drivers/event/cnxk/cn10k_worker.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(C) 2021 Marvell.
   3  */
   4
   5 #ifndef __CN10K_WORKER_H__
   6 #define __CN10K_WORKER_H__
   7
   8 #include <rte_vect.h>
   9
  10 #include "cn10k_cryptodev_ops.h"
  11 #include "cnxk_ethdev.h"
  12 #include "cnxk_eventdev.h"
  13 #include "cnxk_worker.h"
  14
  15 #include "cn10k_ethdev.h"
  16 #include "cn10k_rx.h"
  17 #include "cn10k_tx.h"
  18
  19 /* SSO Operations */
  20
  21 static __rte_always_inline uint8_t
  22 cn10k_sso_hws_new_event(struct cn10k_sso_hws *ws, const struct rte_event *ev)
  23 {
  24         const uint32_t tag = (uint32_t)ev->event;
  25         const uint8_t new_tt = ev->sched_type;
  26         const uint64_t event_ptr = ev->u64;
  27         const uint16_t grp = ev->queue_id;
  28
  29         rte_atomic_thread_fence(__ATOMIC_ACQ_REL);
  30         if (ws->xaq_lmt <= *ws->fc_mem)
  31                 return 0;
  32
  33         cnxk_sso_hws_add_work(event_ptr, tag, new_tt,
  34                               ws->grp_base + (grp << 12));
  35         return 1;
  36 }
  37
  38 static __rte_always_inline void
  39 cn10k_sso_hws_fwd_swtag(struct cn10k_sso_hws *ws, const struct rte_event *ev)
  40 {
  41         const uint32_t tag = (uint32_t)ev->event;
  42         const uint8_t new_tt = ev->sched_type;
  43         const uint8_t cur_tt =
  44                 CNXK_TT_FROM_TAG(plt_read64(ws->base + SSOW_LF_GWS_WQE0));
  45
  46         /* CNXK model
  47          * cur_tt/new_tt     SSO_TT_ORDERED SSO_TT_ATOMIC SSO_TT_UNTAGGED
  48          *
  49          * SSO_TT_ORDERED        norm           norm             untag
  50          * SSO_TT_ATOMIC         norm           norm               untag
  51          * SSO_TT_UNTAGGED       norm           norm             NOOP
  52          */
  53
  54         if (new_tt == SSO_TT_UNTAGGED) {
  55                 if (cur_tt != SSO_TT_UNTAGGED)
  56                         cnxk_sso_hws_swtag_untag(ws->base +
  57                                                  SSOW_LF_GWS_OP_SWTAG_UNTAG);
  58         } else {
  59                 cnxk_sso_hws_swtag_norm(tag, new_tt,
  60                                         ws->base + SSOW_LF_GWS_OP_SWTAG_NORM);
  61         }
  62         ws->swtag_req = 1;
  63 }
  64
  65 static __rte_always_inline void
  66 cn10k_sso_hws_fwd_group(struct cn10k_sso_hws *ws, const struct rte_event *ev,
  67                         const uint16_t grp)
  68 {
  69         const uint32_t tag = (uint32_t)ev->event;
  70         const uint8_t new_tt = ev->sched_type;
  71
  72         plt_write64(ev->u64, ws->base + SSOW_LF_GWS_OP_UPD_WQP_GRP1);
  73         cnxk_sso_hws_swtag_desched(tag, new_tt, grp,
  74                                    ws->base + SSOW_LF_GWS_OP_SWTAG_DESCHED);
  75 }
  76
  77 static __rte_always_inline void
  78 cn10k_sso_hws_forward_event(struct cn10k_sso_hws *ws,
  79                             const struct rte_event *ev)
  80 {
  81         const uint8_t grp = ev->queue_id;
  82
  83         /* Group hasn't changed, Use SWTAG to forward the event */
  84         if (CNXK_GRP_FROM_TAG(plt_read64(ws->base + SSOW_LF_GWS_WQE0)) == grp)
  85                 cn10k_sso_hws_fwd_swtag(ws, ev);
  86         else
  87                 /*
  88                  * Group has been changed for group based work pipelining,
  89                  * Use deschedule/add_work operation to transfer the event to
  90                  * new group/core
  91                  */
  92                 cn10k_sso_hws_fwd_group(ws, ev, grp);
  93 }
  94
  95 static __rte_always_inline void
  96 cn10k_wqe_to_mbuf(uint64_t wqe, const uint64_t mbuf, uint8_t port_id,
  97                   const uint32_t tag, const uint32_t flags,
  98                   const void *const lookup_mem)
  99 {
 100         const uint64_t mbuf_init = 0x100010000ULL | RTE_PKTMBUF_HEADROOM |
 101                                    (flags & NIX_RX_OFFLOAD_TSTAMP_F ? 8 : 0);
 102
 103         cn10k_nix_cqe_to_mbuf((struct nix_cqe_hdr_s *)wqe, tag,
 104                               (struct rte_mbuf *)mbuf, lookup_mem,
 105                               mbuf_init | ((uint64_t)port_id) << 48, flags);
 106 }
 107
 108 static __rte_always_inline void
 109 cn10k_process_vwqe(uintptr_t vwqe, uint16_t port_id, const uint32_t flags,
 110                    void *lookup_mem, void *tstamp, uintptr_t lbase)
 111 {
 112         uint64_t mbuf_init = 0x100010000ULL | RTE_PKTMBUF_HEADROOM |
 113                              (flags & NIX_RX_OFFLOAD_TSTAMP_F ? 8 : 0);
 114         struct rte_event_vector *vec;
 115         uint64_t aura_handle, laddr;
 116         uint16_t nb_mbufs, non_vec;
 117         uint16_t lmt_id, d_off;
 118         struct rte_mbuf *mbuf;
 119         uint8_t loff = 0;
 120         uint64_t sa_base;
 121         uint64_t **wqe;
 122
 123         mbuf_init |= ((uint64_t)port_id) << 48;
 124         vec = (struct rte_event_vector *)vwqe;
 125         wqe = vec->u64s;
 126
 127         nb_mbufs = RTE_ALIGN_FLOOR(vec->nb_elem, NIX_DESCS_PER_LOOP);
 128         nb_mbufs = cn10k_nix_recv_pkts_vector(&mbuf_init, vec->mbufs, nb_mbufs,
 129                                               flags | NIX_RX_VWQE_F, lookup_mem,
 130                                               tstamp, lbase);
 131         wqe += nb_mbufs;
 132         non_vec = vec->nb_elem - nb_mbufs;
 133
 134         if (flags & NIX_RX_OFFLOAD_SECURITY_F && non_vec) {
 135                 mbuf = (struct rte_mbuf *)((uintptr_t)wqe[0] -
 136                                            sizeof(struct rte_mbuf));
 137                 /* Pick first mbuf's aura handle assuming all
 138                  * mbufs are from a vec and are from same RQ.
 139                  */
 140                 aura_handle = mbuf->pool->pool_id;
 141                 ROC_LMT_BASE_ID_GET(lbase, lmt_id);
 142                 laddr = lbase;
 143                 laddr += 8;
 144                 d_off = ((uintptr_t)mbuf->buf_addr - (uintptr_t)mbuf);
 145                 d_off += (mbuf_init & 0xFFFF);
 146                 sa_base = cnxk_nix_sa_base_get(mbuf_init >> 48, lookup_mem);
 147                 sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 148         }
 149
 150         while (non_vec) {
 151                 struct nix_cqe_hdr_s *cqe = (struct nix_cqe_hdr_s *)wqe[0];
 152                 uint64_t tstamp_ptr;
 153
 154                 mbuf = (struct rte_mbuf *)((char *)cqe -
 155                                            sizeof(struct rte_mbuf));
 156
 157                 /* Translate meta to mbuf */
 158                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 159                         const uint64_t cq_w1 = *((const uint64_t *)cqe + 1);
 160
 161                         mbuf = nix_sec_meta_to_mbuf_sc(cq_w1, sa_base, laddr,
 162                                                        &loff, mbuf, d_off);
 163                 }
 164
 165                 cn10k_nix_cqe_to_mbuf(cqe, cqe->tag, mbuf, lookup_mem,
 166                                       mbuf_init, flags);
 167                 /* Extracting tstamp, if PTP enabled*/
 168                 tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)cqe) +
 169                                            CNXK_SSO_WQE_SG_PTR);
 170                 cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, tstamp,
 171                                         flags & NIX_RX_OFFLOAD_TSTAMP_F,
 172                                         flags & NIX_RX_MULTI_SEG_F,
 173                                         (uint64_t *)tstamp_ptr);
 174                 wqe[0] = (uint64_t *)mbuf;
 175                 non_vec--;
 176                 wqe++;
 177         }
 178
 179         /* Free remaining meta buffers if any */
 180         if (flags & NIX_RX_OFFLOAD_SECURITY_F && loff) {
 181                 nix_sec_flush_meta(laddr, lmt_id, loff, aura_handle);
 182                 plt_io_wmb();
 183         }
 184 }
 185
 186 static __rte_always_inline uint16_t
 187 cn10k_sso_hws_get_work(struct cn10k_sso_hws *ws, struct rte_event *ev,
 188                        const uint32_t flags, void *lookup_mem)
 189 {
 190         union {
 191                 __uint128_t get_work;
 192                 uint64_t u64[2];
 193         } gw;
 194         uint64_t tstamp_ptr;
 195         uint64_t mbuf;
 196
 197         gw.get_work = ws->gw_wdata;
 198 #if defined(RTE_ARCH_ARM64) && !defined(__clang__)
 199         asm volatile(
 200                 PLT_CPU_FEATURE_PREAMBLE
 201                 "caspl %[wdata], %H[wdata], %[wdata], %H[wdata], [%[gw_loc]]\n"
 202                 "sub %[mbuf], %H[wdata], #0x80                          \n"
 203                 : [wdata] "+r"(gw.get_work), [mbuf] "=&r"(mbuf)
 204                 : [gw_loc] "r"(ws->base + SSOW_LF_GWS_OP_GET_WORK0)
 205                 : "memory");
 206 #else
 207         plt_write64(gw.u64[0], ws->base + SSOW_LF_GWS_OP_GET_WORK0);
 208         do {
 209                 roc_load_pair(gw.u64[0], gw.u64[1],
 210                               ws->base + SSOW_LF_GWS_WQE0);
 211         } while (gw.u64[0] & BIT_ULL(63));
 212         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 213 #endif
 214         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 215                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 216                     (gw.u64[0] & 0xffffffff);
 217
 218         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 219                 if ((flags & CPT_RX_WQE_F) &&
 220                     (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 221                      RTE_EVENT_TYPE_CRYPTODEV)) {
 222                         gw.u64[1] = cn10k_cpt_crypto_adapter_dequeue(gw.u64[1]);
 223                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 224                            RTE_EVENT_TYPE_ETHDEV) {
 225                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 226
 227                         if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 228                                 struct rte_mbuf *m;
 229                                 uintptr_t sa_base;
 230                                 uint64_t iova = 0;
 231                                 uint8_t loff = 0;
 232                                 uint16_t d_off;
 233                                 uint64_t cq_w1;
 234
 235                                 m = (struct rte_mbuf *)mbuf;
 236                                 d_off = (uintptr_t)(m->buf_addr) - (uintptr_t)m;
 237                                 d_off += RTE_PKTMBUF_HEADROOM;
 238
 239                                 cq_w1 = *(uint64_t *)(gw.u64[1] + 8);
 240
 241                                 sa_base =
 242                                         cnxk_nix_sa_base_get(port, lookup_mem);
 243                                 sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 244
 245                                 mbuf = (uint64_t)nix_sec_meta_to_mbuf_sc(
 246                                         cq_w1, sa_base, (uintptr_t)&iova, &loff,
 247                                         (struct rte_mbuf *)mbuf, d_off);
 248                                 if (loff)
 249                                         roc_npa_aura_op_free(m->pool->pool_id,
 250                                                              0, iova);
 251                         }
 252
 253                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 254                         cn10k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 255                                           gw.u64[0] & 0xFFFFF, flags,
 256                                           lookup_mem);
 257                         /* Extracting tstamp, if PTP enabled*/
 258                         tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)
 259                                                             gw.u64[1]) +
 260                                                    CNXK_SSO_WQE_SG_PTR);
 261                         cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf,
 262                                                 ws->tstamp,
 263                                                 flags & NIX_RX_OFFLOAD_TSTAMP_F,
 264                                                 flags & NIX_RX_MULTI_SEG_F,
 265                                                 (uint64_t *)tstamp_ptr);
 266                         gw.u64[1] = mbuf;
 267                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 268                            RTE_EVENT_TYPE_ETHDEV_VECTOR) {
 269                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 270                         __uint128_t vwqe_hdr = *(__uint128_t *)gw.u64[1];
 271
 272                         vwqe_hdr = ((vwqe_hdr >> 64) & 0xFFF) | BIT_ULL(31) |
 273                                    ((vwqe_hdr & 0xFFFF) << 48) |
 274                                    ((uint64_t)port << 32);
 275                         *(uint64_t *)gw.u64[1] = (uint64_t)vwqe_hdr;
 276                         cn10k_process_vwqe(gw.u64[1], port, flags, lookup_mem,
 277                                            ws->tstamp, ws->lmt_base);
 278                 }
 279         }
 280
 281         ev->event = gw.u64[0];
 282         ev->u64 = gw.u64[1];
 283
 284         return !!gw.u64[1];
 285 }
 286
 287 /* Used in cleaning up workslot. */
 288 static __rte_always_inline uint16_t
 289 cn10k_sso_hws_get_work_empty(struct cn10k_sso_hws *ws, struct rte_event *ev)
 290 {
 291         union {
 292                 __uint128_t get_work;
 293                 uint64_t u64[2];
 294         } gw;
 295         uint64_t mbuf;
 296
 297 #ifdef RTE_ARCH_ARM64
 298         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 299                      "          ldp %[tag], %[wqp], [%[tag_loc]]        \n"
 300                      "          tbz %[tag], 63, done%=                  \n"
 301                      "          sevl                                    \n"
 302                      "rty%=:    wfe                                     \n"
 303                      "          ldp %[tag], %[wqp], [%[tag_loc]]        \n"
 304                      "          tbnz %[tag], 63, rty%=                  \n"
 305                      "done%=:   dmb ld                                  \n"
 306                      "          sub %[mbuf], %[wqp], #0x80              \n"
 307                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 308                        [mbuf] "=&r"(mbuf)
 309                      : [tag_loc] "r"(ws->base + SSOW_LF_GWS_WQE0)
 310                      : "memory");
 311 #else
 312         do {
 313                 roc_load_pair(gw.u64[0], gw.u64[1],
 314                               ws->base + SSOW_LF_GWS_WQE0);
 315         } while (gw.u64[0] & BIT_ULL(63));
 316         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 317 #endif
 318
 319         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 320                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 321                     (gw.u64[0] & 0xffffffff);
 322
 323         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 324                 if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 325                     RTE_EVENT_TYPE_ETHDEV) {
 326                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 327
 328                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 329                         cn10k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 330                                           gw.u64[0] & 0xFFFFF, 0, NULL);
 331                         gw.u64[1] = mbuf;
 332                 }
 333         }
 334
 335         ev->event = gw.u64[0];
 336         ev->u64 = gw.u64[1];
 337
 338         return !!gw.u64[1];
 339 }
 340
 341 /* CN10K Fastpath functions. */
 342 uint16_t __rte_hot cn10k_sso_hws_enq(void *port, const struct rte_event *ev);
 343 uint16_t __rte_hot cn10k_sso_hws_enq_burst(void *port,
 344                                            const struct rte_event ev[],
 345                                            uint16_t nb_events);
 346 uint16_t __rte_hot cn10k_sso_hws_enq_new_burst(void *port,
 347                                                const struct rte_event ev[],
 348                                                uint16_t nb_events);
 349 uint16_t __rte_hot cn10k_sso_hws_enq_fwd_burst(void *port,
 350                                                const struct rte_event ev[],
 351                                                uint16_t nb_events);
 352 uint16_t __rte_hot cn10k_sso_hws_ca_enq(void *port, struct rte_event ev[],
 353                                         uint16_t nb_events);
 354
 355 #define R(name, flags)                                                         \
 356         uint16_t __rte_hot cn10k_sso_hws_deq_##name(                           \
 357                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 358         uint16_t __rte_hot cn10k_sso_hws_deq_burst_##name(                     \
 359                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 360                 uint64_t timeout_ticks);                                       \
 361         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_##name(                       \
 362                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 363         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_burst_##name(                 \
 364                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 365                 uint64_t timeout_ticks);                                       \
 366         uint16_t __rte_hot cn10k_sso_hws_deq_ca_##name(                        \
 367                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 368         uint16_t __rte_hot cn10k_sso_hws_deq_ca_burst_##name(                  \
 369                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 370                 uint64_t timeout_ticks);                                       \
 371         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_##name(                    \
 372                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 373         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_burst_##name(              \
 374                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 375                 uint64_t timeout_ticks);                                       \
 376         uint16_t __rte_hot cn10k_sso_hws_deq_seg_##name(                       \
 377                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 378         uint16_t __rte_hot cn10k_sso_hws_deq_seg_burst_##name(                 \
 379                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 380                 uint64_t timeout_ticks);                                       \
 381         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_seg_##name(                   \
 382                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 383         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_seg_burst_##name(             \
 384                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 385                 uint64_t timeout_ticks);                                       \
 386         uint16_t __rte_hot cn10k_sso_hws_deq_ca_seg_##name(                    \
 387                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 388         uint16_t __rte_hot cn10k_sso_hws_deq_ca_seg_burst_##name(              \
 389                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 390                 uint64_t timeout_ticks);                                       \
 391         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_seg_##name(                \
 392                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 393         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_seg_burst_##name(          \
 394                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 395                 uint64_t timeout_ticks);
 396
 397 NIX_RX_FASTPATH_MODES
 398 #undef R
 399
 400 #define SSO_DEQ(fn, flags)                                                     \
 401         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 402                               uint64_t timeout_ticks)                          \
 403         {                                                                      \
 404                 struct cn10k_sso_hws *ws = port;                               \
 405                 RTE_SET_USED(timeout_ticks);                                   \
 406                 if (ws->swtag_req) {                                           \
 407                         ws->swtag_req = 0;                                     \
 408                         cnxk_sso_hws_swtag_wait(ws->base + SSOW_LF_GWS_WQE0);  \
 409                         return 1;                                              \
 410                 }                                                              \
 411                 return cn10k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);  \
 412         }
 413
 414 #define SSO_DEQ_SEG(fn, flags)    SSO_DEQ(fn, flags | NIX_RX_MULTI_SEG_F)
 415 #define SSO_DEQ_CA(fn, flags)     SSO_DEQ(fn, flags | CPT_RX_WQE_F)
 416 #define SSO_DEQ_CA_SEG(fn, flags) SSO_DEQ_SEG(fn, flags | CPT_RX_WQE_F)
 417
 418 #define SSO_DEQ_TMO(fn, flags)                                                 \
 419         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 420                               uint64_t timeout_ticks)                          \
 421         {                                                                      \
 422                 struct cn10k_sso_hws *ws = port;                               \
 423                 uint16_t ret = 1;                                              \
 424                 uint64_t iter;                                                 \
 425                 if (ws->swtag_req) {                                           \
 426                         ws->swtag_req = 0;                                     \
 427                         cnxk_sso_hws_swtag_wait(ws->base + SSOW_LF_GWS_WQE0);  \
 428                         return ret;                                            \
 429                 }                                                              \
 430                 ret = cn10k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);   \
 431                 for (iter = 1; iter < timeout_ticks && (ret == 0); iter++)     \
 432                         ret = cn10k_sso_hws_get_work(ws, ev, flags,            \
 433                                                      ws->lookup_mem);          \
 434                 return ret;                                                    \
 435         }
 436
 437 #define SSO_DEQ_TMO_SEG(fn, flags)    SSO_DEQ_TMO(fn, flags | NIX_RX_MULTI_SEG_F)
 438 #define SSO_DEQ_TMO_CA(fn, flags)     SSO_DEQ_TMO(fn, flags | CPT_RX_WQE_F)
 439 #define SSO_DEQ_TMO_CA_SEG(fn, flags) SSO_DEQ_TMO_SEG(fn, flags | CPT_RX_WQE_F)
 440
 441 #define SSO_CMN_DEQ_BURST(fnb, fn, flags)                                      \
 442         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 443                                uint16_t nb_events, uint64_t timeout_ticks)     \
 444         {                                                                      \
 445                 RTE_SET_USED(nb_events);                                       \
 446                 return fn(port, ev, timeout_ticks);                            \
 447         }
 448
 449 #define SSO_CMN_DEQ_SEG_BURST(fnb, fn, flags)                                  \
 450         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 451                                uint16_t nb_events, uint64_t timeout_ticks)     \
 452         {                                                                      \
 453                 RTE_SET_USED(nb_events);                                       \
 454                 return fn(port, ev, timeout_ticks);                            \
 455         }
 456
 457 static __rte_always_inline struct cn10k_eth_txq *
 458 cn10k_sso_hws_xtract_meta(struct rte_mbuf *m,
 459                           const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT])
 460 {
 461         return (struct cn10k_eth_txq *)
 462                 txq_data[m->port][rte_event_eth_tx_adapter_txq_get(m)];
 463 }
 464
 465 static __rte_always_inline void
 466 cn10k_sso_tx_one(struct rte_mbuf *m, uint64_t *cmd, uint16_t lmt_id,
 467                  uintptr_t lmt_addr, uint8_t sched_type, uintptr_t base,
 468                  const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT],
 469                  const uint32_t flags)
 470 {
 471         uint8_t lnum = 0, loff = 0, shft = 0;
 472         struct cn10k_eth_txq *txq;
 473         uintptr_t laddr;
 474         uint16_t segdw;
 475         uintptr_t pa;
 476         bool sec;
 477
 478         txq = cn10k_sso_hws_xtract_meta(m, txq_data);
 479         cn10k_nix_tx_skeleton(txq, cmd, flags);
 480         /* Perform header writes before barrier
 481          * for TSO
 482          */
 483         if (flags & NIX_TX_OFFLOAD_TSO_F)
 484                 cn10k_nix_xmit_prepare_tso(m, flags);
 485
 486         cn10k_nix_xmit_prepare(m, cmd, flags, txq->lso_tun_fmt, &sec);
 487
 488         laddr = lmt_addr;
 489         /* Prepare CPT instruction and get nixtx addr if
 490          * it is for CPT on same lmtline.
 491          */
 492         if (flags & NIX_TX_OFFLOAD_SECURITY_F && sec)
 493                 cn10k_nix_prep_sec(m, cmd, &laddr, lmt_addr, &lnum, &loff,
 494                                    &shft, txq->sa_base, flags);
 495
 496         /* Move NIX desc to LMT/NIXTX area */
 497         cn10k_nix_xmit_mv_lmt_base(laddr, cmd, flags);
 498
 499         if (flags & NIX_TX_MULTI_SEG_F)
 500                 segdw = cn10k_nix_prepare_mseg(m, (uint64_t *)laddr, flags);
 501         else
 502                 segdw = cn10k_nix_tx_ext_subs(flags) + 2;
 503
 504         if (flags & NIX_TX_OFFLOAD_SECURITY_F && sec)
 505                 pa = txq->cpt_io_addr | 3 << 4;
 506         else
 507                 pa = txq->io_addr | ((segdw - 1) << 4);
 508
 509         if (!sched_type)
 510                 roc_sso_hws_head_wait(base + SSOW_LF_GWS_TAG);
 511
 512         roc_lmt_submit_steorl(lmt_id, pa);
 513 }
 514
 515 static __rte_always_inline void
 516 cn10k_sso_vwqe_split_tx(struct rte_mbuf **mbufs, uint16_t nb_mbufs,
 517                         uint64_t *cmd, uint16_t lmt_id, uintptr_t lmt_addr,
 518                         uint8_t sched_type, uintptr_t base,
 519                         const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT],
 520                         const uint32_t flags)
 521 {
 522         uint16_t port[4], queue[4];
 523         uint16_t i, j, pkts, scalar;
 524         struct cn10k_eth_txq *txq;
 525
 526         scalar = nb_mbufs & (NIX_DESCS_PER_LOOP - 1);
 527         pkts = RTE_ALIGN_FLOOR(nb_mbufs, NIX_DESCS_PER_LOOP);
 528
 529         for (i = 0; i < pkts; i += NIX_DESCS_PER_LOOP) {
 530                 port[0] = mbufs[i]->port;
 531                 port[1] = mbufs[i + 1]->port;
 532                 port[2] = mbufs[i + 2]->port;
 533                 port[3] = mbufs[i + 3]->port;
 534
 535                 queue[0] = rte_event_eth_tx_adapter_txq_get(mbufs[i]);
 536                 queue[1] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 1]);
 537                 queue[2] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 2]);
 538                 queue[3] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 3]);
 539
 540                 if (((port[0] ^ port[1]) & (port[2] ^ port[3])) ||
 541                     ((queue[0] ^ queue[1]) & (queue[2] ^ queue[3]))) {
 542                         for (j = 0; j < 4; j++)
 543                                 cn10k_sso_tx_one(mbufs[i + j], cmd, lmt_id,
 544                                                  lmt_addr, sched_type, base,
 545                                                  txq_data, flags);
 546                 } else {
 547                         txq = (struct cn10k_eth_txq *)
 548                                 txq_data[port[0]][queue[0]];
 549                         cn10k_nix_xmit_pkts_vector(txq, &mbufs[i], 4, cmd,
 550                                                    base + SSOW_LF_GWS_TAG,
 551                                                    flags | NIX_TX_VWQE_F);
 552                 }
 553         }
 554
 555         mbufs += i;
 556
 557         for (i = 0; i < scalar; i++) {
 558                 cn10k_sso_tx_one(mbufs[i], cmd, lmt_id, lmt_addr, sched_type,
 559                                  base, txq_data, flags);
 560         }
 561 }
 562
 563 static __rte_always_inline uint16_t
 564 cn10k_sso_hws_event_tx(struct cn10k_sso_hws *ws, struct rte_event *ev,
 565                        uint64_t *cmd,
 566                        const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT],
 567                        const uint32_t flags)
 568 {
 569         struct cn10k_eth_txq *txq;
 570         struct rte_mbuf *m;
 571         uintptr_t lmt_addr;
 572         uint16_t ref_cnt;
 573         uint16_t lmt_id;
 574
 575         lmt_addr = ws->lmt_base;
 576         ROC_LMT_BASE_ID_GET(lmt_addr, lmt_id);
 577
 578         if (ev->event_type & RTE_EVENT_TYPE_VECTOR) {
 579                 struct rte_mbuf **mbufs = ev->vec->mbufs;
 580                 uint64_t meta = *(uint64_t *)ev->vec;
 581
 582                 if (meta & BIT(31)) {
 583                         txq = (struct cn10k_eth_txq *)
 584                                 txq_data[meta >> 32][meta >> 48];
 585
 586                         cn10k_nix_xmit_pkts_vector(
 587                                 txq, mbufs, meta & 0xFFFF, cmd,
 588                                 ws->tx_base + SSOW_LF_GWS_TAG,
 589                                 flags | NIX_TX_VWQE_F);
 590                 } else {
 591                         cn10k_sso_vwqe_split_tx(
 592                                 mbufs, meta & 0xFFFF, cmd, lmt_id, lmt_addr,
 593                                 ev->sched_type, ws->tx_base, txq_data, flags);
 594                 }
 595                 rte_mempool_put(rte_mempool_from_obj(ev->vec), ev->vec);
 596                 return (meta & 0xFFFF);
 597         }
 598
 599         m = ev->mbuf;
 600         ref_cnt = m->refcnt;
 601         cn10k_sso_tx_one(m, cmd, lmt_id, lmt_addr, ev->sched_type, ws->tx_base,
 602                          txq_data, flags);
 603
 604         if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
 605                 if (ref_cnt > 1)
 606                         return 1;
 607         }
 608
 609         cnxk_sso_hws_swtag_flush(ws->tx_base + SSOW_LF_GWS_TAG,
 610                                  ws->tx_base + SSOW_LF_GWS_OP_SWTAG_FLUSH);
 611         return 1;
 612 }
 613
 614 #define T(name, sz, flags)                                                     \
 615         uint16_t __rte_hot cn10k_sso_hws_tx_adptr_enq_##name(                  \
 616                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 617         uint16_t __rte_hot cn10k_sso_hws_tx_adptr_enq_seg_##name(              \
 618                 void *port, struct rte_event ev[], uint16_t nb_events);
 619
 620 NIX_TX_FASTPATH_MODES
 621 #undef T
 622
 623 #define SSO_TX(fn, sz, flags)                                                  \
 624         uint16_t __rte_hot fn(void *port, struct rte_event ev[],               \
 625                               uint16_t nb_events)                              \
 626         {                                                                      \
 627                 struct cn10k_sso_hws *ws = port;                               \
 628                 uint64_t cmd[sz];                                              \
 629                 RTE_SET_USED(nb_events);                                       \
 630                 return cn10k_sso_hws_event_tx(                                 \
 631                         ws, &ev[0], cmd,                                       \
 632                         (const uint64_t(*)[RTE_MAX_QUEUES_PER_PORT]) &         \
 633                                 ws->tx_adptr_data,                             \
 634                         flags);                                                \
 635         }
 636
 637 #define SSO_TX_SEG(fn, sz, flags)                                              \
 638         uint16_t __rte_hot fn(void *port, struct rte_event ev[],               \
 639                               uint16_t nb_events)                              \
 640         {                                                                      \
 641                 uint64_t cmd[(sz) + CNXK_NIX_TX_MSEG_SG_DWORDS - 2];           \
 642                 struct cn10k_sso_hws *ws = port;                               \
 643                 RTE_SET_USED(nb_events);                                       \
 644                 return cn10k_sso_hws_event_tx(                                 \
 645                         ws, &ev[0], cmd,                                       \
 646                         (const uint64_t(*)[RTE_MAX_QUEUES_PER_PORT]) &         \
 647                                 ws->tx_adptr_data,                             \
 648                         (flags) | NIX_TX_MULTI_SEG_F);                         \
 649         }
 650
 651 #endif