->dgram_cksum = 0;
 839         }
 840 }
 841
 842 static inline uint8_t hinic_analyze_l3_type(struct rte_mbuf *mbuf)
 843 {
 844         uint8_t l3_type;
 845         uint64_t ol_flags = mbuf->ol_flags;
 846
 847         if (ol_flags & PKT_TX_IPV4)
 848                 l3_type = (ol_flags & PKT_TX_IP_CKSUM) ?
 849                           IPV4_PKT_WITH_CHKSUM_OFFLOAD :
 850                           IPV4_PKT_NO_CHKSUM_OFFLOAD;
 851         else if (ol_flags & PKT_TX_IPV6)
 852                 l3_type = IPV6_PKT;
 853         else
 854                 l3_type = UNKNOWN_L3TYPE;
 855
 856         return l3_type;
 857 }
 858
 859 static inline void hinic_calculate_tcp_checksum(struct rte_mbuf *mbuf,
 860                                         struct hinic_tx_offload_info *off_info,
 861                                         uint64_t inner_l3_offset)
 862 {
 863         struct rte_ipv4_hdr *ipv4_hdr;
 864         struct rte_ipv6_hdr *ipv6_hdr;
 865         struct rte_tcp_hdr *tcp_hdr;
 866         uint64_t ol_flags = mbuf->ol_flags;
 867
 868         if (ol_flags & PKT_TX_IPV4) {
 869                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_ipv4_hdr *,
 870                                                    inner_l3_offset);
 871
 872                 if (ol_flags & PKT_TX_IP_CKSUM)
 873                         ipv4_hdr->hdr_checksum = 0;
 874
 875                 tcp_hdr = (struct rte_tcp_hdr *)((char *)ipv4_hdr +
 876                                                  mbuf->l3_len);
 877                 tcp_hdr->cksum = hinic_ipv4_phdr_cksum(ipv4_hdr, ol_flags);
 878         } else {
 879                 ipv6_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_ipv6_hdr *,
 880                                                    inner_l3_offset);
 881                 tcp_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_tcp_hdr *,
 882                                                   (inner_l3_offset +
 883                                                    mbuf->l3_len));
 884                 tcp_hdr->cksum = hinic_ipv6_phdr_cksum(ipv6_hdr, ol_flags);
 885         }
 886
 887         off_info->inner_l4_type = TCP_OFFLOAD_ENABLE;
 888         off_info->inner_l4_tcp_udp = 1;
 889 }
 890
 891 static inline void hinic_calculate_udp_checksum(struct rte_mbuf *mbuf,
 892                                         struct hinic_tx_offload_info *off_info,
 893                                         uint64_t inner_l3_offset)
 894 {
 895         struct rte_ipv4_hdr *ipv4_hdr;
 896         struct rte_ipv6_hdr *ipv6_hdr;
 897         struct rte_udp_hdr *udp_hdr;
 898         uint64_t ol_flags = mbuf->ol_flags;
 899
 900         if (ol_flags & PKT_TX_IPV4) {
 901                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_ipv4_hdr *,
 902                                                    inner_l3_offset);
 903
 904                 if (ol_flags & PKT_TX_IP_CKSUM)
 905                         ipv4_hdr->hdr_checksum = 0;
 906
 907                 udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr +
 908                                                  mbuf->l3_len);
 909                 udp_hdr->dgram_cksum = hinic_ipv4_phdr_cksum(ipv4_hdr,
 910                                                              ol_flags);
 911         } else {
 912                 ipv6_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_ipv6_hdr *,
 913                                                    inner_l3_offset);
 914
 915                 udp_hdr = rte_pktmbuf_mtod_offset(mbuf, struct rte_udp_hdr *,
 916                                                   (inner_l3_offset +
 917                                                    mbuf->l3_len));
 918                 udp_hdr->dgram_cksum = hinic_ipv6_phdr_cksum(ipv6_hdr,
 919                                                              ol_flags);
 920         }
 921
 922         off_info->inner_l4_type = UDP_OFFLOAD_ENABLE;
 923         off_info->inner_l4_tcp_udp = 1;
 924 }
 925
 926 static inline void
 927 hinic_calculate_sctp_checksum(struct hinic_tx_offload_info *off_info)
 928 {
 929         off_info->inner_l4_type = SCTP_OFFLOAD_ENABLE;
 930         off_info->inner_l4_tcp_udp = 0;
 931         off_info->inner_l4_len = sizeof(struct rte_sctp_hdr);
 932 }
 933
 934 static inline void hinic_calculate_checksum(struct rte_mbuf *mbuf,
 935                                         struct hinic_tx_offload_info *off_info,
 936                                         uint64_t inner_l3_offset)
 937 {
 938         uint64_t ol_flags = mbuf->ol_flags;
 939
 940         switch (ol_flags & PKT_TX_L4_MASK) {
 941         case PKT_TX_UDP_CKSUM:
 942                 hinic_calculate_udp_checksum(mbuf, off_info, inner_l3_offset);
 943                 break;
 944
 945         case PKT_TX_TCP_CKSUM:
 946                 hinic_calculate_tcp_checksum(mbuf, off_info, inner_l3_offset);
 947                 break;
 948
 949         case PKT_TX_SCTP_CKSUM:
 950                 hinic_calculate_sctp_checksum(off_info);
 951                 break;
 952
 953         default:
 954                 if (ol_flags & PKT_TX_TCP_SEG)
 955                         hinic_calculate_tcp_checksum(mbuf, off_info,
 956                                                      inner_l3_offset);
 957                 break;
 958         }
 959 }
 960
 961 static inline int hinic_tx_offload_pkt_prepare(struct rte_mbuf *m,
 962                                         struct hinic_tx_offload_info *off_info)
 963 {
 964         uint64_t inner_l3_offset;
 965         uint64_t ol_flags = m->ol_flags;
 966
 967         /* Check if the packets set available offload flags */
 968         if (!(ol_flags & HINIC_TX_CKSUM_OFFLOAD_MASK))
 969                 return 0;
 970
 971         /* Support only vxlan offload */
 972         if (unlikely((ol_flags & PKT_TX_TUNNEL_MASK) &&
 973             !(ol_flags & PKT_TX_TUNNEL_VXLAN)))
 974                 return -ENOTSUP;
 975
 976 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
 977         if (rte_validate_tx_offload(m) != 0)
 978                 return -EINVAL;
 979 #endif
 980
 981         if (ol_flags & PKT_TX_TUNNEL_VXLAN) {
 982                 off_info->tunnel_type = TUNNEL_UDP_NO_CSUM;
 983
 984                 /* inner_l4_tcp_udp csum should be set to calculate outer
 985                  * udp checksum when vxlan packets without inner l3 and l4
 986                  */
 987                 off_info->inner_l4_tcp_udp = 1;
 988
 989                 if ((ol_flags & PKT_TX_OUTER_IP_CKSUM) ||
 990                     (ol_flags & PKT_TX_OUTER_IPV6) ||
 991                     (ol_flags & PKT_TX_TCP_SEG)) {
 992                         inner_l3_offset = m->l2_len + m->outer_l2_len +
 993                                           m->outer_l3_len;
 994                         off_info->outer_l2_len = m->outer_l2_len;
 995                         off_info->outer_l3_len = m->outer_l3_len;
 996                         /* just support vxlan tunneling pkt */
 997                         off_info->inner_l2_len = m->l2_len - VXLANLEN -
 998                                                  sizeof(struct rte_udp_hdr);
 999                         off_info->tunnel_length = m->l2_len;
1000
1001                         hinic_analyze_outer_ip_vxlan(m, off_info);
1002
1003                         hinic_get_outer_cs_pld_offset(m, off_info);
1004                 } else {
1005                         inner_l3_offset = m->l2_len;
1006                         hinic_analyze_tx_info(m, off_info);
1007                         /* just support vxlan tunneling pkt */
1008                         off_info->inner_l2_len = m->l2_len - VXLANLEN -
1009                                                  sizeof(struct rte_udp_hdr) -
1010                                                  off_info->outer_l2_len -
1011                                                  off_info->outer_l3_len;
1012                         off_info->tunnel_length = m->l2_len -
1013                                                   off_info->outer_l2_len -
1014                                                   off_info->outer_l3_len;
1015                         off_info->outer_l3_type = IPV4_PKT_NO_CHKSUM_OFFLOAD;
1016
1017                         hinic_get_pld_offset(m, off_info);
1018                 }
1019         } else {
1020                 inner_l3_offset = m->l2_len;
1021                 off_info->inner_l2_len = m->l2_len;
1022                 off_info->tunnel_type = NOT_TUNNEL;
1023
1024                 hinic_get_pld_offset(m, off_info);
1025         }
1026
1027         /* invalid udp or tcp header */
1028         if (unlikely(off_info->payload_offset > MAX_PLD_OFFSET))
1029                 return -EINVAL;
1030
1031         off_info->inner_l3_len = m->l3_len;
1032         off_info->inner_l4_len = m->l4_len;
1033         off_info->inner_l3_type = hinic_analyze_l3_type(m);
1034
1035         /* Process the pseudo-header checksum */
1036         hinic_calculate_checksum(m, off_info, inner_l3_offset);
1037
1038         return 0;
1039 }
1040
1041 static inline bool hinic_get_sge_txoff_info(struct rte_mbuf *mbuf_pkt,
1042                                             struct hinic_wqe_info *sqe_info,
1043                                             struct hinic_tx_offload_info
1044                                             *off_info)
1045 {
1046         u16  i, total_len, sge_cnt = mbuf_pkt->nb_segs;
1047         struct rte_mbuf *mbuf;
1048         int ret;
1049
1050         memset(off_info, 0, sizeof(*off_info));
1051
1052         ret = hinic_tx_offload_pkt_prepare(mbuf_pkt, off_info);
1053         if (unlikely(ret))
1054                 return false;
1055
1056         sqe_info->cpy_mbuf_cnt = 0;
1057
1058         /* non tso mbuf */
1059         if (likely(!(mbuf_pkt->ol_flags & PKT_TX_TCP_SEG))) {
1060                 if (unlikely(mbuf_pkt->pkt_len > MAX_SINGLE_SGE_SIZE)) {
1061                         /* non tso packet len must less than 64KB */
1062                         return false;
1063                 } else if (unlikely(HINIC_NONTSO_SEG_NUM_INVALID(sge_cnt))) {
1064                         /* non tso packet buffer number must less than 17
1065                          * the mbuf segs more than 17 must copy to one buffer
1066                          */
1067                         total_len = 0;
1068                         mbuf = mbuf_pkt;
1069                         for (i = 0; i < (HINIC_NONTSO_PKT_MAX_SGE - 1) ; i++) {
1070                                 total_len += mbuf->data_len;
1071                                 mbuf = mbuf->next;
1072                         }
1073
1074                         /* default support copy total 4k mbuf segs */
1075                         if ((u32)(total_len + (u16)HINIC_COPY_MBUF_SIZE) <
1076                                   mbuf_pkt->pkt_len)
1077                                 return false;
1078
1079                         sqe_info->sge_cnt = HINIC_NONTSO_PKT_MAX_SGE;
1080                         sqe_info->cpy_mbuf_cnt = 1;
1081                         return true;
1082                 }
1083
1084                 /* valid non tso mbuf */
1085                 sqe_info->sge_cnt = sge_cnt;
1086         } else {
1087                 /* tso mbuf */
1088                 if (unlikely(HINIC_TSO_SEG_NUM_INVALID(sge_cnt)))
1089                         /* too many mbuf segs */
1090                         return false;
1091
1092                 /* check tso mbuf segs are valid or not */
1093                 if (unlikely(!hinic_is_tso_sge_valid(mbuf_pkt,
1094                              off_info, sqe_info)))
1095                         return false;
1096         }
1097
1098         return true;
1099 }
1100
1101 static inline void hinic_sq_write_db(struct hinic_sq *sq, int cos)
1102 {
1103         u16 prod_idx;
1104         u32 hi_prod_idx;
1105         struct hinic_sq_db sq_db;
1106
1107         prod_idx = MASKED_SQ_IDX(sq, sq->wq->prod_idx);
1108         hi_prod_idx = SQ_DB_PI_HIGH(prod_idx);
1109
1110         sq_db.db_info = SQ_DB_INFO_SET(hi_prod_idx, HI_PI) |
1111                         SQ_DB_INFO_SET(SQ_DB, TYPE) |
1112                         SQ_DB_INFO_SET(SQ_CFLAG_DP, CFLAG) |
1113                         SQ_DB_INFO_SET(cos, COS) |
1114                         SQ_DB_INFO_SET(sq->q_id, QID);
1115
1116         /* Data should be written to HW in Big Endian Format */
1117         sq_db.db_info = cpu_to_be32(sq_db.db_info);
1118
1119         /* Write all before the doorbell */
1120         rte_wmb();
1121         writel(sq_db.db_info, SQ_DB_ADDR(sq, prod_idx));
1122 }
1123
1124 u16 hinic_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, u16 nb_pkts)
1125 {
1126         int free_wqebb_cnt, wqe_wqebb_cnt;
1127         u32 queue_info, tx_bytes = 0;
1128         u16 nb_tx;
1129         struct hinic_wqe_info sqe_info;
1130         struct hinic_tx_offload_info off_info;
1131         struct rte_mbuf *mbuf_pkt;
1132         struct hinic_txq *txq = tx_queue;
1133         struct hinic_tx_info *tx_info;
1134         struct hinic_sq_wqe *sq_wqe;
1135         struct hinic_sq_task *task;
1136
1137         /* reclaim tx mbuf before xmit new packet */
1138         if (HINIC_GET_SQ_FREE_WQEBBS(txq) < txq->tx_free_thresh)
1139                 hinic_xmit_mbuf_cleanup(txq);
1140
1141         /* tx loop routine */
1142         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1143                 mbuf_pkt = *tx_pkts++;
1144                 queue_info = 0;
1145
1146                 /* 1. parse sge and tx offlod info from mbuf */
1147                 if (unlikely(!hinic_get_sge_txoff_info(mbuf_pkt,
1148                                                        &sqe_info, &off_info))) {
1149                         txq->txq_stats.off_errs++;
1150                         break;
1151                 }
1152
1153                 /* 2. try to get enough wqebb */
1154                 wqe_wqebb_cnt = HINIC_SQ_WQEBB_CNT(sqe_info.sge_cnt);
1155                 free_wqebb_cnt = HINIC_GET_SQ_FREE_WQEBBS(txq);
1156                 if (unlikely(wqe_wqebb_cnt > free_wqebb_cnt)) {
1157                         /* reclaim again */
1158                         hinic_xmit_mbuf_cleanup(txq);
1159                         free_wqebb_cnt = HINIC_GET_SQ_FREE_WQEBBS(txq);
1160                         if (unlikely(wqe_wqebb_cnt > free_wqebb_cnt)) {
1161                                 txq->txq_stats.tx_busy += (nb_pkts - nb_tx);
1162                                 break;
1163                         }
1164                 }
1165
1166                 /* 3. get sq tail wqe address from wqe_page,
1167                  * sq have enough wqebb for this packet
1168                  */
1169                 sq_wqe = hinic_get_sq_wqe(txq, wqe_wqebb_cnt, &sqe_info);
1170
1171                 /* 4. fill sq wqe sge section */
1172                 if (unlikely(!hinic_mbuf_dma_map_sge(txq, mbuf_pkt,
1173                                                      sq_wqe->buf_descs,
1174                                                      &sqe_info))) {
1175                         hinic_return_sq_wqe(txq->nic_dev->hwdev, txq->q_id,
1176                                             wqe_wqebb_cnt, sqe_info.owner);
1177                         txq->txq_stats.off_errs++;
1178                         break;
1179                 }
1180
1181                 /* 5. fill sq wqe task section and queue info */
1182                 task = &sq_wqe->task;
1183
1184                 /* tx packet offload configure */
1185                 hinic_fill_tx_offload_info(mbuf_pkt, task, &queue_info,
1186                                            &off_info);
1187
1188                 /* 6. record tx info */
1189                 tx_info = &txq->tx_info[sqe_info.pi];
1190                 tx_info->mbuf = mbuf_pkt;
1191                 tx_info->wqebb_cnt = wqe_wqebb_cnt;
1192
1193                 /* 7. fill sq wqe header section */
1194                 hinic_fill_sq_wqe_header(&sq_wqe->ctrl, queue_info,
1195                                          sqe_info.sge_cnt, sqe_info.owner);
1196
1197                 /* 8.convert continue or bottom wqe byteorder to big endian */
1198                 hinic_sq_wqe_cpu_to_be32(sq_wqe, sqe_info.seq_wqebbs);
1199
1200                 tx_bytes += mbuf_pkt->pkt_len;
1201         }
1202
1203         /* 9. write sq doorbell in burst mode */
1204         if (nb_tx) {
1205                 hinic_sq_write_db(txq->sq, txq->cos);
1206
1207                 txq->txq_stats.packets += nb_tx;
1208                 txq->txq_stats.bytes += tx_bytes;
1209         }
1210         txq->txq_stats.burst_pkts = nb_tx;
1211
1212         return nb_tx;
1213 }
1214
1215 void hinic_free_all_tx_mbufs(struct hinic_txq *txq)
1216 {
1217         u16 ci;
1218         struct hinic_nic_dev *nic_dev = txq->nic_dev;
1219         struct hinic_tx_info *tx_info;
1220         int free_wqebbs = hinic_get_sq_free_wqebbs(nic_dev->hwdev,
1221                                                    txq->q_id) + 1;
1222
1223         while (free_wqebbs < txq->q_depth) {
1224                 ci = hinic_get_sq_local_ci(nic_dev->hwdev, txq->q_id);
1225
1226                 tx_info = &txq->tx_info[ci];
1227
1228                 if (unlikely(tx_info->cpy_mbuf != NULL)) {
1229                         rte_pktmbuf_free(tx_info->cpy_mbuf);
1230                         tx_info->cpy_mbuf = NULL;
1231                 }
1232
1233                 rte_pktmbuf_free(tx_info->mbuf);
1234                 hinic_update_sq_local_ci(nic_dev->hwdev, txq->q_id,
1235                                          tx_info->wqebb_cnt);
1236
1237                 free_wqebbs += tx_info->wqebb_cnt;
1238                 tx_info->mbuf = NULL;
1239         }
1240 }
1241
1242 void hinic_free_all_tx_resources(struct rte_eth_dev *eth_dev)
1243 {
1244         u16 q_id;
1245         struct hinic_nic_dev *nic_dev =
1246                                 HINIC_ETH_DEV_TO_PRIVATE_NIC_DEV(eth_dev);
1247
1248         for (q_id = 0; q_id < nic_dev->num_sq; q_id++) {
1249                 if (eth_dev->data->tx_queues != NULL)
1250                         eth_dev->data->tx_queues[q_id] = NULL;
1251
1252                 if (nic_dev->txqs[q_id] == NULL)
1253                         continue;
1254
1255                 /* stop tx queue free tx mbuf */
1256                 hinic_free_all_tx_mbufs(nic_dev->txqs[q_id]);
1257                 hinic_free_tx_resources(nic_dev->txqs[q_id]);
1258
1259                 /* free txq */
1260                 kfree(nic_dev->txqs[q_id]);
1261                 nic_dev->txqs[q_id] = NULL;
1262         }
1263 }
1264
1265 void hinic_free_all_tx_mbuf(struct rte_eth_dev *eth_dev)
1266 {
1267         u16 q_id;
1268         struct hinic_nic_dev *nic_dev =
1269                                 HINIC_ETH_DEV_TO_PRIVATE_NIC_DEV(eth_dev);
1270
1271         for (q_id = 0; q_id < nic_dev->num_sq; q_id++)
1272                 /* stop tx queue free tx mbuf */
1273                 hinic_free_all_tx_mbufs(nic_dev->txqs[q_id]);
1274 }
1275
1276 int hinic_setup_tx_resources(struct hinic_txq *txq)
1277 {
1278         u64 tx_info_sz;
1279
1280         tx_info_sz = txq->q_depth * sizeof(*txq->tx_info);
1281         txq->tx_info = rte_zmalloc_socket("tx_info", tx_info_sz,
1282                         RTE_CACHE_LINE_SIZE, txq->socket_id);
1283         if (!txq->tx_info)
1284                 return -ENOMEM;
1285
1286         return HINIC_OK;
1287 }
1288
1289 void hinic_free_tx_resources(struct hinic_txq *txq)
1290 {
1291         if (txq->tx_info == NULL)
1292                 return;
1293
1294         rte_free(txq->tx_info);
1295         txq->tx_info = NULL;
1296 }
1297
1298 int hinic_create_sq(struct hinic_hwdev *hwdev, u16 q_id,
1299                         u16 sq_depth, unsigned int socket_id)
1300 {
1301         int err;
1302         struct hinic_nic_io *nic_io = hwdev->nic_io;
1303         struct hinic_qp *qp = &nic_io->qps[q_id];
1304         struct hinic_sq *sq = &qp->sq;
1305         void __iomem *db_addr;
1306         volatile u32 *ci_addr;
1307
1308         sq->sq_depth = sq_depth;
1309         nic_io->sq_depth = sq_depth;
1310
1311         /* alloc wq */
1312         err = hinic_wq_allocate(nic_io->hwdev, &nic_io->sq_wq[q_id],
1313                                 HINIC_SQ_WQEBB_SHIFT, nic_io->sq_depth,
1314                                 socket_id);
1315         if (err) {
1316                 PMD_DRV_LOG(ERR, "Failed to allocate WQ for SQ");
1317                 return err;
1318         }
1319
1320         /* alloc sq doorbell space */
1321         err = hinic_alloc_db_addr(nic_io->hwdev, &db_addr);
1322         if (err) {
1323                 PMD_DRV_LOG(ERR, "Failed to init db addr");
1324                 goto alloc_db_err;
1325         }
1326
1327         /* clear hardware ci */
1328         ci_addr = (volatile u32 *)HINIC_CI_VADDR(nic_io->ci_vaddr_base, q_id);
1329         *ci_addr = 0;
1330
1331         sq->q_id = q_id;
1332         sq->wq = &nic_io->sq_wq[q_id];
1333         sq->owner = 1;
1334         sq->cons_idx_addr = (volatile u16 *)ci_addr;
1335         sq->db_addr = db_addr;
1336
1337         return HINIC_OK;
1338
1339 alloc_db_err:
1340         hinic_wq_free(nic_io->hwdev, &nic_io->sq_wq[q_id]);
1341
1342         return err;
1343 }
1344
1345 void hinic_destroy_sq(struct hinic_hwdev *hwdev, u16 q_id)
1346 {
1347         struct hinic_nic_io *nic_io;
1348         struct hinic_qp *qp;
1349
1350         nic_io = hwdev->nic_io;
1351         qp = &nic_io->qps[q_id];
1352
1353         if (qp->sq.wq == NULL)
1354                 return;
1355
1356         hinic_free_db_addr(nic_io->hwdev, qp->sq.db_addr);
1357         hinic_wq_free(nic_io->hwdev, qp->sq.wq);
1358         qp->sq.wq = NULL;
1359 }