drivers/event/cnxk/cn9k_worker.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(C) 2021 Marvell.
   3  */
   4
   5 #ifndef __CN9K_WORKER_H__
   6 #define __CN9K_WORKER_H__
   7
   8 #include <rte_eventdev.h>
   9 #include <rte_vect.h>
  10
  11 #include "cnxk_ethdev.h"
  12 #include "cnxk_eventdev.h"
  13 #include "cnxk_worker.h"
  14 #include "cn9k_cryptodev_ops.h"
  15
  16 #include "cn9k_ethdev.h"
  17 #include "cn9k_rx.h"
  18 #include "cn9k_tx.h"
  19
  20 /* SSO Operations */
  21
  22 static __rte_always_inline uint8_t
  23 cn9k_sso_hws_new_event(struct cn9k_sso_hws *ws, const struct rte_event *ev)
  24 {
  25         const uint32_t tag = (uint32_t)ev->event;
  26         const uint8_t new_tt = ev->sched_type;
  27         const uint64_t event_ptr = ev->u64;
  28         const uint16_t grp = ev->queue_id;
  29
  30         rte_atomic_thread_fence(__ATOMIC_ACQ_REL);
  31         if (ws->xaq_lmt <= *ws->fc_mem)
  32                 return 0;
  33
  34         cnxk_sso_hws_add_work(event_ptr, tag, new_tt,
  35                               ws->grp_base + (grp << 12));
  36         return 1;
  37 }
  38
  39 static __rte_always_inline void
  40 cn9k_sso_hws_fwd_swtag(struct cn9k_sso_hws_state *vws,
  41                        const struct rte_event *ev)
  42 {
  43         const uint32_t tag = (uint32_t)ev->event;
  44         const uint8_t new_tt = ev->sched_type;
  45         const uint8_t cur_tt = CNXK_TT_FROM_TAG(plt_read64(vws->tag_op));
  46
  47         /* CNXK model
  48          * cur_tt/new_tt     SSO_TT_ORDERED SSO_TT_ATOMIC SSO_TT_UNTAGGED
  49          *
  50          * SSO_TT_ORDERED        norm           norm             untag
  51          * SSO_TT_ATOMIC         norm           norm               untag
  52          * SSO_TT_UNTAGGED       norm           norm             NOOP
  53          */
  54
  55         if (new_tt == SSO_TT_UNTAGGED) {
  56                 if (cur_tt != SSO_TT_UNTAGGED)
  57                         cnxk_sso_hws_swtag_untag(
  58                                 CN9K_SSOW_GET_BASE_ADDR(vws->getwrk_op) +
  59                                 SSOW_LF_GWS_OP_SWTAG_UNTAG);
  60         } else {
  61                 cnxk_sso_hws_swtag_norm(tag, new_tt, vws->swtag_norm_op);
  62         }
  63 }
  64
  65 static __rte_always_inline void
  66 cn9k_sso_hws_fwd_group(struct cn9k_sso_hws_state *ws,
  67                        const struct rte_event *ev, 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, CN9K_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
  73                                      SSOW_LF_GWS_OP_UPD_WQP_GRP1);
  74         cnxk_sso_hws_swtag_desched(tag, new_tt, grp, ws->swtag_desched_op);
  75 }
  76
  77 static __rte_always_inline void
  78 cn9k_sso_hws_forward_event(struct cn9k_sso_hws *ws, 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(plt_read64(ws->tag_op)) == grp) {
  84                 cn9k_sso_hws_fwd_swtag((struct cn9k_sso_hws_state *)ws, ev);
  85                 ws->swtag_req = 1;
  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                 cn9k_sso_hws_fwd_group((struct cn9k_sso_hws_state *)ws, ev,
  93                                        grp);
  94         }
  95 }
  96
  97 /* Dual ws ops. */
  98
  99 static __rte_always_inline uint8_t
 100 cn9k_sso_hws_dual_new_event(struct cn9k_sso_hws_dual *dws,
 101                             const struct rte_event *ev)
 102 {
 103         const uint32_t tag = (uint32_t)ev->event;
 104         const uint8_t new_tt = ev->sched_type;
 105         const uint64_t event_ptr = ev->u64;
 106         const uint16_t grp = ev->queue_id;
 107
 108         rte_atomic_thread_fence(__ATOMIC_ACQ_REL);
 109         if (dws->xaq_lmt <= *dws->fc_mem)
 110                 return 0;
 111
 112         cnxk_sso_hws_add_work(event_ptr, tag, new_tt,
 113                               dws->grp_base + (grp << 12));
 114         return 1;
 115 }
 116
 117 static __rte_always_inline void
 118 cn9k_sso_hws_dual_forward_event(struct cn9k_sso_hws_dual *dws,
 119                                 struct cn9k_sso_hws_state *vws,
 120                                 const struct rte_event *ev)
 121 {
 122         const uint8_t grp = ev->queue_id;
 123
 124         /* Group hasn't changed, Use SWTAG to forward the event */
 125         if (CNXK_GRP_FROM_TAG(plt_read64(vws->tag_op)) == grp) {
 126                 cn9k_sso_hws_fwd_swtag(vws, ev);
 127                 dws->swtag_req = 1;
 128         } else {
 129                 /*
 130                  * Group has been changed for group based work pipelining,
 131                  * Use deschedule/add_work operation to transfer the event to
 132                  * new group/core
 133                  */
 134                 cn9k_sso_hws_fwd_group(vws, ev, grp);
 135         }
 136 }
 137
 138 static __rte_always_inline void
 139 cn9k_wqe_to_mbuf(uint64_t wqe, const uint64_t mbuf, uint8_t port_id,
 140                  const uint32_t tag, const uint32_t flags,
 141                  const void *const lookup_mem)
 142 {
 143         const uint64_t mbuf_init = 0x100010000ULL | RTE_PKTMBUF_HEADROOM |
 144                                    (flags & NIX_RX_OFFLOAD_TSTAMP_F ? 8 : 0);
 145
 146         cn9k_nix_cqe_to_mbuf((struct nix_cqe_hdr_s *)wqe, tag,
 147                              (struct rte_mbuf *)mbuf, lookup_mem,
 148                              mbuf_init | ((uint64_t)port_id) << 48, flags);
 149 }
 150
 151 static __rte_always_inline uint16_t
 152 cn9k_sso_hws_dual_get_work(struct cn9k_sso_hws_state *ws,
 153                            struct cn9k_sso_hws_state *ws_pair,
 154                            struct rte_event *ev, const uint32_t flags,
 155                            const void *const lookup_mem,
 156                            struct cnxk_timesync_info *const tstamp)
 157 {
 158         const uint64_t set_gw = BIT_ULL(16) | 1;
 159         union {
 160                 __uint128_t get_work;
 161                 uint64_t u64[2];
 162         } gw;
 163         uint64_t tstamp_ptr;
 164         uint64_t mbuf;
 165
 166         if (flags & NIX_RX_OFFLOAD_PTYPE_F)
 167                 rte_prefetch_non_temporal(lookup_mem);
 168 #ifdef RTE_ARCH_ARM64
 169         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 170                      "rty%=:                                    \n"
 171                      "          ldr %[tag], [%[tag_loc]]        \n"
 172                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 173                      "          tbnz %[tag], 63, rty%=          \n"
 174                      "done%=:   str %[gw], [%[pong]]            \n"
 175                      "          dmb ld                          \n"
 176                      "          sub %[mbuf], %[wqp], #0x80      \n"
 177                      "          prfm pldl1keep, [%[mbuf]]       \n"
 178                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 179                        [mbuf] "=&r"(mbuf)
 180                      : [tag_loc] "r"(ws->tag_op), [wqp_loc] "r"(ws->wqp_op),
 181                        [gw] "r"(set_gw), [pong] "r"(ws_pair->getwrk_op));
 182 #else
 183         gw.u64[0] = plt_read64(ws->tag_op);
 184         while ((BIT_ULL(63)) & gw.u64[0])
 185                 gw.u64[0] = plt_read64(ws->tag_op);
 186         gw.u64[1] = plt_read64(ws->wqp_op);
 187         plt_write64(set_gw, ws_pair->getwrk_op);
 188         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 189 #endif
 190
 191         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 192                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 193                     (gw.u64[0] & 0xffffffff);
 194
 195         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 196                 if ((flags & CPT_RX_WQE_F) &&
 197                     (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 198                      RTE_EVENT_TYPE_CRYPTODEV)) {
 199                         gw.u64[1] = cn9k_cpt_crypto_adapter_dequeue(gw.u64[1]);
 200                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 201                            RTE_EVENT_TYPE_ETHDEV) {
 202                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 203
 204                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 205                         cn9k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 206                                          gw.u64[0] & 0xFFFFF, flags,
 207                                          lookup_mem);
 208                         /* Extracting tstamp, if PTP enabled*/
 209                         tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)
 210                                                             gw.u64[1]) +
 211                                                    CNXK_SSO_WQE_SG_PTR);
 212                         cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, tstamp,
 213                                                 flags & NIX_RX_OFFLOAD_TSTAMP_F,
 214                                                 flags & NIX_RX_MULTI_SEG_F,
 215                                                 (uint64_t *)tstamp_ptr);
 216                         gw.u64[1] = mbuf;
 217                 }
 218         }
 219
 220         ev->event = gw.u64[0];
 221         ev->u64 = gw.u64[1];
 222
 223         return !!gw.u64[1];
 224 }
 225
 226 static __rte_always_inline uint16_t
 227 cn9k_sso_hws_get_work(struct cn9k_sso_hws *ws, struct rte_event *ev,
 228                       const uint32_t flags, const void *const lookup_mem)
 229 {
 230         union {
 231                 __uint128_t get_work;
 232                 uint64_t u64[2];
 233         } gw;
 234         uint64_t tstamp_ptr;
 235         uint64_t mbuf;
 236
 237         plt_write64(BIT_ULL(16) | /* wait for work. */
 238                             1,    /* Use Mask set 0. */
 239                     ws->getwrk_op);
 240
 241         if (flags & NIX_RX_OFFLOAD_PTYPE_F)
 242                 rte_prefetch_non_temporal(lookup_mem);
 243 #ifdef RTE_ARCH_ARM64
 244         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 245                      "          ldr %[tag], [%[tag_loc]]        \n"
 246                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 247                      "          tbz %[tag], 63, done%=          \n"
 248                      "          sevl                            \n"
 249                      "rty%=:    wfe                             \n"
 250                      "          ldr %[tag], [%[tag_loc]]        \n"
 251                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 252                      "          tbnz %[tag], 63, rty%=          \n"
 253                      "done%=:   dmb ld                          \n"
 254                      "          sub %[mbuf], %[wqp], #0x80      \n"
 255                      "          prfm pldl1keep, [%[mbuf]]       \n"
 256                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 257                        [mbuf] "=&r"(mbuf)
 258                      : [tag_loc] "r"(ws->tag_op), [wqp_loc] "r"(ws->wqp_op));
 259 #else
 260         gw.u64[0] = plt_read64(ws->tag_op);
 261         while ((BIT_ULL(63)) & gw.u64[0])
 262                 gw.u64[0] = plt_read64(ws->tag_op);
 263
 264         gw.u64[1] = plt_read64(ws->wqp_op);
 265         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 266 #endif
 267
 268         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 269                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 270                     (gw.u64[0] & 0xffffffff);
 271
 272         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 273                 if ((flags & CPT_RX_WQE_F) &&
 274                     (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 275                      RTE_EVENT_TYPE_CRYPTODEV)) {
 276                         gw.u64[1] = cn9k_cpt_crypto_adapter_dequeue(gw.u64[1]);
 277                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 278                            RTE_EVENT_TYPE_ETHDEV) {
 279                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 280
 281                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 282                         cn9k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 283                                          gw.u64[0] & 0xFFFFF, flags,
 284                                          lookup_mem);
 285                         /* Extracting tstamp, if PTP enabled*/
 286                         tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)
 287                                                             gw.u64[1]) +
 288                                                    CNXK_SSO_WQE_SG_PTR);
 289                         cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf,
 290                                                 ws->tstamp,
 291                                                 flags & NIX_RX_OFFLOAD_TSTAMP_F,
 292                                                 flags & NIX_RX_MULTI_SEG_F,
 293                                                 (uint64_t *)tstamp_ptr);
 294                         gw.u64[1] = mbuf;
 295                 }
 296         }
 297
 298         ev->event = gw.u64[0];
 299         ev->u64 = gw.u64[1];
 300
 301         return !!gw.u64[1];
 302 }
 303
 304 /* Used in cleaning up workslot. */
 305 static __rte_always_inline uint16_t
 306 cn9k_sso_hws_get_work_empty(struct cn9k_sso_hws_state *ws, struct rte_event *ev)
 307 {
 308         union {
 309                 __uint128_t get_work;
 310                 uint64_t u64[2];
 311         } gw;
 312         uint64_t mbuf;
 313
 314 #ifdef RTE_ARCH_ARM64
 315         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 316                      "          ldr %[tag], [%[tag_loc]]        \n"
 317                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 318                      "          tbz %[tag], 63, done%=          \n"
 319                      "          sevl                            \n"
 320                      "rty%=:    wfe                             \n"
 321                      "          ldr %[tag], [%[tag_loc]]        \n"
 322                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 323                      "          tbnz %[tag], 63, rty%=          \n"
 324                      "done%=:   dmb ld                          \n"
 325                      "          sub %[mbuf], %[wqp], #0x80      \n"
 326                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 327                        [mbuf] "=&r"(mbuf)
 328                      : [tag_loc] "r"(ws->tag_op), [wqp_loc] "r"(ws->wqp_op));
 329 #else
 330         gw.u64[0] = plt_read64(ws->tag_op);
 331         while ((BIT_ULL(63)) & gw.u64[0])
 332                 gw.u64[0] = plt_read64(ws->tag_op);
 333
 334         gw.u64[1] = plt_read64(ws->wqp_op);
 335         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 336 #endif
 337
 338         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 339                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 340                     (gw.u64[0] & 0xffffffff);
 341
 342         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 343                 if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 344                     RTE_EVENT_TYPE_ETHDEV) {
 345                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 346
 347                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 348                         cn9k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 349                                          gw.u64[0] & 0xFFFFF, 0, NULL);
 350                         gw.u64[1] = mbuf;
 351                 }
 352         }
 353
 354         ev->event = gw.u64[0];
 355         ev->u64 = gw.u64[1];
 356
 357         return !!gw.u64[1];
 358 }
 359
 360 /* CN9K Fastpath functions. */
 361 uint16_t __rte_hot cn9k_sso_hws_enq(void *port, const struct rte_event *ev);
 362 uint16_t __rte_hot cn9k_sso_hws_enq_burst(void *port,
 363                                           const struct rte_event ev[],
 364                                           uint16_t nb_events);
 365 uint16_t __rte_hot cn9k_sso_hws_enq_new_burst(void *port,
 366                                               const struct rte_event ev[],
 367                                               uint16_t nb_events);
 368 uint16_t __rte_hot cn9k_sso_hws_enq_fwd_burst(void *port,
 369                                               const struct rte_event ev[],
 370                                               uint16_t nb_events);
 371
 372 uint16_t __rte_hot cn9k_sso_hws_dual_enq(void *port,
 373                                          const struct rte_event *ev);
 374 uint16_t __rte_hot cn9k_sso_hws_dual_enq_burst(void *port,
 375                                                const struct rte_event ev[],
 376                                                uint16_t nb_events);
 377 uint16_t __rte_hot cn9k_sso_hws_dual_enq_new_burst(void *port,
 378                                                    const struct rte_event ev[],
 379                                                    uint16_t nb_events);
 380 uint16_t __rte_hot cn9k_sso_hws_dual_enq_fwd_burst(void *port,
 381                                                    const struct rte_event ev[],
 382                                                    uint16_t nb_events);
 383 uint16_t __rte_hot cn9k_sso_hws_ca_enq(void *port, struct rte_event ev[],
 384                                        uint16_t nb_events);
 385 uint16_t __rte_hot cn9k_sso_hws_dual_ca_enq(void *port, struct rte_event ev[],
 386                                             uint16_t nb_events);
 387
 388 #define R(name, f6, f5, f4, f3, f2, f1, f0, flags)                             \
 389         uint16_t __rte_hot cn9k_sso_hws_deq_##name(                            \
 390                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 391         uint16_t __rte_hot cn9k_sso_hws_deq_burst_##name(                      \
 392                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 393                 uint64_t timeout_ticks);                                       \
 394         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_##name(                        \
 395                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 396         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_burst_##name(                  \
 397                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 398                 uint64_t timeout_ticks);                                       \
 399         uint16_t __rte_hot cn9k_sso_hws_deq_ca_##name(                         \
 400                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 401         uint16_t __rte_hot cn9k_sso_hws_deq_ca_burst_##name(                   \
 402                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 403                 uint64_t timeout_ticks);                                       \
 404         uint16_t __rte_hot cn9k_sso_hws_deq_seg_##name(                        \
 405                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 406         uint16_t __rte_hot cn9k_sso_hws_deq_seg_burst_##name(                  \
 407                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 408                 uint64_t timeout_ticks);                                       \
 409         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_seg_##name(                    \
 410                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 411         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_seg_burst_##name(              \
 412                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 413                 uint64_t timeout_ticks);                                       \
 414         uint16_t __rte_hot cn9k_sso_hws_deq_ca_seg_##name(                     \
 415                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 416         uint16_t __rte_hot cn9k_sso_hws_deq_ca_seg_burst_##name(               \
 417                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 418                 uint64_t timeout_ticks);
 419
 420 NIX_RX_FASTPATH_MODES
 421 #undef R
 422
 423 #define R(name, f6, f5, f4, f3, f2, f1, f0, flags)                             \
 424         uint16_t __rte_hot cn9k_sso_hws_dual_deq_##name(                       \
 425                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 426         uint16_t __rte_hot cn9k_sso_hws_dual_deq_burst_##name(                 \
 427                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 428                 uint64_t timeout_ticks);                                       \
 429         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_##name(                   \
 430                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 431         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_burst_##name(             \
 432                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 433                 uint64_t timeout_ticks);                                       \
 434         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_##name(                    \
 435                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 436         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_burst_##name(              \
 437                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 438                 uint64_t timeout_ticks);                                       \
 439         uint16_t __rte_hot cn9k_sso_hws_dual_deq_seg_##name(                   \
 440                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 441         uint16_t __rte_hot cn9k_sso_hws_dual_deq_seg_burst_##name(             \
 442                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 443                 uint64_t timeout_ticks);                                       \
 444         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_seg_##name(               \
 445                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 446         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_seg_burst_##name(         \
 447                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 448                 uint64_t timeout_ticks);                                       \
 449         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_seg_##name(                \
 450                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 451         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_seg_burst_##name(          \
 452                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 453                 uint64_t timeout_ticks);
 454
 455 NIX_RX_FASTPATH_MODES
 456 #undef R
 457
 458 static __rte_always_inline void
 459 cn9k_sso_txq_fc_wait(const struct cn9k_eth_txq *txq)
 460 {
 461         while (!((txq->nb_sqb_bufs_adj -
 462                   __atomic_load_n(txq->fc_mem, __ATOMIC_RELAXED))
 463                  << (txq)->sqes_per_sqb_log2))
 464                 ;
 465 }
 466
 467 static __rte_always_inline const struct cn9k_eth_txq *
 468 cn9k_sso_hws_xtract_meta(struct rte_mbuf *m,
 469                          const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT])
 470 {
 471         return (const struct cn9k_eth_txq *)
 472                 txq_data[m->port][rte_event_eth_tx_adapter_txq_get(m)];
 473 }
 474
 475 static __rte_always_inline void
 476 cn9k_sso_hws_prepare_pkt(const struct cn9k_eth_txq *txq, struct rte_mbuf *m,
 477                          uint64_t *cmd, const uint32_t flags)
 478 {
 479         roc_lmt_mov(cmd, txq->cmd, cn9k_nix_tx_ext_subs(flags));
 480         cn9k_nix_xmit_prepare(m, cmd, flags, txq->lso_tun_fmt);
 481 }
 482
 483 #if defined(RTE_ARCH_ARM64)
 484
 485 static __rte_always_inline void
 486 cn9k_sso_hws_xmit_sec_one(const struct cn9k_eth_txq *txq, uint64_t base,
 487                           struct rte_mbuf *m, uint64_t *cmd,
 488                           uint32_t flags)
 489 {
 490         struct cn9k_outb_priv_data *outb_priv;
 491         rte_iova_t io_addr = txq->cpt_io_addr;
 492         uint64_t *lmt_addr = txq->lmt_addr;
 493         struct cn9k_sec_sess_priv mdata;
 494         struct nix_send_hdr_s *send_hdr;
 495         uint64_t sa_base = txq->sa_base;
 496         uint32_t pkt_len, dlen_adj, rlen;
 497         uint64x2_t cmd01, cmd23;
 498         uint64_t lmt_status, sa;
 499         union nix_send_sg_s *sg;
 500         uintptr_t dptr, nixtx;
 501         uint64_t ucode_cmd[4];
 502         uint64_t esn, *iv;
 503         uint8_t l2_len;
 504
 505         mdata.u64 = *rte_security_dynfield(m);
 506         send_hdr = (struct nix_send_hdr_s *)cmd;
 507         if (flags & NIX_TX_NEED_EXT_HDR)
 508                 sg = (union nix_send_sg_s *)&cmd[4];
 509         else
 510                 sg = (union nix_send_sg_s *)&cmd[2];
 511
 512         if (flags & NIX_TX_NEED_SEND_HDR_W1)
 513                 l2_len = cmd[1] & 0xFF;
 514         else
 515                 l2_len = m->l2_len;
 516
 517         /* Retrieve DPTR */
 518         dptr = *(uint64_t *)(sg + 1);
 519         pkt_len = send_hdr->w0.total;
 520
 521         /* Calculate rlen */
 522         rlen = pkt_len - l2_len;
 523         rlen = (rlen + mdata.roundup_len) + (mdata.roundup_byte - 1);
 524         rlen &= ~(uint64_t)(mdata.roundup_byte - 1);
 525         rlen += mdata.partial_len;
 526         dlen_adj = rlen - pkt_len + l2_len;
 527
 528         /* Update send descriptors. Security is single segment only */
 529         send_hdr->w0.total = pkt_len + dlen_adj;
 530         sg->seg1_size = pkt_len + dlen_adj;
 531
 532         /* Get area where NIX descriptor needs to be stored */
 533         nixtx = dptr + pkt_len + dlen_adj;
 534         nixtx += BIT_ULL(7);
 535         nixtx = (nixtx - 1) & ~(BIT_ULL(7) - 1);
 536
 537         roc_lmt_mov((void *)(nixtx + 16), cmd, cn9k_nix_tx_ext_subs(flags));
 538
 539         /* Load opcode and cptr already prepared at pkt metadata set */
 540         pkt_len -= l2_len;
 541         pkt_len += sizeof(struct roc_onf_ipsec_outb_hdr) +
 542                     ROC_ONF_IPSEC_OUTB_MAX_L2_INFO_SZ;
 543         sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 544
 545         sa = (uintptr_t)roc_nix_inl_onf_ipsec_outb_sa(sa_base, mdata.sa_idx);
 546         ucode_cmd[3] = (ROC_CPT_DFLT_ENG_GRP_SE_IE << 61 | sa);
 547         ucode_cmd[0] = (ROC_IE_ONF_MAJOR_OP_PROCESS_OUTBOUND_IPSEC << 48 |
 548                         0x40UL << 48 | pkt_len);
 549
 550         /* CPT Word 0 and Word 1 */
 551         cmd01 = vdupq_n_u64((nixtx + 16) | (cn9k_nix_tx_ext_subs(flags) + 1));
 552         /* CPT_RES_S is 16B above NIXTX */
 553         cmd01 = vsetq_lane_u8(nixtx & BIT_ULL(7), cmd01, 8);
 554
 555         /* CPT word 2 and 3 */
 556         cmd23 = vdupq_n_u64(0);
 557         cmd23 = vsetq_lane_u64((((uint64_t)RTE_EVENT_TYPE_CPU << 28) |
 558                                 CNXK_ETHDEV_SEC_OUTB_EV_SUB << 20), cmd23, 0);
 559         cmd23 = vsetq_lane_u64((uintptr_t)m | 1, cmd23, 1);
 560
 561         dptr += l2_len - ROC_ONF_IPSEC_OUTB_MAX_L2_INFO_SZ -
 562                 sizeof(struct roc_onf_ipsec_outb_hdr);
 563         ucode_cmd[1] = dptr;
 564         ucode_cmd[2] = dptr;
 565
 566         /* Update IV to zero and l2 sz */
 567         *(uint16_t *)(dptr + sizeof(struct roc_onf_ipsec_outb_hdr)) =
 568                 rte_cpu_to_be_16(ROC_ONF_IPSEC_OUTB_MAX_L2_INFO_SZ);
 569         iv = (uint64_t *)(dptr + 8);
 570         iv[0] = 0;
 571         iv[1] = 0;
 572
 573         /* Head wait if needed */
 574         if (base)
 575                 roc_sso_hws_head_wait(base + SSOW_LF_GWS_TAG);
 576
 577         /* ESN */
 578         outb_priv = roc_nix_inl_onf_ipsec_outb_sa_sw_rsvd((void *)sa);
 579         esn = outb_priv->esn;
 580         outb_priv->esn = esn + 1;
 581
 582         ucode_cmd[0] |= (esn >> 32) << 16;
 583         esn = rte_cpu_to_be_32(esn & (BIT_ULL(32) - 1));
 584
 585         /* Update ESN and IPID and IV */
 586         *(uint64_t *)dptr = esn << 32 | esn;
 587
 588         rte_io_wmb();
 589         cn9k_sso_txq_fc_wait(txq);
 590
 591         /* Write CPT instruction to lmt line */
 592         vst1q_u64(lmt_addr, cmd01);
 593         vst1q_u64(lmt_addr + 2, cmd23);
 594
 595         roc_lmt_mov_seg(lmt_addr + 4, ucode_cmd, 2);
 596
 597         if (roc_lmt_submit_ldeor(io_addr) == 0) {
 598                 do {
 599                         vst1q_u64(lmt_addr, cmd01);
 600                         vst1q_u64(lmt_addr + 2, cmd23);
 601                         roc_lmt_mov_seg(lmt_addr + 4, ucode_cmd, 2);
 602
 603                         lmt_status = roc_lmt_submit_ldeor(io_addr);
 604                 } while (lmt_status == 0);
 605         }
 606 }
 607 #else
 608
 609 static inline void
 610 cn9k_sso_hws_xmit_sec_one(const struct cn9k_eth_txq *txq, uint64_t base,
 611                           struct rte_mbuf *m, uint64_t *cmd,
 612                           uint32_t flags)
 613 {
 614         RTE_SET_USED(txq);
 615         RTE_SET_USED(base);
 616         RTE_SET_USED(m);
 617         RTE_SET_USED(cmd);
 618         RTE_SET_USED(flags);
 619 }
 620 #endif
 621
 622 static __rte_always_inline uint16_t
 623 cn9k_sso_hws_event_tx(uint64_t base, struct rte_event *ev, uint64_t *cmd,
 624                       const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT],
 625                       const uint32_t flags)
 626 {
 627         struct rte_mbuf *m = ev->mbuf;
 628         const struct cn9k_eth_txq *txq;
 629         uint16_t ref_cnt = m->refcnt;
 630
 631         /* Perform header writes before barrier for TSO */
 632         cn9k_nix_xmit_prepare_tso(m, flags);
 633         /* Lets commit any changes in the packet here in case when
 634          * fast free is set as no further changes will be made to mbuf.
 635          * In case of fast free is not set, both cn9k_nix_prepare_mseg()
 636          * and cn9k_nix_xmit_prepare() has a barrier after refcnt update.
 637          */
 638         if (!(flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) &&
 639             !(flags & NIX_TX_OFFLOAD_SECURITY_F))
 640                 rte_io_wmb();
 641         txq = cn9k_sso_hws_xtract_meta(m, txq_data);
 642         cn9k_sso_hws_prepare_pkt(txq, m, cmd, flags);
 643
 644         if (flags & NIX_TX_OFFLOAD_SECURITY_F) {
 645                 uint64_t ol_flags = m->ol_flags;
 646
 647                 if (ol_flags & RTE_MBUF_F_TX_SEC_OFFLOAD) {
 648                         uintptr_t ssow_base = base;
 649
 650                         if (ev->sched_type)
 651                                 ssow_base = 0;
 652
 653                         cn9k_sso_hws_xmit_sec_one(txq, ssow_base, m, cmd,
 654                                                   flags);
 655                         goto done;
 656                 }
 657
 658                 if (!(flags & NIX_TX_OFFLOAD_MBUF_NOFF_F))
 659                         rte_io_wmb();
 660         }
 661
 662         if (flags & NIX_TX_MULTI_SEG_F) {
 663                 const uint16_t segdw = cn9k_nix_prepare_mseg(m, cmd, flags);
 664                 if (!CNXK_TT_FROM_EVENT(ev->event)) {
 665                         cn9k_nix_xmit_mseg_prep_lmt(cmd, txq->lmt_addr, segdw);
 666                         roc_sso_hws_head_wait(base + SSOW_LF_GWS_TAG);
 667                         cn9k_sso_txq_fc_wait(txq);
 668                         if (cn9k_nix_xmit_submit_lmt(txq->io_addr) == 0)
 669                                 cn9k_nix_xmit_mseg_one(cmd, txq->lmt_addr,
 670                                                        txq->io_addr, segdw);
 671                 } else {
 672                         cn9k_nix_xmit_mseg_one(cmd, txq->lmt_addr, txq->io_addr,
 673                                                segdw);
 674                 }
 675         } else {
 676                 if (!CNXK_TT_FROM_EVENT(ev->event)) {
 677                         cn9k_nix_xmit_prep_lmt(cmd, txq->lmt_addr, flags);
 678                         roc_sso_hws_head_wait(base + SSOW_LF_GWS_TAG);
 679                         cn9k_sso_txq_fc_wait(txq);
 680                         if (cn9k_nix_xmit_submit_lmt(txq->io_addr) == 0)
 681                                 cn9k_nix_xmit_one(cmd, txq->lmt_addr,
 682                                                   txq->io_addr, flags);
 683                 } else {
 684                         cn9k_nix_xmit_one(cmd, txq->lmt_addr, txq->io_addr,
 685                                           flags);
 686                 }
 687         }
 688
 689 done:
 690         if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
 691                 if (ref_cnt > 1)
 692                         return 1;
 693         }
 694
 695         cnxk_sso_hws_swtag_flush(base + SSOW_LF_GWS_TAG,
 696                                  base + SSOW_LF_GWS_OP_SWTAG_FLUSH);
 697
 698         return 1;
 699 }
 700
 701 #define T(name, f6, f5, f4, f3, f2, f1, f0, sz, flags)                         \
 702         uint16_t __rte_hot cn9k_sso_hws_tx_adptr_enq_##name(                   \
 703                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 704         uint16_t __rte_hot cn9k_sso_hws_tx_adptr_enq_seg_##name(               \
 705                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 706         uint16_t __rte_hot cn9k_sso_hws_dual_tx_adptr_enq_##name(              \
 707                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 708         uint16_t __rte_hot cn9k_sso_hws_dual_tx_adptr_enq_seg_##name(          \
 709                 void *port, struct rte_event ev[], uint16_t nb_events);
 710
 711 NIX_TX_FASTPATH_MODES
 712 #undef T
 713
 714 #endif