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
 122         mbuf_init |= ((uint64_t)port_id) << 48;
 123         vec = (struct rte_event_vector *)vwqe;
 124         wqe = vec->u64s;
 125
 126         nb_mbufs = RTE_ALIGN_FLOOR(vec->nb_elem, NIX_DESCS_PER_LOOP);
 127         nb_mbufs = cn10k_nix_recv_pkts_vector(&mbuf_init, vec->mbufs, nb_mbufs,
 128                                               flags | NIX_RX_VWQE_F, lookup_mem,
 129                                               tstamp, lbase);
 130         wqe += nb_mbufs;
 131         non_vec = vec->nb_elem - nb_mbufs;
 132
 133         if (flags & NIX_RX_OFFLOAD_SECURITY_F && non_vec) {
 134                 mbuf = (struct rte_mbuf *)((uintptr_t)wqe[0] -
 135                                            sizeof(struct rte_mbuf));
 136                 /* Pick first mbuf's aura handle assuming all
 137                  * mbufs are from a vec and are from same RQ.
 138                  */
 139                 aura_handle = mbuf->pool->pool_id;
 140                 ROC_LMT_BASE_ID_GET(lbase, lmt_id);
 141                 laddr = lbase;
 142                 laddr += 8;
 143                 d_off = ((uintptr_t)mbuf->buf_addr - (uintptr_t)mbuf);
 144                 d_off += (mbuf_init & 0xFFFF);
 145                 sa_base = cnxk_nix_sa_base_get(mbuf_init >> 48, lookup_mem);
 146                 sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 147         }
 148
 149         while (non_vec) {
 150                 struct nix_cqe_hdr_s *cqe = (struct nix_cqe_hdr_s *)wqe[0];
 151                 uint64_t tstamp_ptr;
 152
 153                 mbuf = (struct rte_mbuf *)((char *)cqe -
 154                                            sizeof(struct rte_mbuf));
 155
 156                 /* Translate meta to mbuf */
 157                 if (flags & NIX_RX_OFFLOAD_SECURITY_F) {
 158                         const uint64_t cq_w1 = *((const uint64_t *)cqe + 1);
 159
 160                         mbuf = nix_sec_meta_to_mbuf_sc(cq_w1, sa_base, laddr,
 161                                                        &loff, mbuf, d_off);
 162                 }
 163
 164                 cn10k_nix_cqe_to_mbuf(cqe, cqe->tag, mbuf, lookup_mem,
 165                                       mbuf_init, flags);
 166                 /* Extracting tstamp, if PTP enabled*/
 167                 tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)cqe) +
 168                                            CNXK_SSO_WQE_SG_PTR);
 169                 cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, tstamp,
 170                                         flags & NIX_RX_OFFLOAD_TSTAMP_F,
 171                                         flags & NIX_RX_MULTI_SEG_F,
 172                                         (uint64_t *)tstamp_ptr);
 173                 wqe[0] = (uint64_t *)mbuf;
 174                 non_vec--;
 175                 wqe++;
 176         }
 177
 178         /* Free remaining meta buffers if any */
 179         if (flags & NIX_RX_OFFLOAD_SECURITY_F && loff) {
 180                 nix_sec_flush_meta(laddr, lmt_id, loff, aura_handle);
 181                 plt_io_wmb();
 182         }
 183 }
 184
 185 static __rte_always_inline uint16_t
 186 cn10k_sso_hws_get_work(struct cn10k_sso_hws *ws, struct rte_event *ev,
 187                        const uint32_t flags, void *lookup_mem)
 188 {
 189         union {
 190                 __uint128_t get_work;
 191                 uint64_t u64[2];
 192         } gw;
 193         uint64_t tstamp_ptr;
 194         uint64_t mbuf;
 195
 196         gw.get_work = ws->gw_wdata;
 197 #if defined(RTE_ARCH_ARM64) && !defined(__clang__)
 198         asm volatile(
 199                 PLT_CPU_FEATURE_PREAMBLE
 200                 "caspl %[wdata], %H[wdata], %[wdata], %H[wdata], [%[gw_loc]]\n"
 201                 "sub %[mbuf], %H[wdata], #0x80                          \n"
 202                 : [wdata] "+r"(gw.get_work), [mbuf] "=&r"(mbuf)
 203                 : [gw_loc] "r"(ws->base + SSOW_LF_GWS_OP_GET_WORK0)
 204                 : "memory");
 205 #else
 206         plt_write64(gw.u64[0], ws->base + SSOW_LF_GWS_OP_GET_WORK0);
 207         do {
 208                 roc_load_pair(gw.u64[0], gw.u64[1],
 209                               ws->base + SSOW_LF_GWS_WQE0);
 210         } while (gw.u64[0] & BIT_ULL(63));
 211         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 212 #endif
 213         ws->gw_rdata = gw.u64[0];
 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                         ws->gw_rdata = cnxk_sso_hws_swtag_wait(                \
 409                                 ws->base + SSOW_LF_GWS_WQE0);                  \
 410                         return 1;                                              \
 411                 }                                                              \
 412                 return cn10k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);  \
 413         }
 414
 415 #define SSO_DEQ_SEG(fn, flags)    SSO_DEQ(fn, flags | NIX_RX_MULTI_SEG_F)
 416 #define SSO_DEQ_CA(fn, flags)     SSO_DEQ(fn, flags | CPT_RX_WQE_F)
 417 #define SSO_DEQ_CA_SEG(fn, flags) SSO_DEQ_SEG(fn, flags | CPT_RX_WQE_F)
 418
 419 #define SSO_DEQ_TMO(fn, flags)                                                 \
 420         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 421                               uint64_t timeout_ticks)                          \
 422         {                                                                      \
 423                 struct cn10k_sso_hws *ws = port;                               \
 424                 uint16_t ret = 1;                                              \
 425                 uint64_t iter;                                                 \
 426                 if (ws->swtag_req) {                                           \
 427                         ws->swtag_req = 0;                                     \
 428                         ws->gw_rdata = cnxk_sso_hws_swtag_wait(                \
 429                                 ws->base + SSOW_LF_GWS_WQE0);                  \
 430                         return ret;                                            \
 431                 }                                                              \
 432                 ret = cn10k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);   \
 433                 for (iter = 1; iter < timeout_ticks && (ret == 0); iter++)     \
 434                         ret = cn10k_sso_hws_get_work(ws, ev, flags,            \
 435                                                      ws->lookup_mem);          \
 436                 return ret;                                                    \
 437         }
 438
 439 #define SSO_DEQ_TMO_SEG(fn, flags)    SSO_DEQ_TMO(fn, flags | NIX_RX_MULTI_SEG_F)
 440 #define SSO_DEQ_TMO_CA(fn, flags)     SSO_DEQ_TMO(fn, flags | CPT_RX_WQE_F)
 441 #define SSO_DEQ_TMO_CA_SEG(fn, flags) SSO_DEQ_TMO_SEG(fn, flags | CPT_RX_WQE_F)
 442
 443 #define SSO_CMN_DEQ_BURST(fnb, fn, flags)                                      \
 444         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 445                                uint16_t nb_events, uint64_t timeout_ticks)     \
 446         {                                                                      \
 447                 RTE_SET_USED(nb_events);                                       \
 448                 return fn(port, ev, timeout_ticks);                            \
 449         }
 450
 451 #define SSO_CMN_DEQ_SEG_BURST(fnb, fn, flags)                                  \
 452         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 453                                uint16_t nb_events, uint64_t timeout_ticks)     \
 454         {                                                                      \
 455                 RTE_SET_USED(nb_events);                                       \
 456                 return fn(port, ev, timeout_ticks);                            \
 457         }
 458
 459 static __rte_always_inline struct cn10k_eth_txq *
 460 cn10k_sso_hws_xtract_meta(struct rte_mbuf *m, const uint64_t *txq_data)
 461 {
 462         return (struct cn10k_eth_txq
 463                         *)(txq_data[(txq_data[m->port] >> 48) +
 464                                     rte_event_eth_tx_adapter_txq_get(m)] &
 465                            (BIT_ULL(48) - 1));
 466 }
 467
 468 static __rte_always_inline void
 469 cn10k_sso_tx_one(struct cn10k_sso_hws *ws, struct rte_mbuf *m, uint64_t *cmd,
 470                  uint16_t lmt_id, uintptr_t lmt_addr, uint8_t sched_type,
 471                  const uint64_t *txq_data, const uint32_t flags)
 472 {
 473         uint8_t lnum = 0, loff = 0, shft = 0;
 474         struct cn10k_eth_txq *txq;
 475         uintptr_t laddr;
 476         uint16_t segdw;
 477         uintptr_t pa;
 478         bool sec;
 479
 480         txq = cn10k_sso_hws_xtract_meta(m, txq_data);
 481         cn10k_nix_tx_skeleton(txq, cmd, flags, 0);
 482         /* Perform header writes before barrier
 483          * for TSO
 484          */
 485         if (flags & NIX_TX_OFFLOAD_TSO_F)
 486                 cn10k_nix_xmit_prepare_tso(m, flags);
 487
 488         cn10k_nix_xmit_prepare(m, cmd, flags, txq->lso_tun_fmt, &sec);
 489
 490         laddr = lmt_addr;
 491         /* Prepare CPT instruction and get nixtx addr if
 492          * it is for CPT on same lmtline.
 493          */
 494         if (flags & NIX_TX_OFFLOAD_SECURITY_F && sec)
 495                 cn10k_nix_prep_sec(m, cmd, &laddr, lmt_addr, &lnum, &loff,
 496                                    &shft, txq->sa_base, flags);
 497
 498         /* Move NIX desc to LMT/NIXTX area */
 499         cn10k_nix_xmit_mv_lmt_base(laddr, cmd, flags);
 500
 501         if (flags & NIX_TX_MULTI_SEG_F)
 502                 segdw = cn10k_nix_prepare_mseg(m, (uint64_t *)laddr, flags);
 503         else
 504                 segdw = cn10k_nix_tx_ext_subs(flags) + 2;
 505
 506         cn10k_nix_xmit_prepare_tstamp(txq, laddr, m->ol_flags, segdw, flags);
 507         if (flags & NIX_TX_OFFLOAD_SECURITY_F && sec)
 508                 pa = txq->cpt_io_addr | 3 << 4;
 509         else
 510                 pa = txq->io_addr | ((segdw - 1) << 4);
 511
 512         if (!CNXK_TAG_IS_HEAD(ws->gw_rdata) && !sched_type)
 513                 ws->gw_rdata = roc_sso_hws_head_wait(ws->base);
 514
 515         roc_lmt_submit_steorl(lmt_id, pa);
 516 }
 517
 518 static __rte_always_inline void
 519 cn10k_sso_vwqe_split_tx(struct cn10k_sso_hws *ws, struct rte_mbuf **mbufs,
 520                         uint16_t nb_mbufs, uint64_t *cmd, uint16_t lmt_id,
 521                         uintptr_t lmt_addr, uint8_t sched_type,
 522                         const uint64_t *txq_data, const uint32_t flags)
 523 {
 524         uint16_t port[4], queue[4];
 525         uint16_t i, j, pkts, scalar;
 526         struct cn10k_eth_txq *txq;
 527
 528         scalar = nb_mbufs & (NIX_DESCS_PER_LOOP - 1);
 529         pkts = RTE_ALIGN_FLOOR(nb_mbufs, NIX_DESCS_PER_LOOP);
 530
 531         for (i = 0; i < pkts; i += NIX_DESCS_PER_LOOP) {
 532                 port[0] = mbufs[i]->port;
 533                 port[1] = mbufs[i + 1]->port;
 534                 port[2] = mbufs[i + 2]->port;
 535                 port[3] = mbufs[i + 3]->port;
 536
 537                 queue[0] = rte_event_eth_tx_adapter_txq_get(mbufs[i]);
 538                 queue[1] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 1]);
 539                 queue[2] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 2]);
 540                 queue[3] = rte_event_eth_tx_adapter_txq_get(mbufs[i + 3]);
 541
 542                 if (((port[0] ^ port[1]) & (port[2] ^ port[3])) ||
 543                     ((queue[0] ^ queue[1]) & (queue[2] ^ queue[3]))) {
 544                         for (j = 0; j < 4; j++)
 545                                 cn10k_sso_tx_one(ws, mbufs[i + j], cmd, lmt_id,
 546                                                  lmt_addr, sched_type, txq_data,
 547                                                  flags);
 548                 } else {
 549                         txq = (struct cn10k_eth_txq
 550                                        *)(txq_data[(txq_data[port[0]] >> 48) +
 551                                                    queue[0]] &
 552                                           (BIT_ULL(48) - 1));
 553                         cn10k_nix_xmit_pkts_vector(txq, (uint64_t *)ws,
 554                                                    &mbufs[i], 4, cmd,
 555                                                    flags | NIX_TX_VWQE_F);
 556                 }
 557         }
 558
 559         mbufs += i;
 560
 561         for (i = 0; i < scalar; i++) {
 562                 cn10k_sso_tx_one(ws, mbufs[i], cmd, lmt_id, lmt_addr,
 563                                  sched_type, txq_data, flags);
 564         }
 565 }
 566
 567 static __rte_always_inline uint16_t
 568 cn10k_sso_hws_event_tx(struct cn10k_sso_hws *ws, struct rte_event *ev,
 569                        uint64_t *cmd, const uint64_t *txq_data,
 570                        const uint32_t flags)
 571 {
 572         struct cn10k_eth_txq *txq;
 573         struct rte_mbuf *m;
 574         uintptr_t lmt_addr;
 575         uint16_t ref_cnt;
 576         uint16_t lmt_id;
 577
 578         lmt_addr = ws->lmt_base;
 579         ROC_LMT_BASE_ID_GET(lmt_addr, lmt_id);
 580
 581         if (ev->event_type & RTE_EVENT_TYPE_VECTOR) {
 582                 struct rte_mbuf **mbufs = ev->vec->mbufs;
 583                 uint64_t meta = *(uint64_t *)ev->vec;
 584
 585                 if (meta & BIT(31)) {
 586                         txq = (struct cn10k_eth_txq
 587                                        *)(txq_data[(txq_data[meta >> 32] >>
 588                                                     48) +
 589                                                    (meta >> 48)] &
 590                                           (BIT_ULL(48) - 1));
 591
 592                         cn10k_nix_xmit_pkts_vector(txq, (uint64_t *)ws, mbufs,
 593                                                    meta & 0xFFFF, cmd,
 594                                                    flags | NIX_TX_VWQE_F);
 595                 } else {
 596                         cn10k_sso_vwqe_split_tx(
 597                                 ws, mbufs, meta & 0xFFFF, cmd, lmt_id, lmt_addr,
 598                                 ev->sched_type, txq_data, flags);
 599                 }
 600                 rte_mempool_put(rte_mempool_from_obj(ev->vec), ev->vec);
 601                 return (meta & 0xFFFF);
 602         }
 603
 604         m = ev->mbuf;
 605         ref_cnt = m->refcnt;
 606         cn10k_sso_tx_one(ws, m, cmd, lmt_id, lmt_addr, ev->sched_type, txq_data,
 607                          flags);
 608
 609         if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
 610                 if (ref_cnt > 1)
 611                         return 1;
 612         }
 613
 614         cnxk_sso_hws_swtag_flush(ws->base + SSOW_LF_GWS_TAG,
 615                                  ws->base + SSOW_LF_GWS_OP_SWTAG_FLUSH);
 616         return 1;
 617 }
 618
 619 #define T(name, sz, flags)                                                     \
 620         uint16_t __rte_hot cn10k_sso_hws_tx_adptr_enq_##name(                  \
 621                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 622         uint16_t __rte_hot cn10k_sso_hws_tx_adptr_enq_seg_##name(              \
 623                 void *port, struct rte_event ev[], uint16_t nb_events);
 624
 625 NIX_TX_FASTPATH_MODES
 626 #undef T
 627
 628 #define SSO_TX(fn, sz, flags)                                                  \
 629         uint16_t __rte_hot fn(void *port, struct rte_event ev[],               \
 630                               uint16_t nb_events)                              \
 631         {                                                                      \
 632                 struct cn10k_sso_hws *ws = port;                               \
 633                 uint64_t cmd[sz];                                              \
 634                 RTE_SET_USED(nb_events);                                       \
 635                 return cn10k_sso_hws_event_tx(                                 \
 636                         ws, &ev[0], cmd, (const uint64_t *)ws->tx_adptr_data,  \
 637                         flags);                                                \
 638         }
 639
 640 #define SSO_TX_SEG(fn, sz, flags)                                              \
 641         uint16_t __rte_hot fn(void *port, struct rte_event ev[],               \
 642                               uint16_t nb_events)                              \
 643         {                                                                      \
 644                 uint64_t cmd[(sz) + CNXK_NIX_TX_MSEG_SG_DWORDS - 2];           \
 645                 struct cn10k_sso_hws *ws = port;                               \
 646                 RTE_SET_USED(nb_events);                                       \
 647                 return cn10k_sso_hws_event_tx(                                 \
 648                         ws, &ev[0], cmd, (const uint64_t *)ws->tx_adptr_data,  \
 649                         (flags) | NIX_TX_MULTI_SEG_F);                         \
 650         }
 651
 652 #endif