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 = CNXK_TT_FROM_TAG(ws->gw_rdata);
  44
  45         /* CNXK model
  46          * cur_tt/new_tt     SSO_TT_ORDERED SSO_TT_ATOMIC SSO_TT_UNTAGGED
  47          *
  48          * SSO_TT_ORDERED        norm           norm             untag
  49          * SSO_TT_ATOMIC         norm           norm               untag
  50          * SSO_TT_UNTAGGED       norm           norm             NOOP
  51          */
  52
  53         if (new_tt == SSO_TT_UNTAGGED) {
  54                 if (cur_tt != SSO_TT_UNTAGGED)
  55                         cnxk_sso_hws_swtag_untag(ws->base +
  56                                                  SSOW_LF_GWS_OP_SWTAG_UNTAG);
  57         } else {
  58                 cnxk_sso_hws_swtag_norm(tag, new_tt,
  59                                         ws->base + SSOW_LF_GWS_OP_SWTAG_NORM);
  60         }
  61         ws->swtag_req = 1;
  62 }
  63
  64 static __rte_always_inline void
  65 cn10k_sso_hws_fwd_group(struct cn10k_sso_hws *ws, const struct rte_event *ev,
  66                         const uint16_t grp)
  67 {
  68         const uint32_t tag = (uint32_t)ev->event;
  69         const uint8_t new_tt = ev->sched_type;
  70
  71         plt_write64(ev->u64, ws->base + SSOW_LF_GWS_OP_UPD_WQP_GRP1);
  72         cnxk_sso_hws_swtag_desched(tag, new_tt, grp,
  73                                    ws->base + SSOW_LF_GWS_OP_SWTAG_DESCHED);
  74 }
  75
  76 static __rte_always_inline void
  77 cn10k_sso_hws_forward_event(struct cn10k_sso_hws *ws,
  78                             const struct rte_event *ev)
  79 {
  80         const uint8_t grp = ev->queue_id;
  81
  82         /* Group hasn't changed, Use SWTAG to forward the event */
  83         if (CNXK_GRP_FROM_TAG(ws->gw_rdata) == grp)
  84                 cn10k_sso_hws_fwd_swtag(ws, ev);
  85         else
  86                 /*
  87                  * Group has been changed for group based work pipelining,
  88                  * Use deschedule/add_work operation to transfer the event to
  89                  * new group/core
  90                  */
  91                 cn10k_sso_hws_fwd_group(ws, ev, grp);
  92 }
  93
  94 static __rte_always_inline void
  95 cn10k_wqe_to_mbuf(uint64_t wqe, const uint64_t mbuf, uint8_t port_id,
  96                   const uint32_t tag, const uint32_t flags,
  97                   const void *const lookup_mem)
  98 {
  99         const uint64_t mbuf_init = 0x100010000ULL | RTE_PKTMBUF_HEADROOM |
 100                                    (flags & NIX_RX_OFFLOAD_TSTAMP_F ? 8 : 0);
 101
 102         cn10k_nix_cqe_to_mbuf((struct nix_cqe_hdr_s *)wqe, tag,
 103                               (struct rte_mbuf *)mbuf, lookup_mem,
 104                               mbuf_init | ((uint64_t)port_id) << 48, flags);
 105 }
 106
 107 static __rte_always_inline void
 108 cn10k_process_vwqe(uintptr_t vwqe, uint16_t port_id, const uint32_t flags,
 109                    void *lookup_mem, void *tstamp, uintptr_t lbase)
 110 {
 111         uint64_t mbuf_init = 0x100010000ULL | RTE_PKTMBUF_HEADROOM |
 112                              (flags & NIX_RX_OFFLOAD_TSTAMP_F ? 8 : 0);
 113         struct rte_event_vector *vec;
 114         uint64_t aura_handle, laddr;
 115         uint16_t nb_mbufs, non_vec;
 116         uint16_t lmt_id, d_off;
 117         struct rte_mbuf *mbuf;
 118         uint8_t loff = 0;
 119         uint64_t sa_base;
 120         uint64_t **wqe;
 121         int i;
 122
 123         mbuf_init |= ((uint64_t)port_id) << 48;
 124         vec = (struct rte_event_vector *)vwqe;
 125         wqe = vec->u64s;
 126
 127         rte_prefetch_non_temporal(&vec->ptrs[0]);
 128 #define OBJS_PER_CLINE (RTE_CACHE_LINE_SIZE / sizeof(void *))
 129         for (i = OBJS_PER_CLINE; i < vec->nb_elem; i += OBJS_PER_CLINE)
 130                 rte_prefetch_non_temporal(&vec->ptrs[i]);
 131
 132         nb_mbufs = RTE_ALIGN_FLOOR(vec->nb_elem, NIX_DESCS_PER_LOOP);
 133         nb_mbufs = cn10k_nix_recv_pkts_vector(&mbuf_init, vec->mbufs, nb_mbufs,
 134                                               flags | NIX_RX_VWQE_F, lookup_mem,
 135                                               tstamp, lbase);
 136         wqe += nb_mbufs;
 137         non_vec = vec->nb_elem - nb_mbufs;
 138
 139         if (flags & NIX_RX_OFFLOAD_SECURITY_F && non_vec) {
 140                 mbuf = (struct rte_mbuf *)((uintptr_t)wqe[0] -
 141                                            sizeof(struct rte_mbuf));
 142                 /* Pick first mbuf's aura handle assuming all
 143                  * mbufs are from a vec and are from same RQ.
 144                  */
 145                 aura_handle = mbuf->pool->pool_id;
 146                 ROC_LMT_BASE_ID_GET(lbase, lmt_id);
 147                 laddr = lbase;
 148                 laddr += 8;
 149                 d_off = ((uintptr_t)mbuf->buf_addr - (uintptr_t)mbuf);
 150                 d_off += (mbuf_init & 0xFFFF);
 151                 sa_base = cnxk_nix_sa_base_get(mbuf_init >> 48, lookup_mem);
 152                 sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 153         }
 154
 155         while (non_vec) {
 156                 struct nix_cqe_hdr_s *cqe = (struct nix_cqe_hdr_s *)wqe[0];
 157                 uint64_t tstamp_ptr;
 158
 159                 mbuf = (struct rte_mbuf *)((char *)cqe -
 160                                            sizeof(struct rte_mbuf));
 161
 162                 /* Translate meta to mbuf */
 163                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 164                         const uint64_t cq_w1 = *((const uint64_t *)cqe + 1);
 165
 166                         mbuf = nix_sec_meta_to_mbuf_sc(cq_w1, sa_base, laddr,
 167                                                        &loff, mbuf, d_off);
 168                 }
 169
 170                 cn10k_nix_cqe_to_mbuf(cqe, cqe->tag, mbuf, lookup_mem,
 171                                       mbuf_init, flags);
 172                 /* Extracting tstamp, if PTP enabled*/
 173                 tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)cqe) +
 174                                            CNXK_SSO_WQE_SG_PTR);
 175                 cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, tstamp,
 176                                         flags & NIX_RX_OFFLOAD_TSTAMP_F,
 177                                         (uint64_t *)tstamp_ptr);
 178                 wqe[0] = (uint64_t *)mbuf;
 179                 non_vec--;
 180                 wqe++;
 181         }
 182
 183         /* Free remaining meta buffers if any */
 184         if (flags & NIX_RX_OFFLOAD_SECURITY_F && loff) {
 185                 nix_sec_flush_meta(laddr, lmt_id, loff, aura_handle);
 186                 plt_io_wmb();
 187         }
 188 }
 189
 190 static __rte_always_inline uint16_t
 191 cn10k_sso_hws_get_work(struct cn10k_sso_hws *ws, struct rte_event *ev,
 192                        const uint32_t flags, void *lookup_mem)
 193 {
 194         union {
 195                 __uint128_t get_work;
 196                 uint64_t u64[2];
 197         } gw;
 198         uint64_t tstamp_ptr;
 199
 200         gw.get_work = ws->gw_wdata;
 201 #if defined(RTE_ARCH_ARM64) && !defined(__clang__)
 202         asm volatile(
 203                 PLT_CPU_FEATURE_PREAMBLE
 204                 "caspal %[wdata], %H[wdata], %[wdata], %H[wdata], [%[gw_loc]]\n"
 205                 : [wdata] "+r"(gw.get_work)
 206                 : [gw_loc] "r"(ws->base + SSOW_LF_GWS_OP_GET_WORK0)
 207                 : "memory");
 208 #else
 209         plt_write64(gw.u64[0], ws->base + SSOW_LF_GWS_OP_GET_WORK0);
 210         do {
 211                 roc_load_pair(gw.u64[0], gw.u64[1],
 212                               ws->base + SSOW_LF_GWS_WQE0);
 213         } while (gw.u64[0] & BIT_ULL(63));
 214 #endif
 215         ws->gw_rdata = gw.u64[0];
 216         if (gw.u64[1]) {
 217                 gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 218                             (gw.u64[0] & (0x3FFull << 36)) << 4 |
 219                             (gw.u64[0] & 0xffffffff);
 220                 if ((flags & CPT_RX_WQE_F) &&
 221                     (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 222                      RTE_EVENT_TYPE_CRYPTODEV)) {
 223                         gw.u64[1] = cn10k_cpt_crypto_adapter_dequeue(gw.u64[1]);
 224                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 225                            RTE_EVENT_TYPE_ETHDEV) {
 226                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 227                         uint64_t mbuf;
 228
 229                         mbuf = gw.u64[1] - sizeof(struct rte_mbuf);
 230                         rte_prefetch0((void *)mbuf);
 231                         if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 232                                 struct rte_mbuf *m;
 233                                 uintptr_t sa_base;
 234                                 uint64_t iova = 0;
 235                                 uint8_t loff = 0;
 236                                 uint16_t d_off;
 237                                 uint64_t cq_w1;
 238
 239                                 m = (struct rte_mbuf *)mbuf;
 240                                 d_off = (uintptr_t)(m->buf_addr) - (uintptr_t)m;
 241                                 d_off += RTE_PKTMBUF_HEADROOM;
 242
 243                                 cq_w1 = *(uint64_t *)(gw.u64[1] + 8);
 244
 245                                 sa_base =
 246                                         cnxk_nix_sa_base_get(port, lookup_mem);
 247                                 sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 248
 249                                 mbuf = (uint64_t)nix_sec_meta_to_mbuf_sc(
 250                                         cq_w1, sa_base, (uintptr_t)&iova, &loff,
 251                                         (struct rte_mbuf *)mbuf, d_off);
 252                                 if (loff)
 253                                         roc_npa_aura_op_free(m->pool->pool_id,
 254                                                              0, iova);
 255                         }
 256
 257                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 258                         cn10k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 259                                           gw.u64[0] & 0xFFFFF, flags,
 260                                           lookup_mem);
 261                         /* Extracting tstamp, if PTP enabled*/
 262                         tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)
 263                                                             gw.u64[1]) +
 264                                                    CNXK_SSO_WQE_SG_PTR);
 265                         cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf,
 266                                                 ws->tstamp,
 267                                                 flags & NIX_RX_OFFLOAD_TSTAMP_F,
 268                                                 (uint64_t *)tstamp_ptr);
 269                         gw.u64[1] = mbuf;
 270                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 271                            RTE_EVENT_TYPE_ETHDEV_VECTOR) {
 272                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 273                         __uint128_t vwqe_hdr = *(__uint128_t *)gw.u64[1];
 274
 275                         vwqe_hdr = ((vwqe_hdr >> 64) & 0xFFF) | BIT_ULL(31) |
 276                                    ((vwqe_hdr & 0xFFFF) << 48) |
 277                                    ((uint64_t)port << 32);
 278                         *(uint64_t *)gw.u64[1] = (uint64_t)vwqe_hdr;
 279                         cn10k_process_vwqe(gw.u64[1], port, flags, lookup_mem,
 280                                            ws->tstamp, ws->lmt_base);
 281                 }
 282         }
 283
 284         ev->event = gw.u64[0];
 285         ev->u64 = gw.u64[1];
 286
 287         return !!gw.u64[1];
 288 }
 289
 290 /* Used in cleaning up workslot. */
 291 static __rte_always_inline uint16_t
 292 cn10k_sso_hws_get_work_empty(struct cn10k_sso_hws *ws, struct rte_event *ev)
 293 {
 294         union {
 295                 __uint128_t get_work;
 296                 uint64_t u64[2];
 297         } gw;
 298         uint64_t mbuf;
 299
 300 #ifdef RTE_ARCH_ARM64
 301         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 302                      "          ldp %[tag], %[wqp], [%[tag_loc]]        \n"
 303                      "          tbz %[tag], 63, done%=                  \n"
 304                      "          sevl                                    \n"
 305                      "rty%=:    wfe                                     \n"
 306                      "          ldp %[tag], %[wqp], [%[tag_loc]]        \n"
 307                      "          tbnz %[tag], 63, rty%=                  \n"
 308                      "done%=:   dmb ld                                  \n"
 309                      "          sub %[mbuf], %[wqp], #0x80              \n"
 310                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 311                        [mbuf] "=&r"(mbuf)
 312                      : [tag_loc] "r"(ws->base + SSOW_LF_GWS_WQE0)
 313                      : "memory");
 314 #else
 315         do {
 316                 roc_load_pair(gw.u64[0], gw.u64[1],
 317                               ws->base + SSOW_LF_GWS_WQE0);
 318         } while (gw.u64[0] & BIT_ULL(63));
 319         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 320 #endif
 321
 322         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 323                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 324                     (gw.u64[0] & 0xffffffff);
 325
 326         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 327                 if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 328                     RTE_EVENT_TYPE_ETHDEV) {
 329                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 330
 331                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 332                         cn10k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 333                                           gw.u64[0] & 0xFFFFF, 0, NULL);
 334                         gw.u64[1] = mbuf;
 335                 }
 336         }
 337
 338         ev->event = gw.u64[0];
 339         ev->u64 = gw.u64[1];
 340
 341         return !!gw.u64[1];
 342 }
 343
 344 /* CN10K Fastpath functions. */
 345 uint16_t __rte_hot cn10k_sso_hws_enq(void *port, const struct rte_event *ev);
 346 uint16_t __rte_hot cn10k_sso_hws_enq_burst(void *port,
 347                                            const struct rte_event ev[],
 348                                            uint16_t nb_events);
 349 uint16_t __rte_hot cn10k_sso_hws_enq_new_burst(void *port,
 350                                                const struct rte_event ev[],
 351                                                uint16_t nb_events);
 352 uint16_t __rte_hot cn10k_sso_hws_enq_fwd_burst(void *port,
 353                                                const struct rte_event ev[],
 354                                                uint16_t nb_events);
 355 uint16_t __rte_hot cn10k_sso_hws_ca_enq(void *port, struct rte_event ev[],
 356                                         uint16_t nb_events);
 357
 358 #define R(name, flags)                                                         \
 359         uint16_t __rte_hot cn10k_sso_hws_deq_##name(                           \
 360                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 361         uint16_t __rte_hot cn10k_sso_hws_deq_burst_##name(                     \
 362                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 363                 uint64_t timeout_ticks);                                       \
 364         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_##name(                       \
 365                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 366         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_burst_##name(                 \
 367                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 368                 uint64_t timeout_ticks);                                       \
 369         uint16_t __rte_hot cn10k_sso_hws_deq_ca_##name(                        \
 370                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 371         uint16_t __rte_hot cn10k_sso_hws_deq_ca_burst_##name(                  \
 372                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 373                 uint64_t timeout_ticks);                                       \
 374         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_##name(                    \
 375                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 376         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_burst_##name(              \
 377                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 378                 uint64_t timeout_ticks);                                       \
 379         uint16_t __rte_hot cn10k_sso_hws_deq_seg_##name(                       \
 380                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 381         uint16_t __rte_hot cn10k_sso_hws_deq_seg_burst_##name(                 \
 382                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 383                 uint64_t timeout_ticks);                                       \
 384         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_seg_##name(                   \
 385                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 386         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_seg_burst_##name(             \
 387                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 388                 uint64_t timeout_ticks);                                       \
 389         uint16_t __rte_hot cn10k_sso_hws_deq_ca_seg_##name(                    \
 390                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 391         uint16_t __rte_hot cn10k_sso_hws_deq_ca_seg_burst_##name(              \
 392                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 393                 uint64_t timeout_ticks);                                       \
 394         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_seg_##name(                \
 395                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 396         uint16_t __rte_hot cn10k_sso_hws_deq_tmo_ca_seg_burst_##name(          \
 397                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 398                 uint64_t timeout_ticks);
 399
 400 NIX_RX_FASTPATH_MODES
 401 #undef R
 402
 403 #define SSO_DEQ(fn, flags)                                                     \
 404         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 405                               uint64_t timeout_ticks)                          \
 406         {                                                                      \
 407                 struct cn10k_sso_hws *ws = port;                               \
 408                 RTE_SET_USED(timeout_ticks);                                   \
 409                 if (ws->swtag_req) {                                           \
 410                         ws->swtag_req = 0;                                     \
 411                         ws->gw_rdata = cnxk_sso_hws_swtag_wait(                \
 412                                 ws->base + SSOW_LF_GWS_WQE0);                  \
 413                         return 1;                                              \
 414                 }                                                              \
 415                 return cn10k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);  \
 416         }
 417
 418 #define SSO_DEQ_SEG(fn, flags)    SSO_DEQ(fn, flags | NIX_RX_MULTI_SEG_F)
 419 #define SSO_DEQ_CA(fn, flags)     SSO_DEQ(fn, flags | CPT_RX_WQE_F)
 420 #define SSO_DEQ_CA_SEG(fn, flags) SSO_DEQ_SEG(fn, flags | CPT_RX_WQE_F)
 421
 422 #define SSO_DEQ_TMO(fn, flags)                                                 \
 423         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 424                               uint64_t timeout_ticks)                          \
 425         {                                                                      \
 426                 struct cn10k_sso_hws *ws = port;                               \
 427                 uint16_t ret = 1;                                              \
 428                 uint64_t iter;                                                 \
 429                 if (ws->swtag_req) {                                           \
 430                         ws->swtag_req = 0;                                     \
 431                         ws->gw_rdata = cnxk_sso_hws_swtag_wait(                \
 432                                 ws->base + SSOW_LF_GWS_WQE0);                  \
 433                         return ret;                                            \
 434                 }                                                              \
 435                 ret = cn10k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);   \
 436                 for (iter = 1; iter < timeout_ticks && (ret == 0); iter++)     \
 437                         ret = cn10k_sso_hws_get_work(ws, ev, flags,            \
 438                                                      ws->lookup_mem);          \
 439                 return ret;                                                    \
 440         }
 441
 442 #define SSO_DEQ_TMO_SEG(fn, flags)    SSO_DEQ_TMO(fn, flags | NIX_RX_MULTI_SEG_F)
 443 #define SSO_DEQ_TMO_CA(fn, flags)     SSO_DEQ_TMO(fn, flags | CPT_RX_WQE_F)
 444 #define SSO_DEQ_TMO_CA_SEG(fn, flags) SSO_DEQ_TMO_SEG(fn, flags | CPT_RX_WQE_F)
 445
 446 #define SSO_CMN_DEQ_BURST(fnb, fn, flags)                                      \
 447         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 448                                uint16_t nb_events, uint64_t timeout_ticks)     \
 449         {                                                                      \
 450                 RTE_SET_USED(nb_events);                                       \
 451                 return fn(port, ev, timeout_ticks);                            \
 452         }
 453
 454 #define SSO_CMN_DEQ_SEG_BURST(fnb, fn, flags)                                  \
 455         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 456                                uint16_t nb_events, uint64_t timeout_ticks)     \
 457         {                                                                      \
 458                 RTE_SET_USED(nb_events);                                       \
 459                 return fn(port, ev, timeout_ticks);                            \
 460         }
 461
 462 static __rte_always_inline struct cn10k_eth_txq *
 463 cn10k_sso_hws_xtract_meta(struct rte_mbuf *m, const uint64_t *txq_data)
 464 {
 465         return (struct cn10k_eth_txq
 466                         *)(txq_data[(txq_data[m->port] >> 48) +
 467                                     rte_event_eth_tx_adapter_txq_get(m)] &
 468                            (BIT_ULL(48) - 1));
 469 }
 470
 471 static __rte_always_inline void
 472 cn10k_sso_txq_fc_wait(const struct cn10k_eth_txq *txq)
 473 {
 474         while ((uint64_t)txq->nb_sqb_bufs_adj <=
 475                __atomic_load_n(txq->fc_mem, __ATOMIC_RELAXED))
 476                 ;
 477 }
 478
 479 static __rte_always_inline void
 480 cn10k_sso_tx_one(struct cn10k_sso_hws *ws, struct rte_mbuf *m, uint64_t *cmd,
 481                  uint16_t lmt_id, uintptr_t lmt_addr, uint8_t sched_type,
 482                  const uint64_t *txq_data, const uint32_t flags)
 483 {
 484         uint8_t lnum = 0, loff = 0, shft = 0;
 485         struct cn10k_eth_txq *txq;
 486         uintptr_t laddr;
 487         uint16_t segdw;
 488         uintptr_t pa;
 489         bool sec;
 490
 491         txq = cn10k_sso_hws_xtract_meta(m, txq_data);
 492         cn10k_nix_tx_skeleton(txq, cmd, flags, 0);
 493         /* Perform header writes before barrier
 494          * for TSO
 495          */
 496         if (flags & NIX_TX_OFFLOAD_TSO_F)
 497                 cn10k_nix_xmit_prepare_tso(m, flags);
 498
 499         cn10k_nix_xmit_prepare(m, cmd, flags, txq->lso_tun_fmt, &sec);
 500
 501         laddr = lmt_addr;
 502         /* Prepare CPT instruction and get nixtx addr if
 503          * it is for CPT on same lmtline.
 504          */
 505         if (flags & NIX_TX_OFFLOAD_SECURITY_F && sec)
 506                 cn10k_nix_prep_sec(m, cmd, &laddr, lmt_addr, &lnum, &loff,
 507                                    &shft, txq->sa_base, flags);
 508
 509         /* Move NIX desc to LMT/NIXTX area */
 510         cn10k_nix_xmit_mv_lmt_base(laddr, cmd, flags);
 511
 512         if (flags & NIX_TX_MULTI_SEG_F)
 513                 segdw = cn10k_nix_prepare_mseg(m, (uint64_t *)laddr, flags);
 514         else
 515                 segdw = cn10k_nix_tx_ext_subs(flags) + 2;
 516
 517         cn10k_nix_xmit_prepare_tstamp(txq, laddr, m->ol_flags, segdw, flags);
 518         if (flags & NIX_TX_OFFLOAD_SECURITY_F && sec)
 519                 pa = txq->cpt_io_addr | 3 << 4;
 520         else
 521                 pa = txq->io_addr | ((segdw - 1) << 4);
 522
 523         if (!CNXK_TAG_IS_HEAD(ws->gw_rdata) && !sched_type)
 524                 ws->gw_rdata = roc_sso_hws_head_wait(ws->base);
 525
 526         cn10k_sso_txq_fc_wait(txq);
 527         roc_lmt_submit_steorl(lmt_id, pa);
 528 }
 529
 530 static __rte_always_inline void
 531 cn10k_sso_vwqe_split_tx(struct cn10k_sso_hws *ws, struct rte_mbuf **mbufs,
 532                         uint16_t nb_mbufs, uint64_t *cmd, uint16_t lmt_id,
 533                         uintptr_t lmt_addr, uint8_t sched_type,
 534                         const uint64_t *txq_data, const uint32_t flags)
 535 {
 536         uint16_t port[4], queue[4];
 537         uint16_t i, j, pkts, scalar;
 538         struct cn10k_eth_txq *txq;
 539
 540         scalar = nb_mbufs & (NIX_DESCS_PER_LOOP - 1);
 541         pkts = RTE_ALIGN_FLOOR(nb_mbufs, NIX_DESCS_PER_LOOP);
 542
 543         for (i = 0; i < pkts; i += NIX_DESCS_PER_LOOP) {
 544                 port[0] = mbufs[i]->port;
 545                 port[1] = mbufs[i + 1]->port;
 546                 port[2] = mbufs[i + 2]->port;
 547                 port[3] = mbufs[i + 3]->port;
 548
 549                 queue[0] = rte_event_eth_tx_adapter_txq_get(mbufs[i]);
 550                 queue[1] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 1]);
 551                 queue[2] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 2]);
 552                 queue[3] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 3]);
 553
 554                 if (((port[0] ^ port[1]) & (port[2] ^ port[3])) ||
 555                     ((queue[0] ^ queue[1]) & (queue[2] ^ queue[3]))) {
 556                         for (j = 0; j < 4; j++)
 557                                 cn10k_sso_tx_one(ws, mbufs[i + j], cmd, lmt_id,
 558                                                  lmt_addr, sched_type, txq_data,
 559                                                  flags);
 560                 } else {
 561                         txq = (struct cn10k_eth_txq
 562                                        *)(txq_data[(txq_data[port[0]] >> 48) +
 563                                                    queue[0]] &
 564                                           (BIT_ULL(48) - 1));
 565                         cn10k_nix_xmit_pkts_vector(txq, (uint64_t *)ws,
 566                                                    &mbufs[i], 4, cmd,
 567                                                    flags | NIX_TX_VWQE_F);
 568                 }
 569         }
 570
 571         mbufs += i;
 572
 573         for (i = 0; i < scalar; i++) {
 574                 cn10k_sso_tx_one(ws, mbufs[i], cmd, lmt_id, lmt_addr,
 575                                  sched_type, txq_data, flags);
 576         }
 577 }
 578
 579 static __rte_always_inline uint16_t
 580 cn10k_sso_hws_event_tx(struct cn10k_sso_hws *ws, struct rte_event *ev,
 581                        uint64_t *cmd, const uint64_t *txq_data,
 582                        const uint32_t flags)
 583 {
 584         struct cn10k_eth_txq *txq;
 585         struct rte_mbuf *m;
 586         uintptr_t lmt_addr;
 587         uint16_t lmt_id;
 588
 589         lmt_addr = ws->lmt_base;
 590         ROC_LMT_BASE_ID_GET(lmt_addr, lmt_id);
 591
 592         if (ev->event_type & RTE_EVENT_TYPE_VECTOR) {
 593                 struct rte_mbuf **mbufs = ev->vec->mbufs;
 594                 uint64_t meta = *(uint64_t *)ev->vec;
 595
 596                 if (meta & BIT(31)) {
 597                         txq = (struct cn10k_eth_txq
 598                                        *)(txq_data[(txq_data[meta >> 32] >>
 599                                                     48) +
 600                                                    (meta >> 48)] &
 601                                           (BIT_ULL(48) - 1));
 602
 603                         cn10k_nix_xmit_pkts_vector(txq, (uint64_t *)ws, mbufs,
 604                                                    meta & 0xFFFF, cmd,
 605                                                    flags | NIX_TX_VWQE_F);
 606                 } else {
 607                         cn10k_sso_vwqe_split_tx(
 608                                 ws, mbufs, meta & 0xFFFF, cmd, lmt_id, lmt_addr,
 609                                 ev->sched_type, txq_data, flags);
 610                 }
 611                 rte_mempool_put(rte_mempool_from_obj(ev->vec), ev->vec);
 612                 return (meta & 0xFFFF);
 613         }
 614
 615         m = ev->mbuf;
 616         cn10k_sso_tx_one(ws, m, cmd, lmt_id, lmt_addr, ev->sched_type, txq_data,
 617                          flags);
 618
 619         return 1;
 620 }
 621
 622 #define T(name, sz, flags)                                                     \
 623         uint16_t __rte_hot cn10k_sso_hws_tx_adptr_enq_##name(                  \
 624                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 625         uint16_t __rte_hot cn10k_sso_hws_tx_adptr_enq_seg_##name(              \
 626                 void *port, struct rte_event ev[], uint16_t nb_events);
 627
 628 NIX_TX_FASTPATH_MODES
 629 #undef T
 630
 631 #define SSO_TX(fn, sz, flags)                                                  \
 632         uint16_t __rte_hot fn(void *port, struct rte_event ev[],               \
 633                               uint16_t nb_events)                              \
 634         {                                                                      \
 635                 struct cn10k_sso_hws *ws = port;                               \
 636                 uint64_t cmd[sz];                                              \
 637                 RTE_SET_USED(nb_events);                                       \
 638                 return cn10k_sso_hws_event_tx(                                 \
 639                         ws, &ev[0], cmd, (const uint64_t *)ws->tx_adptr_data,  \
 640                         flags);                                                \
 641         }
 642
 643 #define SSO_TX_SEG(fn, sz, flags)                                              \
 644         uint16_t __rte_hot fn(void *port, struct rte_event ev[],               \
 645                               uint16_t nb_events)                              \
 646         {                                                                      \
 647                 uint64_t cmd[(sz) + CNXK_NIX_TX_MSEG_SG_DWORDS - 2];           \
 648                 struct cn10k_sso_hws *ws = port;                               \
 649                 RTE_SET_USED(nb_events);                                       \
 650                 return cn10k_sso_hws_event_tx(                                 \
 651                         ws, &ev[0], cmd, (const uint64_t *)ws->tx_adptr_data,  \
 652                         (flags) | NIX_TX_MULTI_SEG_F);                         \
 653         }
 654
 655 #endif